doi
stringlengths 17
24
| transcript
stringlengths 305
148k
| abstract
stringlengths 5
6.38k
|
|---|---|---|
10.5446/20179 (DOI)
|
This talk is with my co-worker, Natal, who should be here for this talk, but he has some problems to come from France, so I will be this alone. I want to tell you our story about the birds in SoulSack. Anyone who had already used a system configuration management here? Okay, that's good. So, about me, my name is Paolo. As you can see, I came from Argentina, so this is why I have a very bad English accent. Also, I've been living in France for the last five years, and I have a very bad French accent too. So, I work as system administrator developers for the last ten years, and as Python developer for almost five years. So, our story is, it starts at PeopleDoc. PeopleDoc is a French startup created five years ago. It has offices in Paris and New York. The IT department has more than 40 people across different teams, locations, different locations in Paris or France. At PeopleDoc, we contribute a lot of open source projects, so it's a good place to start with SoulSack. To describe a little more technical level of the company, here is a list of the teams. We have two Python Django teams, and they work with the same stack. We use ready, static search, post-res, and we have a list of different teams. Each team has different goals, and we have different goals, and we work on different projects of product. We have the Java Scala team. They create the backend services and some APIs used for the Django applications. We also have a new brand single-page application created with MBRJS. We use Django for the API on the backend. And the last one is the DevOps team. This team is responsible for the infrastructure and the different environments like the staging, QA, and production systems. So we will talk about tools. To simplify, we use SoulSack for server provisioning and configuration management. Also, we have open stack for the cloud instances. We have a lot of different tools, and we use the same tools for the cloud. We also have open stack for the cloud instances. We have a local cluster to test some integrations on the QA servers. And for the production system, we have an open stack provider. So we have the same tools, and we have the same tools, and we have the same tools, and we have the same tools for all the environments, local, or private, or production environments. Another important tool is Neuralic. We use it for alert and monitoring. It's a very good tool because it allows to share information between different teams. But sometimes we need to use a different tool, and we need to use a different tool to understand the different systems. And we also have a different tool for the cloud. We also have a different tool for the cloud. We also have a different tool for the cloud. It allows to share information between different teams. But sometimes we need to use more specific tools like SABIC, Sentry, or Munini. So this is the main tool we use every day for the production environment. The next question is why we choose Soul Stack? I think as you know, there are other solutions like Chef, Puppet, and Seville. So why Soul Stack? For us, there are two main reasons. The first one is because Soul Stack is open source. So people do we use a lot of open source projects. So it's a very important reason to use Soul Stack. And the second reason is because Soul Stack is open source. So people do we use a lot of open source projects. So it's a very important reason. The second one is because Python. Because Soul Stack is quite in Python. The DevOps team uses a lot of Python. So it's a logical choice. And this leaves out of the list Chef and Puppet. So the last one is Asuall. Asuall meets these two conditions, open source Python. But Asuall is more recent than Soul Stack. When we choose, we take the decision to use Soul Stack. And Seville was at the very early stage of development. So Soul Stack wins. There are many other reasons to choose Soul Stack. If you want, you can see the list of the features of Soul Stack in the project page. There is more technical details if you want. Before to continue, I will show you some main concepts about Soul Stack. So basically, there are the same concepts in all the configuration management systems. And Seville, I think, has the same concept. For example, the first one is master minion. Such who work in a client server model. So it's very simple. The master, the soul master, push the command to the minion and the minion sends the result back to the master. This is the first most important thing to understand about Soul Stack. Then states. Soul to states describe the state in which the system should be in. By default, the states are the EML file. I will show you an example. Here, for example, we have a file called packages.sls. With beam and a function to install the packet. The second block shows the command to install this package or this list of package in the minions called server one. So this is very simple example of Soul Stack. Another important concept are pillars. Pillars are structures of data defined on the soul master that they are used to the global variables or sensitive data that will be shared by some minions or all the minions at the same time. Here is an example. Here you can see I define a list of users with the user name and user ID. So with this kind of data, I can create all this user or all the minions everywhere. So another concept is grains. Grains are static information stored in the minion. It's not like pillar where the information is shared between all the minions. It's static information only specific to the minion. For example, you can have things like the running code version, the operating system, or I don't know, many other information about the minion. So those were some basic concepts about to understand salt. Now I will show you the workflow we use at People Duck and how we use salt. This is a very simple version of the workflow. But when each team delivers a new release of their product, the DevOps teams use salt stack to deploy these products into the production system. So the important thing to see here is the DevOps team needs to create every site for each product for each team. So you have just one place to describe all the infrastructure for all your environment. And in your case, we have just one Git repository with all the states to deploy every product of People Duck. It works like this. It works pretty well. But not for longer. Why? The first one, the first problem is the bottleneck. As you can see in this slide, the DevOps team can become the bottleneck of your system. Because if you have many teams, many releases or many teams, all the time it will quickly have a problem to centralize all the deployment for only just one team. Another problem may be the complexity. If you have many different stacks, remember there are Java, Scala, Python, last excel, Postgres, Redis, you have a lot of complex relations and you can manage. So it can be a problem very hard to change. Another problem is, as I said, we have only just one Git repository to store all the state. So when you start, it can be practical to have all the state in the same place. But when everything grows, it may be hard to manage all these states together without avoiding the conflict or different versions. Maybe one product needs a version of Postgres, another project needs another version. So it's not easy to manage this kind of problem. So we have seen all these problems at PeopleDoc. So we have some ideas that can help. You build it, you run it. This is a quote from Bernhard Bögel. The idea is to give the operational responsibilities to developers. These developers are in contact with the day-to-day operation of their product. So this can avoid the problem of the bottleneck because each team will be responsible of the deployment of their product. So this is the first idea. Second one is a feature of a source tag called formula. The formula is pre-written state. So it allows you to reuse a state into your projects to do many different tasks like install packages, start a service, or many other common tasks. With the formula, you can have a modular design of your project. You can avoid the problem of monolithic states. So also this allows you or the team to write their own own code. So it's a good idea, a good practice to use the formula. Finally, test. For each change you do in your factory code, you need to test this change somewhere. Because if you think about the factory code, the code needs to be tested. So you need to test. It's not easy to test because you need to simulate all your environment in your local computer. But it's very complex to simulate the same factory tool because in the production system, you can have different, I don't know, server hardware, internet connection, or I don't know, many things can change. So it's not easy. But we start to use Jenkins and LXE, Linux containers. So this allows you to create the same services or containers that you can have in an OpenStack cluster. But you can run this in the same Jenkins job. So it's not easy because we need to think about all these kind of relations of interaction between different products, like the states, grains, PILAR, PILAR is different for each environment. So it's not easy. So also if you test in your local environment, you can never be sure that that will be exactly the same in the production environment because you don't have the same hardware and factory tool in the production system. But you need to approach or detect the problems before that touch the production system. So this idea is worth, we use it for, I think, five months ago. So it's a good idea. But there is no much information about this subject on Solstack. How do you test your state? So we create our own formula to test with LXE and Jenkins. If you go to the formula repository of Solstack, there is a repository called TankistFarmula. Right there, you can find many useful things to test your state. That is all. So thank you. Do you have any questions? Hi. One of the reasons why I haven't used Salt recently is a couple of years ago I was doing DevOps for building development environments. One of the problems was the primary thing I was debugging was Salt was being very unstable. Has that somewhat calmed down a bit? It was actually versions of Salt. Has Salt gotten a lot more stable than it was two years ago, as far as my question? Salt is more stable. Solstack has grown very much the last one to two years. So now I think it's very, very stable. How different are you using Salt in building your development environments as well? Is it just local and then you deploy it and then you use a separate setup? Sorry, I don't know. So one of the things I was doing with Salt was using Inside Vagrant to build the near identical development environments to what we're deploying in live. Is that what you're doing or are you still using Salt in development as well? If we use Salt in development? Yeah. Yes, yes, I. The developer team uses Salt stack to deploy its products in the local environment. Okay. Thanks. Any more questions? Okay, so thank you.
|
Pablo Seminario - The Salt Route An introduction to the devops culture by sharing our experience at PeopleDoc Inc. a successfully French start-up. The salt route talk presents some best practices and common mistakes that arise in everyday teamwork between developers and sysadmins using SaltStack for configuration management, server provisioning, orchestration and Django web applications deployment. As an introductory talk there is no prerequisites required.
|
10.5446/20177 (DOI)
|
So, this is the outline of my talk today. First of all, an introduction about multi-body simulation. Next, yeah, okay, some background information. Then I would like to show you some assemblies. So I would fill this talk, at least half of it, to show you some assemblies so that you get an impression what is in fact possible with this package in the end. This package is not in public already. It will be published maybe in September. And in the end, I will give you a short note of future work and some backup. The next name is multi-body symbolic. So in fact, there are two different ways to approach this problem. You can use symbolic equations and you can stick more to the numerical side. Most of the industrial products stick to the numerical side. But we decided to approach it analytically. Our aim is to provide on basis of existing Python packages to provide once in a while a complete multi-body simulation tool. So why is this important for us? You can guess it. We want to be independent from market leaders. For example, SimPak or also VL or also Adams costs a whole lot of money, the license, and all of these companies are now, they are now, yeah, bought by very big companies. So it is not possible to stick in this development process much easily. Second scripting ability is included. Of course, it's in Python. So it's, of course, included. And third educational purposes. So what is multi-body simulation? Multi-body simulation deals with systems which can be described by these equations. These are the Newtonian Euler equations. They are differential equations. So they are well known since several hundred years. What is the problem writing them down and integrating them? Well, the problem here is, as you will see in my talk, this F is not only an expression which you can write down. It happens that most of multi-body simulation assemblies include constraints. And constraints are like a ball which runs on a table or sliders or whatsoever. So these constraint forces are in fact the difficult thing in multi-body simulations. We look back at the development process of 30 years more or less, not us, but in general. So the community, the scientific community is working on this problem since more or less 30 years. And you can imagine they come up a whole tons of papers out of there. And you cannot expect to climb up within half a year on top. But we are very enthusiastic and we think we can go ahead this way. So use cases of this multi-body simulation is mechanical engineering, ground vehicle dynamics, robotics and biomechanics. Each of these branches, each of these branches is modern and relevant. So I'm working in ground vehicle dynamics since more or less seven years. And I'm doing write and handling simulation for another company right now. And I know what the difficulties are in this field. So what package are we using? SimPy, it's more or less the basis of this one. SimPy, these guys who produced SimPy, they were doing a really, really, really great job. And I would like to thank you, those people. It's a symbolic algebra package. So it replaces more or less these three products. And it's on a very, it's right at the moment, it's on an advanced level. So it's not that you can just do some algebra derivatives or integrals. It's on a high advanced level and it also includes already advanced mechanics. So why do not parcel these together, stick these loose ends together and produce something with which you can do your multi-body simulation? So fine, we try to do this. Yeah, I would like not to forget these packages which are really, really helpful and maybe the core of numerical and scientific Python programming, NumPy and SimPy. And without these, it would not happen anything about of these. For linear algebraic servers, we use NumPy. They are very, very advanced and not forget well tested. So these packages are so well tested that you can trust on this. And this is good. There are some ODE solvers in SimPy available. So in the end, we need some graphics. We did with vPy, it's a medium advanced. It's really, really nice. But it has some primitives called primitives, some rods, springs and so on. And you can put these together and make out of these your graphics simulation. In fact, you need this to make sure that your assembled equations behave well. So just a visual check of your solution. I mean, you cannot just solve your equations and then put some graphs and then you cannot judge if this is done really nice or really good. You have some visual check up. Now some background theory. I will try to make this short. Building blocks of a mechanical system. You have your bodies. Everybody has six degrees of freedom, translation rotational, including the time derivatives. You end up with 12 degrees of freedom. So each body should lead to 12 lines in your system of ordinary differential equations. If there wouldn't be some possibilities to boil it down somehow. Each of these bodies has mass moments of inertia in trincy. So this is the figure citation. This is not mine. This is also not my figure. So I will cite it. You can boil down. You can boil down the numbers of degrees of freedom by joints. Toins you can think about as two bodies are connected. We are joints like in your skeleton. And in fact, it's technically done more or less the same. Each of these joints can be produced in a technical sense. And in the end, with these joints, you reduce the numbers of degrees of freedom. Here we talk about these types of joints. Cardiac, axis, revolute. For example, if you think about a tire on a car, it's just revolute. Except that the front tire can also steer. But more or less revolute, so it's a revolute joint. In these equations, appears forces and torques. Forces and torques accelerates your masses. So forces most of the time appear pairwise. So if one body is attracting another, it's the same done for the other body. But you have different types of forces. Pairwise forces, so think about sun and earth, for example. The external forces, if you just want to find a system of a mechanical system on earth, you would add the gravitational force, of course. And you wouldn't include the earth as an independent mass. But you just would add your gravitational field as an external force. And the difficult thing is the constrained forces. For example, if you have here a surface and a ball is running on the surface, the surface would act a force on the ball to be on this surface. So in multi-body simulation, you see most many times these kind of drawings, because these kind of drawings show most of the things included. Here, for example, you have a chain of bodies, and here are the joints. And you can see if you switch from Cartesian coordinates to just the angle here, you can reduce it somehow, the number of degrees of freedom. Nobody would write this down in Cartesian coordinates, but in angles. But you know, if the generalized coordinates are minimal, it equals the number of degrees of freedom. But it is not always the case, so you can have more generalized coordinates, so you would have add on top of it some equations, which gives you the constraints. Even the generalized coordinates are not unique. So you can, for example, measure the angle towards the set axis or towards the axis of the previous body. So it is not at all unique, the set of generalized coordinates. One problem, which is really maybe the main or one of the core problems in multi-body simulation is this one. So if you have, this is called a constrained loop. If you do not just have your joints, but if you have another constraint in the end, which builds a constrained loop, you can say, okay, these angles here on top, they are somehow connected to each other. They cannot be independent at all. And this is called constrained loop, and this constrained loop produces another equation, most of the time, algebraic equation. And this algebraic equation, you can take care of in just several possibilities to solve these kind of problems. Most of the time, there are differential algebraic equation methods, but they are very costly, and in Python, maybe too slow. So we propose here a solution according to Lagrange one with additional drawback force. For linearization, you can always put this algebraic equation into your set of equations. Here this is the set of methods we can use to generate our equations of motions. In this package, they produce here already Keynes method, which is also called Dallon-Bear's principle of virtual work. And Lagrange method number two, is another possibility but not used. So I will show you how it is used. You just need to have an object which sets up your world, gives you a world coordinate system, and a marker. A marker is always a coordinate system in the language of multi-body simulation. There are methods provided to add bodies, markers, so extra coordinate systems, external forces, extra constraints, and for example, reflective walls. An ODE solver is connected, more or less automatically. A 3D geographical back-end is connected automatically. So each body signs up in the graphical back-end and appears somehow as you want it to be appeared later. Some physical quantities are provided like energy, forces, velocity, and so on. We are inside Python, so you can calculate whatever you want here. And this is the advantage. New and interesting. So what did we put into this work? We would like to have a completeness of joints and tools. And the Jacobian is calculated for linearization analysis. So this is on top. Nice feature, very nice feature. The linearization tool, which is already in Sympy, is kind of completed. You can detect automatically independent and dependent coordinates and so on. Constraints loop are more or less kind of solved here. External models and parameters can be included. This is a very important point here, the external models, which I will show you in one example. We have also some B-splines, which can be used if you don't have an analytical expression for your force. You can also include some kind of B-splines for having a representative of your force function. So sometimes measurements don't give you analytical expressions. Okay, coding style, set up the system. Here this is your multi-body simulation world. You add your bodies, your markers, your forces, your force models, your external force model, your geometric constraint. Solving, assembling and solving is just this one. Here you have maybe some constant, like the gravitational constant. You give it a number. You produce the equations of motion and integrate them. Force processing is calculation of linear analysis, linearization, so stability analysis. You prepare and animate your result. If you would like to be a developer once in a time or if you would like to work with SimPy once in a time, you should be aware of these three things. I mean you should be aware of many things, but these three especially never use SimPy for numerics. Do not try to use, for example, to use SimPy to solve for eigenvectors. It will never happen. It will never be as fast as NumPy. So find the right step between SimPy to NumPy. So I got designed for 10 minutes right now, so I will hurry a little bit up. LambdaPy is one of the core functions of SimPy. I mean if you have an algebraic expression, what can you do with it in a computer? You would like to use it as a function and to use it as a function, you use this word which is called LambdaPy. You lambdaPy expressions into Python functions and this is maybe the most impressive invention here. Not our from SimPy. And ODE solvers don't use your own, so even if it looks like fun to produce one, it is not never as good as those which are around. Use those. We use LS ODR solver. This is also called for sundials. If you are interested in these topics, you know that sundials is an open-source solver which is out there but not connected to Python at the moment. It is connected but it stopped at Python 2.6 or so. So I would like to connect this and use this sundial solvers which is really good in my opinion. Okay, I will show you examples and for these examples, not only the pictures but also the results of the movies. Okay, let's try. Okay, this is bad because now it appears here on my screen and not on the screen. I wanted it to be appeared. So this is what comes out more or less. This is a crank slide, something called crank slide. And this is a simple example for a constrained loop. Here this goes round and round and goes linear. So this is rotating linear. You see we put in also some forces and here an extra constrained force. So this, if we don't put this constrained force here, it would be just a pendulum. So right now I will give you another. So I will skip some of these because I have 12 and this is too much. Maybe there is one. Yeah, some spring and it's a reflective wall here at this position. There is a reflective wall. It's interesting because it's not so fluent but I don't know why because it's in the middle of two screens maybe. So this is a reflective wall example. I will show you this one. This is a simple car. And this is non-trivial. I mean maybe you can see this that assembling these equations takes around 60 seconds. Let's take a look at the assembly here. So these equations, they just come up by this mechanism. And this mechanism, as I told you, is quite easy. So we just add bodies, markers, special forces and so on. And in the end here, this is the steering more or less. And in the end, in the end, here this is the starting vector. So you need starting vectors. And in the end, we just call here canify and then these equations are assembled. And now it's around 60 seconds assembling equations of motion. And why is this non-trivial? Because it is not only the mechanics which you can see here. It's also kind of an external model. It's a tire model which we plugged in. And the tire model gives you, I mean, if you think about tires, I know, yeah, if you think about tires, it's a non-trivial external model. And I'm working in ground vehicle dynamics. So it matters a lot to plug in to your system non-trivial external models. And the nice thing here is that it is possible to do this. Now it's integrating. We have around 27 degrees of freedom, in fact, all together. No, 28. 28. So it's okay. It's fast. It's not real time at the moment, but it's fast. So it's not bad. I mean, if you would like to make it real time, you would have to export it in a C or a Fortran code. Then it would be real time, but at this level, it is not. But it's okay. I hope so. Somehow it's a little bit slower, I don't know why, because maybe it's some resources here out. Yeah, I can tell you this. We used here a Pacheco model. And Pacheco, maybe some of you may know this, Pacheco is a professor. He's working on tire models, and he's really, really famous in this area of research. And tires, they are quite interesting to model, because it's rubber. And rubber is always difficult. So for example, if you think about rolling, rolling tires or so, you sometimes include a no slip, a no slip constraint, but the tires, this would never work, because tires produce the longitudinal and vertical force only with slip. So you have to include slip in your calculation. So now here on top of this, we have calculated in the end the eigenvalues of the Jacobian. So you can see this. And now I will go into the animation. In this step, all of these degrees of freedom has to be prepared in Cartesian coordinates, that it works. So now here, let's go. This is our car model. It looks simple, but it's okay. It is in fact more or less fixed on zero, so that it, this is a sign steering. It's a simple maneuver, sign steering. But it's nice that it's working well. You can see it's breaking in the end here on top. I can show you just this. We have also some outputs calculated for the tires, special values for the tires and put these into these graphs. Okay. So I would skip the examples and go to some future work. This is not so easy. And I would be thought about making this assembly as a persistent. This is non-trivial. Because the assemblies are not, they are quite complex objects. So ZODB, which I really like, but it's based on serialization. And very complex objects cannot be serialized this way. And I would like to do this to make it persistent because to skip the assembly, which may take a little bit of time. So graphics always some improvements to be done, model validation. And just in time compilation is another nice idea to speed up. Post-processing with pandas, maybe you know this package. Okay. So future work. And the basics until September and full vehicle simulation. So to end up with a nice full vehicle simulation, better than that one I showed to you October until December 2015. And thank you for your attention. And I would like to invite you to ask questions. Thank you very much. You talked about transitioning from SIM-Pi to NumPi to do numerical computations. Is there automated path from SIM-Pi to NumPi perhaps? At the moment I'm not aware that they are closely connected. And this is a kind of a pitfall. I think you have to, yeah, except for this Lambda-Pi you need sometimes, if you plug your numbers in, just convert your SIM-Pi matrices to NumPi matrices one by one. This is how I did it and this works fine. Thanks. Okay. That's all the time we've got. So let's have another round. Thank you. Thank you very much. Thank you. Thank you very much. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you.
|
Oliver Braun - Multibody Simulation using sympy, scipy and vpython The talk is about the implementation of multibody simulation in the scientific python world on the way to a stage usefull for engineering and educational purposes. Multibody simulation (MBS) requires two major steps: first the formulation of the specific mechanical problem. Second step is the integration of the resulting equations. For the first step we use the package sympy which is on a very advanced level to perform symbolic calculation and which supports already Lagrange's and Kane's formalism. The extensions we made are such that a complex mechanical setup can be formulated easily with several lines of python code. The functionality is analogous to well known MBS-tools, with that you can assemble bodies, joints, forces and constraints. Also external forces even in a cosimulation model can be added on top. The second step, the integration is done via ode- integrators implemented in scipy. Finally for visual validation the results are visualized with the vpython package and for further analytics with matplotlib. Conclusion: not only highly constrained pendulums with many rods and springs but also driving simulation of passenger cars an be performed with our new extension using python packages off the shelf.
|
10.5446/20176 (DOI)
|
Thank you. Thank you. Thank you. Thank you. Hello, everyone. We are incredibly excited to stand in front of you today and we are really, really grateful that so many of you came here so early in the morning to see our talk. It's hard to imagine that just a year ago we both attended our first European session in the European Union and we had a great opportunity to talk to you. We met three years ago when by a very fortunate accident we end up organizing a conference together and it was DjangoCon Europe 2013. It was famous circus edition. And we had amazing opportunity to host 450 DjangoNOTs in our lovely country. And after DjangoCon Europe, we knew that we are an amazing team and we can basically do all crazy things together and make it work. And I must say that it's a very interesting way to learn about person by setting a circus stand in the middle of horse racing track with planes flying above your head. We come from the same district of Poland called Silesia. Now we both live in London and we work as Django developers for an awesome company called Potato. We've got the same first name, the same number of letters in our full names and our last names all differ by only four of characters. Years ago we actually became members of the DjangoCon team together. Some people actually think there is only one Ola and they were sending emails to the wrong person and the most hilarious are the threads where a person is responding once to me and once to Ola without even realizing that. So a year ago here at Europe Python, we had a very interesting opportunity to find out that something very magical happened. And Django Girls was born. If you don't know what Django Girls is, you should not worry because you will find out very soon. But for now, we are going to go back in time to the imaginary world when all of the magic has started. Once upon a time in the middle of magical forest lived a little squirrel named List. She was happy and trusting and believed that the world has no limits, that nothing could ever stop her. This was super curious of all of the things. She did physics experiments with her parents and discussed newly read book with her friends. She attended art lessons, danced and solved math problems. Everything was so interesting to her. She knew that she is capable of doing anything she can dream of. And one day List discovered the forest of technology. She was amazed. At the bottom of her tiny heart, she knew this is something for her. Something she always desired, perfect combination of art and logic. Mom, look how awesome this is. She screamed with awe. Can I do the same? Can I? Of course, Sweety responded to her mom. I don't know how to do this thing. I sure you can figure it out. From that moment, List spent days and nights browsing and exploring the wonderful places in the Internet. She was trying so hard to learn all these interesting things, asking millions of questions to the strangers. Sometimes she struggled, but she also had so much fun when something finally worked as she wanted. At first, she didn't know anything. But then she can make one little thing to work, then another and another, and then she understood something and she could do things. She was truly amazed with all the possibilities computers and Internet can give someone. She knew it's something important. When the time came for List to decide about which forest she would go to learn more, she knew it needs to be the forest of technology. It's so interesting and so exciting. She couldn't wait to meet more squirrels who love technology and share her knowledge with them. List moved to another forest and was ready for a challenge. But she also felt a little bit scared. She didn't know any wonder. Her friends and family were far, far away. And it was not easy for List this time to fit in. She was incredibly excited to go for a Thursday at school. She was running, jumping, dancing through the forest. She couldn't wait to make friends, to meet other squirrels interested in technology. On her first day, she started to discover that there is not as many squirrels interested in technology as she felt. Every time she entered the room, everyone was at her. Although many were interested in technology, she realized that most of them look kind of the same. It's so much different from her, the squirrel. There were budgers everywhere. She had nothing against budgers because most of the time, they were very friendly to her. But even though she really, really liked to blend in and feel included, she couldn't find a place for herself. Whenever she met a group, someone had to ask, why did she come here? Or is she here because of her budger friend? Sometimes she also felt that she's invisible. And she started to realize that she could not fit in and make friends easily anymore. And she didn't understand why. Days passed, but she was still not her confident self. And even though she made some friends, she still felt those eyes on her every time she entered the room. And she saw that very often, others are embarrassed and don't know what to say to her. Even those friendly ones kept repeating the line, you are great, you are not like other squirrels. You are pretty good at technology for a squirrel, they complimented her with a big smile. At first, this felt special and started to believe. If everyone says that, it means it must be true, right? All of them were so clever and they knew so much about technology. They must be right about squirrels, too. So maybe if I would act more like a badger and laugh at their face, they would accept me, she wondered. For next year, she tried very, very hard and learned as much as she could. Early when she saw another squirrel who was talking about technology, some budger started to peek on them. It was so scary, so scary that the shy side of Liz is preventing her from speaking about things she loved. She was the most quiet one in the class. She was the only one who was not careful enough, she spoke her mind in public. There were always some budgers who knew better and were happy to point her out how wrong she is. She was so afraid to fight back. One day, she was asked a question by a newly met budger. She answered that correctly, but the budger said, 90% of squirrels are not good at speaking. He made his point and Liz didn't know what to say to him. Another time, she had organized some nice technology event and was asked if she was there by accident or for social reasons. She was speechless again. Why am I most proof that I'm worthy to be there, to be here to almost every single budger? She asked herself with sadness. Why are I so afraid of being there? Why am I so afraid of being able to speak freely? I can't. What's wrong with me? In this moment, she was unsure if she can continue her journey with the technology. She was clearly too quiet, not confident enough, not experienced enough. And the word around her was telling her, this is not your place. Despite of that, she managed to make some friends. She discovered some truly wonderful budgers out there. Actually, they just seemed like budgers, but if you took a closer look, you would discover the true nature. There was a kangaroo with the widest smile in the world. There was a cheerful chinchilla and cat and bear, fox and lynx and more, more wonderful mates. And they convinced her that she's capable of doing anything. And they never questioned their skills. And thanks to their support, Liz decided that maybe she will try to hang there a little bit longer. And gradually, Liz began to see that maybe there is nothing wrong with her, but in fact, there is something wrong with the world of technology. She spent days thinking why there is not as many squirrels in her beloved forest of technology. Why am I the only squirrel who likes programming? Is it possible that technology is truly not a squirrel thing? She asked the same questions in the Internet. And then, bitrates. Bitrates appeared. It's obvious. Squirrel brain are smaller than budgers. Squirrels are not meant to code. This is how nature works. Accept this fact. Yes, squirrels can program. A washing machine. Python? You're not a real programmer. Squirrels are too emotional. They can't think logically. She tried her hardest doing million projects and helping everyone around. Yet she still felt it's not enough. She felt like she needs to do so much more than any other budger to receive the same recognition. And when she tried to speak up about that, budgers did not believe her, or even worse. They accused her to spoil everything and make everyone else feel bad. Liz fell alone and hopeless for years. She knew she's just a little squirrel with absolutely no power to change anything. Liz was so afraid of all these internet bitrates. She was terrified that she will not be able to deal with them if she will be their target. And even though she loved to code, once again, she started to think that maybe this is not the best place for her. She was basically too tired to fight every single day. But then, one day, thanks to a series of fortunate events, she met another squirrel. They had so much in common that they even shared the same name. They became friends. And Liz finally had someone who shared with her the same feelings and experiences. She realized that she's not the only squirrel out there that feels alone, intimidated, and scared living in a budger's world. And it's not only her who had to prove herself over and over again. One day, she said to her friend, imagine how awesome it would be to know even more squirrels who code. Can we somehow convince others to join us and show them that programming is fun? They both liked the idea, and in a short time, they came up with a crazy plan of teaching squirrels how to code. They wanted to share their passion to programming and get squirrels excited about all the possibilities that technology gives them. They had no idea how to do it, but they couldn't sleep out of excitement. They told all their friends about their idea, and they loved it. With their support and many, many others who just came and joined them, the crazy idea started to come to our reality. And a year later, thousands of squirrels have learned how to code just because in the right time, in the right place, two little squirrels met and got excited about some idea. And today, Liz is standing right here on this very large stage to tell you a magical story, how one dream turned into something bigger, something that is beyond one single squirrel or two of them. This dream is called Django Girls. A year ago at Europe Python 2014 in Berlin, we organized the very first Python and Django workshop for complete beginners, and we never wrote a single line of code in their lives, except all of the more women. We started this quite unexpectedly just two months before the conference, and we only planned this to do this one single time because with our daily jobs and everything we've been already doing for the community, our schedules were already really tight. What we didn't expect is that the next year we'll turn our lives upside down, and we will have amazing opportunity to teach Python and Django to 1,600 women, literally all over the world. Thank you. So it all started with an email. Just two months before Europe Python, I met with one of the organizers of the conference in Berlin. After that breakfast, I came up with the idea to make Europe Python more diverse and inclusive. The very same day, I sent an email to Ola, hey, do you want to organize a workshop for women at Europe Python this year? Back then, I also wrote, I don't think it's going to be a lot of work. We didn't need to find 20 attendees and six coaches and given the size of Europe Python to be hard. I don't think I could be any more wrong on that. We jumped on board with excitement and started working on this immediately, on the same day. It was just the two of us and we had less than two months to make that happen. And honestly, most of the time we had absolutely no idea what we were doing. And very soon we found out that this very first workshop were going to take a little bit more time and work than we first anticipated. For example, we came up with the idea of providing financial aid for attendees that meant we need to find slightly more sponsors, then plan to review more applications from attendees and find a way to transfer money to them. It also turned out that we don't like any tutorials that are out there. We are hugely dissatisfied with the status quo and that every tutorial available assumed that you are a programmer, that you programmed before, you know what a web framework is, you know what's a server, a URL, a host, IP and so on and so on. And honestly, people just don't know these things. So we ended up writing our own tutorial, which I think ended up as a small book, or it was like 90 pages of text. At the same time, we also tried to promote the event everywhere possible. We set up the website, we spammed almost 70 universities around Europe. We also passed information through all the social media. We could. At the very end, we also spent more than 20 hours just reading 300 applications for the workshop and we convinced Europe to give us a bigger room so we could feed 45 people instead of just 20. And now, one year later, the tutorial we written has been read by almost 90,000 people. That's one fourth of population of the world. And this number is unbelievable. Today we are also proud to say that during the last year, 1,646 women learned about Python and Django at our workshops. Many, many more did this in home just by reading our free tutorial. Django girls madness has been spread to six continents and 34 countries. Seventy workshops have been planned in all different parts of the world. From Canada to Australia, Django girls has been everywhere. And maybe except Antarctica. We are working on that too. Some events even happen in places we didn't know exist. We grew so big, we recently had a certain legal entity. Django girls foundation has been formed just a month ago. And if someone would say that to us one year ago, that this is what's going to happen, we would never believe them. We also registered it as a fun fact. We registered it under a second name, Django girls foundation of awesome cupcakes and emoji. This is actually quite handy because the acronym of that is dog face, which means that you can represent Django girls foundation in just a single emoji. And we think it's very cool. And all of this was possible thanks to hundreds of organizers, tutorial contributors and an army of coaches who dedicated their free time to share their passion to programming with others. And we also came up to the point where the amount of work to coordinate all of the volunteers was basically too huge to two of us. And we decided that we couldn't do that alone. And not getting run out. And we asked for help for amazing people. Anya, Kasia, Baptiste and Jeff. And since then they are helping us stay sane. But it's still not enough. And we realized that we need to make some big decision to keep it going. That is why Django girls foundation is now hiring for our first position. We are looking for a ambassador of awesomeness who will help us make sure that we are always doing the best work we can. Honestly, it was a pretty huge deal for us. And this decision was both extremely terrifying and incredibly exciting. And we still don't know what we are really doing. And we are making up a lot of things as we go. But we believe that this is the right decision to make sure that Django girls is sustainable and can grow beyond this. And most importantly, we are doing this because we don't want to stop here. And we know that this is only the beginning. And we believe that we can make even bigger impact. And today we would like to share one of our huge plans with you. I'm getting super nervous whenever all starts this last paragraph. But well, here it is. So it all started with the tutorial. It turned out that we really enjoyed writing it and it also turned out that to our surprise, apparently we did a pretty good job and people loved it. It's one of the most recommended Django tutorials that currently exist. People send us tweets and emails that they did many tutorials before Django girls. But the one we did is the one that finally helped them understand what Django is all about. As you've seen just a moment ago, the tutorial was read by almost 100,000 people. And we got many requests for more advanced tutorials or tutorials covering different parts of computer science. And when we were writing the tutorial, we had this feeling that we were very limited by the form of the workshop. You can only teach so much in one day. So basically we skipped through the material and just explained things that are absolutely necessary. But there are so many wonderful and beautiful things you can learn and you can teach, both the internet and computers. We want to make learning computer science fun, beautiful and exciting, to empower people to continue journey on their own. We want to make accessible for everyone, easy to understand and beautiful, because we believe that the tiny little details matter. So without further ado, we would like to announce, now I'm nervous, that we're in the process of writing a book. Now there is no going back all alone. And this book is called Ye Python and we're now working to make that happen. So it's going to be very much in the Django girl's spirit, as you can see. With a good dose of emoji, beautiful illustrations and funny little quirks and storytelling. We believe that the form of a book will let us expand the material to cover and include important things we missed in the tutorial. And will also allow us to make an even bigger impact on a broader audience. We want every person, every woman out there who enters the bookstore and sees this book among other technical books and immediately think, hey, this one is different. The same way they think about Django girls. This is the book that you always wanted to have when your family or a member or a friend asked you if they can learn how to program, if you can explain them what you're doing the whole day. Or if you can just have them understand computers. Ye Python is all about the exciting, incredible and beautiful word of technology. This book is still a tutorial that will teach you how to do Django website. But it will also include chapters where we explain how computers work, exercises which you can connect dots to see how Django goes from request to response. Where you can learn about open source community and people who do that. Ye Python is going to be beautifully crafted book that's going to introduce people to technology showing its best parts, the community and the people. You want to tell people a story about our favorite things about the Internet so they can fell in love with it the same way we did. We believe it will be very unique. I hope so. And we literally can't wait to show it to you. We're just at the very, very beginning of this huge, huge challenge. We just started to write first chapters. But if you want to follow our journey, you can visit yepyton.com today and sign up for our newsletters and we plan to share stories from behind the curtains. So when we think about the book now, it reminds us how excited we were just the years ago to announce Django girls to the world for the first time. We started with a very small idea with just a team of two. And now there are thousands of people involved and Django girls turned it into something much bigger than just the two of us. None of that was planned. None was expected. I'm not sure what Ola would reply to me when I would tell her a year ago in this email that this is what's going to actually happen. I have no idea. I would probably think twice. But here we are now, a year later, and we couldn't possibly be more grateful for everything that happened during last year. And if there is one thing that we knew from the very beginning and that didn't change during the last year, is that we wanted to do Django girls our way. We made one rule and one rule only, no matter what, always go the extra mile for people. I think you can see that in almost everything we do. You can see it in all these little wonderful details at our workshop. We truly believe it makes a real difference and that people notice it since the very first contact with us. For example, when you send us an email to our Django girls account, instead of boring best regards, you will have signature hugs, cupcakes, and sunsigns or something like that. And we always try to be as friendly towards people as possible. And they are friendly to us in return. I think this signature was actually started by Mikey, who is sitting here. The spirit of the Django girls is carried through your whole experience. You can tell anything about our workshops, but that they are ordinary. Instead of boring classrooms, we decorate the room with balloons and table clothes and flowers. We do silly things like photo booths and cupcake tasting or even yoga classes during the programming workshop. So everyone feels like it's something extraordinary fun and exciting. And Django girls workshop will provide you with a very unique experience. We start with forming very small groups of free attendees per one coach during the workshop and ends with group hugs. We believe that every person matters. And we want to make something unforgettable for people. I still remember after a workshop in Warsaw where you received this long letter from one of the attendees and she said to us that we gave her the best day she had this year. I think that was quite amazing. And to ensure that organizers, although the words provide the same kind of experience to their attendees, we open sourced basically everything. It took us weeks to write down every single little step and document it. But that also let us achieve one important thing, making Django girls, making sure that Django girls can be bigger than just two of us. We won't be around forever, but Django girls can shine and strive even without us. So every person out there who is not sure what it takes to organize the workshop can simply go and see one of our manuals at our website, read it and do that on their own. I think it removes one of the biggest barriers for people who never did anything like that because there is every single step from finding sponsor to providing balloons or things like that, so there is everything there. Most importantly, we are doing all of that with a huge amount of enthusiasm and excitement. We always look on the right side of things and we love laughing and we try to keep Django girls fun and friendly for everyone. For us too. We got these posters with huge you go girl. We buy tacky gold balloons that take up most of the room. And we are always jumping off excitement because we know it's contagious and it's going to spread to everyone in the room. When me and Ola are excited about something, it's going to get passed to organizers who are going to get excited and they are going to pass that on out on these two. And I think this power of enthusiasm is really huge and I think this is what helped us involve so many people in Django girls in such a short time. But as you probably suspect, all of these things don't happen on their own. And making that happen is a huge, huge amount of work, time and energy. We really care about people we work with and people who attend our workshop. And because we care, we put a lot of feelings, a lot of empathy. And we try to make sure that everyone is happy. It's very, very hard. Every labor of love comes with a huge emotional cost. We always think like we don't do enough yet and because there is always something to do, we very often guilt trip ourselves for not being productive enough. And sometimes it really feels like there is no end to this. And the days come when we are simply just tired of working two full-time jobs. It's frustrating at times where it turns out that your free time now consists of trying to catch up with your inbox or bookkeeping or wasting energy to resolve personal conflicts of people in your community. It's very often not the glamorous work, but yet the work we spend our free time on because we believe it matters. And there is one thing in particular that keeps us going. The stories of amazing women who change their lives because of Jungle Girls. Like the story of Dory. Dory is actually sitting here somewhere in this room, attending her second Europe item. So last year she attended the first workshop at Europe item and she spent half of the day just trying to start Jungle Server that didn't want to play nice with her Hungarian keyboard. And there was like six coaches around her. We thought it was like the worst experience ever. She's going to give up, hate it and never do this again. And today Dory works as Jungle Developer in Budapest where she also co-organizes a monthly meetup for Python developers. She organized two Jungle Girls events teaching 60 other women how to call. Dory also coached on many Jungle Girls events all over Europe. A very similar story happened to Agapta. She attended Jungle Girls together with Dory and now she works as Python developer in Brotswaf. She's also involved in Brotswaf Python meetups and also she put together two amazing Jungle Girls workshop there. Josie was only 13 when she attended workshop with her mom. And after that she got on Europe item stage in front of more than thousands of people. And with no fear she gave a lightning talk about her experience and her plan to organize Pixie Dust, a programming workshop for girls her age. This workshop actually happened last fall in Zakarpak, Croatia. Whenever I open my inbox and I see 15 new emails to respond to and I feel like I just want to die. I think of stories like the story of Hustina. Hustina applied as an attendee but we convinced her to become a coach and since then she coached at five Jungle Girls events and organized them twice. Linda organized workshop in Nairobi in Kenya just two months after our pilot workshop in Berlin. She blazed a trail for many other women who organized workshops in Africa after that. When I have a work day I also remember about Lucy who started learning programming a year ago. She attended Jungle Girls workshop but this didn't stop her from giving a wonderful talk at JungleCon Europe this year and organizing Jungle Girls in Paris. Ana is running Jungle Story on our blog. She's also involved in JungleCon US DSF and she's in the PSF board of directors. Linda from attendee became a serial Jungle Girls coach and organizer. She's always going around Europe visiting her Jungle Girls friends and teaching people how to program everywhere. All of these stories are only for women who attended the very first workshop that happened at Eurobite. We had 45 more workshops like that since then. We could easily fill this whole talk with two stories like that. Knowing that women who attend Jungle Girls workshops actually change their lives is what keeps us going and makes us believe that countless hours, energy and feelings we spent on Jungle Girls are simply worth it. But absolutely none of this would happen if it wasn't for these amazing people, amazing girls and the generous Python community from the very beginning, the amount of support, help and love we received from strangers all over the world left us completely speechless. And there is absolutely no way that Jungle Girls could be this big and wonderful if it wasn't for hundreds of people who said, hey, how can I help and immediately jump on board? So from this place, we would like to say a huge thank you for absolutely every person who helped us along the way. We are grateful, humbled and we want you to know that Jungle Girls is your success too. There wouldn't be Jungle Girls without you. So thank you for putting all of your faith in us. Thank you for all the words of support, advice and encouragement and all of the hacks. So thank you for time, energy and commitment. We couldn't possibly be more grateful. Starting Jungle Girls made us realize how many amazing, badass and successful women who go to Jungle are out there. We knew only a few before it all started and now we can count them almost in hundreds. And the most surprising fact is how many of them just needed this tiny little spark to start doing something amazing for others. And there is so many people out there who are capable of doing amazing and big things. Yet they are waiting for some imaginary permission and we couldn't imagine that doing something so simple as just saying, hey, here's how we did it. You can do it if you want. And this was all what's needed to empower them to start. Everyone who works making Jungle Girls happen doesn't expect anything in return. We do it because we believe that together we can actually change something. In fact, our dream is for Jungle Girls to not even exist in five years. If we have no reason to be here in five, ten, maybe 15 years, it means we achieved everything we wanted. That's I think our ultimate goal of Jungle Girls, to not be needed anymore. And the biggest lesson we learned last year is that doing good things, helping others and not expecting anything in return creates real magic. It turns out that one way or another, the universe is going to find a way to pay it all back to you. And the number of crazy, amazing, wonderful things that happened to both of us in our personal and professional lives during the last year is simply unbelievable. We can't even begin to explain how exhausting but wonderfully rewarding this year has been for both of us. In fact, there is no way we would stand here today in front of you opening EuroPyton 2015 if it wasn't for Jungle Girls. So for a long time, we thought that there is absolutely no way that we can change or influence anything. You worked believing that this is how it is. But in the end, it turns out that all it takes are two little squirrels with a big, big dream and a generous, wonderful community to make something amazing happen. So whenever you think you are just a tiny little drop in the ocean and that all of the things you do don't matter, just like our list up. And take a minute, look around. You are not alone. You are among friends. Thank you very much. Thank you. Thank you. So there is one more thing. So can you still hear me? Yes. I was supposed to come back about this, but all of the illustrations have been actually hand drawn by her. So let's give her maybe one word of applause for that. Great. Thank you so much for this amazing talk. Unfortunately for the 15, 20 minutes of standing ovation, we are a bit late. So maybe have time for a couple of questions. And then we'll have the next talk starting five minutes late so we can deal a little bit with the delay. Questions or comments? Okay. Thank you so much. Thank you. Thank you, all of us. Thank you.
|
Ola Sitarska/Ola Sendecka - Keynote: It's Dangerous To Go Alone, Take This: The Power of a Community In this keynote, Ola and Ola will take you on a fantastic journey to the magical world of little Liz, who is totally enchanted by technology. The story of Liz will show that with a little bit of magic, curiosity, courage and hard work, you can defeat all the obstacles standing in your way. You'll discover with her that making big and scary things is easier when you're not doing them alone. Because sometimes, one magical spell, the helpful hand of a friend or this shiny sparkle is all it takes to make a dent in one's universe.
|
10.5446/20174 (DOI)
|
Just to say that we are not developers. We are organizing. We are organizing in 5-B-C-N, but not really Python developers. I'm an instrumentation engineer in the CIC in Spain, Russia, and Russia. I'm working in marine technology, but I use Python everywhere. I can. I am an engineer working at DevEx.com, media site about NGO and international development. I'm coming more from the system administration part of the thing. I like to write programs since I was a kid, so I'm kind of a wannabe developer, not very good. I'm a little bit worse than a system administrator, but trying to balance both. By the way, if you ever come to visit Barcelona or something, you can check in 5-B-C-N.org about our activities, or meetups, or dojos, or whatever. So if you are around, you can feel free to come and give a salute. So we are going to give a very introductory talk on what we have been doing for the last nine months, one year, because we started with doing some dojos. Is there anyone who knows what the current dojo is right now? With a few. We will start with a brief story of 5-B-B-C-N dojos, how we started doing them, what we have been doing a little bit. Then we will give a small introduction to what a cutting dojo is and what related concepts are with it. We will do a small recap on the sessions we own, just choosing some of them, not all, because it could be quite long. Then we will take a look back on to see what things we learn doing dojos, about what you need to prepare, what things you need to be aware of. Then we will talk about the dojos sessions we are going to be running at the Euro-Python this year. We will give a very small set of ideas to start with your community. Actually, it's a pair of plans. 5-B-C-N dojos started with, actually, we're not related to 5-B-C-N group before. We started with a small group of system administrators, people who was interested in learning some programming stuff. Then we decided to start using Python because it looked like a very low-varyer and two-level language, easy to learn and start working with. The big problem we had is that we did very frequent sessions. We barely did one every two months. After having these sessions, or during having these sessions in another conference, we were committing with some Python members about what we were doing. They were quite interested in setting up these sessions for the whole Python-Bison community. With them, we started running these sessions monthly, determining the dates every three months and having different kinds of sessions. We felt that there were a lot of different skill levels into the people that were kind of interested in having them. We are going to talk about what has passed for the last nine months. For those who don't know, coding dojos is a hands-on session. The first very important target or goal of these sessions is to have fun with code colleagues and people with similar interests. The second is learn from others. Usually it comes that there is someone attending that says, I'm very new by it. I don't know anything about it. I barely use Python. There are other people that master some subjects on Python or even on just the kind of audience we are proposing. For the people who know more about Python or master some subjects, it's to share the individual knowledge. We get that because we try to use several techniques that I'm going to explain later. Also, we use the coding dojos to practice new technical tools which we don't use to do them in our daily jobs. For example, testing or writing tests or doing TDD or such on. Then the related concepts, more important to be aware of when you are attending or organizing a dojo, are those of the here. The first one is the Cata. The Cata is the trivial exercise. We are going to state for the session to complete. It's trivial because it's usually not very difficult. And it's trivial because it doesn't matter if we come to a solution. We are not trying to cure cancer or to find anything big. And it doesn't matter if individually or within the team we form, we find a solution. What we are trying is to learn from what others are doing within the team or with other teams. We are going to, usually we use paper ramming also because this is a very good technique. We are going to use paper ramming, especially when you use paper ramming with TDD, for example. It can be kind of fun. Also we do iterative development. So when we prepare the Cata, we tend to have some preparation before in order to have the steps, the iterations we are going to have. So we can guide people attending. These iterations are done within a time box, the structure. We use the Pamodoro technique, which is just having 25 to 45 minutes to have every iteration. And, well, yeah, we tend to change the times and it's kind of chaotic sometimes, but it's important to try to enforce these techniques. Also we put a very strong emphasis into differentiating testing versus TDD. This doesn't mean that we need to write tests for everything. This doesn't mean that we use TDD in all the sessions. There are some sessions where we enforce much more TDD. There are some sessions where we only want to see some tests to check that everything works. And even we have some sessions that we are not even looking at the tests in case they are. And also on each session, after all the iterations, we have, well, it should be small. Usually it is not. Retrospective session with all the attendants. So people can explain what they like, what they don't like, what they learn, and what they are taking off from the session. So we can learn what we are doing good, what we are doing bad, and we need to improve, or what we should be taking care of. Sometimes it's very stupid. Sometimes it's just format, the air condition of the room, or the quality of the projector, or something like that. But sometimes people explain, yeah, this kind of problem was very nice, but we hadn't a statement before, or we hadn't a statement to check where we were going. Things like that can be helpful to improve the sessions. So this is a small recap on three of the most cool or interesting sessions we had. Some months ago we had the Pytest versus Unitest session. It was intended to be a hands-on on Pytest with comparing with the Unitest session on the same cat, very simple and small. One of the big problems we had with Pytest on versus Unitest session was that we didn't have so much knowledge of Pytest. We just had the very introductory knowledge. So it was very hard to answer certain doubts that people get. We were trying to, when we did that session, we were trying to provide something more interesting for people with some experience that are not so interested into starting with 3D, but learning about what they can get from using other tools. And this was one of the problems we had there. In the CilinQ and Python session, which was based on a Peter Hinge workshop at EuroPython 2014 last year, it was quite interesting. We had, I think, 18 people attending. It was quite stress, but it was very, very nice. The feedback was very good. So we are quite happy. Maybe we should have some more time to learn more about CilinQ before. We used to be kind of the procrastinating people. And, well, the feedback was very good. We had a lot of people working with engineering, with reinvestigating engineering also and so on. And, yeah, sorry. Then we had another session with mocking to introduce the session of the mock library, which it's not re-included, but it's very useful to find out, to fake things happening around our code in the architecture. This session was kind of bad, at least, because we had it very prepared. This was one of the first new kind of sessions we had. We went with a supposed cat that we invented, which we haven't tried. We will see later. This is one of the very first things you need to do if you want to organize session sessions. Okay. And a little bit of a resume about what we have learned doing this, because, yeah, in these sessions we also learn about people who come. One, the first seems very elementary, but you have to practice your cat before to do other people. And usually we find, we try to think what could be the different solution that people we are going to find just to guide the session to be fun and to be fast, and not to be worried about other people. Obviously, if you guide the session, you have to have a specific knowledge, but you don't have to be a master. You only have to have a main idea, and usually people who come to our dojos, when we guide, usually the solution that they show are almost better than us, or have some points that we, all the people in the session learn something, yeah. Then, okay, if you introduce a new library, probably you have to take care to show the basics of this library. You have to take time to, all the people start with the same knowledge, and to guide the session is really important to have learning skills, to learn teaching skills and these kind of things. I think it's the most important to feel that, to make feel comfortable all the people in the room, and that it's useful, you know? Okay, and then it's open to other programs that come from other languages, but we try to focus on Pythagonic solutions, of course. Okay, and a little bit about the session we are going to do in EuroPython, okay, on Fricing Sessions, we are going to focus, to explain a little bit, baby steps and TDD how it works, and, yeah, practice per programming. I don't know if you ever program it in pairs, but I think it's quite fun, not just to be three hours alone programming, programming alone, okay, sorry. And it's just a game, but programming usually appears, program the test and the other program, the program, the code to solve, and then change the roles, and I think it's, you learn how to do the test, how to program, and one of the important parts is, yeah, if you have a Mac and a PC who manage this session, which editor, which, and sometimes you learn new things sharing with your partner, even more than in the group. Okay, and the Saturday session is going to be a little bit complex, but, okay, these two sessions are related to the typical algorithmic problems, which are going to be classical algorithmic problems, okay, just to practice TDD. Okay, Friday is more for Qs and TDD, Saturday is just texting, and in a more complex scatter, and probably you are going to, if you come, you are going to practice some refactoring about the problem, okay? The seats are really, really limited because we, our sessions in Barcelona are in a small group, and we will, we won't, want to be very personal in every group, know what it's doing, and we think that 20, 25 people in the room, it's how to manage, okay? Then, come soon, we are going to publish in Twitter, or, I don't know, in that way, we are going to make an event right, you are going to sing it up, and, okay, I think, you are going to sing it up, and I want to see you in these sessions. Okay, a little bit about, and, if you are interested in Python Dojos, I think it's a good, it's a good way to spice your community. I'm a Python organizer, and, thanks, two years ago, about, and I feel that, okay, you organized meetups and conference, what is a little bit called, okay? People come, people sit here, and go away, and think it's, with PyDogs, we have, we have a community that you know what, what is developing, what, it's like a family now, we do PyDreamers and these kind of things. I think it's a relationship more, okay, more, it's more wire, okay? I think, if you are interested, if you want to ask, because how to, how to start the other coding Dojo in your community, or if you want some examples or some experience, if you want to share some experience with us, we are here all the week, and we recommend to read this, this book about coding Dojos handbook, my Emily Vange, and of course, you can contact us with, using Twitter, and in person here these days. Any question? Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Do you just go and you ask people what they want to do? Yeah, we know a lot about people who usually come, people repeat, that's an important point. But yeah, we ask people, new libraries, if you want to try new libraries, we try different formats. Yeah, it depends on the people. For this, for the reason, we think that the feedback that we, that we get at the sessions, are very, really, really important for us. Just to know that it's interesting or not, or what to change in the next session. Ah, about the, about the catapeperation, it also depends how much time do you need, because sometimes you can have a very complex cat, and then you need like more than a week to prepare it. That doesn't mean that you need to be careful, because if you prepare the catatür in one week, you can easily have a catatür as long as at the time you have for running the session. But if it's simple, you can have it easily in 23, one hour. More questions? I've got ludes, actually. I've organized the London Pies and Coat. Okay, yeah. We asked, actually, we talked with someone, we talked with Rashid about that. Yes. And we asked, what you were doing, because you were not doing so much today, right? Yeah, I think we're doing quite differently, actually. How does yours work? You have only two persons at the front, and everyone is looking? No, you mean like in a Randoly? Yeah. The Randoly format is that some two people's is working and the rest are looking. No, we have been thinking about trying out different kinds of sessions for the dojos. This is one of the ones we have been considering, but we never did one of those. What we did is to start with pairing people, then commenting the Cata, explaining what the details that the Cata can have for what complex and things, and then with the people working in pairs, start working with that. If they got a simple enough to run just one iteration, then we try to enforce people to switch the teams so they can share with more people in the same session. If they got this more complex, then we just stand with that and go on with the complex iteration. Usually, then we aim the people to share the code and explain the code how it works. Okay, now how many iterations are you doing roughly? Three, four. Three, four in the best cases. She said before, there's always someone attending that masters much more than you. So we had, for example, one guy with one Cata, the guy just solved the first iteration, the second iteration, and the third iteration using the first one. So, yeah, you're going to have to have this kind of... You always have to think more of stuff, more fun the time you have because of the people who are coming. Not the people who try to enjoy all the people. We also started doing something by the end of the year. It was, we prepare the sessions every three months, every quarter. So we use... We do a very simple Cata in the first month. In the second month, we do a more complex Cata with less TDD or less test focused. And then we have a third session with, for example, a CELIN queue. That's a specific library. Can I ask more questions? Yes, please. Okay. How many of you are coming every dojo? It has been... depending also on the lack of STVDs and so on, it's very important, especially if the Barca football team is having a match. We have to... we choose the dates in advance. Yeah, also... Really, three months in advance, sometimes we don't know. But we used to be like 10, 15 people. Okay. So, do you provide beer and pizza or something like this? Well, we had some kind of things. Yeah, cookies. Sometimes I cook pizza and those kind of things. Yeah, but we're planning to talk with some local brewery to get beer and something. And we're paying for this then. We're paying for the cookies and the beers and... And the location. Yeah, we are within the pirates on our hook, so everything comes together, actually. Okay, and is it free to attend? Yes. Well, we have only the... We are publishing the album right with 25 seats. Yes, much more people could be kind of difficult to manage. Thank you very much. I mean... Don't play it out. Okay. Right, now there will be a coffee break. We'll start at quarter to five sharp. Don't forget that you can rate this talk at the guideline. Get booked, sorry. And that's it. See you after the break. Thank you for coming. Thank you.
|
Núria Pujol/Ignasi Fosch - What dojos are and how we run them at pyBCN Coding dojos are a very good way to share coding knowledge among members in a community, and, at the same time, introduce people into the language and community. Sometimes, though, the typical approach to set coding dojos may prevent expert coders to join the session. This is the story of the pyBCN's dojos, so far.
|
10.5446/20171 (DOI)
|
So, as he said, my name is Naomi Cedar. I've been in the Python community for longer maybe than I would care to admit. I am at the moment serving as one of the board of directors for the Python Software Foundation. I am refusing to be rattled by flashing lights in my face. I encourage you to do the same. And among other things, I did found it's the first trans-themed hack day in Europe, transcode. So, my, I guess what I would say, my somewhat unusual, although not completely unique, experiences I talked about a little bit here at EuroPython last year. And my experiences have led me to think a bit about diversity and advocacy and things like that. Before I go any further though, I need to give you a commercial. In a very short time, right after lightning talks today, in the Barilla room, one room, we will be having a meeting of the Python Software Foundation and everyone is invited to attend. For a while I just, you know, foolishly put members meeting on it, but it's really for everyone. We want to update you on the new membership model that we have been putting in place for the Python Software Foundation over the past two years. A short story is that anyone can be a member and more of you, I think, than you would expect are actually qualified to be voting members. So, we would like to tell you about that. Work groups, which are a way that we hope will allow more people to be involved. And, as we always close, such things more. So, I will see you there, I hope. Okay, so to begin, we know that diversity is a good thing. I'm going to take that as an article of faith. If you strongly disagree, then maybe you better go find another talk because this is the assumption. In fact, studies have shown diverse teams solve complicated problems better, diverse teams are more adaptable, diverse teams are more creative, and these studies have been around literally for as long as I have been alive. And that is longer than almost all of you have been alive. Okay, a long time. We also know that in our particular sector, we have a shortage of talent. We need more people. Recruiting, whether it's in my experience in Chicago or in London, is kind of a nightmare to find people. And it's only going to get worse. We need more people, not fewer people. And finally, this probably should not need to be said, but I feel an obligation to say it in any case. We have, I would submit to you in the tech world, a good thing. We have interesting work to do that allows us to make a good living. Not sharing this, it strikes me, is just wrong. Yet, I went looking, this is on Google UK, on the image search for a programmer. I'll give you a second to appreciate that. I'll say it in a little comment here. We've got 15 images, got one duplicate. We've got one rather abstract, although still kind of much looking character there. We've got a cartoon or two. And out of those 15 images, you'll notice that we have one that appears to be a woman and we can't see her face. Okay, we have one person who is black. And if you look closely, he looks both puzzled and angry. Okay, the rest of them, I don't think I need to say anything about. So, to put this another way, Twitter in the US, after a campaign to improve diversity announced not long ago, that they have now 49 black engineers out of their engineering staff that's less than 2%. As Lynn mentioned in her talk earlier today, women are leaving the tech world in mid-career. And this comes from, I think, the same article that Lynn had referenced. They are leaving in droves. They are leaving more than they leave other things. They are leaving far more than men do. So, in other words, we don't just have a pipeline problem. In fact, if you look at the numbers of graduates in computer science, and the data that I could find was US specific, but I have a feeling it's pretty general. The percentages of minorities, women, et cetera, graduating from computer science programs are not matched by the percentages being hired. But, again, it's not just a pipeline problem. On Tuesday, how many of you saw this hashtag trend? Oh, that's a shame. In my circles, it was a big deal. I started trending late on Tuesday. I strongly encourage you to look it up. Real diversity numbers started trending late Tuesday and on through Wednesday. And basically, this was started asking the tech world for let's have some real diversity numbers. Let's find out how many people who are, you know, depending upon gender, age, race, economic status, et cetera. There were things posted like, yes, companies, please tell us how many autistic people had to quit your wonderfully open-planned company because they couldn't stand the sensory overload and do their work. Please tell us how much more work a black female needs to do in your company just to get the same amount of recognition as a white male. Things like this, it's a provocative thing to read. Again, I strongly encourage you to do that. And I need to stop here and say that I wouldn't go along with what Lynn said earlier. The Python community has done a lot. There's very, very much that I'm proud of. Honestly, in many ways, the fact that I'm standing here is a testimony to that. The fact that I am on the board, all of those things. But I do also have to say that when I'm at a conference like this, not just this one, but like on U.S. other ones, I tend to ask myself, who am I not seeing here? And I'm not going to tell you who I'm not seeing here. I'm instead going to invite you, maybe not now, maybe later, whenever, to think a little bit about who it is you don't see at these things. And then maybe the next thought would be, what do we need to do so that we do start seeing those people? So diversity that is getting lots of different people involved is a hard problem. Inclusion, that is making them feel like they really belong so that they stay involved is a harder problem. But it is everybody's problem. And I don't have clear answers about this, but as I get into this, I want it to be clear that I am not talking in this talk about them. And I'm not talking really about you. I'm talking about me as well here. These are things that I particularly also think about because I think we need to be really clear here that being part of a marginalized group, being in some way, quote, oppressed, does not excuse you, it does not give you a free pass for everybody else. Okay, the fact that in the tech world I am an old trans woman that's three strikes already, does not excuse me from worrying about questions, race, economic status, disability, any of those other things. So I'm really, these are things that I have been thinking about as well. These are things that I try to watch myself about here. So I'm not trying to externalize this as me lecturing you or as me pointing a finger at them. This is us that I am talking about. So if what we're doing isn't working, why do we keep doing it? This was what prompted this whole talk. And really, you know, it sort of then led me to think about, you know, as far as technical people, we like to solve problems. We do things. And I think, you know, TDD is kind of an example of that. You know, it's everybody now says, yeah, this is kind of a good idea, unless you actually change your behavior and write the tests, it doesn't work, of course, if somebody says, oh, yeah, TDD is wonderful, but, you know, we just aren't having much success with it. Well, why is that? Well, we haven't actually written any tests yet. Nobody's going to buy that. Okay. So this is what I'm thinking of. And then honestly, just so that I could get a provocative title, it occurred to me anti patterns is what we're talking about here. All of these things that seduce down the wrong path when we're trying to do something, whether that's encoding, or, you know, perhaps it's in management styles, taking care of people. If you actually go look up anti patterns on Wikipedia, they have pages, palm pages, palm pages of the ways that we can mess ourselves up. So these are the things that I'm going to talk about. So first I'd like to mention is, and by the way, this is just my sort of lost the cup grouping of these. If you want to debate the niceties of how these should be classified, then that's fine. I hope you find somebody who wants to debate. So the first one I said called denying the problem or denying that there is a problem. And you do still see this here. And I mean, this I think is the meritocracy thing was like, well, we would hire good people. There just aren't a good people of type X, whatever that might be. And in fact, it seems to me, and I know that for some people, meritocracy is a very sacred idea, but it seems to me meritocracy is a way for the people who are successful and empower to make a story that justifies why they're successful and empower. And similarly, there are people that say, well, I don't see a problem with sexism. You will notice that the people who never see sexism are all male. I'm just observing that you can draw your own conclusion. Or similarly, there's the whole well, but people of type X, whether it's women or minorities or whatever it is, they don't want to do this kind of work. So that's why it's there. So all of these actually are sort of ways to kind of shoo the problem away and say that it doesn't even really exist. And of course, then the last one is kind of a generic description of what a lot of the industry does. That is say, yes, yes, yes, we want it to be diverse. And then just hope nobody notices the actual numbers. So again, refer back to the real diversity numbers hashtag. Similarly, you can deny it. You can say yes, yes, it's a horrible problem, but X, Y and Z is not a fix. There is no fix. In fact, this is also quite common. So if you, the people who blame the pipeline are in fact saying, well, there's nothing we can do. Somebody else has screwed up and now it's just all, you know, we can't fix that. Or it's the education system. Or, you know, just for whatever reason, there's nobody here that even wants this job. It's a problem. And again, this is a way of saying, well, I don't have to do anything about it. This one is perhaps a bit slippery, but I think another thing that we all fall into is assuming that everyone is like ourselves. They aren't in that in some way that we think important like ourselves. And they can't be good at what we do. Right? I mean, we can't have coders who don't want to play ping pong or something like that. I mean, that just wouldn't be right because all of us here in this group play ping pong. No offense, but I hate ping pong. So, you know, it's that sort of thing. So we start talking about then culture fit or I was hired by doing this long grueling whiteboarding programming exercise. Therefore, anybody who is good at programming must be able to do the same thing. All of these sorts of things. This also actually ignores a couple of things that I would like to revisit sometime at another time. And that's the notions of imposter syndrome and stereotype threat. These terms are thrown around a lot. So I'm going to sort of assume that you've heard of them before. Imposter syndrome is, of course, very prevalent amongst very, you know, highly skilled people who look around them, see everybody else and decide that they themselves must be a fraud while everyone else must be doing wonderfully. And it tends to rob us of a lot of good efforts by people who say, oh my God, I don't think I can even try to do this because I'm going to be caught out as a fake. It tends to cause a lot of anxiety. It's a problem. It's not a problem specific to women, but I think it is a problem endemic with women in tech. A stereotype threat, if you haven't read this one, this is the one where if you tell, say, a room full of people about to take a math class, maybe they, you know, mixed composition, you say, oh, by the way, black people will almost always do poorly at math. This very fact will make the black people start thinking, oh crap, nope, I'm not going to fall into that stereotype. I'm not going to do it. And their performance goes down by the very fact that they're now worrying about that rather than what they were supposed to do. And this happens with almost any population. Again, women and minorities particularly get struck by that. If you have processes or ways of sort of bringing people in that don't take this into account, then you're not going to have very much success would be my name. I call this one rigging the game because it is basically these are things that will make it impossible for somebody who is not in the system to really succeed in the system. And microaggressions, this is the first one. And how many of you have heard the term microaggression? Pretty much everybody. Good. How many of you committed one? All lying. Thank you. So microaggressions are things that are done that are micro. They're not necessarily big, but they sort of eat away at you. My particular personal favorite in my life is that I've had a couple of friends tell me, you know, you're really pretty impressive. You're not totally crazy like the other trans people I know. Okay, so if you're thinking about, well, that's a compliment, right? I mean, if you call people, they say, what? I was trying to be nice. In effect, what you're saying is that I probably really am crazy. I'm just good at hiding it. Or the classic one that some women will have heard, sadly I have not, is you're too pretty to be good at coding, right? This manages to kind of tie your appearance to your talent, kind of to the detriment of your talent and things like that. There's a lot of these things out there. And in fact, they happen a lot and they tend to wear people down. My favorite metaphor for this kind of process is pecked to death by ducks. I mean, it's nothing, it's just endless. And this is a reason, honestly, that people will leave if they have to put up with this day after day after day, hour after hour, minute after minute. And if you don't believe this happens, then you should talk to somebody who is sort of in a minority position in an organization for just five minutes to hear their stories. Double standards, this one, again, and this one I think is one that women point out, a lot of course it applies to other minorities as well, where you basically get caught in a double bind for things that don't really apply to everyone else. My favorite story when I transitioned, I went from never having anybody having a problem with my wonderful and sweet personality to being told that I was simultaneously unapproachable and too nice. Okay, you can't win against that. Women are quite often accused of being shrill if they happen to get angry. Even if they do happen to get angry, as a matter of fact, stating a fact can get you in trouble. Not having defined processes, this would particularly apply in a company or something like that, whereas, you know, so how do I get promoted? Oh, don't worry, we'll take care of it. I suppose there's a complaint about harassment. Oh, we'll work it out at the time. All of those things tend to mean the people who are marginalized lose. Okay, and there was in fact a good talk by Katie Hedlestone at PyCon US about this, and there is now a project called No Null Processes to Work on kind of coming up with open source statements of processes. And then finally, and this is the big one that a lot of women are leaving the industry in mid-career, is that they're not given a way that, well, underpaid, this is kind of a fact, and they're not given a path to advancement. A junior female developer can say, well, okay, so I want to go to the next step. Well, you need to have worked on a good project for that. Okay, so I want to work on a good project. Oh, well, we don't think you're ready yet. Well, so how about I get a mentor so that I can get ready, and you're sort of shoved off one way or another or another? Okay, let's see here. This is my one image. I wanted to do more images except that I did a Google search for them yesterday and got so depressed that I had to stop working on my talk for a while. So, but I figured, okay, so this is kind of the example of the double standard that you see. I'll let you digest that on your own. Ignoring intersectionality. The problem with the way that marginalizations can work is that one of them is bad, two of them is more than twice as bad, three of them is more than three times as bad. They add up. If you're a woman, that's not great. If you're a woman of color, that's even worse. If you're a trans woman of color, that's even even worse, and so on, and it builds. And unless you understand that in the people that you're dealing with, you're not equipped to even talk to them intelligently. And then finally, and this one here honestly touches my experience perhaps more than any of the others even, is not listening. And the first part of that is that if you're not actually involved, if you're not involved as a wrong person, if you don't actually experience those things, if you are not the target of those things, it's often quite hard to see this. So I'll go back to the thing where, again, I have heard from many men of absolutely good faith that they just don't see sexism in their industry. And that almost always means, yeah, that's because they're men. They don't see it. It's not there. You know, a white person is not likely to see racism. And these things, and here again, I'm speaking from having spent a large part of my life, as one of my students put it, undercover as a cisgender white male. And when I was getting ready to transition, and I've mentioned this before, I was really worried about some of these things, and I tried to watch for it. This was not a matter of me trying to brush something under the rug. I was actually trying to see what I would be up against. And honestly, even trying from that viewpoint, it was very hard for me to see things that once I was in that position became totally, completely obvious. Okay? You need to, and we all need to be aware of those blind spots. That is, if it is not something that affects me, I may have a hard time seeing it. That means that when somebody who is affected by it tells me something about it, I need to listen, not ignore it. Okay? If a person of color tells me about racism, my saying, I don't see any racism in this industry, does not mean a thing. They're the ones that are experiencing it. I know from experience that they are probably right. This can be a little bit uncomfortable because quite often listening in these situations can feel like you're being accused of something. If someone I know who is black tells me about racism in the US and how I have gotten an advantage that they haven't had, it feels like an accusation. In fact, it isn't. It is a problem. It is the truth, but we need to get past this whole, no, I didn't do it to you. Maybe I didn't, but I benefited from it, so it's my responsibility to help fix it. So, again, as I was saying, diversity is a hard problem. Inclusion is even harder. We don't have any easy answers. We don't have any overnight fixes. But this affects all of us. Again, just because we're part of one minority, just because we're part of one marginalized group, that doesn't give us a pass for everybody else. So we all need to be part of the solution. And this will require having different people join us will require doing different things. We need to come to terms with that. We need to write test if we're going to do TDD. We need to actually make the changes. And I believe that will make us all better. Thank you. Thank you. Thank you, Naomi. Does anybody have any questions? I'm a manager and I sometimes have the opportunity to hire people. And I don't know how to handle the situation because I probably would like to increase the... Most of the people in my group, they are white men, but I cannot tell my boss I'm not going to hire a white man. And that wouldn't... My company wouldn't accept that. So I don't know how to handle that. I don't want to start interviewing white men because these are going to be the majority of the people applying and rejecting them simply because of that. So I don't know how to handle this. Like I promised, I'm going to stick to my promise. I don't have a good answer for that. I have been in those situations where we needed to make a hire and these were the people that we had. And I did not have the time or energy to find anybody else as an alternative. So, you know, I can't say there's a good answer. I try to think of ways that I can get the word out about positions to other communities. So I think making contact with other things, networking with other people is a way to start. But there isn't a necessarily always an easy answer. We have to do what we can where we can find things to do. Okay, that's unfortunately all the time we have. So if you have any more questions, talk to Naomi directly. We'll be around at the PSF meeting, I might ask. Thank you.
|
Naomi Ceder - Antipatterns for Diversity **Stop doing the same thing but expecting different results** As developers we put considerable effort into optimisation. We are always tinkering, trying to make things better, and striving to remove antipatterns from our code and our development processes. Yet for some reason we have not been as good at applying this spirit of optimisation to the problem of increasing diversity, even though most people these days agree that, like good tests, agile methodologies, and virtual environments, diversity is a "good thing". My position is that just as there is no single easy way to write good code there is no single easy way to increasing diversity. There are, however, several things that companies and organisations do which actually work against diversity. This talk will explore these antipatterns for diversity, including uncritical belief in meritocracy, lack of understanding of the realities of marginalisation, null processes, misunderstanding of "culture fit", and an unwillingness to change, as well as some ways that teams, companies, and organisations might work to combat them.
|
10.5446/20166 (DOI)
|
Hi, so I'm Michael Hrysek, but call me Michael because if you are not from Czech Republic and want to try to pronounce my last name, it will be sounds very funny, like I am now talking English. I work for Cysnamt Z, which is a company making web apps for local people like email search maps. And we are actually good at trying to open our maps to the world because when I made a planned trip to Pyrenees, it's actually more detailed than Google ones. So we do web apps, so we have to test them somehow. And we do that and therefore I'm sending it right here to share our knowledge, some best practices, some tips, and point out some pitfalls. So how can we test websites? The easiest thing is unit test and call some API. This is by test. And you should, you really should, does it? Does test like this? Because it's very mandatory, but it's not enough for web apps because you have to test some requests and response. You can do that as well. Flask and Django supports it. It can be very simple like this piece of code, but still it's not testing of the page because there are a lot of client-side scripts, so you have to test also JavaScript. You can do that. This is, for example, Yasmin, or how is it called? And again, it's unit test. Now you can test client-side and backend-side, but you have to somehow connect it, make some integration test of full web page that it's work as a user sees. So how can we do that? You can do that with Selenium. Selenium automates browser. That's it. This is what documentation says. And how does it look like? It sounds pretty hard, but actually it isn't. It looks like that. It's very simple code. You firstly import something, then you open a browser. In this example, it's Chrome. Then you get some page. It's our home page, for example. Then print a title of that page, and after that, just quitting closing the browser. I'm sorry for these comments. It's pretty dumb, and I don't like these comments in code, so I will not say it again in this talk. And what can do Selenium more? It can travel the web. You can check some attributes on the elements. You can click on them. You can check some cookies. You can make skin shots. But that's it. And I will not talk about it, because you can read it in documentation of Selenium. I want to talk about some pitfalls and some best practices, how to test the pages with Selenium. So first pitfall is forms. It's very hard. This code isn't simple, and it just does that type Selenium in search field and submits that form. And that's it. And you can see it's five lines. It's a lot of code. And it's hard to find what actually it does. And you have to think about a lot of things. Like when you have on the client side some on change event, it will not be fired when you send keys to that input. Because when you send keys, the focus is still here. So you have to press stop, like in this example, to lose that focus and fire on change event. Also you have to think about all inputs, like check boxes, selects, text area and more. And every input has to be written differently. Also there is bootstrap, which has shiny check boxes, which hides actually real checkbox and instead shows some other element. And you have to think about it. That's a lot of things you have to do. So I made it simple. With library web driver wrapper, which is on PyPy, you can use it for Python 2 and Python 3. And it's well-documented. And code like this one can be very simple. Just find some form and pass a dictionary with keys that are names of these inputs. And the values are values, Pytonic types. Like there is some text. If it's checkbox, you can just pass through false and doesn't have to think about it, that it's some bootstrap checkbox or that you have to click on it. That library will do it for you. So very simple. Another example is exceptions. Yeah, I can show you some example. I forgot it. Sorry. I have some code here. You can see that it's what slides shows. And when I run it, it will open a home page, print title and grid. That's it. So exceptions. When you can find some exception, some element which is not here on that page, it will raise no such element. But it will not say where or which one. So I made it simple because, for example, when you have, for example, our search engine or Google wants, it doesn't matter. When you go to home page, you see some form. You fill in something, click submit, and then you want to check that some results are on the next page. And Selenium will say just no such element. But I want to know that it's on which page because is there a bug on that page with results or is there a bug with page with form and that page is not fairing the good action. With this exception, you can see that the result count is not on home page. So yeah, it's not here because you have to be on a different page. So I know that the bug is not on page with results but on the home page. Another pitfall is status calls or headers. There is a big discussion on WebBraver, Backtracker, and it's closed. They want to feature, a lot of people want feature to check these attributes, but you cannot because they says that it shouldn't be there. It's just API for browser and if you want some headers or status codes and so just use some library for request. That's true. But if you want to make some own request, you have to know that link, you have to know if it's get or post or another. You have to know if you have to pass the cookies and you have to pass if it's some kind of form, you have to get these attributes and it's pretty hard. It's about ten lines of codes to make it done and I made it and put it to that library so you can just get the element of some link and just download file or page and you can check actually what is there. Another way is my colleague did that, he run a proxy and checks status codes and headers and more from proxy because you can do that with Selenium. State element exception. If you will start write some Selenium code and some tests with Selenium, you will find it pretty soon because when you do with something with DOM on the client side with JavaScript and these elements change or disappear and JavaScript completely changes DOM, you don't have in Python you have just reference to that element in DOM and if something changes, JavaScript changes it or if you go to another page, if there is something, you don't have this element anymore and you have just reference to something that doesn't exist. In that exception is just get me some element which has IDQ and send some keys, JavaScript makes some changes and when you send keys again to that element, state element exception. It's something, how to say, you don't want to use some elements and if you do that you have to be very careful because it happens a lot and actually it's easy to fix it just, you know, you see the difference. In that example is very bad thing and it's time sleep. You don't want anything like this in your test code because what if it will take a longer time, it will fail because it will not be, you want to wait for some action but it will not be there so it will fail. For that, there is wait machinery in the selenium, you can use it like this but you have to wait for elements a lot because there is a lot of JavaScript and you always want to wait for something so I made it simple again. Just wait for element and that's it. I have proof too. I have some code here. You can see this code from slides just a little bit longer and it opens our home page, pass something, then reuse that. Fire some autocomplete, then it wants first element in that autocomplete what was found or autocompeted and then show text and send another keys and show what's there later. If I will run it, it will fail. Now you can see that it failed, still element reference exception. If you want to fix it, you have to just, for example, use lambda or whatever to reuse that's getting a filament and always get the fresh one. It works but you could see that there is time sleep too which waits for two seconds and you could see that action was much faster and it doesn't have to wait for two seconds so you can make it wait for element and it's almost same just a little bit shorter and it works as expected. You can see the change. When I run it, it will be much faster. I will hold for seconds. So use dot. It is searching by text. So then it doesn't support it but it's also commenting because you want to, you will insert something to database and you want to show some table or some contact for something and you want to check that information is here and you don't care about which element or whatever. So you can search by that. It's implemented by Xpath which are very great. You can make a lot of things but I would not recommend it use it a lot. The first line is very simple example how it's done but it's more complicated than that and it's very slow. You can see that in one run but if you use it a lot of times, it will be very slow. It's not good for maintaining test with Xpath. It's better with IDs and classes because it's much easier because if you use classes and someone changes template or whatever, move elements around and still these classes are moving with these elements, you don't have to change a test code. Ideally. Okay. We know how to manipulate browser but how to write test. Well, you can use web-driver test case which is in the library as well. And why is it there? Because it implements a lot of cool things like it creates driver for you and you just implement the driver method like in this example, the default is Firefox. And you can say with constant if you want open new browser for every test or test case or all tests. If you will reuse browser because it's very slow to open a new browser window, then you have another things to keep in mind which are, for example, that in test you can click and suddenly some alert will show up. And when that happens, it's blocking like for user. When some alerts show up, user can't make anything, just click on okay. Somehow dismiss it. So when there is alert, you have to dismiss it and if you will try to do something else, you will get just exceptions. And when some unexpected test, some failed test, will open some alert, it will stay there forever. No, forever but to the time when you close the browser. So you have to close it and that test case does it for you in teardown. Another thing are windows because your web page can open a new tab to some other page, some link to different word. And also you want to go back to your web app which you are testing. So again, this test case does it for you. And it does a more thing. Because you want to check out if your page works just perfectly. And for example, if there is some 500 error or something else like access denied or whatever, you want to know about it. And this test case after every test checks that if there is something like this. By default it just looks for classes error. And you can change that. And you can see that in documentation. Also there are error messages because your app can show some error message to user. For example, if user wants to insert another user with some username or some registration form or whatever, and there is already that username registered, it should show error message. And you want to check that test failed because you want it to insert it and if it's not done, you want to know about it. And you have to write a lot of code in every test to check that everything is all right. So I made some decorators. And that test case always check if there is some error or error page. And if there is something, it's that test will fail. And if you want that it should be access denied, you just put decorator there. And if you go to admin, it should show the night page and that's it. You are done. You don't have to always check everything. Also if there is some form and you want to register another user with admin and you want to check that it will show error message, again you will just put the decorator error message and which one should be there. And if there is this extension, it's okay. If there is not, it's not okay. And if there is different error message, it's also a failed test. So it's very easy to test things like this. And there are more decorators like this. Also for information, if you want to test that after registration, there is some information that everything was fine and you can look in or whatever, you can test that as well. Another thing I have seen JavaScript. As I was showing you can test JavaScript, some unit tests. But when you run JavaScript in browser, there always can be some bug which is not covered in unit test. So just put this code to your page and Selenium always checks for these errors. And if there is some unexpected error in JavaScript, again, it will fail. I can, I have an example for you. It looks like that. It's very simple. Just go to home page. And two tests, one for searching, one for some telescopes. And when I run it. It works. You could see that I used one instance per test, I can also use one instance for full test case or full test. And if I run it, it will be much faster. Because there will be just one browser. Done. Also I can show you some error. You can see that this element is not. So I can run it in that page and that's correct because it's very not here. Also I love PyTest. Also the web drivers. So you can write tests with PyTest with this library. So that's it. It's very simple to write tests with my web driver. But there is still one more thing. And it's you don't have to run it in your laptop with open browser. You don't want to always look at it. And you want to run it on server without X server. It can be done on the BIAN. You can install PyVertual display. It's Python client for backends, X server backends. You can pick one of them which are shown in slide. And for example we are using X3 FB and we are happy with that. For example XVNC are good that you can connect to it. So if it's failing you can connect and see what's happening. So it's nice. And usage is very simple. Just start some display and run your test and then stop it. I can show it to you that it works because I don't have any virtual X server on my machine here on the on server. But it really works. And I would recommend you to use a very big size of that display because there is one thing that you can keep it for that if you are using some fixed elements it will mess up because Selenium will always scroll to that element you want to work with. But if there is some fixed element it will just scroll to that element and if it's fixed something above it will fail because that click action or something else would fire with another element. Therefore if you have some fixed elements you want the biggest size you can. Other things are Selenium server you don't have anything because as you could see I just run browser directly but if you want something more you should check out Selenium. It's written in Java so be careful with it. And it allows you to run it in more browsers with more operating systems also some mobile devices and it will be faster. It works that you have some master Selenium server and there are a lot of nodes on different type of machines you will just run some windows and Android and whatever and install Selenium server there. Configure it like a node and on some master you will register it and with your code you will just connect it to that master and you will say which browser you want. In this example for example Firefox and when it's the same like with Chrome you put remote there and say you want Firefox for example and that master will find node where is Firefox and you can say also that you want Firefox on the BN or that you want Chrome on the Android and that master will find that node and if it's there it will run it and if it's not you will see some exception. So that's it. If you like it check out the documentation. There are a lot more things than I was saying. Thank you. Thank you very much Michael. Do we have any questions? We do. Hello. If I use one instance per test case can I have something like set up and tear down methods for each test? Sure. You can. It doesn't restart the browser but it is fast. You have to think that if some tests messing up something you have to fix it in tear down. Like in some other test if you will run some unit test and this unit test well unit test and it will write something to database it's same here. When you click somewhere and it will do something it's messed up and you have to fix it in tear down but you can use it normally as you are used to. Go back one slide please. Thanks for the talk. You said you had to put big configurations when you are using the display servers like XVNC and other stuff. Is that really mandatory or is it usable to test responsive design? You can test responsive design. If you will make it smaller your web page will adapt. That's all right. I'm just saying if you have some fixed elements and you can't redo it. If the page is broken it won't work. Yeah. Thank you. Exactly. Hi. Thanks for the talk. Could you comment on architecture of your test suite when you have a big application are there some things that you definitely need to test or some that you shouldn't test? Can you comment on that? Well it's a question for 140 minutes but there is always 80% of work is in 20% of code. So focus on what your users do mostly and test that. If this question is good for your answer question. Should you for example test your styles? Sys. Styles. CSS. Yeah CSS. Is it possible? Like how the weather you get. I don't think so. You can check which values of CSS styles are there but I wouldn't recommend it. Thanks. I have one more question. Can you make something like automated screenshot comparisons with this? For example start the test case. Do screenshots at every step. Compare it somehow and then compare it every time you run the test case? Is there a tool to automate that? Well I forgot one thing. Thank you for the question. That when test fails and you use my test case for the only test or my tools for pie test it will automatically make a screenshot. And when you see in Jenkins that it's broken you can just check out the screenshot. It's very fast. Then run it on my computer and check out what's happening. Making screenshots is very simple. You can just driver that make screenshot I think and just pass a parameter where you want to store it and that's it. Then you can write your code which just compares these images. So you can do that. Very simple. But I don't know anything about any library which is doing that. Any other questions? Thanks for the talk. Which version of Selenium are you using with the developer? Which Selenium? Yes, the version. Yeah, there are two versions of Selenium 1 and 2. You don't want to use Selenium 1 because it's called Selenium and Selenium 2 is actually called WebDriver and Selenium 1 just calls some JavaScript. So it's not really user actions. It's Selenium 2.0, it's WebDriver and all browsers implemented it inside the browsers except Chrome. It's binary alongside. You call real API of browser and it's faking almost real user actions. So Selenium 2. Thanks very much, Michael.
|
Michal Hořejšek - Testing web apps with Selenium „Selenium automates browser.“ Selenium can be used as tool for testing web applications. At first it can be pretty hard to start testing with Selenium, but later on it can be even harder. I want to show you that it doesn't have to be true. That it can be easy, actually. But you have to know few things which you have to be careful about and that there is tool webdriverwrapper which can make it easy for you. I will speak about handling pages with JavaScript and which common problem can you have, how to run Selenium on servers without X server, how to deal with tabs, how to test with UnitTest or pytest, and how can webdriverwrapper make things easier for you and more.
|
10.5446/20165 (DOI)
|
Can you hear me? Aha, wonderful. I'll put this one down. So I'm Michael Ford. I work for Canonical on a project called Juju that's written and go. I've been a Python developer for about 13 years now and a Python core developer for about eight of those years. I released or created Unit Test 2 and the Mock Library, which reflects my particular interest and my particular passion for testing. I think good testing practices are the only ways or certainly an important part of keeping developers sane. This is my colleague and actually my boss, Dimita. Hello everybody. Can you hear me? I'm Dimita Rydywf. I work with Michael in Canonical on the hopefully next generation cloud orchestration framework suite system Juju. Here is our talk about it. So the talk is called to the clouds and it's about deploying applications and services to the clouds and why you should probably be deploying to the cloud even if you don't want to. Because we're talking about the cloud, we're going to have a lot of buzzwords. The cloud is a piece of jargon that's had a lot of hype over the last decade really. I'm hoping that in this talk we'll rehabilitate the term a little bit and show how despite the hype there's actually a useful set of technologies and principles behind this. So this photo, rather stretched I'm afraid, is actually one of the original computers that Google used in the early days of their search engine. And part of their genius, which I think was actually forced on the more out of necessity than directly entirely out of genius, was they built their service using cheap commodity hardware rather than the very expensive big iron mainframes that their competitors were using. A consequence of this was that they were able to scale out very rapidly, very easily and very cheaply just by buying new units of inexpensive commodity hardware. But because this hardware wasn't as reliable as the big iron mainframes that other people were using, they built on top of it this false tolerant architecture that as well as enabling them to add new units of hardware to their system also allowed them to take out failing hardware. They eventually released this platform as a service as the Google App Engine. Amazon kind of took this to the next level with their infrastructure that they used for running a giant retail website. They took a slightly different approach providing a host of virtual machines as deployment targets for their services and this infrastructure as a service approach rather than the platform as a service approach is really loved by developers because it just gives them a machine. When they just have a machine, they know what to do with it, they know how to deploy to it. When Amazon made their cloud public, they rapidly became the dominant player in the public cloud market. Other modern infrastructure as a service, public clouds include HP cloud, Microsoft Azure, Joy-Ents, OpenStack based offerings, a whole host of them. There are several problems that using the cloud solves and these include dependency health, resource underutilisation and hardware management. The way the cloud solves these problems is by separating the deployment target, your deployment layer from your hardware layer. When you deploy to the cloud, you deploy to virtual machines without having to care about what physical machines your service is running on. Underutilisation or equally resource overutilisation, a situation you might be familiar with, you have to deploy a new small public service say or an email server, a DevWiki bug tracker, so you need a new machine and you deploy stuff to it and you're using about 10% of its capacity. Alternatively, you're working in a company where getting new hardware is a really slow and painful process and so what you do is the three servers that you already have, you jam everything onto that and everything runs as slow as hell. Dependency health, again situations that you may be familiar with, you have a whole bunch of applications and services, some of them use the same library, they have the same dependency but they use different versions of the same library. So you have these applications that you can't deploy on the same machine, they have to be located on physically different machines. Alternatively, you do the work, you make sure they're all using the same version of the library and then what happens is you want to use some fancy new feature that comes in a newer version of the library, so you deploy that for one application, you forget about the other services and your deployment breaks another application. Obviously, all Python libraries take backwards compatibility very seriously, so this never happens in practice but I've had to do emergency rollbacks of deployments because we upgraded a dependency, we forgot there's some other application on the same box using the same dependency and it just broke. Alternatively, you do the work and you port all your applications at the same time to use the new version of the library and you do your upgrade, you do your new releases all in lockstep and then there's a regression in one of your applications and you have to roll back all of them at the same time. There are various ways of solving this, you can deploy all your services to separate machines and then you're back into resource underutilisation or you can use virtualM to provide isolated deployment environments for all of your application. Again, this is something that I've done but what you then have is you then have the same libraries in multiple different locations in non-standard places on the file system and that's a security problem and system administrators, they tend to hate this solution because they like to understand and preferably be able to control dependencies. So an alternative approach is for every service that you run or even every component of every service that you run to have that in its own virtual machine where you're able to very tightly control and specify the dependencies just for this application. And then hardware management, that's one of the most important benefits of the cloud because we have this separation of our deployment layer from our physical layer, we're able to deploy new services just by acquiring a new virtual machine, we're able to add new machines to the cluster very easily and we're able to take out failing machines, upgrade machines without shutting down running services. OK, next slide I think. So it may be the case that you're already running a bunch of services, you may have hundreds of servers, you may have just a few and you certainly don't want, you feel like you don't want the cloud, you certainly don't want your data or your customers data on someone else's machine, maybe you don't want your data located in America or hosted by an American company. And you can get some of the benefits that I've talked about just by managing virtual machines on your own servers, you get this the isolated environment. But wouldn't it be nice if there was a framework that provided the dynamic server management aspects of this that was able to provide, automatically provide you with new virtual machines, automatically provide you with new deployment targets and make it easier to have machines and take machines out. And that's what you really want is a private cloud and if what you want is a private cloud, that probably means you want open stack. Ah, PDF export, screwed up the bullet points there, sorry about that. So open stack is basically the private cloud, it's not entirely the only option but it's the giant in the world and it's written in Python. It's probably one of the biggest things going on in the Python world right now. So there are lots of companies by big and small hiring Python developers to work on the cloud either directly on open stack or on specific implementations of clouds, both public and private using open stack. Open stack is huge. It's huge in terms of the amount of code. It's huge in terms of the amount of sub projects within open stack. Huge in terms of the functionality it provides and huge in terms of the number of people using it and contributing it. And you can get all the benefits of the cloud but without having to use somebody else's implementation. Oh, let's go back to the last slide. Just to mention that there are alternatives. Eucalyptus certainly used to be an alternative cloud implementation. They got acquired. I think they got acquired by somebody, by a company that has a public cloud offering using open stack so I don't even know if eucalyptus currently exists. And there's an alternative data center technology that I've worked with a bit called Mars, Metal as a Service, and that's another canonical product. And that gives you a lot of the benefits of the cloud, the dynamic server management aspects of it, but with physical hardware rather than virtual machines. And juju, the project that we work on, you can deploy open stack to Mars or you can deploy directly to Mars with juju. And it achieves full density, full resource utilization with Mars through using lexie or KVM containers as deployment targets on the physical machine. So it's an interesting technology, sort of a data center level technology sits below the level of open stack. So the benefits of the cloud, solving the problems of dependency hell, hardware management, resource underutilization. If you have, alongside that, if you have fully automated deployments, then the other big benefits that you get from the cloud are you get the ability to rapidly scale out and easily deploy new services. And you only get those if you have fully automated deployments. And an important principle for fully automated deployments are that we treat our servers as livestock rather than pets. If you have a pet, your pet is unique, your pet has a name. Do you name servers in the company you work for? So they all have names. And if your pet gets ill, you spend a lot of time and money and effort on getting it well again. Whereas livestock, they don't tend to have names, they have numbers. And if your livestock get ill, well, it's a cruel world, you shoot them and you get another one. And this is how we should be treating our servers. Deploying new services, we ought to be able to tear down our application servers, reprovision them with a single command, without caring about what physical machine they're located on, without caring anything about the machine they're located on, virtual machine or physical machine, and preferably without having to worry about machine configuration. Now the trouble is with infrastructure as a service, which has largely beaten platform as a service in the market, by the way, except for some specific platforms like Salesforce. Infrastructure as a service is largely one. Developers like it because the paradigm it provides is you have a machine to deploy to. And what that means is although we leave some problems behind, the problems of well, how do we provision, how do we configure, how do we manage those machines, how do we administer them. We just take those problems with us when we go to the cloud. Now there are lots of tools out there that will help you with machine provisioning, Chef, Puppet, Salt Stack, Ansible. Some people use Docker, although that's really about image based workflows, it's not really. But some people are using it in this way to help with the problem of machine provisioning. But these tools all require you to think about machine provisioning, what's going to live on this machine, how do I configure and administer this machine. And as developers or even DevOps, we don't really want to spend our time worrying about machine configuration or administration, or at least that's not how I think about the services that I deploy. When I deploy an application or I deploy a service stack, what I think about is I think about the components. I have my application server, I have a message queue, I have load balancers, I have a database. And these components are all related to each other, they're all interrelated, they all communicate. This is how I think about my application stack. But the actual deployment, we tend to, we're forced into thinking about our services in different ways. We're forced into thinking in terms of units of machines. And Juju is a tool that takes a slightly different approach. So this diagram shows, this is through the GUI, the GUI, you don't have to use the GUI of course, but it's a great way of visualising services. This shows deployed and related services deployed with Juju. The lines there show the relationships between the components of the services. So Juju takes a different approach. It's about service orchestration, it's a tool that provides a powerful service modelling language that just happens to use virtual machines as a deployment target. The basic unit for deployment for services or service components in Juju is the charm. And the charm no codifies the knowledge required to deploy and configure that application. This is a very important principle of DevOps, a principle of automated deployment, is that the knowledge about how to deploy and configure your service components, rather than living in some system administrator's head and then that system administrator gets a better job somewhere else. And like Guido said in his talk this morning, they have deployed services using binaries and nobody knows how to get those binaries back. If some disaster occurs, presumably they have lots and lots of backups because they can't reproduce, they don't have a repeatable full stack deployment, they can't automate and reproduce that. You need to have your deployment knowledge codified and charms for Juju are the codified knowledge about how to, not just how to deploy the service, but how it communicates with other services. And these charms tend to be written in Python, so when you're using Juju you get to codify your deployment knowledge in Python. But not just that then, through the charm store there's a whole host of common infrastructure components, many of which you're already using, I'm sure, where there are existing charms out there. So Juju is sometimes called apt get for the web. So why Juju has kind of talked about some of these, we get to think in terms of service orchestration, think in terms of service components, how they relate and connect to each other, which is how we think about our application stack anyway. We get to work with Python to do all this. Cloud independence is important as well. Juju will happily deploy using exactly the same configuration to Azure, Joi-Ant, HP Cloud, EC2, Mars, and also using Lexi containers to your local machine, which is not just to your local machine, to your CI server. So you can take exactly the same production deployment configuration, you can deploy that locally for running your own manual test, but your CI server can also take that, can spin up the whole production stack, configure the communication, configure things to talk to each other, and you can run your acceptance tests on your CI server, using effectively your production configuration. Now you are running CI tests, you are running acceptance tests for your applications, I assume. Like I said, I have a particular interest in testing, and I think a good principle with testing is if it isn't tested, it's broken. If it is tested, it might not be broken. Those are the principles I work to, so testing doesn't guarantee that things work. What it does is provide an assurance that things might work. You need that assurance. And through the charm store, there are many charms available, things like Hadoop, Seth, Postgres, Django applications, MySQL, Postgres. All of the standard components, they are all charmed up there and ready to deploy. So let's look at an example deployment of a Django app. We have ten minutes left, time has flown. Demeter is going to show you, it's a Django app called D-Paste, using the Django framework charm. This is one we already have deployed, but this is a full stacker, and I'll let Demeter take over and talk about it. As you can see here, well, the resolution is not very good, but you have services represented as boxes, and this is all you really care about in your deployment. You don't really care that much about machines, you care about your workloads, and also how are they related to each other. As we can see, we have a Django framework charm here, which is a modified version of the default charm available in the charm store. It's configured to run D-Paste, pulling the Git repo from GitHub. There are a few other things like a squid proxy, also a couple of HA proxy charms which are configured as application load balancer and cache load balancer. We have the front-end Apache, and also the back-end Postgres database. This is the Django itself, which is also a charm that you may or may not deploy if you don't want to. This here is the G-Unicorn, which is a subordinate charm to D-Paste. What does that mean? It means that with each and every unit of D-Paste, there is a G-Unicorn deployed alongside it. It can be used, for example, for logging or other closely related services which depend on each other. This is the service view, this is the default view you can see. Here we have the machine view, which shows that we have eight machines, actually lexic containers. What service is there running on? If these were physical machines, this is all running on the laptop using the local provider that I mentioned, so they are lexic containers that these services are running in, which is why hardware details are not available. If these were running on virtual machines, the hardware constraints of the actual machines would be visible here. Also, I'll come back to this, but you can also use the command line to do the same sort of inspection and management. For example, if I do the status format short, that's just the gist of what I have in my deployment. If I want to, I can show a lot more details about it. What sort of workload status it is for each charm, what relations there are, what IP addresses, when it was deployed, and so on. There is even a status history that shows, for example, I deployed this charm, but it didn't come up. It's waiting for some sort of interaction, like I need to enter database details or I need to do something manually in order to enable it to work. Also, we can easily do, well, not quite right now because the network is non-existent, but I would have shown you the same exact deployment than in Amazon. It's actually running. If you can hit this. I can show you whoever is interested later. Any questions? Of course, you're welcome. What you can do with this, you can actually see what are the details of each service. You can see what units are running on it, what's the configuration. There are lots of things that you can tweak for each charm. This is also part of what the charm encapsulates as a best practice. What we could do here is you saw that there was one running unit. If we needed to scale out our app, we would do juju, ad unit, I think Django, whatever this charm is called, five, and that would create five new units. Because the Django charm is related to the Postgres database, so as the new units come up, there's a juju knows these need to be related to the database. Charms are comprised of hooks, which are code that can run. If we go on to the slides, we'll have to fly through these because we have just five minutes. There's a charm that says relation joined. The new unit of the application server is joined to the Postgres database. It knows the configuration information. The Postgres charm, we don't need to generate a password and user for the app, for the database, sorry. This is all done as part of the charm. When on install, on startup, a new database user is created, and then as the relationship is joined to the application server, that information is passed through, that configuration is automatically given to the application server. As we add new units of the application server, that configuration is given to the new units, so these services are automatically added and configured to talk through the relationships that we've defined. They're automatically configured to talk to the rest of the application, so it's not just the deployment information that's codified into the charm, but the configuration information, how to talk to other units of the service, and as we add new units, this reconfiguration, this updated configuration happens automatically for us. This is service orchestration. You can see this isn't the directory of an example charm, it's just a zip archive. You can see here that all of the different hooks that fire when different things happen, they're actually similar to a single Python file. Here's the install hook that actually act installs a bunch of packages. On to the next slide. We're actually using a deployer bundle here, which is a YAML file that you can keep under version control. Deployer is a separate tool that the functionality is now being rolled into juju call, and you can see for the Django charm that we're using that there's a whole bunch of configuration information stored in there. And if we go on to the next slide, the relationships between the charms and the other components of the service that we're using are also defined in the deployer bundle. So as you can see, there's an awful lot to juju, there's an awful lot more we could talk about, but the basic principle of codifying our deployments and configuration information in Python, being able to reuse the existing charms using virtual machines as a deployment target to get us the benefits of the cloud of separating our deployment layer from our physical hardware, and with juju being able to then just retarget that multiple back ends is a very powerful combination. I think we'll... This is just an example of how easy it is to do this deployment. We'll make the slides available as the last... If we go to the last slide, that has a URL, and I'll tweet it and Demeter will tweet it. This is deployer doing its stuff. But I think that's about all we have time to... The slides are available, and there is also documentation and all the different concepts of juju explained on jujucharm.com. There is also a live demo that you can use as a sandbox to do experimental deployments. You can then export that as a YAMO file and then deploy it as we show with the juju deployer to your cloud or to your local machine. OK, thank you very much. Do we have time for some questions? Yes, yes. There's a couple of questions over there. Well, I see juju like something like more top level than PAPED, SOLD or any of this, but I don't really see how complex could it be because when you have some simple applications, you use it as a platform. Usually you can configure them with three parameters or so, but usually the applications that are listed in my company with target are really, really complex, where we need to have like four.yambs, three exams, XMLs, two database connections, et cetera, and they are configured differently. At the end, the edge cases you have are really, really complex. How would juju scale with complexity of an application? So there are quite a few very complex charms that have certainly in the order of tens of configuration options, which are things that you cannot, in the deploy YAMO, obviously they all have same defaults, but there are many tens of configuration options that you can tweak in the deploy YAMO. What we tend to see people doing is for their own applications, they will write their own charm, possibly taking like the Django framework charm, they're taking that as a base if you're deploying a Django app. You can have fat charms that contain the application bundle itself and extra configuration information, so you could merely specify the paths to the configuration files within the charm bundle, and so you would have your own charm that is specific to your application, and then you would reuse existing charms for the other component pieces that are deployed alongside it. A good thing about charms is that we showed an example in Python, but they're actually language agnostic, so you can, for example if you're using Chef or Puppet or Ansible, you could reuse the same language, you can use Bash or whatever you like, just we have a very good support for Python, like nice tools, libraries, unit test support, coverage and benchmarking as well. One of the things to note is that under the hood, of course, for an individual service component, Juju is doing machine provisioning. The difference is that at the modelling layer, you're not thinking in terms of machine provisioning. Juju really works alongside these tools, and you can see that this particular charm has a fab file, it has Ansible.py, this particular Django framework charm here, I don't know if it still is, but at the point that we were creating, this demo was originally created, was actually using Ansible and Fabric for doing the service installation and machine configuration. I can see that Juju can be useful for provisioning services, where the basic building block is like a container or something, or like anything that runs on the top of VM or a machine that's already running, let's say, but is Juju the right tool, or is it intended to be the tool that you can use for actually spinning up old VMs in the cloud and talk to like Google Cloud API or AWS API or any, this sort of task. Well, so if you just want something to manage your cloud, I mean Google, computer, NGIN, AWS, all of these things, you do Juju add machine, it will create a new virtual machine for you, you can then use that, you can specify the machine, you can create containers on those virtual machines and deploy components to those machines. Juju SSH machine number gives you an SSH connection to that specific virtual machine, so if you just want something to spin up virtual machines to users alternative workload targets, there are then Juju actions that allow you to run specific commands on those machines, so you kind of can use it in that way. It's not entirely the intended use case. So it has the integration with some other APIs. It has it with all the cloud providers that are supported, so Google, computer, NGIN, AWS, Joyant, HP Cloud, Azure, Joyant. Yes, it has within its knowledge of how to talk to those APIs, how to create the virtual machines, how to destroy the virtual machines, how to connect to them, some aspects of how to configure networking on them, this kind of stuff. It has to do that because it's doing the machine provisioning for you. A question on, you mentioned that your machine should be treated like livestock. How do you deploy configuration changes? Do you take down the machine as an immutable server pattern? No. Or do you reapply, reapply the changes? Right. So the idea is that you go away from the actual machine configuration and instead codify what configuration your service needs within the charm. And then if you need to change that configuration, you specify either tweak charm config settings or you relate it to something else, which provides the functionality it needs. So the actual configuration happens automatically depending on how the charm is written. So if, for example, your charm is written both as a standalone application and as a clustering solution, if you just add more units to it, then the charm will automatically discover those other units and set up clustering in the same way. And there is a concept of a charm upgrade where there's a new version of a charm that you need to deploy. Individual units can be upgraded to the next version of the charm. More questions? What do you have for autoscaling, like a load balancer, on deploying new servers on demand? So this is kind of a high level feature on top of what Juju currently provides. There are plans to do things like autoscaling, load balancing as part of the set of features that we provide. That's not yet there, but it's being worked on. It's also, we mentioned that Juju's open source, we're accepting contributions, it's on GitHub. I think the standard answer is landscape will do scaling on top of Juju. Right, that's it. Hello, how is it different than OpenStacks Heat and Murano? So it's different that it's not targeting, I'm not quite familiar with both of those in detail. But I think that the thing is Juju is not prescribing a certain sort of cloud that it can orchestrate. For example, I know Heat orchestrates OpenStack, whereas with Juju you can orchestrate OpenStack and other clouds, but you can't orchestrate it as well with the same set of features and steps. More questions? This is just a very quick question. How popular is Juju with respect to other similar services like Chef or Ansible? Juju is fully open source developed by Canonical. We're using it a great deal with our customers. This is how and why Juju exists. We have cloud customers with large deployments, often but not only large deployments of OpenStack, where we either set it up for them or manage the deployment for them. This is how we're using OpenStack. Juju is quite battle tested and is getting refined based on those experiences. There is community uses, particularly community around the charm creation. I don't think in terms of the wider community, it's as widely used as Chef and Prophet. I think the DevOps mindset is still used to thinking in terms of machine provisioning. The idea of service orchestration, although Juju has been around for a while, service orchestration as a concept and a way of thinking matches more directly the way developers think about their application components. It's something that's only gaining traction. Because we're seeing the community of charm creation, those are obviously people who are using Juju. Last year we had a lot of big contributions, like we're working with our partners Cloudbase, which actually ported Cloud in it for Windows and CentOS as well. I think there is a... This is a Romanian company who were using Juju. They had a customer who needed Windows workloads, so they did the work to... Juju now deploys Windows workloads and runs fully on CentOS as well as Ubuntu. For Windows you need an Ubuntu state server to manage the applications, but it will deploy two Windows machines. But for CentOS you can have a fully CentOS Juju installed if you want. That's on the server side. On the client side, the client is available for Windows, Linux and Mac. Pretty much any platform that go runs on because it's written in Go. More questions? Thank you very much.
|
Michael Foord - To the Clouds: Why you should deploy to the cloud even if you don't want to Do you deploy your Python services to Amazon EC2, or to Openstack, or even to HP cloud, joyent or Azure? Do you want to - without being tied into any one of them? What about local full stack deployments with lxc or kvm containers? Even if you're convinced you don't need "the cloud" because you manage your own servers, amazing technologies like Private clouds and MaaS, for dynamic server management on bare metal, may change your mind. Fed up with the cloud hype? Let us rehabilitate the buzzword! (A bit anyway.) A fully automated cloud deployment system is essential for rapid scaling, but it's also invaluable for full stack testing on continuous integration systems. Even better, your service deployment and infrastructure can be managed with Python code? (Devops distilled) Treat your servers as cattle not as pets, for service oriented repeatable deployments on your choice of back-end. Learn how service orchestration is a powerful new approach to deployment management, and do it with Python! If any of this sounds interesting then Juju maybe for you! In this talk we'll see a demo deployment for a Django application and related infrastructure. We'll be looking at the key benefits of cloud deployments and how service orchestration is different from the "machine provisioning" approach of most existing cloud deployment solutions.
|
10.5446/20163 (DOI)
|
So, hello, welcome, everyone. I hope you're enjoying your Python as much as I do. And for the next 45 minutes, you can just sit relax and enjoy the talk about big data with Python and Hadoop. Slides are already at slideshare.net, and I'll give you the link at the end of the talk. And this is our agenda for today. At first, a quick introduction about me and my company. So you get an idea about what do we use Hadoop for. Then a few words about big data, Apache Hadoop and its ecosystem. Next we'll talk about HDFS and third-party tools that can help us to work with HDFS. After that, we'll briefly discuss MapReduce concepts and talk about how we can use Python with Hadoop. What options do we have? What third-party libraries are out there, written in Python, of course, about their pros and cons. Next we'll briefly discuss a thing called PIG. And finally, we'll see the benchmarks of all the things we've talked about earlier. These are freshly baked benchmarks, which I made a week ago just before coming to Euro Python, and they are actually quite interesting. And of course, conclusions. By the way, can you please raise your hands who knows what Hadoop is working with Hadoop or maybe worked with Hadoop in the past? Okay. Okay. Thanks. Not too much. All right. This is me. My name is Max. I live in Moscow, Russia. I'm the author of several Python libraries. There's a link to my GitHub page if you're interested. I also give talks on different conferences from time to time and contribute to other Python libraries. I work for the company called A-Data. We collect and process online and offline user data to get the idea of users' interests, intentions, demography, and so on. In general, we process like more than 70 million users per day. There are more than 2,000 segments in our database, like users who are interested in buying a BMW car or users who like dogs or maybe users who watch porn online, you know. We have partners like Google DBM, Turn App Nexus, and many more. We have quite a big worldwide user coverage and we process data for more than 1 billion unique users in total. We have one of the biggest user coverage in Russia and Eastern Europe. For example, for Russia, it's about 75% of all users. Having said all that, you can see that we have a lot of data to process and we can see ourselves a data-driven company or a big data company like some people like to call it now. What exactly is big data? There is actually a great quote by Dan Ariely about big data. Big data is like teenage sex. Everyone talks about it. Everybody really knows how to do it. Everyone thinks, everyone else is doing it, so everyone claims they are doing it. Nowadays, actually, big data is mostly in marketing term or buzzword. Actually there is even a tendency of arguing like how much data is big data and different people tell different things. In reality, of course, only a few have real big data like Google or CERN, but to keep it simple for the rest of people, big data can be probably considered big if it doesn't fit into one machine or it can't be processed by one machine or it takes too much time to process by one machine. But the last point, though, can also be a sign of big problems in code and not a big data. Now that we figured out that we probably have a big data problem, we need to solve it somehow. This is where Apache Hadoop comes into play. Apache Hadoop is a framework for distributed processing of large data sets across clusters of computers. It's often used for batch processing, and this is a use case where it really shines. It provides linear scalability, which means that if you have twice as many machines, jobs will run twice as fast, and if you have twice as much data, jobs will run twice as slow. It doesn't require super cool, expensive hardware. It is designed to work on unreliable machines that are expected to fail frequently. It doesn't expect you to have the knowledge of inter-process communication or threading in RPC or network programming and so on, because parallel execution of the whole cluster is handled for you transparently. Hadoop has a giant ecosystem which includes a lot of projects that are designed to solve different kinds of problems, and some of them are listed on this slide, more just didn't fit in. HDFS and MapReduce are actually not a part of ecosystem but a part of Hadoop itself, and we'll talk about them on the next slides, and we'll also discuss PIG, which is a high-level language for parallel data processing using Apache Hadoop. I won't talk about the others because we simply don't have time for it, so if you are interested, you can Google this for yourself. So, HDFS, it stands for Hadoop Distributed File System. It just stores files in folders. It chunks files into blocks, and blocks are scattered randomly all over the place. By default, the block is 64 megabytes, but this is configurable, and it also provides a replication of blocks. By default, three replicas of each block are created, but this is also configurable. HDFS doesn't allow to edit files, only create, read, and delete, because it is very hard to implement and edit functionality in distributed system with replication. So, what they did was just, you know, why bother in implementing editing files when we can just make them not editable? Hadoop provides a command line interface to HDFS, but the downside of this is that it is implemented in Java, and it needs to spin up a JVM, which takes up from one to three seconds before a command can be executed, which is a real pain, especially if you are trying to write some scripts and so on. But thankfully, to great guys as Spotify, there is an alternative called Snakebite. It's an HDFS client written in pure Python. It can be used as a library in your Python script or as a command line client. It communicates with Hadoop via RPC, which makes it amazingly fast, much, much faster than native Hadoop command line interface. And finally, it's a little bit less to type to execute a command, so Python for the win. But there is one problem, though. Snakebite doesn't handle write operations at the moment. So while you are able to make meter operations like moving files, renaming them, you can't write a file to HDFS using Snakebite. But it is still in very active development, so I'm sure this will be implemented at some point. This is an example how Snakebite can be used as a library in Python scripts. It's very easy with just import client, connect to Hadoop, and start working with HDFS. It's really amazingly simple. There is also a thing called Hue. Here is a web interface to analyze data with Hadoop. It provides awesome HDFS file browser. This is how it looks like. You can do everything that you can do through native HDFS command line interface using Hue. It also has a job browser, a designer for jobs, so you can develop big scripts and a lot of more stuff. It supports Zuki, Peruzzi, and many more. I won't go into details about Hue, because, again, we don't have time for this, but this is the tool that you love if you don't use it to try it. And by the way, it's made of Python and Django, so, again, Python for the win. So now when we know how Hadoop stores its data, we can talk about MapReduce. It's a pretty simple concept. There are mappers and reducers, and you have to code both of them, because they're actually doing data processing. What mappers basically do is the load data from HDFS, the transform filter, or prepare this data somehow, and output a pair of key and value. Mapper's output then goes to reducers, but before that, some magic happens inside Hadoop, and mappers output is grouped by key. This allows you to do stuff like aggregation, counting, searching, and so on in the reducer. So what you get in the reducer is the key and all values for that key. And after all reducers are complete, the output is written to HDFS. So actually, the workflow between mappers and reducers is a little bit more complicated. There are also shuffle phases, sort, and sometimes secondary sort, to combine as partitionists, and a lot of different other stuff, but we won't discuss that at the moment. It doesn't matter for us. It's perfectly fine to consider that there is just a, only mappers and reducers, and some magic is happening between them. Now let's have a look at the example of MapReduce. We will use the canonical word card example that everybody uses. So we have a text used as an input which consists of three lines. Python is cool, Hadoop is cool, and Java is bad. This text will be processed by, it will be used as an input which consists of three lines. So it will process line by line like this, and inside a mapper, line will be split into words, like this. So for each word in a map function, a map function will return a word and a digit one, and it doesn't matter if we read this word twice or three times, or we just output a word and a digit one. Then some magic happens provided by Hadoop, and inside the reducer we get all values for a word grouped by this word. So we just need to sum up these values in the reducer to get the desired output. This may seem unintuitive or complicated at first, but actually it's perfectly fine, and when you're just starting to do map reduce, you have to make your brain think in terms of map reduce, and after you get used to it, it's all will become very clear. So this is the final result for our job. Now let's have a look at how our previous word card example will look like in Java. So now you probably understand why you earn so much money when you code in Java, because more typing means more money, and can you imagine how much code you should write for a real-world use case using Java? So now after you've been impressed by the simplicity of Java, let's talk about how we can use Python with Hadoop. Hadoop doesn't provide a way to work with Python natively. It uses a thing called Hadoop streaming. The idea behind this streaming thing is that you can supply any executable to Hadoop as a mapper or reducer. It can be standard Unix tools like CAD or Unix or whatever, or Python scripts or Perl scripts or Ruby or PHP or whatever you like. The executable must read from standard in and write to standard out. This is a code for mapper and reducer. The mapper is actually very simple. You just read from standard input line by line and we split it into words and output the word in digit one using a tab as a default separator because it's a default Hadoop separator. You can change it if you like. So one of the disadvantages of using streaming directly is this input to the reducer. I mean, it's not grouped by key. It's coming line by line, so you have to figure out the boundaries between key piece by yourself. And this is exactly what we do here in the reducer. We are using a group by and it groups multiple word count pairs by word and it creates an iterator that returns consecutive keys and the group, so the first item is the key and the values, the first item of the value is also the key, so we just filter it, we use an underscore for it and then we cast a value to sum it up. It's pretty awesome compared to how much you have to type in Java, but it's still maybe like a little bit more, a bit more complicated because of the manual work in the reducer. This is a comment which sends our mapper to Hadoop via Hadoop streaming and we need to specify a Hadoop streaming jar, a path to a mapper and reducer using a mapper and reducer arguments and input and output. One interesting thing here is two file arguments where we specify the path to a mapper and reducer again and we do that to make Hadoop to understand that we wanted to upload these two files to the whole cluster, it's called Hadoop distributed cache. It's a place where it stores all files and resources that are needed to run a job and this is a really cool thing because imagine like you have a small cluster of four machines and you just wrote a pretty cool script for your job and you use the next library which is not installed on your cluster obviously. So if you have like four machines, you can log into every machine and install this library by hand, but what if you have a big cluster like of 100 machines or 1,000 machines, they just won't work anymore. Of course, you could create some bash script or something that will do the automation for you but why do that if Hadoop already provides a way to do that. So you just specify what you want Hadoop to copy to the whole cluster before the job will start and that's it. Also after the job is completed, Hadoop will delete everything and your cluster will be in its initial state again. It's pretty cool. And after our job is complete, we get the desired results. So Hadoop streaming is really cool but it requires you to do a little bit of extra work and though it's still much simpler compared to Java, we can simplify it even more with the help of different Python frameworks for working with Hadoop. So let's do a quick overview of them. The first one is Dumbo. It was one of the earliest Python frameworks for Hadoop but for some reason it's not developed anymore. There's no support, no downloads, so just let's forget about it. There is a Hadoopy or Hadoop.io, I don't know. It's the same situation as with Dumbo. The project seems to be abandoned and there are still some people trying to use it according to PIPI downloads. So if you want, you can also try it, I don't. There is a PyDoop. It's a very interesting project. While other projects are just wrappers around Hadoop streaming, PyDoop uses a thing called Hadoop PIPES which is basically C++ API to Hadoop and it makes it really fast. We'll see this. There's also a Luigi project. It's also very cool. It was developed at Spotify. It is maintained by Spotify. Its distinguishing feature is that it has an ability to build complex pipelines of jobs and support many different technologies which can be used to run the jobs. And there is also a thing called MRJob. It's the most popular Python framework for working with Hadoop. It was developed by Yelp and it's also cool. There are some things to keep in mind while working with it. So we'll talk about PyDoop, Luigi and MRJob in more details in the next slides. So the most popular framework is called MapReduceJob or MRJob or MRJob like some people like to call it. So I also like this. MRJob is a wrapper around Hadoop streaming and it is actively developed by Yelp and maintained by Yelp and used inside Yelp. This is how our work on example can be written using MRJob. It's even more compact. So while a mapper looks absolutely the same as in the raw Hadoop streaming, just notice how much typing we saved in the reducer. But behind the scenes, actually MRJob is doing the same group by aggregation we just saw previously in the Hadoop streaming example. But as I said, there are some things to keep in mind. MRJob uses so-called protocols for data serialization, deserialization between phases. And by default it uses a JSON protocol which itself uses Python's JSON library which is kind of a slow. And so the first thing you should do is to install simple JSON because it is faster. Or starting from MRJob 0.5.0 which I think still in development, it supports UltraJSON library which is even more faster. This is how you can specify this UltraJSON protocol. And again, this is available only starting from 0.5.0. Lower versions use simple JSON which is slower. MRJob also supports a raw protocol which is the fastest protocol available but you have to take care about serialization, deserialization by yourself as shown on this slide. So notice how we cast one to string in a mapper and some to string in a reducer. Also with the introduction of UltraJSON in the next version of MRJob, I don't think there is a need to use these raw protocols because they are not so much faster actually compared to UltraJSON. And at least most of the time, of course, it depends on the job. So you have to experiment for yourself and see what fits best for you. So MRJob processing cons. In my opinion, it has best documentation compared to other Python frameworks. It has best integration with Amazon's EMR which is elastic map reduce and compared to other Python frameworks because Yelp uses, it operates inside EMR so it's understandable. It has very active development, biggest community. It provides really cool local testing without Hadoop which is very convenient while doing development. And it also automatically uploads itself to a cluster. And it supports multi-stab jobs which means that one job that will start only after the second, another one is successfully finished. And you can also use bash utilities or JAR files or whatever in this multi-step workflow. The only downside that I can think of is a slow serialization and deserialization compared to raw Python streaming but compared to how much typing it saves you, we can probably forgive it for that. So this is not really a big con. The next in our list is Luigi. Luigi is also a rapper around Hadoop streaming and it is developed by Spotify. This is how our work count example can be written using Luigi. It is a little bit more verbose compared to Mr. Job because Luigi concentrates mainly on the total workflow and not only on a single job. And it also forces you to define your input and output inside a class and not from a common line interface as for the map and reducer implementation, they are absolutely the same. Four minutes left. Oh, my God. I have so much to say. Five minutes. Okay. So Luigi also has this problem with serialization and deserialization. You also have to use alt adjacent. Just use alt adjacent and everything will be cool. Okay. So we'll probably skip that. It's also cool. Luigi is cool. But it's not so good for local testing. And we'll also skip Hadoop. Okay. Okay. Oh, man. All right. All right. Okay. Benchmarks. This is the most important part. This is probably why a lot of people are there for the benchmarks. So this is a cluster and software that I used to do the benchmarks. So the job was a simple work count on a well-known book about a Python by Mark Lutz. And I multiplied it 10,000 times, which gave me 35 gigabytes of data. And I also used a combiner between a map and reduce phase. So a combiner is basically a local reducer which just runs after a map phase. And it is kind of an optimization. So this is it. This is the table. Java is fastest, of course. No surprise here. So it is used as a baseline for performance. All numbers for other frameworks are just gracious relative to Java values. So for example, we have a job runtime for Java, like 187 seconds, which is three minutes and something. To get the number for PyDup, you need to multiply 187 by 1.86, which will give you 387, 47 seconds, which is almost six minutes. So each job, I ran a job three times, and the best time was taken. And so let's discuss a few things about this performance comparison. So PyDup is the second after Java, because it uses this Hadoop, Hadoop Pipe C++ API. It still takes almost twice as slow compared to the native Java, but another thing that may seem strange is the 5.97 ratio in the reduced input records. So it looks like the combiners didn't run, but there is an explanation to that in PyDup manual. It says the following. One thing to remember is that the current Hadoop pipes architecture runs the combiner under the hood of the executable run by pipes. So it doesn't update the combiner counters of the general Hadoop framework. So this is why we have this. Then comes PyG. I actually thought that PyG should be the second after Java before I ran these benchmarks, but unfortunately I didn't have really time to investigate the reasons, so I just can't say why it is slower, because PyG translates itself into Java, so it should be almost as fast as Java. Then comes raw streaming under Cpython and PyPy, and you probably may be surprised that PyPy, no? Okay. Do you have any questions, or I just can continue? Okay, okay. So, yeah. So it's actually, I'm speaking for a half an hour, and this is a 45-minute talk, so I still have 15 minutes. What is the performance for questions? No questions, you see. Okay, so yeah. Cpython and PyPy. Yeah, you probably may be a bit surprised that PyPy is slower, but actually the thing is that the work count is a really simple job, and PyPy is currently slower than Cpython when dealing with reading and writing from standard in and standard out, so it really depends on the job. In real-world use cases, PyPy is actually a lot more faster than Cpython, so what we usually do, we implement a job, and then we just run it on PyPy and Cpython and see what's the difference, and like I said, in most cases, PyPy wins. So just try for yourself and see what fits best for you. Then comes Mr. Job, and as you see, UltraJSON is just a little bit slower than these raw protocols, but it saves you the pain of dealing with manual work, so just, I think, use UltraJSON. Then finally, Luigi, which is much, much slower even with UltraJSON than Mr. Job, and I don't want even to talk about this terrible performance using its default serialization scheme. So, okay, if we still have a little, like, no, 15 minutes, so I can probably return back. Okay, so we stopped it, I think, this, or this, yeah, this one. So, Luigi, as we just saw, Luigi uses, by default, it uses a serialization scheme, which is really, really slow. So this is how you can switch to JSON, and I didn't really have time to investigate also, but after switching to JSON, I needed to specify and encoding by hand, so I don't know. It's also something to keep in mind. And don't forget to install UltraJSON because, by default, Luigi falls back to the standard libraries, JSON, which is slow. So, okay, pros and cons. Luigi is the only real framework that concentrates on the workflow in general. It provides a central scheduler, which has a nice dependency graph of the whole workflow, and it records all the tasks and all the history, so it can be really useful. It is also in very active development, and it has a big community, not as big as Mr. Job, but still very big. It also automatically uploads itself to Cluster, and this is the only framework that has integration with Snakebite, which is awesome. Just believe me, it provides not so good local testing compared to Mr. Job because you need to mimic and map and reduce functions by yourself in the run method, which is not very convenient, and it has the worst serialization and deserialization performance, even with UltraJSON. The last of Python frameworks that I want to talk about is PyDup. Unlike the others, it doesn't trap Hadoop streaming, but uses Hadoop pipes. It is developed by CRS4, which is a central for advanced studies, research and development in Sardinia, Italy, and this is an example of WordCount in PyDup, which looks very similar to Mr. Job, but unlike Mr. Job or Luigi, you don't need to think about different serialization and deserialization schemes. Just concentrate on your mappers and reduce this on your code, and just do your job. It's cool. Okay, so, pros and cons. Okay, okay. I'll do my best. So PyDup has pretty good documentation. It can be better, but it generally is very good. Due to the use of Hadoop pipes, it is amazingly fast. It also has active development, and it provides an HDFS API based on libHdfs library, which is cool, because it is faster than the native Hadoop HDFS command line client, but it is still slower than Snakebite. I didn't benchmark this, but Spotify guys claims that it's slower. And it is slower because it still needs to spin up an instance of JVM, so I can't believe them that Snakebite is faster. This is the only framework that gives an ability to implement a record reader, a record writer, a petitioner in pure Python. And these are some kind of advanced Kadoop concepts, so we won't discuss them, but the ability to do that is really cool. The biggest con is that PyDup is very difficult to install because it is written in C Python and Java. So you have to have all the needed dependencies, plus you need to correctly set some environmental variables and so on. And I saw a lot of posts on Stack Overflow and on other sites where people just got stuck on installation process. And probably because of that, PyDup has a much smaller community, so the only place where you can ask for help is a GitHub repository of PyDup. But the authors are really very helpful, they are cool guys, so yeah, the answer to all the questions and so on. So PyDup doesn't upload itself to a cluster like other Python frameworks do, so you need to do this manually and it's not so trivial process if you're just starting to work with Hadoop. So this is it. So PyG. PyG is an Apache project. It is a high-level platform for analyzing data. It runs on top of Hadoop, but it's not limited to Hadoop. This is a work-count example using PyG. It will be translated to map and reduce jobs behind the scenes for you and you just, you don't have to think about like what is my mapper, what is my reducer, you just write your pick scripts. And also in most of the time, in real-world use cases, PyG is faster than Python. So this is really cool. It is very easy language which you can learn in a day or two or something. It provides a lot of functions to work with data, to filter it and so on. And the biggest thing is that you can extend pick functionality with Python using Python UDFs. You can write them in Cpython which gives you access to Molybs, but it's slower because it runs as a separate process and sends and receives data we are streaming. And you can also use Jython which is much, much faster because it compiles UDFs to Java and you don't need to leave your JVM to execute your UDF. But you don't have access to libraries like NumPy and Cypy and so on. So yeah. This is an example of PQDF for getting a due data from an IP address using a well-known library from MaxMind. It may seem complicated first, but it's not actually. So in the Jython part, at first, we import stuff, some stuff from Java and the library itself, then we instantiate the reader object and define the UDF which is simple. And it accepts the IP address as the only parameter and then tries to get a country code and see this geo name from a MaxMind database. It is also decorated by the PICS output schema decorator and you need to specify the output of the UDF because PICS is statically typed. And as for the... Then we put this UDF into the file called gyp.py and as for the PIC part, we need to register this UDF first and then we can simply use it as shown like here. So it's really simple concept when you get used to it. Yeah. There is also a thing called embedded PIC, this one. So we already saw benchmarks, so conclusions. So for complex workflow organization, job chaining and HDFS manipulation use Luigi and Snakebyte. This is a use case where they really shine. Snakebyte is the fastest option out there to work with HDFS. But you have to fall back to native Hadoop command line interface, of course, if you need to write something to HDFS. You just don't use Luigi for actual map reduce implementation, at least until performance problems won't be fixed. For writing, lightning, speed, map reduce jobs and if you are not afraid of difficulties in the beginning, you use Spiroop and PIC. These are two fastest options out there except for Java. The problem with PIC is that it's not Python, so you have to learn it. It's a new technology to learn, but it's worth it. And PyDup, while maybe it is very difficult to start using it because of the problems or installation and so on, it is the fastest Python option. So it gives you an ability to implement record reduce and write as in Python, which is priceless. So, for development, local testing or perfect Amazon's EMR integration, use Mr.Job. It provides the best integration with EMR. It also gives you the best local testing development experience compared to other Python frameworks. So in the conclusion, I would like to say that Python has really, really good integration with Hadoop. It provides us with great libraries to work with Hadoop. Well, the speed is not that great, of course, compared to Java, but we love Python not for its speed, but for its simplicity and ease of use. And by the way, if you are wondering what is the most frequently used word in Mark Lod's book learning Python without counting things like prepositions, conjuctions and so on, this word was used 3,979 times and this word is, of course, Python. So this is all I got. You can find slides and code. I used for the benchmarks on slidesharing GitHub. So thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you.
|
Max Tepkeev - Big Data with Python & Hadoop Big Data - these two words are heard so often nowadays. But what exactly is Big Data ? Can we, Pythonistas, enter the wonder world of Big Data ? The answer is definitely "Yes". This talk is an introduction to the big data processing using Apache Hadoop and Python. We'll talk about Apache Hadoop, it's concepts, infrastructure and how one can use Python with it. We'll compare the speed of Python jobs under different Python implementations, including CPython, PyPy and Jython and also discuss what Python libraries are available out there to work with Apache Hadoop. This talk is intended for beginners who want to know about Hadoop and Python or those who are already working with Hadoop but are wondering how to use it with Python or how to speed up their Python jobs.
|
10.5446/20162 (DOI)
|
Thanks. Good afternoon. My name is Max. And today I'm going to talk about why we think that you should stop trying to glue your services together and import Lymph instead. Who are we to get that out of the way first? We're a delivery hero. We're an online food ordering service holding. So what online food ordering is all about is basically the concept is simple. I guess that there's no one around who's really unfamiliar with it. You get hungry and you go online to one of our web pages. You search for restaurants that deliver to where you are. You compile your order. You pay online and then you wait for the food to be delivered. So basically it's like e-commerce but with grumpy customers by definition. And also the fulfillment part is interesting because food needs to be delivered quickly. That's something you need to take into account. We're operating 34 countries and we're headquartered in Berlin. This is our mascot, the hero, therefore, delivery hero. So just a quick show of hands. Who of you attended the talk by Matt about Namiko before lunch? That's a fairly good amount because there are a few things that we're not going to talk about but which Matt nicely introduced. So we're not going to talk about what services really are or as opposed to monoliths or why you should go with a service oriented approach, why you should not do it, how this helps you. Neither are we going to talk about cloud stuff or Docker or why you should call it microservices and not microservices or how micro is micro. But what we're going to talk about is LIMF today. So LIMF is a framework for writing Python services. And to start with, I'd like to justify a little why we wrote another framework because usually developers say, hey, there's something out there already. In this case, there wasn't. So once we have that out of the way, we're going to get our hands dirty with a live demo fingers crossed that that works all right. And that's basically the main section of this talk. And afterwards, we're briefly looking under the hood of LIMF what other features are there which we don't touch today. Briefly touch on things like Namiko and so on, give you a little outlook. And then hopefully there's time for Q&A in the end. So I have to be fair and say, if you want to go over things in your own time, then there's this entire introduction. You can find an article at import minus LIMF.link. Everything is written down there. You can go over things in your own time with even there's even more detail. You will find the exact same examples or services that we're talking about today. You can try things for yourself. There's a vacant box set up that you can use, which we'll use later or just to debrief yourself on what we talked about today. So why do we write another framework? That's pretty simple. Two years or roughly, we've been to the situation where we said we want services in Python and not worry. So let's assume that our decision to go with services was right. We were running with a big jungle monolith, basically a lot of spaghetti code of the legacy variety, the one that no one really likes. So therefore the idea of going with services became increasingly attractive to us. We want you to stick with Python because usually you say, hey, if you do services, the idea is that you don't have to worry which language you're going to run with. But as we like to do Python, every developer should be able to be as productive as they are. And if we would not have stuck with Python, well, then I couldn't be here today and talk about it. So that's good. And we didn't want to worry. That means if you want to run your services and do services, there was nothing really that was helping you a lot back then. So we wanted services in Python and not worry. The first two things are easily ticked off. The third one wasn't. Therefore we came up with LIMP. So we had certain expectations, though. Running and testing your services should be as easy as possible. You should not have to worry about glue. That means I, as an author or operator of a service, you should not have to worry about how to register your service, how to run it, how to configure it. You should not have to worry about any of these glue code at all. Configuration should be simple and flexible. You should get a lot out of your configuration files without having to write a lot of code to pass them and deal with them. Possibly you should take the same service, run it on your local machine, your lab environment, staging live, possibly another country even simply by configuring it differently. Scaling naturally should be easy. So if you need more resources, then you just throw more instances into your cluster, yet the client code should be totally unaware of this. We wanted to have the opportunity or the possibility to speak RPC rather than the HTTP interfaces be able to do to emit events so that we can do asynchronous communication. But we also wanted to easily integrate or expose HTTP APIs. And last but not least, if you want to introduce a new service, then there should be as little boilerplate as possible, yet a fair amount of scaffolding helps you to nicely structure your stuff. So what we came up with was LIMP. You can find LIMP at LIMP.io and we think that it or the idea is that it should satisfy all of these requirements. So I repeat myself here. I think it's a framework for Python services and by default LIMP depends on RabbitMQ as an event system and ZooKeeper for service registry. So just one more quick show of hands who knows what RabbitMQ is. Good. ZooKeeper. Fair enough. So ZooKeeper is a distributed key value store. And that's how we do service registry. But we'll find about that later. So here comes the scary part. People say that I should not animate my slides. I should not show code on slides. And neither should you ever, ever, ever do a live demo because it will go horribly wrong. So I'm going to show you code and it's animated and we're going to do a live demo so there's nothing that could possibly go wrong. So to begin with, and to jump into the thick of it, we're going to write services and we're going to increasingly introduce new services to see how they interact with each other. We're going to run them. We're going to play around with them to explore the tooling that LIMP brings with itself. So we start with the most sophisticated Hello World example you could think of. That's a greeting service. And it's funny because Matt used basically the same example that was not planned, but it's funny though. So this greeting service, you give it a name and what it's supposed to do, it's supposed to return a greeting for that name. So to begin with, we need two files. Sorry, that's the first line already. We need two files. We need the implementation of our service naturally in a Py file and we need a configuration file in a YML file. Sorry, that's YAML. So to begin with, for our service, we start with importing LIMP and this is basically where this talk lives up to its claim. We import LIMP and we want to define a service called greeting and we do so by inheriting from LIMP interface. And like I said, we want to expose one method as its interface. It's called greet. Takes a name and we can expose it easily by decorating it with LIMP RPC. And what we do inside of this method is we simply print the name which we received saying hi to. We are omitting an event to let the world know that we're just greeted someone and lastly we're returning this greeting. The configuration file is rather straightforward as well. We have to tell LIMP which interfaces to run and where they are located on the Python path because what LIMP does is that it imports it at runtime to bring up an instance of this service. So let's get our hands dirty and we'll do this within a prepared vagrant box which is readily accessible for everyone at import minus LIMP.link. It provisions via Ansible and what this vagrant box does is it has ZooKeeper and RebianMQ running inside and to make things even more accessible there are prepared T-Mock sessions which you can easily fire up which then start services and panels and which are nicely labeled using the toilet command and this is what we're going to do now. So let's go there and run our services and play around with it. You don't see the topmost line of my shell which is confusing but you should see everything from now on. So this is where in the box now it greets us very friendly with import LIMP and there is this T-Mock session prepared. We're fired up with Mux, start greeting and what you see now is two panels. One is running an instance of the greeting service on the right hand side. You can see that we are running LIMP instance and we need to point it to the configuration file so LIMP knows which interface to run. On the left hand side we simply have a shell and this is where we'll explore the tooling that LIMP comes with. So to begin with let's say we don't know anything about LIMP at all so it should tell us about the commands that are available. LIMP lists does so. That's a whole lot of text. Don't worry. You don't have to read them one by one. We'll explore things bit by bit. So let's say we have no clue whether there are any services running at all or which there are. LIMP discover should tell us so we can discover services and indeed it tells us there's one instance of the greeting service running as expected. Let's continue to play dump and say we don't know anything about this service. I want to get to know something about its interface. So LIMP inspect greeting should inform us about this service's interface. And this is more than expected actually so the top most function that you see, sorry the top most method that's the greeting method, the one we've just implemented ourselves and below that you see four built in methods which you get by inheriting from LIMP interface. So let's exercise this service. We can do so by issuing LIMP request greeting. Greet will supply the request body that needs to be valid JSON talking and typing at the same time is hard. And I'll greet you guys the Euro Python. So what we'll expect to happen now is the request should hit the instance of the greeting service supposed to print something and we should receive the greeting in the response. So fingers crossed this works and did. On the right hand side you see that it said saying hi to Euro Python and we received the response on our end as expected. That's very nice. On to the next service. So yeah, this is what we just did. So the greeting service is also emitting events every time we're greeting someone. This is something we haven't seen yet and it's emitting events. There's no service that consumes them. So let's write a service that consumes these events. And as creative as we are, we're going to call it listen service and once more we need two files. One where we implement the service and one where we configure it. So we start with importing lymph and we define the service by subclassing lymph interface. We're calling it listen. And like I said, every time an event of type greeted occurs, we want to consume it and we want this method to be invoked. Exactly invoked. Sorry. It's called on greeted and it receives the event that has been emitted and all it does is that it takes the name from the event body and it prints that somebody greeted that name. The configuration is just as straightforward as before. We have to tell lymph that it's supposed to run one interface, the listen interface and we have to point it to where this is located on the Python path. So that can be imported. So let's run them in combination to see how they interact. Therefore, I'm firing up the T-Mock session for that. And we're doing a leap of faith here. We're not only running one instance of each service, but in this case we're running two instances of the greeting service and one of the listen service. But let's assure ourselves whether they have registered correctly with our service registry. This cover should tell us. So as you can see, indeed, there are two instances of the greeting service and one of the listen service. So the listen service is supposed to consume certain events. Let's assert whether that really is the case. So we can event, sorry, emit an event of type greeted with lymph and we have to provide a body, also, once more, it needs to be JSON and the name is EuroPython. So what we're expecting to see now is once we emit it, the listen instance is supposed to consume it and it needs to print something. In fact, that consumed the, you can see this down here, it consumed the event that was emitted before when we were requesting the greeting instance. So we're expecting to see it print again now. Very nice, it printed as expected. So let's request a greeting and see whether they're correctly working together. So we're expecting to see now, once we send this request, we expect it to be handled by one of the greeting instances. It should print, should return to us and the listen service should print once more. And in fact, the second greeting instance handled it. Now if we repeatedly issue this request, the request should be randomly distributed over the greeting instances and fingers crossed, yes, this worked, topmost one handled it and the other one, very good. So this seems to work as expected. But it wouldn't be 2015 if we were not talking about web services. So let's expose our, the functionality that we have within our service cluster, which is the bleeding edge greeting service and we want to expose it via an HTTP interface. What we need to do there is we're going to write a web service and once more, we start out by implementing it in Python and configuring it afterwards. So this case we import from lymph web interfaces, the web service interface and we'll also need some VACSug tooling to deal with URL mappings and since we also want to return a valid response. Business as usual, we define our web service by inheriting from web service interface. We want to expose one URL, which is slash greet, supposed to be handled by the greet method, receives the request and we expect the name to be in the query string. And what we do there is when we receive the request, we pick the name from the query string, we print that we're about to greet someone. We're invoking the greet method of one of the greeting instances. This is RPC basically. Sorry. And then in the end, we're returning the greeting in the response. And we also need to configure it and as our web service is supposed to listen to a port, we have to include this in the configuration. That's the only bit that differs from the two configuration files we've looked at before. So let's run everything together. So what we see here is one instance of each service running, web greeting and listen. And since old habits die hard, let's make sure that they have all registered correctly. Lymph discoverer should tell us. And indeed, there's one instance of each service. Let's exercise the web service now and see whether they're actually working in combination as they should. So we're listening to port 4080 and the name should be in the query string. That's Europe Python once more. So once we issue this request, we expect to receive the greeting in the response and all services, sorry, all instances should print something in order to validate that they were actually being spoken to. Let's issue this request. In fact, all services, all instances printed something and we received the greeting in the response says, hi, Europe Python over here. But there's one thing that you might see already is the more services you run, the more complicated it becomes to develop with them locally. You need more shelves to run the instances. And if you want to run several instances of one service, you need to run them in several shelves that's becoming rather painful. And it has become rather painful here already because we want to run three services. We need three shelves. That's a bit of a pain. But there is LIMF to the rescue. It comes with its own development server. It's the LIMF node command. And what we need to do to get its leverage is within the directory where we want to run our development server, we have to, there needs to be a configuration file called.LIMF YML. And in there, we configure the services that we want to run and we configure how many instances of each. So this is highlighting the important sections. So if you want to configure instances, you basically tell LIMF how to bring up that instance and how often. So we run two web service instances, three greeting service instances and four instances of the listen service. And within the last section, since we have two instances of our web service running and they're listening to a port, we have to configure this one as a shared socket. So let's bring up our node. You won't only see the node on the top right panel, but below that you also see tail. To LIMF tail, you can subscribe to all the logs of any service. So in this case, we subscribe to the web greeting and listen service and it will print all log statements that it receives from the instances. So let's make sure that everything registered correctly because there's no output in the LIMF panel. LIMF discover should tell us that we have indeed two, three and four instances of the services running respectively. And let's hit our service cluster as before with local host. That's great. Name goes into the query string. Once more, Euro Python. And what we expect to happen is once we issue this request, it should be handled by the instances. We should see three print statements now in sequence in the node panel and below that plenty of log output. So fingers crossed this works. Very nice. It did. So you see three print statements up here. Almost reads like a little haiku. About to read Euro Python saying hi to Euro Python. Somebody greeted Euro Python. The response looks good as well. We can see the greeting has been returned as expected. And we see plenty of almost confusing log output below here. But now consider that possibly your instances might be distributed over any number of machines. And if you want to debug something or follow the logs and get information from that, it's hard to tell which log statements belong to each other. How can I relate them to a request? Possibly they well belong to the same request, but the log statements come from several different machines. Let's see overcome this problem with the trace ID. So whenever a request enters the cluster and it does not have a trace ID assigned yet, Lymph assigns a trace ID to the request and this trace ID is being handed forward via every RPC request and event that is being emitted. And then whenever we log, it's being logged with it. So you can see here we hit the web service and returned a header called xtraceid and that's where it included the trace ID. And let me allow to use the iTerms search and highlight function. So you can see the trace ID is appearing in the logs properly and within your own time maybe you can assure yourself that it actually logs correctly with the trace ID and we can correlate all the log statements via the trace ID. That's very good. I managed to successfully go through the demo part and nothing broke. So let's just briefly reason about the communication patterns which we've just observed and I think I went a little bit too far with animating stuff but maybe that's hopefully it's entertaining. So we started with having two, three and four instances of these service running respectively and we issued an HTTP request. It was handled by one of our web instances. It printed something and then we wanted to invoke the greet method via RPC of one of the greeting services. So what happens behind the scenes is that we consult our service registry which is Zookeeper by default and we ask it for all the instances of the greeting service and then we pick one at random to send the requested and in this case it was the lowest one. The request is being sent over, it printed something, emitted the event to our event system which is Rebitem queue by default, return the response and then we had nice output on the shell and on a possibly entirely different timeline, one of the listen instances consumed the event by getting it from the queue and print it naturally. So we see there's RPC available which follows the request reply pattern and it's synchronous communication and we are also emitting events that's the pop-up pattern and that's asynchronous communication. So you've seen, I've jumped slides here already, exactly one instance of all the listen services will consume the event. However, there are situations where you'd like to inform every instance of a service that something occurred and all we need to do is on the lower left you can see that we decorate the service, sorry, the method which is supposed to consume the events as usual but we say that it's broadcast and what happens instead is that when we emit the event we're publishing to three, sorry, to four queues in this case and then it's being consumed four times and that as a repercussion naturally we would have seen four print statements. So these are the communication patterns which are available with LIMP. So what else does LIMP come shipped with? So what's in the box is that what I mentioned already, LIMP manages your configuration files, you can get a lot out of your configuration with very little code. It provides a testing framework so that you can unit test your services following the passion of if I invoke this RPC method is an event being emitted as expected or run some together and exercise them. Its dependencies are basically pluggable so you could exchange ZooKeeper for something else, you could do service registry with something else, you could not use RevitMQ but something else like Kafka for instance. There are service hooks so when your service starts you want to possibly set the stage for it like you want to provision a database connection and then once you shut it down it's supposed to be cleaned up. There are service hooks for this. LIMP allows you to do futures so usually classic RPC is blocking but possibly you're not interested in the reply from the service or you're interested in later so you can do this with a future, you can defer the call. LIMP collects a good amount of metrics when it runs your service already and then it exposes them but you can also collect custom metrics so for instance if your service is talking to third party API and every now and then this request times out and you want to keep track of how often this happens this is what you can easily do. You can also write your own plugins for LIMP, there are even more hooks that you can plug into and get custom code executed whenever something interesting happens. Out of the box there is a New Relic and a Sentry plugin for LIMP. It's easily, sorry the CLI interface is easily extendable, colleague of ours wrote LIMP top which is basically like top but for LIMP services and you can handle, you can receive remote errors, you can get shells on remote services and so on, there's a whole lot more. So how LIMP works under the hood is that anything that is supposed to be sent of the wire is being serialized with message pack. Message pack is, well that's their claim, like JSON but a little smaller and a little faster. RPC is being dealt with with zero MQ. Like I said service registry by default happens with via Zookeeper and events are being, the event system is Rebin MQ. Every LIMP instance or every service instance is a Python process and it handles requests and events within greenlets and this is what we do with G event and everything that is web or HTTP, we use Vectork tooling for that. So since some of you attended the Namiko talk already, as for things that are out there that are similar to what LIMP is, there's one thing that needs to be mentioned of course that's Namiko. Namiko does a lot of things very, very similar, almost startingly similar to how LIMP does things. It naturally does certain things differently but it's very nice and if you haven't attended the talk I suggest that you have a look at Namiko. Also there are other things out there which don't solve the big picture like Namiko or LIMP tried to solve it but they supply solutions for niche problems like zero RPC or other stuff and you would still have to provide a good amount of glue code yourself and this is what Namiko and LIMP both try to avoid. So what we have in mind for the future of LIMP is that we want to have this little ecosystem of libraries for writing special purpose services easily. So there's, we have LIMP storage in mind, LIMP monitor which then collects all the metrics from other services and stores them wherever or does with them whatever you want. LIMP flow which is basically the ideas to write business processing engines which deal with your business process and manage your entities and wherever there is, there is to come. So to sum things up, if you can remember this one thing that LIMP is a framework for writing services in Python, I think I have been successful today. You can find out more at LIMP.io and naturally it's open sourced. Your contribution is very welcome. You can find the docs at readthedocs. Everything is linked there at LIMP.io. Like I said this introduction, it's all written down in more detail following the same narrative as today at import minus LIMP.link. That's where you find all the examples that's where the vagrant boxes, that's basically where you can go and play around with LIMP. Last but not least, if you are Spanish speaker and you'd like to hear this talk again later this week in Spanish, then my colleague Castillo will give the same talk in Spanish. In Spanish, I had to learn this by heart. It's, I don't know whether I made a good effort but I see you are nodding. Very good. So that worked. And here comes the shameless plug that goes with every talk. So we're hiring. If you're interested in working with us in Berlin and you want to work with LIMP, you find that interesting. You possibly have seen this flyer in your attendee back already. Feel free to reach out at deliveryhero.com or see us at our table in the hall. We've brought goodies and gummy bears, most importantly. Thank you. And thanks to the organizers, of course. Questions? Or maybe you can just shout and then I'll repeat the questions so everyone can hear. Sorry if you would like to skip. Your first question, well, theoretically it's possible to talk to LIMP services but you would have to do it yourself. So you would have to basically re-implement the, exactly within any other language. And your last question, I didn't get it. So your question was whether you can use LIMP node. The idea behind LIMP node is to be used in development. It's not the idea is not to replace your production, how you should run it in production. The idea is to help you run stuff locally. Yes, you can spawn anything with the LIMP node. Basically, you just supply a command. I don't know, for instance, to run your Redis server. You could include it there as well and then it runs it for you. Hi. Just to correct. What versions of Python do you support? And if you support Python 3, why didn't you use something like IO HTTP or modern things in place of Verkzerk? So to answer your question, yes, both Python 2 and 3 are supported. I don't know with which Python 3 version it starts. I know that we have had a little trouble with this in the past. But we're supposed to support Python 3 as well. And your other question, IO HTTP? I don't know about that, but to preclude your question already. But other, yeah, sorry. So about message versioning, do you support that or that something somebody will have to do on top of your? Message versioning? Yeah. So you run two different versions of a service in a cluster. You want to, see, you're running two different versions of a service and there's nothing that deals with it out of the box right now. You would have to deal with it yourself, I guess. I mean, it depends on whether the interface is backwards compatible, but there's, you can't, it should, would be another service then. If you want to run two different versions, right now you would have to run two different services or just expand the interface. Yes. Thank you for a great presentation and it worked out. It's amazing to see such a polished software which promise a lot. But what's on your, in your backlog, what's the issues you're working on? What's your roadmap for developing it further? So the, the idea is to further let these special purpose libraries, I'd like to call them that, to let them further mature and then release them as open source at some point. But right now the idea is to simply make Lymph more stable. I mean, we're going to run it in production anyway. So this is something that will naturally, and that will naturally grow and mature within the future. Hi. I don't know if I got it right, but I understood that you're using a zero in queue for the RPC, handling the RPC calls and rapid MQ for handling the events. Have you considered perhaps using rapid MQ for both and then getting rid of one extra dependency or did you experiment but didn't work or? For instance, Namiko uses, uses RPC goes via zero MQ, sorry, via rapid MQ there as well. For us it was a design decision not to, to do RPC over, over something persistent as rapid MQ. Actually my question was going the other way around if it is also possible to replace rapid MQ with a zero MQ for the pub, pub sub. Yeah, definitely. So is it, but you said it is pluggable, but is it already implemented or is it something else? So I was expecting your question, therefore I prepared something. So this is, this is the part of the.lib YML which I actually didn't want to show because it's a little confusing. However, what you can see here is this is actually where things are pluggable. So you could provide another class which does registry or handles events. This is what it looks like by default, but you could provide your own back ends for either. And I see everyone's eyes narrow so yes it is confusing. And just one more quick question. I've seen in the YML files that were, at least maybe in the simple examples, just names of services and class path. So wouldn't it, would it be possible already just to have a decorator that defines the name and then just launch it like getting rid of the YML file and just launch the instance just providing the path to the class? Well in theory, yes, that's possible. But the idea behind this is that you could group several interfaces together and run them as one service. And I think this is just more flexible because then you don't start to mix what's in your configuration, what's not. This way you have everything in your configuration and that's where it is. Thank you. Cheers. Nice. Cool. Thanks guys. Thanks. Thanks guys. Thanks. Thank you.
|
Max Brauer - Stop trying to glue your services together; import lymph What if you could focus on functionality rather than the glue code between services? Lymph is an opinionated framework for writing services in Python. It features pluggable service discovery, request-reply messaging and pluggable pub-sub messaging. As our development teams are growing, we're moving away from our monolithic architecture. We want to write services and not worry about the infrastructure's needs. We want development to be fast, quick and simply work.
|
10.5446/20159 (DOI)
|
Thank you. Welcome everybody. Thanks for coming. This talk is about microservices and Nomeco, which is an open source library that you can use to write them in Python. My name is Matt Bennett. I'm the head of platform engineering at a company that's currently still in stealth mode, so I can't talk about it while I'm on camera. Previously, I was a senior engineer at One Fine Stay, where Nomeco was born. How many people in the room know the phrase microservices? That's quite a lot. How many of you knew about it two years ago? Not that many. Microservices is the hot new buzzword, and suddenly they seem to be everywhere. In November 2014, Martin Fowler and James Lewis published this paper, microservices, which I think is considered to be the seminal paper on the topic. I highly recommend it. It's very accessible. It's not very long, and there's a lot of information in there. It's also very recent. They didn't invent the term microservices, but they gave it a concrete definition and really propelled it into our vocabulary. At One Fine Stay, we discovered this paper when it was published and realised that it described what we'd been building for some time. That was really exciting because suddenly we had a common language with which to share ideas about this stuff. For the uninitiated, what are microservices or more correctly, what is the microservice architecture? This is Martin Fowler's definition. It's an approach to developing a single application as a suite of small services, crucially, each running in its own process and communicating with lightweight mechanisms. I think it's helpful to contrast microservices to a monolith, which is probably your default way of building an application as a single process. Your typical Django site is a good example. You would probably compartmentalise your logic into different apps in Django parlance, but ultimately they would run in the same process and memory space as each other. Whereas in microservices, your apps become entirely separate programs. In essence, this is good old fashioned decoupling and encapsulation, but applied at the process level. What this forces you to do is consider the boundaries of the services or the seams that run through your application. A common response to the hype around microservices is you should be doing this anyway. That's just good design, which is true, but with microservices you can't be lazy and do, say, a cross component import because it's not there to import. There are other benefits to using separate processes as well. Some reasons for adopting microservices is the primary reason for adopting any software architecture is scale, or rather, maintainability at scale. I don't mean scale in terms of serving hundreds of millions of requests a second, but rather in the complexity of the problem that you're trying to solve and the team that is charged with solving it. There's an analogy for this that Alan Kay, who is the inventor of small talk and object-oriented programming, used in a 1997 keynote, which I watched a video of because I was 13 in 1997. It goes like this. If somebody asked you to build a doghouse out of wooden planks and nails, you'd probably be able to do a reasonably good job, reasonably sound structure. If they ask you to then scale that up to 100 times the size using the same equipment and tools, you couldn't do it. The thing would collapse under its own weight. When society started building massive structures like cathedrals, we used stone arches to support the weight of the structure. I had this light bulb moment where I realised that the etymology of the word architecture is literally the application of arches. How can microservices help you achieve maintainability at scale? We've already said it's about decoupling and encapsulation, but what else? As separate programmes, they're independently deployable, which means you can have separate release cycles and separate deployment processes for each part of your application. The Guardian newspaper have written about how they've embraced microservices, and it's allowed them to start using continual delivery and iterate very quickly to deliver one part of the application, but without putting some slower moving or more legacy or more risky parts at risk. Separate programmes are also independently scalable. Now I am talking about serving hundreds of millions of requests a second. To scale a monolith, you have no choice but to replicate it and deploy another instance. You have to replicate the whole thing. Microservices are much more granular and therefore more composable, so if you have a service that is very highly CPU bound, you can deploy more of those across more CPUs without having to drag along the rest of your application as well. There's also a freedom of technology. Being good Pythonistas, I'm sure we all really want to use Python 3 where we can, but sometimes we get stuck using an old library that hasn't been updated, and therefore we can't. We're stuck in Python 2. In a monolith, you have to use the lowest common denominator. Microservices are individually free to use the most suitable interpreter for them, so Python 2, Python 3, Python 3 is up to you. I perhaps shouldn't say this too loudly at a Python conference, but this extends to your choice of language as well. If you want to experiment with something functional like Haskell or Erlang, you can write a service in that language. Forgive the circular reference on this one, but microservices are not monolithic. Outside the realms of software architecture, a monolith is something that's big and imposing and impenetrable. Think that the monolith from 2001, space obviously. Whereas microservices are small and nimble and quick, they have a smaller code base, which means it's easier to bring a new developer on board and understand the whole thing. There's a lower cognitive overhead to understanding how it works, which is inherently more maintainable. And then there's Conway's Law. How many people have heard of Conway's Law? This is something that ThoughtWorks talk about a lot. In 1968, a chap called Melvin Conway said this, organisations to which design systems are constrained to produce designs which are copies of the communication structures of those organisations. In 1968, it doesn't seem like there are any new ideas in software architecture. If you have your regular three-tiered web application, you have a database layer and an application logic layer and you have a user interface layer and you likely employ specialists that work in those areas. I've worked in a team like this. As a member of the middle tier, I would be able to talk to my application developer peers every day and it would be really easy for us to communicate. But then when we went to speak to the UI folks, we used a subtly slightly different language and there was this layer of friction that meant that we made mistakes and it was harder for us to communicate with them. That's Conway's Law in action. What ThoughtWorks recommend instead is you build small, multi-disciplinary teams and then you separate them based on the natural divisions that exist within the organisation that you're serving. As a result, you get an application that better reflects the organisation rather than these somewhat arbitrary technical boundaries. These wonderful benefits are all well and good. What does it cost? It's a grown-up architecture. There are a lot of things that you have to have in place before you can make it work for you if you want to avoid the architectural doghouse. There's a DevOps overhead. If you're increasing by 10 or 20 times or 100 times, the number of things that need to be built and deployed and looked after, that's a massive burden for an operations team. The only way to cope, really, is to leverage automation. For your tests, for your deployment, for your machine management. Another insight from ThoughtWorks is that microservices are a post-CD architecture. What they mean is that it's enabled by automation and without automating your tests and deployment and machine management, that this burden would make microservices completely infeasible. I think this is why microservices all of a sudden seem to surround us. It's the same good ideas of decoupling and encapsulation, but with this new dimension being enabled by DevOps tech. As well as the DevOps overhead, you also have to embrace the domain in which you're operating. You could argue that for a sufficiently complex application you should really be doing this anyway, but I've certainly worked places that didn't. What I mean by domain knowledge is you have to really understand the business requirements, i.e. the problem that you are trying to solve for your organisation. You have to do that so that you know where to draw the lines between your services, how to divide your application up. You can't just build a web app and then tack things on as they become apparent. Microservices forces you to do this up front. Then there's the decentralized aspect. You no longer have a single source of truth like the traditional database layer. You have to relinquish, that means you have to relinquish, asset guarantees and instead embrace base, which stands for basically available soft state, eventually consistent, which is a really awkward background in, but a good chemistry joke. What this means is that you can't apply transactions across calls to multiple services. You have to apply it in one place and then wait for those changes to be propagated and reflected in all the other places. That's eventual consistency. At one fine day we made a mistake in this realm. We built an abstract calendaring service that handled the data, the calendaring data and calls to several other services. Because they were in separate, the calling service and the calendaring service were separate, we couldn't apply transactions across them, which is kind of a rookie error really. What ended up happening was the calling service would write to the call the calendaring service and write something to the calendar, which would succeed or fail. If it succeeded, the calling service would then do something else. If that something else failed, we had to catch it explicitly and then call the calendaring service to say, oh, just please undo the thing that we've just done to you. We couldn't just roll back a transaction to achieve that. That's an unnecessary loop that we forced ourselves to jump through. Of course, there's also the race condition where, while the calendar has something in it that we end up removing later, something else looks at the calendar and they say that it's full, but it should actually be free. The decentralized aspect means you have to think about these things. You have to be aware that you are introducing complexity. A collection of microservices is fundamentally more complex than a monolith. There are more moving parts, and those moving parts are connected by a network, which is inherently less reliable than in memory calls. In a complex system, failures rarely happen for exactly one reason. It's usually a cumulative effect of various soft failures adding up. Your network slows down in one area of your infrastructure, which causes a backlog of requests, which, combined with a recent code change, means you're writing more to disk, which means that you run out of disburs. It's only when you get to the fourth or the fifth of the nth soft failure that you actually fall over. To mitigate this, you need monitoring and telemetry, and you need analysis of the data that that produces so that when something goes wrong, you can figure out what it was, what caused it, or preferably you figure it out before it goes wrong. By now, you may be asking yourself whether microservices are right for you, and if so, here are some questions to consider. Is your code base large enough that no one person understands it? Are your dev and release cycles slow because of changes of dependent changes that need to be made? Do your tests take forever to run? If so, you might be fighting a monolith. And if that's the case, are you ready to support a distributed system? Are you leveraging automation for your tests, deployment, and machine management? Do you have sufficient monitoring and analysis in place to figure out what's going on inside it? If your answers to these questions are yes and no, respectively, they're not. Maybe you can build a multi-lith. So this is a term that I came up with yesterday, so I'm not sure whether it will stick, but it serves the purpose for the presentation. There is a sliding scale between tens or hundreds of microservices at one end and a single monolith at the other, and this is a continuous spectrum. So you may choose to augment your existing monolith with one or two satellite microservices, the multi-lith. And this way, you get some of the benefits, like you could choose to use a different interpreter, or you could try out CD without most of the cost. So assuming we're all emboldened and ready to embrace microservices or a multi-lith, let me talk about Nomeco. So it's an open source Apache 2 project, and it's a framework that is designed for writing microservices. We named it after the Japanese mushroom, which grows in clusters like this, and we thought it kind of looked like microservices with many individuals making up the larger thing. So I asked a botanist friend of mine why they grew like that, and he shrugged and said, because there's not much room. True story. So there's a couple of important concepts that I need to introduce to explain some of the design principles in Nomeco. There are entry points, which are how you interact with a service. So this is how you request something from it or otherwise get it to do something. Entry points are the interface or boundary of a service, and there are dependencies, which is how the service talks to something external to it that it may want to communicate with. So, for example, a database or another service. So if we jump into some code, I put the code in the following examples in a repo on GitHub so you can grab them later if you want. Nomeco service is written as a Python class, so it has a name, which is declared with the name attribute, and it has some methods that encapsulate the business logic of the service. And then the methods are exposed by entry points. So this HTTP decorator here will call the greet method if you make a get request to that URL. So if I expand this example slightly and let's pretend for a minute that string formatting is really expensive and we want to cache the greetings rather than generate them every time. I've also switched out the entry point, so now it's a remote procedural implementation as opposed to HTTP. So the first thing to notice is that the business logic of the method is unchanged by twitching out the entry point. We've added logic to deal with the cache, but it's entirely isolated from anything to do with HTTP or RPC. So, in other words, it's a declarative change that has no impact on the procedural code in the method. And the second thing to point out is that the cache is added as a dependency. So this line here, cache equals cache client, is the declaration of the dependency. So dependencies are special in the maker. You declare them on your service class like this, but the class level attribute is different to the instance level attribute that the method sees when it executes. And that's because the dependency provider, which is our declaration, the dependency provider injects at the instance level attribute at runtime. So if we hacked our method here to print these two attributes when it runs, we'll see that they're different. So the first one here, the top one, cache client is our dependency provider. And the second one is actually an instance of a memcache client object, and that's what the dependency provider injected. So using dependency injection like this means that only the relevant interface gets exposed to the service method and the service developer. All the plumbing of managing a connection pool or handling reconnections is nicely hidden away inside the dependency provider. So this emphasis on entry points and dependencies also makes Nomeco very extensible. All entry points and dependency providers are implemented as extensions to Nomeco, even the ones that we ship with the library, which we include so that it's useful out of the box. But the intention is that you're free to and encouraged to build your own, or maybe through the wonders of open source somebody will have already built it for you. So this is the list of built-in extensions. So the RPC decorator that we saw earlier is an MQP-based RPC implementation that gives you a request response type call over a message bus. There's also a published subscribe implementation that gives you asynchronous messaging over MQP, and there's a timer for cron-like things, and there's experimental web socket support. So I think it's worth explaining why we have this MQP stuff in here. So HTTP is a natural starting place for microservices. There's a lot of great lightweight web frameworks out there. There's great tooling around API exploring and caching, and HTTP is ubiquitous. And you're probably going to need HTTP on the outside of your services so that clients can interact with it. But for service-to-service interaction inside your cluster of microservices where you control both sides, you probably want something other than HTTP. In particular, Microsoft Sub is a killer app for microservices. There are all kinds of patterns for distributed systems that rely on asynchronous messaging with fan-out capabilities and stuff. MQP is really great for that. So that's why we include it out of the box. But you don't have to use it. There are also some really great test helpers in Nomeco. So we've already seen how injecting dependencies keeps the service interface clean and simple. But it also makes it really easy to pluck those dependencies out during testing. So in this snippet here, we're using a helper called the Worker Factory, which is really useful when unit testing services. So you pass it your service class, and it gives you back an instance of that service, but with its dependencies replaced by mock objects. So you don't need a real memcache server, and you can exercise your methods by calling them and then verifying that the mocks get called appropriately. The Worker Factory also has another mode of operation where you can instead provide an alternative dependency. So in this case here, we're providing an alternative dependency, and we're using the mock cache library, which for this test has a much nicer interface. We don't need to set up the return value or anything like that. And there are other helpers in Nomeco that do this kind of thing for integration testing, to help you with your integration tests, that let you run services with mocked out dependencies or disabled entry points so that you can limit the scope of your integration to the service interaction to the things that you actually want to test. So to summarise, in the microservices architecture, you split your application into services, which run as their own processes. And this is a way to achieve maintainability at scale so that you can build cathedrals of software. And it comes with a host of other benefits, too, like freedom of technology, decoupled release cycles, even team structure, if you want, for each component part. But it's a grown-up architecture. You're building a complex distributed system, which means you need to automate your DevOps, you need to monitor it, and you need to analyse the result of that monitoring. And you need to overall be aware that you are building a distributed system on all of those distributed trade-offs. But you can also adopt incrementally, you know, by adding one or two satellite microservices to your existing stat. And if you want to go on this microservices adventure, there's an open source library that can help you with it. It's made for writing services, encourages you to write clean, highly testable code. There are several built-in extensions, so it's useful out of the box, but it's designed to be extended to your use case. So if you want to know more, read the docs, fork the repo, and with that, thank you very much. Thanks for the talk. We finished a bit early, so there's lots of time for questions. My question is, it seems that there is an implication that you migrate stuff towards, you know, the monolithic to a more microservice kind of architecture. But is it a good idea to actually start doing microservices from the beginning? Is it a sound idea to try and start something small? That's a bold move, I would say. So in that paper that Martin Fowler published, he talked about microservices, and then he says, you probably shouldn't start with microservices. I think that probably depends on your prior experience, what your roadmap is, whether you're starting from a blank slate or not. It's kind of a trade-off. Thanks for the good presentation. Is there a big enough open source project that is an amicode that we can look at as a real-world example? Yeah, that's a good question. So Nomeco is heavily in use at One Fine Stay, which is closed source, and there are a number of other smaller London startups that are starting to use it. I don't think there are any public open source applications that are using it. Hi. It seems to me that Lymph that is going to be presented this afternoon is very, very much similar. Do you, I mean, maybe it's a crazy idea, but why not try to take up the best ideas of both and build something similar? Yeah, so I'm excited to talk to the guys from Lymph later. So we, oh, hi, hi. We've had some email exchanges about stuff, and EuroPython is our opportunity to get together and talk about sharing some ideas. So I have a secondary question, which is more technical, but is there a way, I mean, I looked at the API and the documentation this morning, and it seemed pretty nice, but there's one thing that's missing that is just a simple XMLRPC of Python provides is the possibility to do introspection in the methods. You want to know what arguments are expected by a certain method. You want to have access to the doc string of the method. And I couldn't find in the code or in the documentation if there was a way to do this with Nomeco. Well, so the entry point decorators don't, they don't mutate the service methods. So you should be able to take your service class and inspect it like a regular, like a regular class and look at the doc strings of the methods that you've implemented in it. You import your service class and inspect that. So you don't want to import in the, on the client side, you don't want to import a service class because sometimes the service code will depend on many things like, I don't know, database interfaces and whatnot. So on the, on the client, you really do not want the service code. So you want to be introspecting on the client side? Yeah. Like at runtime? Yeah. Okay. So the simple examiner PC allows you to do, by this service.system.list method, for instance, or whatnot. I mean, this is really useful for, to develop a client independently of the service. So one thing that we have, you know, bounced around for a while is the possibility of a client library where you can, from your service, you can export something that the client can then interact with. And otherwise you're talking about shipping schemers over the wire, which is also a possibility. I think that's how actual RPC does it. This, you know, I would put that in the category of fun extensions that you can add. You know, Nomeco is actually quite a, quite a young library, certainly as an open, as an open source project that's being promoted. So there's a whole bunch of possibilities like this that, you know, I hope that we get to. I'd like to ask if there is any ongoing efforts to make a Pai-chi-kaf-ka interface as a message bus. So again, this is part of the extensibility. So we, we, at one place today, we used MQP very heavily. And so we built, and the built in things we built because we needed them. And then we shipped them with the library because we think they're useful. But yeah, using an alternative message bus or using, you know, zero MQ or, you know, any, any alternative communication mechanisms falls squarely into the category of, you know, this is an extension. Let's build it. Let's build an entry point for it. And I hope that that's what, what happens. Okay. So basically right now there's a kind of, are you building right now anything beyond what's been on the slides? Yes. Yeah. Okay. I think that's not, it's not Kafka. It's not Kafka. Okay. Sorry. Hi. Thank you for the talk. Sure. Just have a basic question. So when you go for a Microsoft's architecture, so you need to be sure in advance that two services will never need to share memory in the future. Otherwise, it can be quite a large amount of work to merge them together, isn't it? Sure. Yeah. Thank you. Hi. Thanks for the talk. It looks like one of the hardest things to do is transactions. Do you have any suggestions on how to approach the problem? Not really. These transactions are a wonderful thing that we've, that we have got used to and you don't, you don't lose the ability to have, have atomic transactions in microservices, but it's within the scope of one individual service. So, you know, you need, that's why dividing your application up is difficult because you need to make these decisions about, you know, where do you put these boundaries so that you can have atomic transactions in the places where it matters and fall back to eventual consistency for other things? Hi, Ydyn Mar. Great talk. I'm sure there's a few of us here in the room who are working on monoliths. Do you have any suggestions on how to approach, say, refactoring it into microservices? Yeah. And what to watch out for? Go for the multi-list. So this is exactly what happened at One Fine Stay. We built this, we built this Django app, which is still our front end, and it accumulated all this logic about, you know, bookings and payments and financial stuff. And it just became unwieldy. And so the journey started with, okay, let's take, it was actually, let's take a piece that doesn't yet exist that we know is going to be really hard to add into this gigantic code base. And let's just build that as a separate thing, you know. So the first, a good candidate for the first microservice is a new thing that you need to do, make that separate, and then maybe you can identify another segment within your app that's, you know, reasonably decoupled already, and then you can, you can move that out. You know, there's no, you can't really answer that question in anything other than abstract. Hello. Are there any situations where you wouldn't recommend to use microservices? Any integrations? Any situations where you wouldn't recommend it? Yeah, yeah, because, I mean, you cannot use it for everything, can you? So you probably don't need it if you're, if you're a developer team is, you know, two or three people strong. You definitely don't want, you shouldn't do it if you're not prepared to support the distributed system aspect of it. If you don't have automation in place, you know, for your DevOps, it's kind of, it's kind of a big commitment. So you only really want to start going down this road when you know that you've got the relevant things in place because otherwise you come and start pretty quickly. Hello. Have you tried doing using Namco on platform as a service? That's one thing. And how do you approach like configuration? I see that you declare a service, but where is the services configuration taken from? So on platforms, in platforms of a service, I think there's too many, there's too much recursion there for me to get my head around what we're offering. But let me come back to that. On the config stuff, yeah, so I didn't show it, but you can provide on the command line YAML file that contains your config and then dependencies have access to that config. So you can, it excluded it for simplicity, but you would probably specify the config key to look up for say the memcache location when you declare your dependency. And then it would know to go and read that element of the config file. So if you look at the code on GitHub, I've done that. Do you maybe have a kind of environment variable parser like for 12 factor apps that are configured through environment variables? No, not yet. And last question, how do you run your services like with Ganycorn or with micro with G? So there's a command line interface in Namco. And then we just run that with supervisor. Hi. So I've already been using Namco and I was wondering if there's any interest in doing a sprint this weekend on Namco. Yeah, totally. So I fly out on Saturday night, but all of Saturday. Okay. I'm up for that as well. Hi. I always wonder how micro should be the microservices. It's a bit unfair to compare only against the monolith. And as you said before, it's just a good architecture if you have components. So you said you should do microservices if a single person cannot keep the whole code in his mind. But well, if you split it up into medium sized services that still fit into a human brain, well, so I suppose to having basically every route of your jungle configuration as its own service, I find that I don't know where is the limit. So that doesn't that doesn't seem like a good idea. The term microservices is actually kind of unhelpful because it implies a size, which I don't think really, I don't think really applies. You know, one fine say, we perhaps didn't make all of the decisions correctly about how you know how to divide our application up. But we had some services that were minus skill, you know, just a couple of methods. And we had others that were, you know, could only just fit in somebody's brain, you know, thousands of lines. So it's an unhelpful classification. You probably, you know, it's very unlikely that you'd end up with lots of services that are all the same size, you know. So, yeah, the, the, the granularity deciding where to draw the lines, that's, that's the hard bit really. So, you talk about built-in extensions. Can you extend, can I extend and write my own extension? Like if I want to support in the protocol. Yeah. So, what was your example, what was the example right at the end? What was the suggestion right at the end? The last thing you actually wanted to do? I don't know, like support in the protocol. Okay, yeah. So, yes, you can. You absolutely can. So entry points are harder to write because it's a bit more machinery. But the, but dependencies are pretty easy. So, if you, so, if you want to talk to a different type of database, for example, that's easy. If you want to send a message or put a message in SQS, it's a relatively easy thing to do. Does anybody have any more questions? Oh, there's one. Yeah, you mentioned some of the key points to go into, or the hard points to when you go into microservices and DevOps. And one of the things you said you have to have in place is really good monitoring. If you consider covering some support in Namco for having some sort of approach to monitoring. Right. So, the thing we used at One Fine Stay, which worked extremely well, was we used Logstash and Elasticsearch. So, every time an entry point fired, we would dispatch a message, stick it on a queue that would be ingested by Logstash and put in Elasticsearch, and then we used Kibana to explore the data. So, you could see which methods got called, and then through the call stack, you could see which methods called them and which arguments were, and which ones generated errors, and how long they took, and what size the payload was, and you could build all sorts of cool graphs so that you can see spikes and explore it. And that worked really well. So, that didn't get open sourced before I changed jobs. So, I'm currently in the process of re-implementing that, and that will become one of the first open source things. Thank you. Right. So, if there are no more questions, please thank Matt.
|
Matt Bennett - Nameko for Microservices Microservices are popping up everywhere. This talk will explain what this fashionable new architecture is, including the pros and cons of adopting it, and then discuss an open-source framework that can help you do so. Nameko assists you in writing services with well-defined boundaries that are easy to test. By leveraging some neat design patterns and providing test helpers, it also encourages good service structure and clean code.
|
10.5446/20156 (DOI)
|
Thank you for coming and welcome. Thank you. Hey, thanks, everyone. You all hear me? I think you can. Good. Excellent. Thanks for having me at Python, EuroPython. My last big Python event was Python, PyConJP in Japan last year. I didn't get to speak, though, but it was really fun. Although most of the talks are in Japanese and my Japanese is getting better, it's not so great. My Spanish is really, really bad. But Spanish and Japanese are very similar, so maybe I should learn both together. No, really, seriously, they are. I know it doesn't seem to make sense. We're going to talk about containers and having containers and having lots of containers, because ultimately, everything is going to be containerized and we're going to have lots of containers. We won't know what to do with. And I'll ask you some questions, Ada, and see how far you are along with moving towards containerization. So basically, when we have lots of containers, what do we do then? And this is a problem we face at Google. So this is a data sensor. This is a Google data sensor in Iowa in the US. It's a place called Council Bluffs, and this is one of our bigger data sensors. And if I leave it out for a long enough, you'll probably be able to count all of the machines. How many there are. But this is a cluster. So clusters are one of the constructs we have internally, but these clusters are broken down into cells. So cells are smaller. We have many cells per cluster. And this will probably, a cell we're going to look at today, is going to have about 10,000 machines in it. So they're quite large. And this is a huge amount of compute power. Lots of compute power, but we need to make this available to our engineers, our software engineers, our developers. So how do we go about making this compute power available to our own developers? And it works something like this. This is what a developer does. Well, first in context, the one thing, given what you see there, we don't want the engineers to have to kind of select a rack, select a machine, and say, hey, I'm going to run it on that machine. I'm going to SSH, SFTP a binary over to the machine, SSH into the machine, stand up my process, my server or whatever, maybe log into many machines and do that multiple times. But that's not going to be possible. Huge amounts of machines, huge numbers of engineers, huge amounts of jobs to run. So how does it happen? So basically, we have a configuration file. In this case, it's called a Borg configuration file. I was in India recently, and nobody there had heard of Borg. How many of you are familiar with Borg and Star Trek? Okay, right. So we never used to be able to talk about Borg because Paramount Pictures own it. It was kind of like one of our worst kept secrets that we had this thing called Borg running internally. But now we talk about it all the time. So because it's fun, and it's really good to show this in the context of what we're going to talk about later, which is Kubernetes. So basically, this is a Borg configuration file, and what the developer does is he creates a job. Jason file, calls the job Hello World, says which cell he wants to run it in. Going back to what we said earlier, a cell is a few thousand machines. In this case, he's saying it's called IC, some random cell name we chose. And he specifies what binary to use. In this case, Hello World web server. So he wants to run Hello World on a web server. And this is going to be a fat binary. Statically links all of its dependencies with it. So effectively, we can run it pretty much anywhere without having to worry about the underlying operating system. And that includes the web server as well. So this thing is quite big, probably about 50 megabytes. So he specifies the path to his binary, or her binary. And unfortunately, we have too many male software engineers, not enough female software engineers. So let's encourage women to be software engineers. And arguments. We have to specify some arguments for our binary, pass them in via the environment. In this case, we want to specify what port to run on. This is parameterized. Then we have some requirements in terms of resources. Now, this is important. We'll circle back to this in a minute. So we can specify how much RAM, how much disk, how much CPU. And ultimately, we can say how many we want to run. So in this case, we want to run five replicas of this job, five tasks effectively. And why five? Why not do it at Google scale, 10,000? Makes more sense, right? We have all those machines. We saw how many machines we have. So let's run 10,000 copies of this. So once we finish this, our software engineer, she types in a command on the command line, passes in the config file. And that gets pushed out to somewhere. It gets pushed out to this ball scheduler. And what happens then is this. Over a period of time, in this case, about two minutes, 40 seconds, 10,000 tasks start. 10,000 instances of that job start. And it takes two minutes, 40 seconds, roughly. We do phase the rollout of all of these jobs to make sure we don't do them all at once. One of the key factors here is the size of the binary, 50 megabytes, 10 times 10,000. It's about 20 gigabytes per second of IO. We're going to be cashing that binary quite a lot, but we had to move it around between 10,000 machines. So there's a huge amount of IO going on. But eventually we get to a point where we have 10,000 running, or nearly 10,000. Maybe not quite 10,000. We'll talk about that in a second. And Borg looks like this. This is what Borg is to Google. It's not going to assimilate you, but I think we came up with a name because it's probably going to assimilate everybody eventually. So this is Borg. And Borg runs within a cell. So each cell has its own Borg master, its own Borg configuration. In this case, we have a Borg master, which is highly replicated. We have five copies of it for resilience. And we have lots of other things. These down here are our machines. These are our machines we saw in the racks. They're all running a thing called a Borglet. We have a scheduler. We have some configuration files in the binary. So what happens is the developer, the engineer, creates his or her binary. And they use a massively distributed parallel build system called, well, I won't say what it's called, but it's externally available now called Bazel. So we made this open source. So our own build system is now available open source called Bazel, B-A-Z-E-L, or if you're American, B-A-Z-E-L. It gets very confusing, believe me. If you go to Canada, it's so confusing. Like routes and routes. So basically, he or she creates the binary, pushes it out, and it gets stored in storage for the cell. And then they push their configuration file. Configuration file gets copied to the Borg master. We have a persistent pack source back store, consensus based. And what happens then is this scheduler, looking around, comes along and says, hey, what is the desired state? We should have this running. Do we have this running? And it sees 10,000 new tasks. And it says, hey, they're not running. We should have 10,000 of those. Let's make sure that's happening. Let's fix that. And so it goes about planning the running of these 10,000 tasks. And it creates a plan, and then it starts telling the Borg master what the Borg master makes decisions and tells them the Borglets on these machines to run this particular task. So they get communicated. The task will ultimately run inside a thin container wrapper. So it has a container around it. It's not just running the binary. It is containerized. A very lightweight shim container that's not Docker. It's not standards-based. The Borglet ultimately will pull the binary over from storage, and it will start running. And we'll see this. Lots of Hallowels. All over our data center. So now we're running multiple copies of that. And so that's what we had, 10,000. But if we look at it a little bit closer, we find there's 9,993 running. Not quite the 10,000 we expected. But this is a highly available service. We expect some lessening of the number of tasks we're running over time due to the way we operate. And that's interesting. So let's look at that in a little bit more detail. So failures. Things fail. But failure is kind of more of a generic term here. There are many reasons for failures. And one of the main reasons for failures, particularly for low priority jobs, is preemption. If we look at the top bar, which is our production jobs, we have very few failures. And most of them are down to a machine shutdown, where we've actually scheduled some maintenance on the machine, and we've taken the machine down. That task, any task running on that machine, we don't be rescheduled elsewhere in the cluster. We have a very small number of preemptions. Down here are non-production jobs, which are things like map producers, batch jobs. They get preempted all the time. They're happy to be preempted. And in fact, the calculation generally says that for about 10,000 tasks, about seven or eight of them will be not running at any given time because of preemption. They'll be about to be scheduled somewhere else, but they won't be running at that particular time. And we see other things here. We see, again, the blue bar, which is the machine shutdown, which is pretty much the same as production. And we have some other things as well, out of resources, very small number of machine failures. And for when you have as many clusters, the many machines as we have, machine failures are given. We expect that. We don't panic when machines go down. It's part of the normal running of our business. And another interesting thing is how we try to make efficient use of our resources. So we have CPUs, we have memory, we have disk.io, we have network.io. And sometimes it's quite possible for one task to be using lots of memory for very little CPU or vice versa, lots of CPU and very little memory. If you put one of those on a machine, then you may be wasting one of those resources. It's what's known as resource stranding. And these are the available resources, these white bars here. This example here is actually our virtual machines, which are tasks. Our virtual machines are actually containers, believe it or not. It's a Google computing engine. So these are all virtual machines, these bars, individual bars. And what we can see here is that some of these machines have available capacity, available RAM, available CPU, and if we look over here, we see a different situation where we have maybe some with available CPU and others with no available RAM and vice versa. This here and this here is called resource stranding. It means we're not actually making use of that resource. So we have spare memory capacity or spare CPU capacity that's being wasted effectively. So one of our challenges is like a Tetris puzzle to try to stack these things in a way where we get the best possible utilization out of our clusters. So we will mix and match them to make sure we have low CPU, high memory jobs running with high memory, low CPU jobs. And of course, we run multiple tasks per machine. That's extremely important. That can then come back to all this with Kubernetes shortly. And another interesting thing is this, which is going to be a huge challenge in the future when it comes to Kubernetes, but it's going to be really important to all of us. So we saw earlier that our developer, she specifies what resources she wants to use or he wants to use. 100 megabytes of RAM, 100 megabytes of disk, 0.1 CPU. And that will be this blue line up here. So everything's running. We'll match into this blue line. These are the resources that were requested by these jobs. In reality, though, it's like this. And so we have all of this wasted space, which we can't use because it's been allocated effectively for those running jobs. But we can use it. So what we do is we effectively estimate, based on the run patterns of the current jobs, how much they're going to use. And that's this blue line here. So this is our reservation. So this is how much we reserve specifically for those jobs. And what we can then do is reuse all that space. Now we can reuse that space for very low priority jobs. Again, those batch jobs, those map reduces. Things that we want to run, we want them to finish eventually, but we don't really care when it happens. It could be like running some kind of monthly report that nobody ever looks at that gets logged or running a map reduce across a huge amount of data that may be important at some point or just needs to be done, but we don't really care when it needs to be done. So all of that stuff, we can reuse it and we can run jobs within it. But that's really important. That's how we can get maximum utilization out of all of our machines in that data center. And so moving on to Kubernetes now, gradually, one of the observations is that if you have your developers spending time thinking about machines or thinking in terms of machines, and you're probably doing it wrong, because it's too low a level of abstraction. Now today, maybe it's fine, but in the future, this is not going to be the case. We need people to be thinking in terms of applications and not having to worry about the infrastructure in which they run. I mean, anybody who uses a platform as a service knows how important that is anyway. You don't care about the infrastructure. You want to write your work, configuration file, build a binary, and just say, run this for me. I don't care where you run it. I don't care about how you do it. Just run it for me and make sure it stays running. We get efficiency by sharing our resources and reclaiming unused allocations. And containers, the fact that we containerize everything, allows us to make our users much more productive. So everything we run runs on a container. Two billion containers a week, we estimate. We never really thought that was very important until Docker came along. Containers became the next big thing, right? Alex, then Docker. And Docker became huge. So now one of the things we talk about all the time now is we run containers all the time. We are pretty good at running containers. Which is why we created Kubernetes. If you're interested in more details of what I've just talked about, Borg, there's a paper here. G.gl, one capital C for N U O. That's the white paper in Borg. That's got all of the details, all of the graphics you just saw. It goes into much, much more detail, of course. So let's look in terms of a simple application and how we can do this externally with containers and through Kubernetes. So this is a very simple pattern. Generally when we give this sockets PHP in the middle, MySQL, Memcache, and we have a client. We have many of these pythons running. This could be many instances of flask, it could be some kind of event system, but we have the ability to run many, many concurrent requests. We're probably going to want to scale this thing on demand. We may not want to scale MySQL that much until we get to a point where we have to do replicas and sharding. Memcache, we've probably gone on a scale as well, but we're going to keep it simple for now. Just keep one MySQL, one Memcache, and a few of these pythons that are at the front end. So let's talk about containers. So how many of you are familiar with containers? How many of you have actually spun up a Docker container? Hey, lots of you. It's almost the same number, right? Again, last year we'd asked this, how many of you have heard of containers, lots of hands, how many of you have spun up a Docker container? Lots of many. So things have changed now. Docker is the future. Well, containers are the future. We now have this thing called the open container project. Docker have kindly made what they have into a spec, and we're all going to get behind it. We have this common specification from which we can write containers. Things like CoreOS with a rocket container, they're all going to fall in line and we'll have a common format for containers. Which is going to be great. But just for those of you who are not really familiar with containers, just a few slides, very few slides on containers. Just kind of give you some of the concepts. This is the way we used to do things in the old days. We have a machine, maybe next to our desk, in our bedrooms or in a colo or in a server room. And the machine would run our operating system. It would have all of the packages installed at provided libraries, things like open SSL. On top of that, we would run applications. And how many of you have had a situation where you're running one application and all of the other applications on the machine fail because that one application went mad? Use all of the CPU, use all of the RAM, it crashed the machine and took all of the other applications down. And this may have been a very low priority app, one that you didn't really care about taking down some really important ones. Now, this is never a good idea running multiple applications on one machine because there's no isolation between them. Whatever affects one application will probably affect all of the others. There's no namespacing. They all have one view of the machine in which they're running. They have one view of the CPU, one view of the memory, one view of the file system, one view of the network. They share libraries. And so you're in a situation where maybe one day you update a version of a package, it updates a library, and one of your applications says, hey, I'm not going to run anymore. That library is not compatible with me. So dependency hell. And if it's on Windows, it's DOL hell, and it's probably even worse. Applications are highly coupled to the operating system. This is a problem. And so we created virtual machines. And what we did basically is stuck a layer on top of the hardware called a hypervisor. And we now had an idealized piece of hardware on which we could run multiple operating systems. So now we have this thin layer, it looks like a piece of hardware to be running virtual machines. And that gives us some isolation. So now we can run applications in their own virtual machine. So each application is now isolated. If one application crashes, it doesn't affect the others. But it's extremely inefficient because we have this red bit at the bottom here. We have the operating system, the kernel. And you know when you install a virtual machine, you pretty much have to install the entire devian stack or the entire center stack or the entire Windows stack. So that's not very efficient at all. They're still the same type coupling between the operating system and the application. And as anybody who's tried to manage lots and lots of virtual machines to provide isolation, you know it's hard. So there are new ways containers. In this case, we move up a layer. So we move above the operating system and provide an idealized operating system. No longer idealized hardware and idealized operating system on which we can run apps and their dependent libraries. So the libraries here are part of the container. So the container has an application. It has all of its dependencies. It has its entire environment. So we can move this container around anywhere we want to. We can move it from one machine to another, from one runtime to another, from a laptop to a virtual machine, to a cloud, to a bare metal server, to a set-top box, ultimately, maybe even to a phone when we had Docker on Android and iOS. I'm sure it's going to happen, right? And let's look at the example. So we have our application, PHP in Apache. It should be Python in Apache. Sorry. See? I do apologize. So wherever you see PHP in this deck, read Python. We'll change it before I share the slides. So I'm trying to think of what could offend a Python audience most. It's probably talking about PHP, right? My god. Okay. So we have containers. We want to run these components of our application, Python and Apache, Memcache, MySQL. Not Apache. Obviously, Python and Vlask and Bottle and all of the other things we could potentially use. Memcache, MySQL. And MySQL has its own libraries. It doesn't have any common libraries with the others. So we can stack those libraries with the container in which MySQL runs. And Memcache, PHP, Python and Apache have their own, I keep saying it. And Python and whatever. Unicorn. Anything. They have their own dependencies. But they also have when we install them some shared dependencies as well. So some commonalities. So when we actually create the image, we can actually share some stuff between them. But that's not shared at run time. So when we create the container, they will have their own dependencies packaged together in the container. And underneath that, we have a server. And again, this could be a virtual machine. It could be a laptop. It could be a bare metal server. It could be anything, pretty much. And underneath it, we have the actual hardware. And all of this is being maintained by a Docker engine. So Docker is the thing that runs this. So when we talk about containers, mostly synonymous with Docker nowadays, but again, there are other container formats. And hopefully, they will all comply with a standard. And that's the Nirvana we're all heading towards. So Docker effectively controls the creation of these containers and the management of these containers. So at the end of it, we will have Python, and Flask, Angular, Memcache, MySQL, all running in containers. So why containers? So there's many important reasons for having containers. But you can see, just by looking at what we do, that's the only way we can do it. We can't do it any other way. This is perfect solution for the kind of scale that we want. But it's also perfect for smaller scale as well. Why? Because it's much more performance. It's much more performance in terms of the fact that we don't have to do all of that installation stuff. They are pretty much like they're running on bare metal. So the performance is pretty much the same as a virtual machine, but they're much quicker to get up and running. Which means you can swap them out quicker. You can do upgrades quicker. You can do pretty much everything quicker. Repeatability. So the whole problem where we have development, QA, build, test, production, where we want to have repeatable environments, where we have a situation where when we test something in QA and then run it in prod, it fails in prod where it works in QA. How many people have had that situation? Have you had any hands remembering those days? So what containers give us is the ability to have a consistent environment because the environment is packaged with a container. So basically when we run it in QA and when we run it in prod, it's exactly the same. It's exactly the same environment. So that's one of the great use cases of containers today. But much more is the portability of it, which we're going to talk about in a second. Quality of service. We can now do resource isolation as well. Using things like C groups and Linux and namespaces, we can actually isolate the resources. We can say we only want this to have 100 megabytes of RAM, 100 megabytes of disk, 0.1 CPU. And ultimately accounting. These things are easier to manage. They're easier to trace. They're easier to audit. They're small, composable units that can be tracked very easily. And ultimately, portability. You can move these things around from one cloud provider to another. Images specifically. You can't just pick up a running container and move it, but you can easily run the same container in a different cloud provider on a bare-measure machine on the laptop. You can move them from one machine to another as the shape of your cluster, if you have a cluster of machines, changes. You can move them around to be more efficient. So we can go back to what we had before with the efficient allocation of resources. We can do that if we have containers. And ultimately, this is a fundamentally different way of managing and building applications. So demo. I'm not going to do this demo. I left that slide in by mistake. This would have been a containers docker demo. And I don't think I want to bore you with that. It's very easy to find a tutorial on docker and get up and run them a bit. Now, let's not talk about that. Let's talk about Kubernetes instead. How many of you have heard of Kubernetes? How many of you can say Kubernetes? It's hard word to get your head around. Probably easier if you agree, because it's a Greek word. But if you want to help pronouncing it, I'll be outside in the Google booth after this talk. So I can definitely provide assistance on that. Maybe I'm saying it wrong. Maybe I've been saying it wrong all this time. So I'm happy to be corrected. So Kubernetes, let's talk about that. And we've given you an introduction to what we do at Google. So that should provide the context on why Kubernetes is necessary. What we often miss out, is that we don't provide that kind of context. So I'm hoping that the introduction to Borg has provided that for you. So Kubernetes is Greek word, means helmsman. It's the root of the word governor. So Arnold Schwarzenegger's governor comes from Kubernetes. And it's an orchestrator or scheduler for docker containers. Ultimately for other forms of containers, I think ChoraOS is already using it to schedule orchestrate rocket containers. It supports multiple cloud environments. So Metasphere, I always forget them. VMware, even Microsoft involved. You can run Kubernetes on Amazon. You can run it pretty much anywhere. You can run it on your laptop with Vagrant. So you can just create a four machine cluster, virtual machines with Vagrant up. And you'll have a Kubernetes cluster. And ultimately, eventually, we may have a situation where we can run Kubernetes across multiple cloud providers. It might be difficult. It might be possible. But it may be one day you'll have your fleet of machines will be running in Google, in Amazon, and Microsoft Azure as well. Possible. I'm not sure if it's going to happen. So this is kind of inspired and informed by everything that we saw previously, everything with Borg. And it's basically our experiences. Open source, ridden and go, like many good programs nowadays, but completely respect Python. I love Go. I love Python. I used to be a Java developer. I spent 15 years developing Java. No, 11 years. Now I moved to Google. I haven't wrote a line of Java code since. Now I write... Now I write... It's like Java program is anonymous, right? It's been four years since I wrote my last line of Java code. So now I write in Python, go, and write in Angular, and write in Java scripts, and all of those more interesting, useful languages. Java is getting better. Java rate is a big step forward. And ultimately, we want to be able to talk about managing applications and not machines, which is actually what we talked about earlier. And some very quick concepts. I'm not going to introduce them, but I want to show you the icons so that when you see them, you'll know what they mean. Angular, pod, service, volume, label, replication controller, node, are all of the key concepts. How many of you are familiar with sort stack? How many of you like the terminology in sort stack, like grains and such like it? I think it's really hard to get your head around. And I think one of the dangers about an abstraction is that you get too far away from the terms that are familiar to people. Most of these are familiar to people with service, the idea of a replication controller, a node, a label, a container. The pod is probably the most difficult one to get your head around. So let's talk about pods. Let's talk about nodes first and clusters. So we have a cluster, kind of maps back to what we talked about earlier with Borg, where we have a master. And the master has a scheduler and it has an API, an API server that can be used to talk to nodes. The nodes are all running a thing called a kubelet. And they have these things called pods running containers. We'll talk about pods shortly. They also have a proxy by which we can expose our running containers to the outside world. And we have many nodes. And a cluster, this is an abstraction. So a cluster could be different depending on which cloud provider you're using. And ultimately what you want to have is a fabric of machines that looks like a flat shape, which we can run containers. You don't care about it. You just care that they're all joined together. And it's one big flat space in which we can run stuff. And we'll let this thing, the scheduler, take care of running stuff for us ultimately. And so basically the options for clusters are laptops, multi-node clusters, hosted or even self-managed, on-prem or cloud-based using virtual machines or bare metal in virtual machines. Many, many options. There's a matrix down here, a short link. Hopefully we can share these slides afterwards. And the short link will give you a matrix of how you can run Kubernetes on what you want to run it on, CoreOS on Amazon. We have different ways of doing a networking. The networking is quite tough. Google Compute Engine makes it easy because of IP addressing. But often we have to put this other layer in called Flannel to actually provide that ability to give an IP address in a group of subnets to a running machine or a running pod. So let's talk about pods. How many of you are familiar with the concept of pods? OK, not so many of you. So in the diagram here we have a pod. It has a container. This is a container, this web server is a container. And it has a volume, like Docker containers can have volumes. A little bit different, but very similar. And so we want to run this web server. And the construct we use within Kubernetes is to create this thing called a pod. It's like a logical host. So like if you wanted to run Apache and something else alongside it, you would run it on a host machine. That's the same as a pod. So anything you would run together on the same machine will run in a pod. These are the atomic units of scheduling for Kubernetes. This is what Kubernetes schedules. We talked about jobs earlier when we looked at bulk. Kubernetes schedules pods. And your containers run within the pod. So thin wrapper around them. These are ephemeral. These are like, I've got this analogy. So everybody uses this pets versus cattle analogy. And I don't really like it from a vegetarian. So crops versus flowers. So pods are like crops. You don't care about them. You have a wheat field. You don't care about your individual plants that are growing. When you have flowers, you probably give them names and you water them and you talk to them as well. So you care about them. You don't care about your crops though. So pods are like crops. They can come and go. They can be replaced. They're all absolutely the same. You can take one and replace it with another. And ultimately, to make things simple now, you don't have to worry about a pod if you want to run a single container. You just say, run this container for me. It will create a pod for you. And you still have to think in terms of pods when you're doing monitoring, but you don't have to create a pod. You just say, run the container for me. It will create a pod for you. OK? So pods are an abstraction. Difficult to get ahead around. A little bit more information about them. Imagine this scenario where you want to have something that synchronizes with GitHub. This may be a push-to-deploy type scenario where whenever your developers do a merge into GitHub, you want those changes to be immediately pushed out into production or maybe into a staging service. So you have a thing called a Git synchronizer. And it's talking to Git, monitoring your project in Git. It pulls down any changes. And it writes them to somewhere on the disk. And your web server can then serve that latest content. Those things are tied together. They work together. And it makes sense for them to run side by side. And when one goes away, the other goes away. So we can run them both in the same pod. So now we're saying, on this logical host, this pod thing, let's run two containers. In this case, Git synchronizer and a Node.js app or a Python app. And we have a shared volume, a concept of a volume, which we'll talk about shortly. These are tightly coupled together. So when one, when a pod dies, they die together. It doesn't make any sense to have them running separately. It might do in your, in the way you architect things, but it doesn't have to. They share the network space and port space. They have the same concept of local host. They are completely ephemeral and think in terms of things you would run on a single machine. So a volume, what's a volume? I don't normally talk about volumes, but they are very important, so not talking about them seems a bit stupid, really. So a volume is basically bound to the pod that encloses it. And this is something where we can write data or read data from. Okay? And we have many options when it comes to volumes. Docker already has volumes. This is slightly different, but very similar. So to a container running in the pod, the volume looks like a directory. And what they are, what they're backed by and such like and when amounted is determined by the volume type. So the first type we have is an empty directory. So whenever we create a pod, it creates this space somewhere on disk, on the local disk, and they can basically share that volume between them. But it lives and dies with the pod. It only exists while the pod is there. So it could be your git synchronizer is writing stuff to this volume, being read by the Apache server or whatever server. And you don't care when the pod goes away, if that space goes away. It's just scratch data, just temporary data. There's nothing stored there that's important to you. And it can't even be backed by memory as well. So it could be tempFS file system. And that's great. It's really efficient, much faster as well. So that's what an empty directory is. That's the default you get for a, well, I don't know what it is, a default, actually. That's the specified type it is. So empty there is one of the options. The next one is host path, where we can actually map part of the file system of the node on which the pod is running into the pod. So this volume is actually effectively a snapshot of, not a snapshot, a link into the file system of the actual running machine. That's useful to read configuration data and stuff. But it's also kind of dangerous as well, because it may be that the state on the node may change in such a way that whenever you run a pod on one machine to another, you don't run it. Whenever the scheduler runs the pod on a different machine, it may see a different view of what's happening. So it no longer becomes completely isolated. So it's a kind of dangerous thing to do, but it might work for you. The other one is nfs and other similar services like GlusterFS. I can never say that. And if you're a G on it, I can't say some reason. So again, nfs, we can mount nfs paths on our pod and expose them to our containers as directories. Or we can also use a cloud provider, persistent storage, persistent block storage. Now we call them persistent storage in Google. Amazon call them elastic block storage, that kind of thing. So this is persistent disk. So basically they can write and read from the data from the disk, and it will always be there, whether the pod goes away or whatever. So what we're likely to do in this case is create a volume, a volume in the cloud provider. I call it a disk. We create a disk in the cloud provider which stores data and we'll mount it onto the pod. Whenever that pod goes away, the data is still there. The pod comes along and can mount it as well. And also with Google Cloud Platform, you can actually mount and read only on multiple pods as well. So some patterns for pods. The first one is the sidecar pattern. Because basically it's motorcycled inside car. I guess in this case the Node.js application or the Python app is the... You don't get offended when I say Node.js, right? The Node.js application is a bike and the git synchroniser is the sidecar in this case. That makes a lot of sense, right? Ambassador, in this case, something that acts on behalf of the actual running container. So this is a secondary container, a Redis proxy that effectively allows the PHP application to make calls and then have the Redis proxy call that to Shards. So we can just make... Have one service that the PHP application calls for reads and writes and the Redis proxy can do with the hard work of deciding whether to read from a master or read from a slave or write from a master. And the final one is an adapter pattern where in this case we have Redis running and we want to monitor it. We want to monitor all of our pods but we need a common format for monitoring. So in this case we actually adapt the output from the Redis monitoring using an adapter container. An adapter container will be plugged into the monitoring system. So it kind of adapts what's happening within the container. So these are kind of examples of where it makes sense to have a pod. I'm hoping it does make sense and I'll be interested to hear from you afterwards about whether pods make sense to you. So labels basically the single grouping mechanism within Kubernetes. This allows us to group things that we can build applications like a dashboard. So we have a running pod. We give it a label. Labels are key value pairs. So in this case type equals fe. Completely arbitrary metadata. Some of these things are meaningful to Kubernetes but mostly it can be anything that's meaningful to you. So we've put labels on pods and we can say I can build a dashboard application that uses the API to say give me the pods with this label. And I can show you all the status of that. And we can have different labels for different pods. In this case we have a version two pod. We have a different dashboard application with more different nodes. And that makes a lot more... Pods can have many labels. I surprise myself with my slides sometimes. Makes more sense with replication controllers because replication controllers are things that actually manage the running of pods. Now remember I said before that we created 10,000 tasks and we pushed them out to persistent storage in a ball master. And the scheduler comes along and says, yeah, these should be running but they're not. I'll fix that. So this is the same thing. The replication controller is responsible for managing your desired state. You say this is the way I want it to be. I want to have X number of these pods based on this container template. I want you to maintain that state for me. That is the job of the replication controller. So basically what they do is they work on a constituency of a label type. So in this case version equals V1 is what they select on. This replication controller is responsible for all pods with label version equals V1. And we tell it I want to have two of those. So this job is to make sure there's always two running. In this case we also have another replication controller. That has V2 of our pod. Version equals V2. I only want one of those. So make sure there's always one of those running. And the way it works is this is kind of like a control loop. The replication is one big control loop. Simple as that. It says look at the desired state. How many of we got running? We should have four running. We got four running. We got three running. That's not good. Let's start another one. We have four running. We have five running. That's not good. Let's take one away. It's basically monies the state to make sure we had the ones running. It also works with a template. So we provide a template, which is the pod template, which contains the container image definition. And how many we want to run. We pass that into the replication controller. It doesn't create the pods. But when we create the replication controller and we say we want two of these pods, it says, hmm, there's not two of these running. I should start them. So it starts them. That's how it works. And we can also plug in replication controllers after we've created the pods. And say, you're managing containers with this label. And finally, we get to services. And services are how we actually expose our running stuff. And we do this through this service here, which creates a virtual IP address, which has a constituency of pods based on the label selector. Again, we're going to have labels on here. We'll show it in the next slide. So basically, certain pods with a certain label are the constituency of this service. And when requests come in from clients, it will load balance them across the running pods, regardless of which node they're on. So they could be 10,000 nodes. We could have 10,000 pods run in different nodes. And it would load balance them across the running pods. At the moment, it only works around robbing. But eventually, it will have much more intelligence support for load balancing. This is used for exposing internal services within Kubernetes and also expose mining services to clients externally, which we'll see shortly. It not only provides a virtual IP address, but also a DNS name so we can do service discovery. And we want to move on. So this is a Canary example. So who understands the concept of Canary? A few of you. So basically, when you have a situation where you have a running application, you want to try out a new version of it. You may have one instance or two instances of that running application that are different. So some of your traffic will be pushed to the new versions. Some will go to the old versions. You can then do A, B testing against them to make sure that the new service works. If it doesn't, you can roll it back. If it does, you can push out the change for all of them. This is a similar situation where we have version equals V1, version equals V2, replication controllers and pods. But a service, all it cares about is labels type equals FE. And so the service has its constituency of all three of these pods. But these pods are managed by different replication controllers. So that's how it works. Virtual IP address exposes that to a client. And so we map the Kubernetes. It all looks kind of like this. We have pods. Remember all the symbols? I believe this is why it's important. So that's a pod and a volume and a service. And we have all a memcache vd drop down. A pod, a service and replication controller with a service. How does that look to a developer? So remember how it looks to a developer on Google? So this is how it looks. They specify a name. They can specify the image. This is a Docker image now. It could be a different image format in the future for a different type of container format. I left it in deliberately to stop, so you. Yeah, PHP gets back. You can specify resources. 128 maybe bits. Maybe bits. You can specify how much CPU. And the Kubernetes unfortunately had its own idea of slicing off a CPU. And I'm not going to get into it, but it's like 500 bits of a CPU in terms of Kubernetes. So you have to read the manual for that. Otherwise it won't make any sense. I'd rather have a percentage, but that doesn't work because you can't have a percentage of a core because you don't know how powerful it is. So that's how we specify CPU. The ports, protocol, TTP, and the replica is one or maybe $10,000. We covered that case as well. So that's how it works within a replication controller. There's other configuration files as well for services. Oh, sorry, do you? And scheduling at the moment, we saw the complexity of scheduling at Google. It's a bit simpler for Kubernetes currently. It's based on pod selection. So we want to have the pods run in that are based on selectors, and it's based on node capacity. So how much capacity does that node have? Is it capable of running my pod for me? If I have multiple nodes that can run my pod, I'm going to run it on the one that has the least resources consumed by running pods. And that's the priority. In the future, we'll have resource-aware scheduling. So we can do what we do back in Google where we try to make maximum utilization out of our CPU and memory. Kubernetes is 1.0 as of this week. Now it's on 21st of July at OSCON in Portland, Oregon. It's been open-sourced for over a year now. And we have a product called Google Container Engine, which I'm going to talk about shortly, not so much, but it is a good way of running Kubernetes. But it's not a product pitch. Hosted Kubernetes, I'm going to talk more about Container Engine shortly. And the roadmap for Kubernetes is there. It's kind of sparse at the moment because we've just gone through 1.0. So they're now deciding on the roadmap for the next releases, V1.1. And the one on the roadmap currently is autoscaling. The ability to autoscale your nodes dynamically based on the amount of work you have. Container Engine is a managed version of Kubernetes, and it manages uptime for you. You don't have to worry about the master in this case. It will take care of the master for you. You can't even see the master. You can't connect to the master. So one of the problems we have at the moment with Kubernetes is high availability. So we don't have that replicated master scenario we saw with Borg. So the only way to do it is to have multiple clusters to do high availability. But if we look after your master for you and make sure it's running, then you don't have to worry about it. We will make sure that your cluster is highly available. By making sure your master is always running. We can resize using things called Managed Instance Groups, which we'll look at in a minute. Centralize Login. We can pull all of our logins to one place in the Google Developer Console. And it also supports VPN. So you can actually have your pods inside your own network, your own private network. So demo. Very quickly. And we had to change the setup earlier to make all this work. And this is a cluster. We have kubectl get nodes. So we have two nodes running. So these are machines in our cluster. And I can look at them here. This is the Google Developers Console. And I can probably make that a bit smaller. If I go into VM instances here, I can see my running machines. I have a couple of extra machines as well. But these two in the middle are the nodes for our cluster. I have this thing called an instance group, which has two instances. This is the thing that manages the size of our cluster. And below here, we have container clusters. And we can see we have one cluster. OK. And if we go to... Wait a second. I've got very little screen real estate, so I can't see everything that's going on. So we go here, we can see a representation of what's running currently. So we have... These are pods. So this is a pod. This is a service. This is exposed internally. This is a service. And this is another service. Oh, MySQL is not running, which is a real pain. I'll have to run it. OK. I don't know how that happened. OK. So we have a front-end service. We have a memcache service. And we have MySQL. We don't have a pod running. So we need to have a pod running. So I'm going to start the pod very quickly. And then... OK. So I'm going to start the pod very quickly. OK. That's why it's not running. So we've just gone to 1.0. All my demos break. That's fine, though. Should be fine. I had it all running, but we had to reboot my machine because we were having problems with the display. And then we have a pod. Hey, pod. MySQL. Have you ever spun up MySQL so quickly? I bet you haven't. The next thing we want to do is run some PHPs. Unfortunately, there are PHPs currently. But I'm trying to get time to update them completely. But I had some problems with flask and Angular. So anybody else had problems with flask and Angular? No? OK. I should talk to your... Basically, on my badge here, it says that my Python skills are rated as free stars. So I probably should talk to you guys about doing it. OK. So I'm going to create a control. We have a file already created. And we're going to create that. And now we have pods and a replication controller. I need to make that smaller. So now we have some frontend pods and a frontend RC controller. So the next thing we want to do is actually look at the running application because my windows are all screwed up. We have it running. So this is the IP address of the service, as we can see here. And this is the application running. It's DevOps. Anybody who's been to DevOps, you don't want to go to DevOps. You want to go to the EuroPython. I told them to fix this beforehand. We have an update and we can roll that update out really easily to our cluster. Let's do that. Let's roll out an update to our cluster. And I'm going to close that down so we can see the visualization. And I will go and reverse my history for this. And I'm going to update to v2 of our frontend controller. So what's going to happen now is it creates a new controller. And now it's going to change those pods one by one to roll out our new version. So now we have three pods, the 2.0 and 2.1.0. We're going to get rid of one or 1.0. We're going to have a 1.0 and a 2. And then we're going to create a new 2.0 pod. And then we're going to get rid of the other 1.0 pod. And eventually we only have two 2.0 pods. And we get rid of the 1.0 controllers. We don't need that anymore. And if we go back to our app, nothing's working. And refresh. And we should get... Yay! APPLAUSE I'm hoping it works. I'm hoping my SQL is running properly. Yay! So, OK, that works. Brilliant. The other thing I can do as well, I should mention it. I'm probably getting close to my net time. What is the command? Is it RC or not? I don't remember. I never... Why should I get the SQL command? So I'm going to do B2. And I'm going to scale it to six replicas, the five replicas. And we go back to our biz. So now we want to add replicas to this. We can do that by scaling like that. And then we have five running pods. OK? It's as simple as that. So now we have five running. We can do that also within the developers console. And just to wrap up on the whole thing, just a quick talk about the last bits and pieces. That's how we visualized it. We visualized it using the API and a proxy. So kubicostal supports a proxy. We just pointed at some JSON. The JavaScript is all JS Plum. So if you want to know what we use, JS Plum. In terms of container engine, cluster scale in, we have this thing called a managed instance group. And that runs all of our nodes. And nodes run within the managed instance group. And we have this thing called an instance group manager that creates them and is responsible for making sure they're running. So that's actually monitoring the cluster of nodes. And we have a template by which we can create new nodes on demand so we can resize that managed instance group very easily. And yeah, I think that's about that for cluster scaling. We can also create clusters using tools such as the Google Developers Console, Google Deployment Manager, and Terraform. I was going to give an example, but it's very basic. Terraform will create a cluster for you, but it won't allow you to resize it. If you want to resize it, you have to replace it completely, which isn't really what you want to do. So you can create clusters with various different ways. And oh, that's a visualization. Some frequently asked questions are answered in the documentation. I could spend entire hours on all of these subjects. So if you have questions, I'll be outside all day on the Google booth. Come and see me. And Kubernetes is an open source, so we want your help making it even better. So please contribute to Kubernetes. If you have questions, go to IRC, IRC.free.net on hash Google containers. It's a very popular place. And also on Twitter, Kubernetes.io. You can tweet questions for me. If you don't answer them now, or you can find me on the booth. And ultimately, that's it. Thank you. We have time for one or two questions. At the beginning, you were talking about Borg and like five masters that you run in. And those figures are based on the data center or how does it? Based on the cell. So we break it up into a cell and each cell has its own Borg master. And that's about my limited knowledge of how the complexity of how Borg works. Not being a sweet one, sorry. But yeah, that's how it works. That's why we have multiple. So within Google, you're going to find multiple Borg clusters or Borg cells effectively. Hi. Thank you for the talk. Very interesting. When you compare VMs and containers, even if the user of a VM has root access, it's very difficult to escape from the IPvice or et cetera. How do you see the security in the current containers implementations? It's a work in progress. So this is about security with containers. I'm not really going to comment too much on it, but it's getting better all the time. Initially, we had problems with the kernel level and syscalls and such like being made back into the operating system. But it's getting better. So ultimately Docker and such like becoming more secure all the time. Ultimately, doing multi-tenant maybe currently with multiple customers applications run inside by side may not be the best idea. But we have to tackle that. So ultimately, we have to make sure people are more confident that they can run all of their jobs on containers securely. I don't think we're quite there yet, but we're working on it. So that's one challenge we need to crack. Is that a question? Yeah. One more question? Is that done? No, that's enough. OK. Come and find me outside. Come and find me outside. We can talk about PHP. It's Python, sorry. From our organization, we want to thank Mando to come in. Oh, wow. Give her a present. Thank you. Should I answer it? Should I put it? No. Oh, no. That is wonderful. Ah, fantastic. Exactly what I need. Thank you very much. Thank you. Thanks for having me. Thank you.
|
Mandy Waite - Keynote: So, I have all these Docker containers, now what? You've solved the issue of process-level reproducibility by packaging up your apps and execution environments into a number of Docker containers. But once you have a lot of containers running, you'll probably need to coordinate them across a cluster of machines while keeping them healthy and making sure they can find each other. Trying to do this imperatively can quickly turn into an unmanageable mess! Wouldn't it be helpful if you could declare to your cluster what you want it to do, and then have the cluster assign the resources to get it done and to recover from failures and scale on demand? Kubernetes is an open source, cross platform cluster management and container orchestration platform that simplifies the complex tasks of deploying and managing your applications in Docker containers. You declare a desired state, and Kubernetes does all the work needed to create and maintain it. In this talk, we’ll look at the basics of Kubernetes and at how to map common applications to these concepts. This will include a hands-on demonstration and visualization of the steps involved in getting an application up and running on Kubernetes.
|
10.5446/20155 (DOI)
|
Hello, I'm Maciej and I'm a data scientist at List. We're a fashion company. We basically go all around the internet and we get all sorts of fashion products from all sorts of retailers and all sorts of designers. Put them all in one place so that as a user you can just go to one website and follow your favorite designers and browse all your favorite fashion products and buy them from us. So that's the principle. And I'm going to be talking about nearest neighbor search. So nearest neighbor search is very simple in principle. Basically you have a point and you want to find other points that are close to it. So the most obvious application is obviously maps. So that's something we use every day. You've got, let's say, your location on a map and you want to ask Google or Bing or whoever you want to ask, what are the nearest restaurants to me or what are the nearest cafes. And that's what it does. It figures out where you are. Rather, you give a location and then it looks up other points on the map that are the restaurants that you're looking for and then calculates the distances between where you are and when the restaurant is and tries to give you the closest ones. So that's basically the essence of nearest neighbor search. Given a point, give me other points that are close to it. Now this is the most obvious application. But it's even if you're not building a mapping application, which you may well not be building, we certainly aren't, you may still find it very useful. So what do we use it for? Well we use it for image search and we use it for recommendations. So image search, how does it work? The principle of image search is, well you've got an image. Let's say we had a user and they submit an image to us of a dress and they want us to give them images of similar dresses or something that's a good substitute for the dress that they've submitted to us. So as programmers, it's sort of hard to find similar images just by looking at an image. So the first step you want to do is to transform an image into something that you can more easily work with. So one very naive idea would be to, let's say, well we've got an image which is RGB, like red, green and blue. And these are three numbers and for a given image we could average the values of each color and then try to find other images in our database which have sort of similar color values. And that's a very naive approach and it's definitely not going to work. But it illustrates the principle that after transforming the image into a numerical representation you've got a point. In this case you've got a point in a 3D space and that's the nearest neighbor search. If you want to find similar images to the image the user has just given us, we look up other images, other points which are close in that space. So that is nearest neighbor search. The way we would actually do it is something more complicated. So recently the most fashionable way of doing this and probably the most effective is deep learning. So this is a very simplified diagram of a convolutional neural network. So basically the black square, the far left end of the slide is the image that you start with. And then you basically build a machine learning model which successively detects more and more interesting features of a given image. So let's say at first you detect just simple edges. Is there a line in that part of the image that goes from left to right? That sort of stuff. But then as you progressively build better and better representations you start to learn more and more about an image. So maybe the first layer you just detect edges. In the second layer you detect shapes. It's a square. It's a circle. And in the following layers you detect high and high level concepts. Is it a cat or is it a dog or is it a building or is it a bridge? Now the nice thing about it is that in the final layers you've got a long list of numbers and ordered list of numbers, a vector which represents a point in space in a high dimensional space. And the nice thing about it is, well, the images of cats in that space are going to be close to other images of cats and the images of bridges will be close to other images of bridges. So that's how you can do very good, very good image search. And this is indeed what we do at List. We sort of process, take images and process them into this point representation in a high dimensional space and then we want to use nearest neighbor searches to find similar images. That's useful for two things. One thing is search. You give us an image and we can find you similar images. Or maybe you type in a text phrase and then we convert your text phrase into a point in the same high dimensional space where the images are. And suddenly we can find images which are similar to the text you typed. So that's really cool and that's something we can do. Are there other useful applications, the duplication? So we've got two images and we have no metadata associated with those images. But we know that these are the same products. Like the underlying truth is this is the same product and we can use this nearest neighbor search to find out that these are the same products and present them just as a single thing rather than two disparate things on a website. So that's really useful. Set applications recommendations. So this approach is known as the collaborative filtering approach. So basically you have your products and you have your users and you represent both products and users as point in space. So let's say we take our products and this is a handful of points and we cast them at random into this high dimensional space and we do the same for users. These are our users and we represent them as points and we cast these points at random into a high dimensional space. And then what we do is, well if a given user, a point in our space, interact with a given product, we draw the points together. But if a given user did not interact with a given product, we push them further apart. And the nice thing about it is at the end of this pushing apart and pulling together process, users end up close to the points, to the products that they will like and far apart from the things that they wouldn't like. So when we want to recommend things to a user, well, we look up the point that represents them in the high dimensional space and then we use the newest search to find the products that they would like. So that's also very, very useful. So I'm a data scientist and data scientists are very excited by these sorts of things and we spend a lot of time thinking about them and implementing them and you sort of go away and you work for six months and you come up with this amazing solution where you translate pictures of products into this high dimensional space with similar products close in space and you say, well, as a data scientist, I've done this amazing thing and it now works and it's going to be great. Let's just deploy it on the website and make lots and lots of money or make a user really, really happy. Now this is where you are as a data scientist. You've got your beautiful child, which is going to be great and you sort of go and let's just deploy it. Let's make it work. So how would you do it? Right? You are given a query point or you take a user and you want to find all the newest products and that's simple, right? You go to your database and you find all the points representing products and you compute the distance between your query point and all the points, all the products in the database and you just order that and you give the closest ones. Simple, right? I mean, it couldn't be simpler. So yes, but no. The problem is at least we have 80 million images and we have about 9 million products. So if we wanted to do the simple solution, I mean, you know, calculate distances with all the points and then return the closest, our users would be very, very bored by the time we finished. It would take literally minutes. So that will not work. Okay, so how do we make it work? Look how it is sensitive hashing to the rescue. So we all know about hash tables or dictionaries in Python. So we're going to build a special hash table. So we're going to pick a hash function which unlike normal hash functions, it maps points that are close together in space to the same hash code, which is very different than normal hash functions which are supposed to map things uniformly over the hash space. So this is sort of special. It picks two points that are close together in your space will map to the same hash code and you just build a hash table, a normal hash table using that. So you take your points and you take the hash codes and then you put your points in the hash packets corresponding to your hash codes and then magically all your points are close together will end up in the same bucket. And when you're doing search, you just look up the bucket that you need and you just search within that bucket, which is really nice. So to do this at least, we use random projection forest and I'm going to tell you how it works. So this is our imaginary space of points. So we've got about 100 gray points and one blue point, which is the query point, the point that we want to find nearest neighbors for. And this is how we do it. So if we didn't do locality sensitive hashing, you would have to calculate the distance between the blue point and every other point, which takes too long. We cannot do that. So to make it faster, we draw a random line. That's the beauty of it. We just take a random line and draw it. It has to go through the origin, but otherwise any line will do. And the nice thing about it is, well, if we look at the picture, most of the points that are closer to the query point end up on the same side of the line. And the points that are not close to the query point end up on the other side of the line. So just by drawing a random line, we manage to create two hash packets and suddenly we only have to look through half of our points to find nearest neighbors. So that's already a speed up factor of two, just with one random line. And we didn't have to do anything intelligent to draw that line. It's just a random line. So that's the nice thing about it. If this speed up is not enough for you, you draw another random line. Again, completely random. And the points that end up on the same side of the line end up in the same hash packet. If this again, the speed up is not enough for you, you keep drawing lines until you've got a few enough points in your hash packet. And that's your speed up, right? So you draw enough lines to have small enough hash packets. And then when you need to perform a nearest neighbors query, you take a blue point and you calculate which hash packet it should end up in and then just compute brute force distances between the query point and the points that are in the same hash packet. So that's the principle. You can think of it as building a binary tree as well. So we start with all the points and then we have a split. And the points that are on the left side of the line go into the left subtree. The points on the right side of the line go into the right subtree. And then we follow the right subtree in this example and do another split into a left subtree and right subtree and another split and another split. And you can sort of, with the query point, you can start the root of the tree and then follow the splits until you end up in the right hash packet. So that's the principle of how it works. Now it works really, really well in some cases, but in some cases it doesn't really work very well. So the way we started this at least, we thought, will we going to draw a fixed number of these lines? And hopefully that will give us speed up and hopefully will be accurate enough. So what we did is we decided on dimensionality. Let's say we want to do 100 random splits and then after 100 random splits we stop and then things that end up in the same packet are the nearest neighbors and the rest we discard. That works reasonably well if your points are fairly uniformly distributed in your space, right? Because all regions have fairly equal density. Wherever you draw the lines, your hash packets are going to end up to roughly the same size, so the same number of points and the splits are going to be good enough. But in spaces where some regions, in problems with some regions of your space are of high density but other regions are of low density, what you're going to end up is with some buckets having lots and lots of points and some buckets being completely empty, neither of which is very good. So if you have a bucket with lots and lots of points, you're not going to get a good speed up if you have a bucket with very few points in it, you're not going to get any results back, both of which are horrible. So the first point is keep splitting until the nodes are small enough. So we don't take a fixed number of splits, you just build a binary tree and you split and you split and you split until you've reached a stopping criteria in which this bucket contains X number of points and when that happens, you stop splitting and you take that tree. So that's the first point. The second point is use medium splits. So if you just take random hyperplanes, random lines, you can end up with highly unbalanced trees, so the left subtree will be very short because for some reason there are very few points there, but the right subtree will be very, very, very deep because there are lots of points in that part of the space. That's not really horrible but it's not great either because you spend more time traversing the deep part of the tree and you will be traversing the deeper part of the tree more often because that's where more points are. So medium splits. You take a random line and then you calculate the median distance from the point to that line and then you split on that distance. So it's guaranteed that half of the points will always go to the left tree and half of the points will always go to the right subtree. So that gives you nice balanced trees and faster traversal times. And the final point is build a forest of trees. So you don't just build one tree, you build lots of trees. What's the reason for that? Well, the random projections algorithm and locality sensitive hashing in general is approximate algorithms. They don't give you an exact answer. They're probabilistically correct but in some cases they will be wrong. So if we look at this picture, if you look at the query point, well maybe there are some points to the left of the line that are closer to the query point, the points on the right side of the line. But if you build just one tree, we're never going to surface these points because then in a different part of the tree. So that's a mistake that this algorithm makes. And that's not great. We want to have as good results as we can giving the speedup. So the way we get around this problem is we build lots of trees. And because in each tree, the lines are chosen in random again, each tree will make its own errors but they will not repeat each other's errors. So when you combine the predictions from all the trees, they will end up correcting for each other and the accuracy score of the aggregate will be higher than the accuracy score of any single tree. So that's why you build lots of trees. Another nice property of this approach is if you build more trees, you're going to get more accurate results at the expense of having to traverse more trees. So it's going to take slower. But this is something you can control. You can pick basically the trade-off between the number of trees you build. It draws your performance curve. And you can pick a point on that performance curve. And speed versus accuracy, that's appropriate for your application. So that's very nice. If you want pure speed, build few trees. You get pure accuracy but you'll be very quick. If you want something accurate but don't care about speed that much, there are lots of trees. It will be accurate, maybe not so fast. So that's the principle of the algorithm. Does anybody have any questions at this point? Because I'm happy to clarify. So the question is, how do you generalize it to higher dimensions? Like if you, well, I assume you will use hyperplanes, right? But how do you split the feature space with a hyperplane? Right. So the algorithm, actually, one of the main reasons of the existence of the algorithm is high dimensional spaces. And the reason why it's in 2D on the slides is because 2D is easy to visualize and high dimensional spaces aren't. So that's why it's in 2D. But basically what you do is in high dimensions, you draw a random hyperplane, which is, you know, whatever your dimensionality is, and then you do the same calculation. So it's exactly the same principle, which is just in high dimensions. And it works very well for very high dimensions. I can, we can talk about that afterwards as well. Okay. So that's the principle of the algorithm. How do you do it in Python? Since we're at a Python conference, it's sort of useful to give an idea of Python packages. So there are several Python packages for doing this. One of them is annoy, which is a very cleverly named package, A&N approximate nearest neighbors. Very cleverly named package from Spotify. It's very nice. It's a Python wrapper for C++ code. It's pip installable. It's very, very fast. It's actually very nice. Another package is LSH forest, which is in scikit-learn. Those of you who are data scientists or play with ML, you probably already have scikit-learn computers. So it's really easy to use it because it's already there. And it's also quite easy to use. And then you've got Flan, which is, I believe, C++ code. And it's sort of gnarly and hard to deploy. The nice thing about it is you give it your data and then it takes a long time to train, but it figures out the optimal structure for your problem. And that gives you high performance in the end. And there's a Python wrapper for it, which works, I am told. There are some bits that we don't like as much. So LSH forest itself is in scikit-learn. You can read the source. It's fairly readable, but it's actually very slow. So if you want to develop for a high performance application, maybe not the best solution. Flan is a pain to deploy. It's C++ code. You need to have C make all sorts of dependencies. And all is great. I recommend you use it. But for us, it didn't really fulfill two important requirements. One is it doesn't allow you, once you build a forest of trees, you can't add any more points to it, which for us was a no-no. We need to add new products to the index as they come in. So this is something we needed. And secondly, you cannot do this out of core. You have to keep all the vectors in memory all the time. So yeah. So like any engineer, we wrote our own. It's called RP forest, which speaks to the algorithm. It's available on GitHub. And it's pip installable as well. So please go forth and try it out and break it in all sorts of novel ways. And I'll try to fix them. It's quite fast. It's not as fast as annoying. But it's fast enough for us. It's certainly much, much faster than LSH forest, which is built into scikit-learn. Allows adding new items to the index. And does not require us to store all the points in memory, which is really, really, really very nice. So how do we use it? We use it in conjunction with Postgres QL. So basically, we have a lightweight service that has the ANM indexes in it, the RP forests. So we send a query point there. And what it gives us back is these are the product IDs or these are the image IDs you are going to be interested in. So we get these IDs back. And then we push them to Postgres and go, dear Postgres, here are the IDs. Please apply the following business rules or the filtering or all your worst statements and so on and so forth. And then do final distance calculations also in Postgres using C extensions. So we store the actual point locations, the actual vectors, as arrays in Postgres. And we've written some C extensions to Postgres that allow us to do the distance calculations in Postgres, which is quite nice. Side note, Postgres is awesome. If you're doing all sorts of numerical stuff, you have arrays in Python. You have arrays in Postgres, and you can write custom functions to do anything you want in C. So if you really want to, you can write your stochastic gradient descent, machine learning algorithms in Postgres and random in Postgres. I'm not sure you should do that, but it's definitely possible. So the whole combination, the algorithm and the implementation and Postgres as a data, as a backing store, gives us a fast and reliable ANM service that we've deployed in production, gives us approximately 100 times speedup with a 60% precision at 10. If we get 10 nearest neighbors, we get the 6 out of 10 actual nearest neighbors using the approximate approach at 100 times speedup. So I think that's reasonably good. Speedup over brute force. Not particularly demanding speedup baseline, but that's where we start. So we've got that. It allows us, and the speedup that we gained, it allows us to serve real-time results. So we don't have to pre-compute. We can just serve it real-time, which is also very nice. And it's all built on top of a real database. So stuff like concurrency and updating the database that's all taken care of by smarter people than us. So it works well. Anyway, thank you, and I'll be very happy to take any questions. Yeah, hi. I just did something like that for chemistry with distance, and I'm doing brute force. But I would like to ask you for an estimate of what number of entries this is going to be hard. So when is my application going to need trees like in the future? I don't know, Ben. I mean, it's sort of an empirical question, right? So it depends on your requirements as well. So let's say if you're doing offline processing and you want to, you don't really mind that it's taking 10 seconds or 20 seconds with the data you have. That's fine. But when you, let's say one lookup takes 10 seconds, but you need to do your lookup 100,000 times, then maybe that's the point where you really want to look into these solutions. If it's fast enough for you for what you're doing, you don't need it. So what was your pain point in your application? What was too long for you? So let's say if we want to serve web requests, like 100 milliseconds is too long for us. And doing this brute force would take anywhere up to three seconds and it would completely destroy the database. So getting from three seconds to under 100 milliseconds, that was the difference for us. Hello, and thanks. I have a question about clustering. So are there any algorithms like KAMINs and so on, which could be fast and precision, have nice precision? Thanks. Yes. So the question is, can you use clustering algorithms to achieve the same sort of effect? And yes, you can. And there's an approach called hierarchical clustering, which again is sort of building, let's say a binary tree, for example, where you have all your data points and you put them into two clusters and you build two clusters and then these are your two clusters on the first level of the tree. And then you go down into each cluster and you recursively build more clusters and keep splitting the tree. That's also an approach. It's not something I have investigated. So I cannot give you a good answer on what the performance tradeoffs are. Okay. Any more questions? Cool. Thank you. Thank you, Matt. Thank you. Thank you, Matt. Yeah. Hi. Michael. Crystal. Hi everybody. Welcome. Welcome.
|
Maciej Kula - Speeding up search with locality sensitive hashing Locality sensitive hashing (LSH) is a technique for reducing complex data down to a simple hash code. If two hash codes are similar than the original data is similar. Typically, they are used for speeding up search and other similarity comparisons. In this presentation I will discuss two ways of implementing LSH in python; the first method is completely stateless but only works on certain forms of data; the second is stateful but does not make any assumptions about the distribution of the underlying data. I will conclude the presentation by describing how we apply LSH to search at Lyst.
|
10.5446/20154 (DOI)
|
Thank you. Hey, folks. It's not quite good morning yet, so I guess good afternoon. I'm still stuck on San Francisco time. I've had a lot of coffee, though, in some sleep, so I'm a little jittery, and I have to pee a lot, but I'm mostly awake. So who am I? I'm a back-end engineer for Spotify. I am based in San Francisco, as I said. I am also vice chair of the PSF, along with Naomi Cedar. Side note, if you are a PSF member, quick announcement, there is a PSF members meeting Thursday at 6 p.m. Forget the room, but it's on the online schedule. And if you're interested in becoming a PSF member, you should come talk to me. Last bit, I'm also the founder of Pied Ladies of San Francisco, and I'm one of the main, like, lead organizers of the global organization. And thank you. Since I have the stage, who here has actually heard of Pied Ladies? Maybe it's better to go, who has not heard of Pied Ladies? Okay, those hands will be down at the end of this paragraph. So Pied Ladies is a mentorship group for women in the Python community. It's open to women and friends, so it's open to everyone. Essentially we're a loose group of meet-ups in various locations. One on every continent except Antarctica, which is my new mission, to go to Antarctica and start Pied Ladies. But what we do is we host Python workshops, speaker events, hack nights, coffee meet-ups, everything around Python and learning Python and development in general. And we welcome all experience levels. So I think there's one here in Spain, I think based in Barcelona, that I highly encourage you to check out. I also have a talk, another one, on Thursday afternoon at 2.30, I figure which room, I think the Education Summit room. And that talk will be about Pied Ladies and more in-depth talk about what we're doing and how the work has actually been going and the effects that we've been seeing and the work that we still have to do. So who here has not heard of Pied Ladies? Come on, I see two hands. You need more coffee like me? All right, so I gave my spiel. All right, so this talk, I will first give a quick introduction to Spotify over an overview of how we use data. I'll go into sort of how we use metrics and how we came about to implement them, the agile way and essentially what was learned along the way when my team implemented them, sort of the bigger picture. And you can sit back. I'll give a link at the end of this presentation with the blog post and the slides. So you can just watch me rather than your computer or tablet or whatever. All right, so what I basically want you to take away is metrics and tracking is super fun, but should you track all of it, everything that moves? And we as developers, we track everything like website visitors, referrals, how folks use our services, if our servers are even up. We have a lot of tools at our disposal like New Relic, Graphite, Google Analytics, Sentry, PagerDuty, a bunch of other things. We even track ourselves like steps and sleeping patterns, exercise, calories consumed, breathe everything, hair growth rate, I don't know. Everything that you can, we track, right? Maybe hope to get some insights or just to feel better about ourselves. But if should you measure everything? And it's very easy to get lost in the forest, right? It's easy to lose the meeting and easy to lose the understanding of why you're measuring it in the first place. So to start, some background information about Spotify. We're just all on the same page. And also how we use data. So Spotify, streaming music service. We've updated our logo. This is the not the correct green, but I haven't gotten a new shirt yet. So we beta launched in 2007. I think we came to Spain and some other parts of Europe in 2008. And in 2011, we came to the US. We're in about 58 countries. We have over 20 million paid subscribers and 75 million active, monthly active users. We have over 30 million unique songs, not including compilation albums and such. And we add about 20,000 songs a day. We also pay about 70 to 80% of our income to rights holders, totaling about $3 billion to date. While I work in a very small office in San Francisco with about five other developers, our main engineering offices are in Stockholm and in New York with a lot of data and machine learning in Boston. So as you can imagine, at Spotify, data is quite important. These numbers that you see are only about a month old. And I have to check every time I do a presentation like this because they are always growing and growing fast. So we track user data like signups and logins, activity within the application itself, even tweets like the good and the bad and the ugly. We also track server-generated data, including requests to various services, response times, response codes, among a million other things. Each squad owns what they want to collect, how and when, and how they will consume such of data. We have analysts that run thousands of Hadoop jobs a day to glean insight from user activity, answering questions like how many paid subscribers do we have at this moment in time, or was this partnership financially beneficial to us? Engineers behind the platform watch usage rates to like our web APIs, we watch login failures, feature usage, et cetera. We also have data scientists and machine learning, analyzing listening behavior, music, metadata, and trends that power the recommendations behind our features. Teams have actually started to analyze actual audio signals and the sound of a song to pick up genres, beats per minute or whatever, and instruments played. Yeah, it's actually quite difficult to pick up a few things from the audio signal itself, like the mood, how do you classify and define a mood of a song? But that's the stuff that we're doing at Spotify. And this only scratches the surface of what we collect and what we pay attention to. We do use various technologies related to data, including Hadoop as well as Cassandra, Postgres, and Elasticsearch. All of our user data sits in Hadoop, which we run jobs against using our own Python library or Crunch, Scalding, or Hive. We also use Spark, Tes, and Flink. I've heard actually a lot of people use iPython with Scikit Learn and Pandas. And I've also discovered recently that we have our own iPython notebook server set up, so that's pretty cool. On the back end side, some of our service activity actually gets parked in Elasticsearch where we have Kibana set up. But the majority of service activity is actually in a homegrown system, of course. But we have open sourced it. It's called Fast Forward or FFWD. And it's written in Ruby. Sorry. Yet, with all this setup, with all this technology, I'm really embarrassed to say that my team did a lot of development in the dark. We were not tracking anything. We did not know how successful our future integrations were. We had no clue how our back end services that we maintained, how they were holding up. I do want to make note that a lot of squads in Spotify do track a lot of data to do, pay attention to all of this. We're just sort of the black sheep. I think it's partly because we were nine hours behind Stockholm and three hours behind New York. So this story, it's about self-discovery, how we became a better, more effective team. And we did this by capitalizing on understanding our own data. Not everyone can be data scientists, mathematicians, statisticians, analysts, whatever. But everyone can grasp why it's important when 70% of our users can't log in for whatever reason. And so this is a story of how our team finally developed and adapted the use of logging in metrics. So you might know that Spotify, very public about using Agile. We actually have a few videos on YouTube that are very, very awesome, very entertaining and awesome, nicely done that I highly encourage you to check out. But one key aspect of Agile is iteration. And we certainly iterate over our product. You might be as annoyed as I am when you open up the Spotify client on your desktop, it has a blue banner asking you to update. And it comes pretty much every single day. So that's our Agile approach. But we also iterate over ourselves as a team, as an individual, as a squad. I'm trying to find what works for us as a company, us as a squad, and everything in between. Late last year, my squad began participating in an internal program. It's a very corporate speak called Effective Squad Circle. And its purpose was to hone in on the squad itself, how we can be more effective. And I actually found that to be very beneficial for myself and for the team. So what it was is essentially monthly challenges to figure out the team's current condition, essentially not tracking anything, not knowing what's going on, and comparing it to the desired condition in terms of delivering the product, the feature, service, whatever that we were meant to. The following explanation might sound very like project manager-y, business-oriented-y, but I found it very useful when implementing metrics. And I also find myself definitely incorporating this thought process when talking to other teams non-tech and tech about diversity and stuff like that, like measuring our diversity initiatives. So it's very widely applicable. But so the main goal was to find our target condition as a squad. Where do we want to be? And certainly difficult to establish a goal without context, without an understanding of where we are now. And so to figure out our baseline, we sat down and answered a few questions as a group. So the first question, what do we deliver? Seemingly easy question, right? Yet myself and the squad initially struggled to answer this right away. It definitely didn't roll off our tongues. So we looked at our past and listed out integration projects that we delivered and services that we currently maintain. Includes Uber, Last FM, Soundhound, Twitter, Hashtag, Music, among others. The most critical is certainly our Facebook login and signup registration. As I've hinted before, about 70% of our user base logs in via Facebook. The rest is an email login which my team does not own. The next question is, for whom do we produce said product or service? And who actually defines our work? At Spotify, we believe that leadership is meant to convey a vision and the squad is meant to implement that vision in a manner that they choose. There isn't micro management. There's a lot of trust, actually. But our lead team defines the direction that our squad takes. So they're certainly one of our customers. Also there are many integrations we've done, a lot of external partners. Thankfully the squad's a bit shielded from direct communication. But that makes our business development team and indirectly the partners themselves, our customers as well. But who depends on us? Who actually uses our work, product, or service or whatever? So yes, 70% of users log in via Facebook. It's safe to say it's a pretty integral system to the Spotify platform. We certainly can't fuck it up when Facebook makes breaking changes to their login protocol or API, which they often do unannounce. And I've had to live patch our servers because of Facebook. But there are other teams within the company that plug in to our system for social aspects like sharing to Facebook within the platform. Moving on, the next question is about expectations. What do our customers actually expect from us? When trying to answer this question, it occurred to us that we never actually asked them what their expectations were. And so we did. We wanted to know exactly what was important to them with what we deliver. Was it on time delivery? Was it predictive versus productive? Do they expect solutions to problems that they don't even know existed? What were their expectations on quality, on usability, other measures? Were there expectations with how we work as a squad? Did they want updates on progress and problems? So we couldn't ask all our customers, right? We have 75 million customers. And expectations would be different among our various customers for the team. Internal teams expected our Facebook service to be reliable and scalable. Business development wanted us to be clear on what we can feasibly implement, which is definitely hard for a web developer to adequately say, appropriately say, how long something will take, right? And it's safe to assume that users will want to log in or sign up via Facebook if they so choose and for it to just work. So moving on, the last question was simply, did we meet those expectations? How do we know we've met those expectations? This is where we sort of stopped in our tracks, right? We didn't know if our services could handle the extra load or if slash when users couldn't log in. Or how many users have activated Spotify with Uber and of those, does the experience actually work? So being people that have an affinity for tech and automation, we naturally implemented a technical solution. So feedback loop, very generic term, not just tech. But to understand the feedback given for our squad, one of our main feedback loops that we chose was metrics. We all wanted those snazzy looking dashboards, eye candy graphs and visuals using the latest technology that will probably be obsolete tomorrow. But in all seriousness, we wanted an immediate visual of various metrics. But what do we want to see? What questions do we want to answer? So in line with the idiom, to throw spaghetti on the wall to see what sticks, the squad brainstormed for a while trying to come up with questions that we would like to answer. So some ideas included sign up or off-flow abandonment, Facebook connected users, the percentage of toll users and the trend over time, percent of users that signed up through Facebook per hour day week, we didn't even know what the frequency should be, Facebook related errors, which is a lot, daily active users by partner feature, registration, subscription rate, referrals by partner, web API usage by partner. We wanted a squad focused Twitter when you search for Uber and Spotify to see what people are complaining about that neither Uber or our team could see in our logs. We wanted to know outstanding JIRA issues, request count by internal requesting service or team. So you group these metrics into buckets like usage, system health, business performance. These buckets eventually are becoming their own dashboards, cycled through our big office monitors like everyone sort of has. But we also created a few new processes based on these questions. One of the process reviews our progress as a squad. Every retrospective we will look at a couple of metrics that deals with the squad performance like how many bugs were closed in the past sprint period. We will also judge if this metric, if we want to continue seeing this metric, if we can actively improve upon it, like maybe we only closed two bugs this week, but it's because it took us two days to acknowledge one bug, right? And what if new, if any, measurable items should we look at for our next retrospective? Another is to have goal targets for every integration project that we do. For example, we will know we're successful when we have, with this integration, we have X amount of new users within the first two weeks. It's true that this sort of goal line can only be judged based on historical user acquisition numbers. So we basically, so we definitely have to do some work, but this will feed into our retrospectives, especially once the project is complete, like how did we do? We also have a few post integration questions for business development folks to ask of our external partners on behalf of the squad, like understanding our responsiveness, how our developer tools are, if they met their goals. And we may think an integration was super successful, but on there and not so much, which has definitely happened to us before. So we've only been caring about metrics for, since the beginning of the year. So this certainly is only the beginning for us, but it allowed us to iterate and give us a hard look at what we track and why. You can track everything that moves, but will you get inundated? Certainly so if you count every leaf of the tree in the forest. So how can you tell what's important and what's just noise? And so this goes back to understanding your customer's expectation and essentially boils down to business value. How can you maintain and improve upon business value of your server as a product? How does counting every Facebook connected user help us better ourselves? So when thinking about implementing various metrics of our feedback loops, I came across various questions to help me see the forest for the trees. So when creating a new metric, how do metrics map to business goals? Do we lose money or lose so much money if the Facebook's login service isn't up? How do you prioritize different goals of what you want to drive? What's most important? Does it mean that you're going to neglect others or just allocate time by priority? Is this new integration project more important to pay attention to than other ones? That's fine if it is, but you're just going to have to prioritize. How can we create dashboards that are actually actionable? What is the goal and more importantly, how can we drive that goal? Are we just going to say, oh, look, Facebook signup service is down. Let's go have lunch. When representing metrics, how do we correctly measure what we care about? We have to break out the old static book to understand how to best represent all the metrics that you take. We have so many tools to help us create gauges, counters, meters, histograms, timers, but what representation is best for that question or metric? When actually consuming them, how often do you check on metrics? I've never looked at, which is a common problem I found in my team. They just become background noise. How do you make dashboards more visible, more in your face? Should someone be responsible once a week to a goalie? Do you make them more visible by slapping them up on the TV monitor, which I found does not entirely work, especially if it's right in front of me. I just kind of ignore it. Or perhaps you have email snapshots sent out to the team, but maybe they're automatically filtered away, or you're like me, and auto archive all unread emails. Being a bit introspective, for things that we don't reach 100% of our goals, we need to assess the difference. Why does it exist? Is it even solvable? If you look at dashboards, what actions are you actually going to take? Do you even create a dashboard if a goal or an alert isn't set up or if no action will be taking? Probably not. What about unknowns? What is a known? We know that X amount of IOS users have connected their accounts to Uber, but how many don't use it because the driver has an Android phone or the driver just isn't aware of the service? How do you approach those unknowns? Are you comfortable with them? Is it even worth it to explore them? Bringing back to this slide, ultimately the goal in us answering these questions is to give us both a short decision making cycle as well as more informed decisions about strategy and partnerships. It's super easy to get lost in the forest and it doesn't help us that we can get all this instant feedback that all these visualizations just look awesome. But in essence, we're in a place in current values in historical context in order to see patterns developing. How long on average does it take for the team to implement a new integration? Do our customers or ourselves expect a shorter turnaround time? Do we wish to just be able to appropriately estimate the time and work on the work that takes for such a project? Or maybe which internal team do we have to educate about rate limits on our service? The one here with these feedback loops, these thoughtfully implemented metrics, we can use these goal lines in the alerts to create a more efficient team. We'll deliver higher quality software since we'll get immediate feedback on any bugs that we introduce, any system that fails and the like. Before I answer this question, before I wrap up, I do need to be a good friend. With all these questions and contexts in your mind, you should go to Hennick's talk tomorrow. It's 11.45 and it's about practical, it's title, practical logging and metrics. It's basically the tentacle complement to this talk. So all right, to answer this question, should you track everything? Very anticlimactic answer? Probably. Only if you define a goal, you can define an action if you haven't met that goal and if you actually pay attention to it. I know, anticlimactic. But within reason, right? So thank you. I hope you took away something. And I think I have one minute for questions. Or how about we just go out and get some wine and you can find me if you have questions. Yeah? All right, thank you. Thank you.
|
Lynn Root - Metrics-driven development At Spotify, my team struggled to be awesome. We had a very loose understanding of what product/service our squad was responsible for, and even less so of the expectations our internal and external customers had for those services. Other than “does our Facebook login work?”, we had no understanding of how our services we’re responsible for were doing. How many users actually sign up or log in with Facebook? How many users have connected their Spotify account with their Uber account? Do folks even use Spotify with Uber? With a 2-month challenge period, my squad and I focused inward to establish those unanswered questions and to establish feedback loops and always-on dashboards. This talk will tell the story of how we chose which metrics are important for us to focus on, what technologies we have used and are using, and how we’ve iterated over our feedback loops to fine-tune what metrics we care about.
|
10.5446/20151 (DOI)
|
So we will try to make it short, like around 35 minutes. So what I want to do is first, five minutes, just give you a few information about what is Odu. And after we'll do a small demonstration about the product, and then I will show you a little bit of code to see how it works behind the scenes. So let's start with some information about Odu. Odu is actually a big project, but it's not very well known among the Python community. It started 10 years ago in 2005. And actually, the reason why we did it in Python, so I was at the university around 2000 in Belgium, and there is a guy, and I was coding in C in PHP. We were doing a lot of websites. And one day, there was a guy that came, and his name is Denis Frère. I just found an email from that era. You see, people were talking about hosting some Perl web application, and he says, you should do it in Python. And he came to the university, to the dorms, and he showed us what was Python. And we started to learn Python, and since that time, Python is my main programming language. So I've been programming in Python for 15 years. And okay, let's go back to Odu itself. So Odu is a framework and also a set of applications built on that framework. The total code base of Odu is around 140K lines of code. 40K lines of code are the frameworks, and the rest is the 30 main apps. So we are a company. The company is named Odu. We edit the software, the software is also named Odu, and that software is made, is actually 30 different business apps. And I will show you a demonstration of a few of them after. And then after, you have a lot of people that are doing other apps and using the framework or extending the apps that we ship. And so you have a very vivid community around Odu. We have around 400 contributors in the core with a truck factor of 11. So a truck factor is the number of people you have to kill if you want to kill the project. And if you compare to Django, for example, it's five percent. If you kill five percent Django, we'll probably stop. And Rails is seven. Odu is 11. The Linux kernel is around 150. And if you take account of the people that doesn't contribute to the core, but to other modules, then there is around 2,000 contributors. So it's a very big project. And there is around 500 companies that do their living because of Odu, and some of them are in Spain. We have 2,800 stars on GitHub. And we have 2 million of users. So by users, I mean somebody who every day sits at these computers and uses Odu. Maybe not the whole day, but at least some part of the day. So every day, we have 2 million people that would log and use Odu. Why is it different? I think it's superior to many frameworks, actually, to almost every framework. Maybe there are some frameworks that I don't know, but because it's very modular. And all frameworks say they are modular. But Odu is modular in a different way, and I will show you some examples after. It's business-oriented. So it has all the features you need to do business apps, like the security mechanism to people in the company might not see some information that other people can see. So all that kind of thing is built in. And it's only since one year that it's fully web-based. Before that, it was not fully web-based. It was, you know, at first started as a client and server application. So you had to run a client application on every PC and you were running a server. Since one year, it's fully web-based. And it uses a special templating engines that nobody else uses, but I will show you why I think it's superior. And also, it has a small JavaScript framework built on standard libraries. And so it has a rich client. It's a full, like a native application in JavaScript. And also, it is very simple because you see, it's only 40K lines of code for the framework. And the API is around 30 functions. So if you know those 30 functions, you know how to consume and program everything you know. And the power of Odo comes from what is already available, from its add-ons. And you have add-ons for everything, everything related to business. So when you do, you need to make invoices, payment, manage product, deal with customers, with physical products, accounting, those kind of things, they are built in Odo. So I think many people who develop new web applications, they spend most of their time reinventing the wheel. Because when you do a new, for example, a new startup, at some point, you have to invoice customers. You have to do payment. You have to manage people. So those sets of features, they are built in Odo and you don't waste your time doing that again and again. So why isn't it more popular? I think there are a few reasons. One of them is because it's business-oriented and business is not fun. People prefer to do games instead of business. It's wrong actually doing business apps. It's very fun. Also it was not web-based before. So now you can build web applications using Odo. Before it was not possible. It was just a tool to manage companies. So it's very recent. And also we had no documentation or bad documentation. It's only since December that we have a good documentation. And also Odo is not a good Python citizen. It's not packaged on PIP, for example. So you cannot do PIP install Odo. Because Odo is quite a big project. And it's actually also lots of JavaScript, lots of other things. It uses PostgreSQL as a database storage. It's not just a Python library. And also there were a lot of ugly code, ugly quirks, Odo. And we fixed most of them in the V8 API that we released in September last year. So no, it's much more cleaner, much more Python-ing that it was before. Because we started at the time when ORM didn't exist. SQL object was the first in Python that I used. But it was much later after we started Odo. Yeah, so Odo was named OpenERP before. And before that it was also named TinyERP. So we changed the name two times. So I will show you what it looks like. Let's go. So I will show you the new UI. The new UI is still alpha. So sometimes I might switch to the stable tree because some things might not work in the development version. That's the master tree. So basically, at least at the beginning, Odo was just a tool to manage company. And I will show you just a simple flow of what happens every day in every, not every but many companies. So we will look at the CRM. The CRM is a tool to organize your sales. So how it works is like this. And it's the same in many, many companies. So you first have a contact with the customers. And this is this, this customer that wants to buy some keyboards. So that's the first contact you have with the person. And what will happen is that you will call the customers, discuss with them, qualify his needs, then probably you will do a concrete proposition about selling him something. And then there is sometimes a negotiation phase. And then after we get a deal, we lost the deal. So opportunities are just one of the simple objects that Odo managed. And you saw here, that's the view with the flows of opportunities. And I can show you what an opportunity looks like. It's just a few information with some fields like, OK, who's the customers? This is Agro-Lay. And what's the revenue we expect from that deal? What's the probability? OK. And there, so it's just crud. So it's just information that you save. And you can see the flow there. And you see at the bottom of the document, we have what we call the chatter. It's just like a Facebook thread. And it's available on every business document in Odo. So when I look at the opportunity, I can discuss with my customers using this here. I need to add this email address. OK, it's already done. So I send in a message. What are you interested in? Blah, blah, blah, blah. It will send him an email. If the person's reply by email, it will come back on the thread below. So I can manage all my communication from the opportunity itself. So at what point, the customer will say, OK, I want to buy this or this. So what I do, I will create a quotation. A quotation is just a set of things that I propose to sell to our customers, like maybe some keyboard here. And after, I can send the quotation by email. Oops. Oops. OK. I send the quotation by email, the person can reply on the quotation, it will appear below. And then after, there are two ways you can close the deal. One of them would be, OK, the customers buy phones say, yes, I want to buy this, send me the goods. Or you can also use an online version where I click here, I can see the quotation here. And the person can accept the quotation online, sign, it's difficult with the mouse. My name, OK. And here, I accepted the quotation. I go back into sales. After the quotation is done, what we will do is to create an invoice. I create an invoice. And also, it will create a delivery order. So the delivery order, it's just a document that goes to the warehouse and the people will deliver the goods to the customers. And the invoice is the order part, it's the customers that has to pay the money. And after the person's pay, I won't go into all the details because I want to make it quick, but you get the ID. After the customers will pay. And we can reconcile, that's the old UI. I just wanted to show you here, it's the way you reconcile invoices with the payments. So you get some money on the bank account, you have some invoices, and then you reconcile them. So the deal is completely finished because we've delivered the goods and we've got the money from the customers. Very simple flow. And what's available also in Odo, we have a... Oh. OK. I've lost the Wi-Fi. OK. I wanted to show you, I had another version online with more data and more things installed, but I think I don't have internet anymore. So I want to show... OK, I can show it here. So you have also tools to do reports. For example, you can know... You want to know the sales made by every country, by every sales team. OK, it's always the same here. And you can also get nice... I will just install something. OK. So what you see here are all the applications of Odo, the basic ones, the ones that we as the company Odo edit, and there are plenty more outside that are made by other companies. And I'm installing the CRM so that we get more data here. Let's go back to... OK. So what we saw was just a simple sales flow. But there are other ways to sell your stuff. One other way would be to have a shop. And Odo also has a point of sale. I hope it's open. Yeah. OK. Here. So here, it's touch base. So you can click here, click, click, click, and pay. So that's what the person will use in his shop at the... I don't know. In French, we say à la caisse, a casa, maybe in Spanish. So here I pay some amount. I can validate. OK. I need to pay at least the amount. OK. Here. OK. And then I go to the next order. You see? So that's another sales channel. And I told you recently we added the web layer on top of Odo. Odo was already using HTTP because the client server protocol was XML RPC. So we had a small web layer, but we decided that because we had almost everything that the company needs to manage itself, we were just lacking one step was the website part of the company. So we say, OK, let's do a CMS. And I know Plone was supposed... Plone or Zope was supposed to be the killer app at the time I've started with Python, but didn't happen. So I will show you what the web layer of Odo looks like. Let's go now on the website. So we say, OK. OK. Because we have information about many things that happens in the company, and the company needs to have a... Needs to be public, needs to have a presence on the web. So we wanted to make a simple tool to allow people to create their own website, and there are many CMS available, maybe hundreds at least of them. So we wanted to make something different, something really easy to use. And we made this website builder. The website builder works with blocks. So what you do is you drag and drop different kinds of blocks like this. OK. And then you edit these blocks. And that's the basic blocks that you see, but I will show you some team after. And then you can edit the content here and change it. And because you use blocks, you're not creating all the layout yourself. So you can have beautiful looking websites because some designers created some very fancy blocks for you. So for example, if you want to compare price, you can use this block to compare price and say, OK, we have three offers with that price, that price, et cetera. OK. And that's just static pages. Here because we use the same system, for example, here I have a contact form. And when I fill my contact form here, it's linked with the back end I showed you before. So when I will fill the contact form, if I go on to sales, I can see that I have a new opportunity. Actually, it's a lead. Here, it's here. So each time I fill a form, I arrive in my CRM and then the flow of sales begins. OK. Pretty straightforward. And after, we said, OK, we have all the information about the product. So let's do an e-commerce. And we just publish the product online here. And we have the add to cart button. And when you do add to cart, you actually do a sales order, like I did before. I did the sales order myself. Now, it's the customer itself that creates the sales order. The object behind is the same. So for the whole e-commerce that you can see here, I need to fill all the information. Blah, blah, blah. OK, shipping to the same address. And then I can pay. Here, I only install wire transfer, but we have integration with many payment providers. OK. And the whole e-commerce flow that you saw here takes only 1,000 lines of code because we had all the built-ins in Odu already. And I only showed you the easy stuff because you can add more and more features and get very complex flow in the companies by adding some options. So the forms I showed you when I do a sales order, it's just one or two fields. Let's go back to a sales order. You see? It's very simple, name of the customer's address, things to sell. But you have many cases where you need more. It's OK. You go in configuration, in settings, and then you enable new stuff. For example, I want to display margin on the sales order. So I want to know how much I get when I sell something 100 euro, how much do I gain on it? And I go back to my sales order. And here you see there is a new field. And you can add many, many things. And at the end, you get 20 fields. And it's very complex. But the way it's designed is so that people can start easy, but if they have complex things, they can manage it. OK. So let's go back to my web website. And what's displayed on the website, what you can see here, if I edit, I can edit everything. So here, if I edit this, it will change the name of the product. OK. If I change the price, it will change the price. So everything that appears on the screen, you can edit it. And we do that because we have a special template engines. And we know when we display something on the screen where it comes from in the database. And so you can edit everything. And the same system to edit static page can be used to edit products. So I can use also my building blocks here. And here I save the description of the product. OK. So what I showed you is just two applications, the CRM and invoicing. And we also host it online for people. So we have a cloud offer. That's the pricing of the cloud offer. But you can see there are lots of more applications like managing, manufacturing, accounting, project management, inventory, point of sale events. So if you want to manage an event like this, you can use also a do. Also yeah, I wanted to say the people, the guy who introduced me to Python, that guy Denis Frey was the guy who started Europe Python. So he started the first year in Charleroi. I remember we're organizing the event with him. So if you are all here today, it's because also of him. Let's go back to my, yeah. So I finished my demonstration. It's not, I don't want to be complete. I just want to tease your curiosity so that after you might go and say, that Python project looks interesting. So let's dive into the code and look what it looks like and why I think it's superior to many frameworks. OK. Hello. I think it's big enough so that people can read. Do you define object like I think in many frameworks like Django or Rails by defining the fields. For example, we will check an invoice. That's a journal. Let's take an invoice here. OK. Yeah. It's an invoice. So there are many types of fields. Odo stores its data into PostgreSQL. So you have chart fields, selection fields, integer, Boolean, dates, and relational fields. So you have many to one when it's a relation to one of the objects and many to many when you have multiple relations between two objects. Pretty common. And what you have a special in Odo is the compute attribute. So when you say, for example, amount tax or maybe amount total, it's easier. Amount total is the amount of the invoice and it depends on the order data. So it's a compute field. So you say, OK, this field, it's not a real field that is simply stored. It's a field that you compute when other value changes. So I will check. So this function is called each time a dependency fields of the amount is changed. So each time, for example, you add a line on the invoice, you have to compute the total again. So it's just a function here that you define and it will compute the value. And then you have two different ways to use compute fields. One of them is to say, OK, each time I need the field, just compute it. Or you can store it. So you say store to. So it means that we compute it when it changed, but then we store it in database. And by storing in the database, it's much more easier to after search for the value of the fields. Because if you want to search for a value, if it's not stored, you have to compute all the values to know which one is the correct. So you use it like a regular field, but it's computed. And many of the business logic in Odo goes into those compute fields. After you have a few functions, business functions that are linked to the different buttons on the object. So for example, when I click confirm an invoice, what should it do? So that's all the code that we see here. And after you have the view. So how to display the invoice on the screen. And I show you the back end view. It's in XML. So let's check a continuous. So that's the view description. So it's just the list of fields to display with some layouting. Like I want to see different tabs or group the fields together. And that's it. After you have also web pages, but we don't have web pages for invoice. For the sales order or the product page, I will show you what the page looks like. I want to show you that. I will show you the extension mechanism. But it's also just an XML based template like Genshi or Kidd or Zop template. Actually it was inspired by Zop template. And what I want to show you is how to add new stuff on Odo. And that's the strength of Odo. You can say, okay, I show you that I installed product margin. It's just a module that adds a field and shows you the margin. How does it do that? You can inherit object and override methods. But everybody knows how to. It's common in Python. So it's very easy. What you can do is add new columns on those objects. You can add new fields. Once you have added those fields, here, I have the compute fields. That's the old syntax. And I added the two functions to compute the fields. After I want to display the field somewhere. So I go in the view and I say, okay, I want to extend the view. And after this part of the view, I want to display this. And everything is constructed this way. That's why I showed you at the beginning that my form view was very simple. Where is it? Here. I show you my form view is very simple. But the more and more module I install, the more complex it gets. And it's really small Lego bricks that you can build on up to and you get a very big and very complex system and everything works together. I think I will ask if you have any questions. I will take questions. And I hope that I teased your curiosity and that you will look at Odo. If you have any questions after the talk, feel free to come to see me. I will be there this afternoon and also I will be there at the social event. So don't hesitate to come and ask me a question. Thank you. And let's proceed with the question if somebody has questions. Hello. I would like to know how you compute the read access and write access for documents. So we have two mechanisms of security. One is the access control list. So you can say this person from this group doesn't have the right to read, write, create things. It's almost like ACL in Unix. And you have a second mechanism. It's record rules. It's more precise. It can give you visibility. So you say every time somebody do this operation, apply this set of criteria. I didn't show you. But we have a filter mechanism here where you know you can say, I want to know. You can see the sales order that are to these customers. So we have a syntax to express filters. And it's called domains. And you can apply domains based on the person who use it. So for example, you can say every salesman can only see his invoices or that kind of things. And it was built into Odo since the very beginning and it didn't change. How do you deal with large data sets? Because I'm familiar with the approach. The biggest deployment of Odo is 50 terabytes of data. We don't do nothing. It's just Postgres is very cool. And Postgres can deal with a lot of data. So we rely on Postgres. And also, when you have blobs, attachments, image, they are stored outside the database. So only the business data are stored in Postgres. I didn't perform my question precisely enough. How do you deal with situations where a subset of users in these really large data sets, when a particular user only has access to a very small amount of documents? For instance, you have 100,000 documents and the user can only see 15. And what's the problem? Well, there's no, your implementation bypasses that. Yes, no, you can use this syntax. But what's the criteria? You have thousands of documents. Why does he see only 15? It's for whatever reasons. So you have to define the, once you define the reason, you get the system work. You just define why and it just applies. Okay? I can show you example if you want. Hi. I have two quick questions. One is if you start with a managed solution, is there a possibility to go to a hosted solution? Yes. So, you want to know that you retain all the data? Actually, yes. So, if you start with what you have for the 10 euros, I don't know, a month and then want to host it yourself, if you can do it and retain all the data? So in the cloud solution, we only ship the basic apps, the one that, you know, the 30 basic one. When you, if you want to use external modules, you need to be self-hosted. So you need to be on premise. So the cloud solution is just for people, for small companies, because for simple companies. When you have complexity, probably you need to go self-hosted. Okay? And the other question is when you have saved fields. When you have what? Saved fields. The fields where they... Yeah, yeah. You can add custom fields and actually, I didn't show you, but you can do it from the UI. You can add fields and customize everything from the UI. You can do it in Python modules or you can do it in the database itself. When do they know when to update? When you are on the cloud, we migrate database from version to version. So we keep the customization and sometimes it's manual work because we have to make sure everything works. And that's part of our offer. If you are on your own, you have to be careful with your customization when you migrate from version to version. Is it clear? Okay. Just come to see me after. Okay. Any other question? Yes. Hey, do you know about any company using Odo in Brazil? Yeah. We... I don't know the names, but there is a partner named Acrecion. That's a company that integrates Odo there and they have plenty of reference. I don't have internet access, but I could show you... No problem. Just curious because in Brazil we have some business logic with this particular to the country. Oh, yeah. I know they told me that it's a hell about accounting and the paperwork you have to do. So they have a lot of modules to deal with that. And I know it's very complex. So this is already done. So if I install Odo and I can just use it in Brazil? Yeah. Yeah. Okay, cool. Thank you. And I forgot to mention also in Portugal, it's not in Spain, but in Portugal, the biggest installation of Odo is 500,000 people. It's all the teachers in Portugal that have to use Odo because they use Odo to schedule, to assign teachers to the school and that kind of work. So the Ministry of Education uses Odo for everything now. They replace I think 1,000 different applications, no, 100 different applications with Odo. Yeah. Cool. Thank you. Okay. We still have time for a very short question. Okay. Which one is the shortest? I have a short question. Yeah, okay. The answer may be long. When you go into price, one of the common challenges is integration. You have legacy apps, some of those will go away, some of those need to stay forever. Yeah. So you need at some point to consider how to bring in data, either live or migrating into and out of Odo. So you can do it both ways. Would you think this is a good framework for that? Yeah. You can do it both ways. You access the Odo API using Gson RPC or XML RPC from the other apps or you do it from within Odo using any Python library, you access the database or the API of other systems. So there's a lot of work and connectors for other systems. Yeah. Well, thank you again, Anthony. Thank you. Thank you. The lunch is happening now, so just try to get some food.
|
Antony Lesuisse - Odoo the underdog python killer app. A python framework for web based business apps. Odoo is used by 2 millions of users, although relatively unknown in the python community, it has a vibrant community and is one of the most active python open source project. I will present you the Odoo framework and how it can help to be more productive when building web based business apps. I will highlight its advantages compared to more popular framework such as django.
|
10.5446/20148 (DOI)
|
Yes, hello Welcome to my talk. I'm here as a private entity and obvious. I'm not here for a company But at least now my project got a logo So which you can see on the right side so I Have a quick overview of the topics. I've did previous talks at Europe hyphens Two and three years ago and these were longer talks this time. I was going for a faster talk. I Have some actual good news and I have some problems to share with you. So I'm going to talk about who I am My name is Kai I'm a professional from the ATC industry and I do this as my hobby. It's my spare time effort and my spare time effort the title of a talk kind of disclosed it already is a The Python compiler. I got a bit preposterous about this but basically if you see the goals That makes sense Show you what it takes takes basically nothing. We are going to compile a simple program and the full-blown program material but There's not going to be a lot of time to look at this so That's going to be fast. I'm going to present you with my goals and The plan how we get there The last point on the slide join me is the most important because this project is really high potential and It's limited by the amount of contributors so far. I am mostly on my own I have a few people who helped me and sustained parts of the project but this is not enough So it's going slower than it could be. Although if you as you will see it's pretty It's progressing pretty well Then I will have a look at some details of newtka newtka as you know, Python is very dynamic and complex language and I Have taken steps to reduce that problem and there's some common complaints here Everybody knows that pydons highly dynamic. How could a compiler even work? Then we look at optimizations what we have so far that list got longer recently and Actually, I'm demoing now here and I when I wrote the title of the talk This didn't work and it didn't work until maybe last week the inlining of function codes which is a IC this as a breakthrough to the compiler Technically practically it's probably not but technically is very Good achievement and what else is going to come? so Maybe let me start with a name. It's named after my wife Anna and like it was she gested Seconds ago. I could have named it like this. I named it after her in Russia's Russian and in Russian she's called a newtka and short form newtka Which is tricky because it's pronounced differently than it is written. So but that is your name and I started it with after mingling with other projects Pi Pi and Cyton To be a fully compatible compiler that doesn't have to make any compromises and doesn't have to invent a new language and so on I was thinking out of a box so Most people see Python as a very powerful tool for some part of the language landscape But not all and I wanted to take it to Also where performance critical stuff happens. So sorry I Don't see any time pressure. So I do this the right way and the right way means that I can do this so it carries all all the weight of Python all the time It's licensed very liberally and you can use it with everything so it's free software of the most free kind Most major milestones are now achieved. It's basically working if you want all you want is an acclurator It's going to work on all the operating systems Android and iOS need some work, but in theory they should work and I know that some people have done some things but It's still future work. Obviously the mobile space and Python Could see some help and maybe newt car can provide that So what it does it use it uses older Python versions and new ones alike Even the latest 3.5 meter Anticipating a question from you. I added support for that. So it passes the CPI from 3.4 test you'd Running a compiled code And it takes a C++ compiler. I will cover that issue more on later slides And it takes your Python code and that is it. So it's really just a C plus plus compiler and newt car and you can compile So having a new language that is separate from Python means I lose all the things I like I put them all on the slides here and I'm trying to be a bit fast about the presentation, but you know There's lots of things that you are used to and if it's not Python, but for example Something else Then we just lose this so I put a kind of stop sign Below there so very important to me is if we have a fast Python It should be a Python much like pi pi tries to be one or Jiton tries to be a Java dialect I can switch back and forth. So The thing I'm trying to do is if you start using newt car you are not going to have a price attached It doesn't mean that your project is Bound to using it well means if you encounter a bug in newt car and it stops working You can just use something else instead So My ideas here for performance and these are very old ideas. I've not done anything actually in this direction yet And I know that Edo is running around and presenting ideas for type hints and Everybody asked me will I support them and the answer is yes Although technically I would like something That also works during the runtime of Python so in his proposal. It's just something that Python very exquisitely ignores and doesn't use and I don't like that at all I want it to be code that actually improves the quality and makes these actual checks and then the compiler just gets to Benefit from the knowledge extracted from such checks So the first goal and one which I met a couple of years ago was feature parity with see Python It's compatible with all the language constructs and it's also compatible with runtime so Qt Lxml Whatever extension object there are You can use them The compatibility that I have achieved and that I have increased since it's amazingly high Basically, my first attempt at newt car was to make a demonstration That something like a Python compiler actually can fit into things without having a price and This is now what I consider a true statement so from there on well On to the next thing Some of these projects I mentioned need patches so pi qt Pi side and so on sometimes they make too tight check on what is a function so I have a compiled function type and They were not tolerant about this without patches So the next thing is to generate efficient code from that as You will see of a pi stone benchmark. I achieved a number Two and a half fold speed up. So this is something I looked at but it was only a concept It was only to show If we don't have bytecode, but have compilation, what can we gain? It's not really worth it So I think this sort of speed up is unimportant So What we got new is these code generation is now starting To remove code that is not used and it's using traces to determine if objects need releases and as we will see later exceptions are now fast. I have a slide about this So constant propagation, which is basically just people optimization so Identify as money values and push it forward So if you assign a constant into a variable and use that later on you generate a efficient code I have just recently achieved that What I haven't got yet and which will be an important part to getting any actual improvement that is worthwhile for anybody Is to make type inference and treat strings integers lists and so on differently that's only starting to exist Then interfacing with C code the so-called bindings I had a discussion with with a site and guy this morning Newtka will and should be able to understand C types and CSC if I and make direct calls I have a slide about that too and And hence type hints doesn't exist. So not this year type hints and So on so I have here a outside view of Newtka Where you can see that on on the top left you have your code when you put that through a newtka Can be multiple files on newtka recurses according to your Python path and just find their code and produces from that a bunch of C++ files and Put it in a directory and then it runs scons And what typically happens is that people tell me for some reasons that I do not really understand that scons is somehow bad. I Don't think it is it does the job and I have a scons file in newtka Which then can be used to produce a module? So if you were to deliver extension module from your Python code that's feasible even whole packages or you can produce an executable So from a user standpoint Newtka and your code that is basically it scons does handle the C++ details and I get very Nice emails from people who said it even found my my Microsoft compiler and just worked It's very easy to use so I have a very low barrier of entry When we look inside You will find that I have a couple of faces so based on the Abstract syntax tree the same one that Python uses So in a sense, I'm reusing the Python parser, which is one of the benefits of not having a separate language I Enter a step called reformulations so for example in Python 2.6 with statement got added and it's well, I could have a with note and generate code from that and actually First versions of newtka did that So I had a C++ template and I generated code which just happened to do the proper thing the compatible thing But that's not how it's done anymore We now have reformulations and with these reformulations the with statement ends up a simpler Python We are going to see a few examples of that So I'm speaking very fast and I try to be fast the idea is also that you can have your questions asked So if you have a question just raise your hand and ask questions whenever you think you have one, please do so Then we go into optimization which is basically an endless loop because Optimizing a Python program you cannot have a single or two-parse approach because after every optimization Any other optimization may become feasible again. So it's an endless loop But it finishes at some point and then finalization is entered which just annotates the code a bit more This then final tree receives a code generation and then the directory we were seeing So that's very typical. What's probably special is that there is this reformulation step Which tries to make a baby python out of things So Time for demo This is a Python function and It has a nested function and it does local variable assignments And then it makes this call which can be inlined and actually You and I as a human we can We can see what happens and the thing which I'm very proud of is that I now have variable tracing and SSH sufficiently strong To justify that on a global scale Newtka will be able to understand that sort of code and produce a Simpler result So it has a verbose mode and here we see a look into the inside What happens there? When it runs so there are tried blocks Which is sort of true Because of a reformulation for example this statement here does an unpacking and Secretly that involves try finally semantics So if you get interrupted while unpacking you get to release something so but the static analysis Finds out that the tried blocks can be reduced it finds out That the assignment to G can be propagated Entirely and therefore be dropped the one to X one to Y the value is then actually propagated and then In line Nine here. We have a constant tuple Constant result we can replace the call to G with a direct call and we can inline the function We can discover that previously there was a variable G But it now is no longer used and it's so it's not assigned and then it's not initialized anymore so this uninitialized variable G it can be Well the releasing of it It should there have been an exception in a Python function your locals get released. That's What's behind this and then we propagate the inline variables and the variable build a tuple and so on remove all the try handlers and Ultimately, we are done with that so For example This is a even simpler program, but it will help me to make a better demonstration Better in the sense that right now the unpacking I don't I started to have analysis years ago for that, but it's not yet sufficient so I Cannot have tuple unpacking and show a full reduction So when we run this as you can see outputs and a lot of findings and now for easier debugging I Have invented a XML representation of the notary and I use this to test that something is Entirely optimized and as you can see here we have a statement return that is just a constant to two so This function f all it does is a lot of churn around the notion of producing one constant Obviously your code is not going to be like this but It could be if for example the X were an input value of some sort and then If this was already a partially optimized function for some reason then these things make sense so We got this any questions about this The question was is it storing the reduced Python code anywhere actually That's a cool idea for a project that I have is to generate Python code from the reduced One right now. I only generate C. I would love for somebody to Take the internal representation of the optimization and generate Python code from that Python code That is just faster than the other Python code but since We are making a Python compiler for a reason. I'm going to see directly and in C I'm outputting this but basically technically The internal final representation is not entirely Python anymore So as we will see in the reformulation parts For example wire loops and for loops they don't exist So it is a reduced set, but it would be for example feasible to create Python code out of that so Maybe quickly here I don't do this Because I made XML and because I'm easily confused I removed the code that is not used and I had opened it already But here we go again So This is what The generated code for example looks like so we have a local Variable return value initialize it to nothing then we initialize it to the result which is a premade constant And we go to the function return exit which checks that it's actually return value and then returns it So this is how it's become in Python world the most efficient Code that you can have obviously There's more to it. So we can also do mercurial This is something that worked two years ago. It's passing the test suite with mercurial So we could compile this now and actually I was doing it It took 13 35 minutes to compile all of mercurial which is A huge body of of Of code Right now newtka is not making enough Optimizations and discovering enough dead code, but half an hour is pretty okay on this laptop without power so Generated code works like this. I will be quick so I will be quick I will be quick. So now it's c code When I initially started out, I was aware that this is a very ambitious project So the only reason I did even started was because c++ 11 The new c++ language was having so much cool new features that convinced me that code generation would be relatively simple now So the gap between c++ 11 and Python was relatively small it turns out That for example c++ exceptions suck and in place operations of Python are Optimizable, but that doesn't fit into one object only one thing so I went to c++ 0 3 and then to See ish c++ which is basically just see with some c++ elements, but no class no types And it's going to be c99 soon enough So I'm going to skip something so as an evolution three years ago. I'm talking now about the C++ 11 one the blue part that was the code generation So I had achieved something phenomenal and that was a compiler which Was capable of integrating with all that Python landscape and make things faster Which was tremendous, but the other part they are so small you can barely see them parser Optimization there was basically only loophole people optimization three years ago two years ago I went to c++ 0 3 and the code generation got a lot more dumped and Reformulation Started to appear and optimization become bigger and right now code generation has become really stupid and Optimization is carrying the day so now These reformulations. I'm making some overhead there using temporary variables and so on I can now optimize them away So availability I have a high focus on correctness So it's available in a stable and a developed form the developed form is also Better than other stable projects I content and I have a factory Where I publish things that are not finished yet. For example, the inlining code is right now on a factory branch There's not just good. There's RPMs and so on lots of people are already using newtka This is my most important slide. So I want you to join the project help me I will guide you and One thing I have to cover for correctness. You see the Oracle of Delphi Delphi means I can use C Python and compare with Python So testing for correctness. It's it's a dream. It's very easy for performance It's much harder It's a race and I have ideas and what I would like you to do is to help me Come up and develop with a tool that will help us Give a user feedback for performance because if you now compile your code, it may not be faster at all It may even be slower and we wouldn't know why There's no feedback. There's no idea which function is slower or faster and how much there needs to be a tool I need somebody with an interest to help me out with this and rescue us So This is the most important things I meant to say I would leave the rest of the time. I hope it's still 10 minutes For questions if you have them I think Yeah, okay in my opinion which Python language constructs were we're making code generation for hardest I Think technically Once we are able to inline meta classes and very effects, I think they will not be an issue and they are very very easy technically classes and instances Need a lot of babysitting Especially under Python to to be correct. So that's that was an issue and I had huge amount of difficulties with in place operation and and Exceptions exceptions are totally a nightmare. So I waste and yeah reference counting is no fun Which is why I develop a compiler too. So you do not have to write c code ever again Yeah, so next question For what The default type to me right now is object. So the question was how to handle if something doesn't have a type Right now in Utkir is basically using node type information at all yet What it will do and let me Show you this In the future it will be able to understand for example C types and they make direct calls But right now everything is an object be it a list string integer. I'm barely own not I'm not using the knowledge yet I will start to make now that I have this tracing capability. I can produce proper traces of Python I will be able to trace the lists and make optimizations dedicated to types, but I don't have it yet I'm integrating with lip Python and its pi object asterisks It's it's a standard Python object. It's it's like you wrote a C extension code Another question yes I tried various compilers Obviously on the C level there's still always something to gain and I'm trying to be clever and smart about code I generate and I find the Microsoft compiler to be terrible and intro will probably be better But technically new to guy should be in a position to understand Python So if the example that I showed you it gains by an order of magnitude performance by just inlining a function call I Don't have a slide time to show the slide now But if I'm if I just avoid a function call of Python I can have speedups in the domain of 20 fold and And and so on and maybe on top of that with an C compiler. You can edit a few percent again Yeah, that's also there was a Python talk 2014 and the presenter was also said it just works. I throw things at it What you just tried is also something called it's a standalone mode I'm not mentioning that because I'm not interested in it, but it means It would also pack all the things together and allow your distribution to other machines Something people also expect from a compiler is to be able to take the code to another machine I'm my interest is mostly in accoloration and I'm solving this as a sidekick Yeah, so but the feedback is it just works and So when I'm it's at the point where I'm surprised if something doesn't work for standalone I'm not surprised if something doesn't work because it's very hairy with extension what you're doing Yes Yes, the bytecode is incredibly smaller definitely It's there's no need to talk about it a binary which contains the Python Implementation and uses it it's larger, but I don't think it's it's anywhere near important issue and Obviously you will be much faster with a smaller you are but Yeah, it's it's it's larger, but it's still small binaries. So if we can have a look at Hgx It's sort 31 megabytes, I think Yes to confuse users That's a very important part of my project So suppose the program you compile is named Hg What is going to be the resulting name without overwriting it? I want to put it alongside and I've made good experience with dot X it's relatively rare But the Python program exists as dot X So but I'm getting questions and please take note this one will run on linux despite its name And you can rename it it's will still work. So if you despise the name Yes Or I think I'm using stack memory allocation. Yes And I would definitely if what I will need is some sort of list implementation That is not malloc driven if I know the size and if I know enough Sufficient things I will use stack memory. Yes try to be faster Yeah, the question is if I take advantage of rev counting I I I I try to do this So I'm not always taking a rev count when part C Python does but it's a very marginal game The real direction must be to avoid Python objects wherever we can and We we will see how far this gets but of course where I can I will have his analysis and know that I don't have to Take another rev count because I will be holding one already No, there is no he's asking about standalone and the standalone distributions if you want to copy it to another machine basically due to the Incompleteness of code removal includes the standard library and all the libraries that standard library uses and so you end up with a largest set So the distribution I don't think it's huge, but it's it's like a Python installation. I suppose Yes, I Don't have real world benchmarks because that defies the purpose of benchmarks. I know that Piper is really cool now with presenting Real world programs and how Piper stacks up to this I I Have this idea about valiant all my benchmarking I'm doing with valiant and valiant gives me ticks so I don't have to run many times and I can make Analysis directly and I would want your help or I will do it myself But I want your help to create a tool which will run the program and Python and run the program in newtka and compare the tool and make a highlighting of what parts are fast and slow and so I can Get you as a user can get an image. How much speed up do I get in my program and And it should be simple enough to just run your program under under another tool and get a report so It's all laudable to have these benchmarks with synthetic Code but I would actually like to empower us Me and the developers of newtka and the users I don't know if we can make this tool and the sad truth is right now There's nothing basically. I just have random number numbers of something and I'm working very hard on getting to somewhere but like I said I have a time and I don't have a panic to be fast in everything tomorrow and Right now I am only starting to wonder about actual performance. So this is now Where I want to know how good I am Okay, thank you so much. Thank you for being here and for the good questions
|
Kay Hayen - The Python Compiler The Python compiler Nuitka has evolved from an absurdly compatible Python to C++ translator into a **statically optimizing Python compiler**. The mere peephole optimization is now accompanied by full function/module level optimization, with more to come, and only increased compatibility. Witness local and module **variable value propagation**, **function in-lining** with suitable code, and graceful degradation with code that uses the full Python power. (This is considered kind of the break through for Nuitka, to be finished for EP.) No compromises need to be made, full language support, all modules work, including extension modules, e.g. PyQt just works. Also new is a plugin framework that allows the user to provide workarounds for the standalone mode (create self contained distributions), do his own type hinting to Nuitka based on e.g. coding conventions, provide his own optimization based on specific knowledge. Ultimately, Nuitka is intended to grow the Python base into fields, where performance is an issue, it will need your help. Scientific Python could largely benefit from future Nuitka. Join us now.
|
10.5446/20141 (DOI)
|
Okay, good afternoon. Thank you for coming to my talk. I'm going to tell you something about how you can try to get the plot lip a little bit faster. Maybe you remember my last year's talks about the GR framework, where I promised that I had the intention to write a backend for my plot lip for our software. And I made a promise to use our software in iPython notebooks. And these are the things which I want to introduce to you this afternoon. So I don't have to talk about visualization needs. I think most of us have the same needs when it comes to the point that you want to make plots or diagrams. But for our scientific world, we have some more requirements. For example, we need real-time graphics. And we want to visualize very large data sets. And that might be challenging. So we began to think about another solution. And there are a lot of visualization solutions in the Python world. The most popular ones are mentioned here on this slide. And I think the basic workhorse, which is also part of the scientific stack of Python, is met plot lip. And it has some solutions where you can use met plot lip output in the browser. Maybe you have heard this morning about the BookA software or plotly, which are browser solutions which allow to redirect the met plot lip output to a browser window. There are even other tools which make 3D graphics. They are very powerful, like MyAvi, VDK, or VisP. And they are very fast. But the problem is they are low-level. And there are other problems. And for example, in principle, you have three different things which have to be mentioned. We need interoperability, speeds, and quality. And you can't have all of these three things. So we tried to get another solution. We wanted to combine the 2D and the 3D, the hardware-accelerated 3D world. And we wanted to create a back-end which can not only produce figures, but also can stream data. So this was our primary attention. And at that point, I started to write a back-end format plot lip. Because I didn't know another method to get all these graphics stuff faster. Scython didn't help. It doesn't make sense to improve or speed up the numerical code segments. And hardware acceleration is a very nice feature. But in most cases, it cannot be applied to visualization software. Because you have your existing codes, and you want probably to mix it with other code snippets. So there was the idea, whether it would be possible to write a back-end format plot lip, which would be faster. And as our software is completely written in C, it is capable of presenting continuous data streams. And with a module called GR3, which has been written by a colleague of mine, which is also here in the audience, Florian Reem, we can also mix 2D and 3D graphics into one canvas or one window. There's one important point that the software right now is also interoperability, that there's interoperability with the graphical user interfaces or web frameworks like iPython or in the meantime called Jupyter. So what can we do with our framework? You can combine the power of Matplotlib and GR, so you can mix the output. For example, output where you create real-time plots of signal analysis or whatever you can think of. And you can, for example, produce video content on the fly. You don't have to put your frames into a PNG and then put them together and render them as an MPEG file. You can do it on the fly. And one important point is that you can really mix 2D and 3D graphics elements. So how does this all work? I don't want to go into the detail because I want to show you some demos later. But due to the layer structure of GR, which contains logical device drivers for nearly every technique we have today, we are not there. There's no dependency on third-party software, so we can really mix components together. And we have a very good 3D software which is capable of producing HTML5 outputs and all these things. So this is the output of the Matplotlib GR backend. You can see that there's no difference to the original output. The only advantage is that it's a little bit faster. And here you can see the GR framework in action. You can really produce a very fast output. For example, you can take an audio signal, calculate Fourier transform of this audio signal and display it in real time. This is done in the middle section here. In the right section you could see that you can produce a molecule, that you can visualize a molecule sequence while on the right side creating a histogram with Matplotlib. So they can be combined in one plot and you don't have to change any line of code in your Matplotlib examples. So another feature I mentioned, right now you can use the GR framework in notebooks, for example in Jupyter, which is a follower of Ipython. It can be used both with Python and Julia. I was very excited about the performance, but I had expectations and when I saw the first results I was a little bit disappointed. You can see the left two bars that when using the GR backend the performance improvement is only factor two or three. I had expected much more because in the right bar you can see that with a GR framework as a standalone software in Jupyter you get much higher results. So we have to explain what's the reason for that. You can see in those codes, in those log files here which I have produced with the Python profiler that Matplotlib is wasting too much time in Python itself. It doesn't send enough output to the graphics backend but it's organizing plot data and data files. So at this point I was a little bit disappointed. But then we had another new feature, interoperability. And on this side maybe it's a little bit too small, you can see how you can mix different code segments. I did not change any line in the Matplotlib code in this example and I did not change any line in the GR3 code and I simply put them one after the other and this can then be displayed as a sequence in one canvas or in one web canvas. There's another advantage, if you have such a sequence you can create an MPEG file for example on the fly without adding any animation code. If you know Matplotlib you always have to define some animation functions and then create a loop and put your scenes together. That's not required with this GR framework so you have a big advantage here. Also it's possible to use inline graphics both with Matplotlib and GR. It's not a problem to produce inline figures with Matplotlib but if it comes to the point that you want to generate streams it's much easier with a GR framework and I will show you later how this works. Although I have to say that with the GR software itself it's again ten times faster than Matplotlib. At this point I would like to show you some demos. Let's start with the animation example. Can you see? Okay. Should be big enough. Okay, first of all we define some numpy array. We tell Matplotlib to use inline graphics and we create a figure which right now is empty and then we have to do an animation loop. We have to define animation function. Then we start the loop and in the final step we have to save this animation and this takes a lot of time and once this is done Matplotlib will give us an mpack file which can be displayed here in the browser by this HTML macro. Hopefully it will. Okay, here it is. But we have seen it's a little bit complicated because we have to write this callback function. We have to write, we trigger this animation loop and all these things are not very comfortable because we have to write extra code. So let's try this with the Matplotlib backend. Unfortunately I have to restart the kernel at this time because Matplotlib only checks once whether there is an external backend available. So I have to restart the kernel. I also have to redefine my numpy arrays and now I can tell Matplotlib to use a gfrain work and generate a movie on the fly. Again I import Matplotlib. I create the plot and I can immediately show it. He has to render it like Matplotlib has to do it too. But we have seen we didn't, there was no need to write a callback function, an animation function. You could make the animation on the fly without any change of the Matplotlib code. So finally let me show you how this works with the gfrain work. You have a very simple loop. I also already told the gfrain work to generate inline graphics and as you can see here there are only three lines of code and you have the same animation here. The speed is the same because the frame rate is set to the exact same value. So let's take another example. I want to show you that there are advantages concerning interoperability. In this example I use a Matplotlib package to draw a histogram of some angles. I use a GR3 module called Mocly for molecular dynamics visualizations to read and visualize a sequence of molecules. Then I add a 2D plot as shown in one of the earlier slides using GR. And so let's look whether this works. Again I use a Matplotlib package. I use a 3D package Mocly to read the data file which contains the coordinates of the atoms. Then I use a GR framework until the GR framework to generate the movie on the fly and to generate the output. That is important because Matplotlib should not create the output too early. So now we start the loop. This should take some time because we have to render about 100 scenes. Quite slow this afternoon. Now we are done. And finally we show what has been rendered. And here you can see we have both GR output. We have Matplotlib output and we have output from GR framework. I do it again so you can see what happens. And we have this line output on the bottom here. And all this is done in real time. It has been rendered in real time when you can produce it in your browser. So there's one nice feature which I also want to show. You can export this scene from our GR3 software and then even rotate it in the browser. Create something which Florian has written and which we will probably make available for tool de-graphics in the next release. So let's open the next one. I talked about inline graphics so I have to speed up a little bit. Again we read some data here and you see that inline graphics with Matplotlib is a little bit slow and it's flickering and there's a function I never knew about. It's called clear output which is part of IPython and you can use this function to re-create or re-draw your plot. This is very useful for our scientists to generate sequences of plots in the browser. So I make the same now with GR and you can see it's much faster and it's less code. So we have a speedup of 10 here and you can even do this in JavaScript. So with the next version we can create this output as a JavaScript, HTML file which can then be displayed in real time which will be much faster than our own version. Right now we have finished writing a JavaScript and with the next version we will make it available in the open source. So let's take the last example. This is an example taken from a sci-fi talk from Matplotlib guy which has produced this graph here and I just want to show that with our software. You can visualize this in real time in 3D. It's flickering at this time but I have to mention that all this data has to be in-transferred from the kernel to the browser and back again so that's much traffic here. Maybe this demo makes no sense. I only want to show you that there's a lot of room for performance. Okay, that's for the demos. So let's finalize. So I already mentioned that we are planning to make a JavaScript, that we have written a JavaScript logical device driver for our software which can then be used for example to embed JavaScript code in your UP browser or you could write your own JavaScript code and take the and visualize a display list which has been generated by our software and then fill it with your own JavaScript code. Here on the right side you can see that they are nearly the same commands or that are the same commands which are used in C or in Python or in Julia. Okay, what can GRLs be used for? Here are some examples. PyMolden is a new development of a colleague of mine and which will be made available for the open source very soon. And then we have Nicos, an instrument control system which has been written by two colleagues of mine which will demonstrate it in the poster session tomorrow. I don't know. And so what are the conclusions? Well, you can use the Map.Lip.Lar backend as a GR logical device driver. But the speedups are not as expected. It's mostly factor times two or three faster. But I think the second feature that you can mix 2D and 3D graphics and that you can create movies on the fly is still something which could be interesting for most of you especially for scientists. And you can produce plots and figures much faster with a GR framework. And that's the reason why we are really planning to write a complete Map.Lip implementation, PyLap implementation in C because I think that's the future. And well, these are our plans for the future and I hope I can fulfill my promises next year and show you what we have done. So thank you for your attention. We have five minutes for questions if you have any. I have one. No questions. So. Quick question. This seems quite popular in the scientific community as you showed the PyMol version. Where else do you see this for non-scientific community applications? There are not so much users. For example, with Map.Lip you can't compare those software packages. The package itself is very old because it's C code. It has already written years ago. But now we have written both for backends, both for Julia and for Python, Map.Lip and I think that it will be more popular in the near future because especially with Julia you can get even more performance than I showed today. As Julia can call Python modules, that might be very challenging. I hope that there will be more users in the future. Any other questions from the crowd? Okay. Well, let's thank Joseph again for his presentation. Okay. Okay. Thank you.
|
Josef Heinen - Getting more out of Matplotlib with GR Python is well established in software development departments of research and industry, not least because of the proliferation of libraries such as _SciPy_ and _Matplotlib_. However, when processing large amounts of data, in particular in combination with GUI toolkits (_Qt_) or three-dimensional visualizations (_OpenGL_), Python as an interpretative programming language seems to be reaching its limits. In particular, large amounts of data or the visualization of three- dimensional scenes may overwhelm the system. This presentation shows how visualization applications with special performance requirements can be designed on the basis of _Matplotlib_ and _GR_, a high-performance visualization library for Linux, OS X and Windows. The lecture focuses on the development of a new graphics backend for _Matplotlib_ based on the _GR_ framework. By combining the power of those libraries the responsiveness of animated visualization applications and their resulting frame rates can be improved significantly. This in turn allows the use of _Matplotlib_ in real- time environments, for example in the area of signal processing. Using concrete examples, the presentation will demonstrate the benefits of the [GR framework] as a companion module for _Matplotlib_, both in _Python_ and _Julia_. Based on selected applications, the suitability of the _GR framework_ will be highlighted especially in environments where time is critical. The system’s performance capabilities will be illustrated using demanding live applications. In addition, the special abilities of the _GR framework_ are emphasized in terms of interoperability with graphical user interfaces (_Qt/PySide_) and _OpenGL_, which opens up new possibilities for existing _Matplotlib_ applications.
|
10.5446/20139 (DOI)
|
So first to present myself, I'm Jun Santu. I'm a software engineer at Salando as part of the continuous delivery and deployment team. And today I'm going to present the part we took at Salando from the dark cases of subversion to the enlightened age of autonomous work with EGIT. So some quick facts about Salando. So you understand the context of my work. We are, we have our headquarters in Berlin. We have also other offices in other places like Dublin, Helsinki, Dortmund, airport. We have more than 700 actually nowadays almost 1000 engineers. And we do lots of every almost everything we do is done in house. So we do lots of development. Okay. So the first question when talking about GITUX is why do we need GITUX in or subversion groups or other kinds of commit so in the first place? The answer to this is to enforce rules. This raises the second question. Why do we need rules? Any large group organization needs rules. Some rules are internal to the groups, some are external. And they are very important to achieve an efficient workflow to ensure quality and transparency to stakeholders. And in case of Salando, stakeholders go from the colleague at next desk to the customer that as trusted as his credit card to the investor who spent lots of his money and he trusted and believed in us. Second question is how we enforce rules. Not our rules are enforced in GITUX. Some rules are enforced before deployments. Some rules are only monitored are not enforced. And so you have to look at your workflow and see what rules can be enforced a priori or with checks and what rules can be checked after you have moved on your workflow. And it's vital to keep a proper balance between two types of enforcement because if you have too many blockers on your workflow, you prevent your developers from being productive. But if you have to, few checks and blockers, you risk that bad software goes live and you break the trust all stakeholders that are in you. So now starting point. So this is a version. This was the place where we were when I arrived at Salando two years ago. We had one batch script with some basic checks. A precommit took. Then we were growing as a company and we needed more rules because it was not possible for someone to keep track of everything that was happening. It was not possible to keep a batch script anymore. It was not efficient. It was not maintainable. So we ported the GITUX subversion from Pesh to Python. Around December 2013, we started moving from GIT because the subversion was not good enough for guys anymore and we had to do something. So we started moving to GIT. We still had same rules in places before and we knew we would still keep subversion for some time because we already had lots of teams and some teams wanted to move faster and others slower. Not everyone wanted to move to GIT so we had to convince everyone and we decided to do migration in waves. So the first decision we took was to adapt the SVN commit took to work on both SVN and GIT. So we implemented GITUX as a per-gC book because it's invoked at the same point on the workflow as commit took on subversion. This is when user sense changes to the remote server and because it allowed us to reject the reference or branch without rejecting full push. The other alternative would be an update took but in that one if you return an error code it will block the full push even if most of it is following the rules. Doing this had some problems. This approach had some problems. The first problem in this plan was that GIT is not subversion. They are different in fundamental ways and trying to support both would mean support for both of them. The other problem is that being distributed GIT is more flexible than subversion which means that people use GIT in the same way they use subversion and don't use GIT in the same way that other people use GIT. The first things that I learned to migrate to GIT start experimenting different workflows and this was a good thing because they were very different projects and different projects have very different requirements. So it makes sense for them to have a different set of best practices and slightly different rules and different pricing strategy but this also forces us to rethink how we check the rules and how we implement the rules and we had to rethink the GIT rules. So we moved to plan B where before we had one common look for GIT and subversion. Now we decided to have two different looks. In practice this meant we forked to look into...the subversion commit look was aggregated so no more changes there because we don't want to really support the subversion anymore and we continued to work on to get priority look because different teams wanted to use different workflows. So we moved away from one size fits all strategy and allowed teams to specify set rules for themselves. So this is the way we allowed teams to set rules for themselves. We have a configuration format based on YAML where Stash projects and repositories could mesh several times by name. This is similar to CSS so for example you can see there that one team could have one general rule for themselves for their project but for a specific repository they could change part of the rules and not the whole set and a second team could have a completely different rule set because they have different needs. The problem with this second approach and this was a reflection we did after using the GIT looks for one year was that we had a centralized configuration which became a bottleneck because it could only be changed by a small number of people and sometimes we were all on vacation and it was really bad. Another more technical problem we had is that we were trying to check all the commits on a push because some rules apply to every single commit. For example we checked that one of the rules we had at Zalendo was that every commit message has to have a ticket and for this we have to go through all the commits and this was problematic for a number of reasons. For example GIT's history is not linear so if the user merges or rebases from a branch and we try to check all the messages in the commit range for ticket ID and see if it matches to a branch, a commit set came from other branches by the rebase or the merge would fail the check. In the end we tried to several solutions, some worked way better than others, in the end the solution for this was filtering out commits that were already in a branch because this meant they were already checked before and were not originated on this branch. The second problem we had is that when you merge to or from a long-lived branch, some teams had some very long-lived projects and they had branches for it and sometimes they deferred from master for two, three months and when they merged from master or to master or when they rebased they had to check. Because of some of our checks, sometimes we had pushes that took more than half an hour to check and this was not a good experience for our developers. Related to this, one of the slowest checks we had is checking code style, especially for Java code because we use Jallopee to validate code style and it runs on JVM and the architecture of our code validator meant we had to spawn one JVM for every changed file. So if someone pushed 200 changed files we would spawn up 200 JVMs and other people were pushing at the same time and it was constantly freezing and slowing down our server. We were able to mitigate this problem by using NailGant that is a kind of server that keeps JVM alive and you can run JVM programs there and we could run Jallopee there but we had to do some work around to run Jallopee in parallel and we still had some performance issues for some reason there were more pushes than usual. The last problem of this is that the system was inflexible. Sometimes people have very good reasons to ignore rules. Sometimes they pushed something that they didn't want to push and they want to do a forced push and we didn't allow that. Sometimes they, sometimes some push was accepted by mistake and it was a bag and it has to be removed because it should not be there. Sometimes there's a bag and we reject it. Something that we should have accepted and because this was on our server and people didn't have access to the server, they had no way to go around this. The other way we were inflexible is that because it's a remote Github and it's living on the server, we only supported our internal Github server. For example, you could not use our Github for our projects on github.com and so our open source projects were completely unsupported. In the end, I could summarize all these problems as we tried to centralize Github. While Github is distributed, we still saw our source control management system as centralized because we came from subversion and we were still thinking like subversion. But it was clear that we would have to rethink our approach. This process of thinking about the Github's coincided with the use change that's the ZLANDU as part of something we call ready calligility. ZLANDU was getting rid of all the error keys and giving teams more autonomy and one of the new motors was autonomy instead of control. And this presented a clear path for us and in the end was the solution that we are still implementing. So we decided to move away from remote Github's to a set of local Github's. One advantage of this is that Github's see one commit at a time. So we avoided the issues where we had a set, a list of commits to check and we didn't know what came from a merge or a rebase and what was created in the branch. And it's better adjusted to the distributed nature of Github because they are distributed by themselves and people run it on their local machines. We decided that they should be optional because we don't trust people. This also means that Github's are no longer responsible for enforcing the rules. Instead, we are creating mechanisms to ensure that a level of development follows the rules before the code goes live but we don't block anyone from pushing anything right now. We did this because we believe Github's should be seen as a tool to help autonomous developers and not as a barrier that makes their work harder. We decided to make them extensible because as I said before, one size doesn't fit all different teams have different needs and sometimes we only have a small group of developers working on this more or less full time but sometimes people have other ideas and they can implement new stuff for themselves. Sometimes they don't want one of our checks and they don't want to install the dependencies of our checks so by being extensible, they can opt out of those dependencies. And we decided to make it open source because by only being and thinking open source, it could fully support autonomous teams and this also allowed us to support our teams that want to open source their code and this also includes my team. So our new set of hooks that we call turnstile are available on Github and by PI so everyone is free to use them. So if you want to install it on your machine, you can just use pip install turnstile core. This will just install core without any of our extensions. We also have some extensions on Zalendo's Github account. So now you want to use it on your repository. What do you have to use? Well, the first step is add a turnstile.eml file in the root of your repository. If you use something like Travis or something similar, the same process you just added a file to your repository and it looks something like this. You have a list of checks you want to use. So you just name them and you can give options to all checks that they support by themselves. And every check, even if you are implementing your own checks as an extension, you can read this file and you have access to it. So and because the way local Github works, you need to see link to on your.kit folder inside the repository and to make this easy, turnstile supports a subcomment that's turnstile installed that will automatically see link to looks for you. And because I was already adding support for subcomments, I added several more. We have to config subcomments that right now the only thing it does is set, allow you to set verbosity. So if you want only to see errors, you can only see the errors. If you want to see everything that turnstile is doing behind since you can also do it, you have to install command to add to looks to a specific repository. We have to remove if you decide to, you don't want to looks anymore for the repository. The specification subcomment is used for check if all your projects contain a valid specification of all your commits contain a valid specification. And right now this means if they start with a new URL, but I'm planning to support more stuff in the near future. The upgrades of command will check by PI to see if turnstile and all your extensions are updated and if they are not, it will offer to update everything for you. And version just brings the version and it's not that interesting. So now if you want to create extensions to turnstile, how can you do it? The way you do it is with set of tools and entry points. So you just have, if you want to add a command, you just go to the end, use the entry point turnstile.comments. You give the name of your subcomment and this will be used to call it from the command line and you just provide the module where your subcomment is and you can see more on documentation later. It's saving things for checks. So if you want to check something with the precommit hook, you also give it a name that will be used on the configuration and you provide the module that is going to be used. So you have three entry points right now, one for commands, one for commit message hook, the checks commit message and the other for precommit hook that we use for checking code style, for example. So what did we learn from this process? First, don't, and this is the most important, don't get stuck in the past. It's very easy to make mistakes because it was always done like this. What was the best solution yesterday can be the wrong solution today and this will be mostly, will mostly likely be true if you change technologies because if you change technologies, you will have different limitations and different opportunities and you should rethink what you're doing and adapt to it. One other thing we learned was developing the open because by being more transparent about what we wanted, where we were going, we avoided assuming too much and we avoided backing ourselves into a corner and we got early feedback that also avoided lots of issues. We also learned that we should build tools, not barriers. Software is meant to help people and to be more productive. If your software is making people lose time instead of saving time, you should rethink it. By rethinking our GIT hooks, we were able to move from a position where sometimes our software was entering the productivity of our engineering teams to a position where they are useful tools to help them work autonomously. So, any questions? Thank you for the talk. My question is about the cost and the effort of any change. We're talking about your company changing. Can you speak louder? I cannot hear you. Sorry. No? Yes. Yes, thank you. I was asking about the effort of changing in each step because sometimes, okay, leaving subversion may be a good thing, but changing for centralized hooks to local hooks and everything, out the effort. Yes. Yes, it's an art process and we are still working on it. Sometimes, and actually, we are doing things where we are disabling checks on the remote hook in phases and telling people, now start using this and we are trying to make it compatible. So, I think that's not working in the same way on the remote GIT hooks and local GIT hooks specification because before, we only allowed people to use gyro tickets and now we are autonomous. We want to allow them to use whatever solution they want. So, now we support more than that. Fortunately, at Zalendo, we are already used to always question everything and it's one of our philosophy. If you are doing the same thing all the time, you should question yourself and see if there's no better way to do it. It's part of our corporate culture and that doesn't mean that people don't complain that we are changing again and that we have to manage it. But at least in relationship to this, most feedback I have is that people are happy to be able to control when they use the hooks, how they use it and they are very happy that they don't have to rely on me and my team to change the step for them. You can provide any samples of the kind of rules you are enforcing with that? Yes. So, for example, this is actually from one of our open source repositories. I don't remember exactly what. This is a very minimum set. We have more than that. But for example, we have the specification check. What it does right now is checking if your commit message starts with a URL and you have the option allowed schemes right now and I'm saying that I only accept HTTPS URLs because on this case, all specifications are done using GitHub issues. In the near future, I want also to support GitHub references so you can just put the text that I support by GitHub and also the Jira ticket IDs because we still use it at Solando and we wanted to make it easy for people to use it. We have the branch release check that in this case uses that regex expression on the bottom to check if a branch that starts with release slash something, the second part has too much that one. In this case, it has to start with a V and follow some rules that's format to use at Solando but you can use whatever you want and to protect master branch in this case forbids you to commit directly to master because I want to enforce pull requests. Any other questions? Okay, just one more thing. If you want to find more information about Solando, we have a tech blog and tech.solando.com. We have our GitHub page, GitHub.com slash Solando. We have lots of open source projects there and we are going to have much more in the future. We also have Twitter, Instagram account and a jobs page and one of my colleagues will give a presentation in the recruiting session later today. Thank you.
|
João Santos - Using Git Hooks to Help Your Engineering Teams Work Autonomously In this talk, Software Engineer Joao Santos will describe how the engineering team at Zalando has been migrating to local Git hooks to ensure that engineers can work autonomously and flexibly. Zalando--- Europe’s leading online fashion platform for men, women and children-- began shifting from SVN to Git in late 2013. Santos and his colleagues used Python to create a Git update hook that enabled the team to reject changes to a branch while still allowing changes to other branches. He’ll explain why his team chose Python for this job instead of a bash script, point out mistakes made during the process (and solutions his team used to fix them), and the benefits generated by this migration. He’ll also talk about turnstile: a set of open-source, configurable, optional local Git hooks, created by the Zalando team, that enables engineers to abide by internal rules for committing code while following their own coding style and workflow preferences.
|
10.5446/20136 (DOI)
|
Hi, everyone. Thanks for having me. This is last track for Herobiton. My name is Jean-Philippe Casey. Can reach me on Twitter at JP Casey. I work at Shopify in Montreal. I'm French-Canadian. I'm also an organizer of the Montreal Python user group. We do monthly meetups with little conferences and Project Nice also. If you're ever in Montreal, talk to our website, see if you have anything on. I would be glad to see you there. Today, I'm here to talk to you about types and study type checking in Python. Before we go in depth, I want to bring some theory in it. The first thing I'm going to want to talk to you about is type systems. It seems boring, but it's fun. You'll see. First thing we're going to check is what is a type system, of course. The Wikipedia definition is quite hands-on on it. It says in programming languages, a type system is a collection of rules, a set of rules that assign a property called a type to various constructs composing a computer program such as variables, expressions, functions, or modules. So it's basically just a set of rules. But also, it has lots of purposes. Of course, the first one is we have type systems in place to help us reduce and identify potential bugs we have in our programs. It also is going to give us meanings to a sequence of bits. Because if I just give you this eight bits, we have no idea what it is. So it could be either 72 if it's an integer, it could be the letter H if it's an ASCII character set. It could also mean that it's a true Boolean value for C for instance, because it doesn't equal to zero. So let's dig quickly to some fundamentals. So with the type systems, it's going to be studied by type theories. It's lots of math, lots of computer science also. And also, so a programming language is going to need a type checking in place. And typing was just going to be meaning that it's assigning a type to a value. Type checks could be done either at runtime, it could be done at compile time. It could also be manually noted in the source code. So you're going to declare the type before a variable. Or the language, the type system could also automatically infer. So without having to declare what the type of a variable it is, the type system can deduce what it is. And like I just said, typing will give meaning to a sequence of bits. So with type systems, we're going to have of course type checking. Type checking or type safety, it's the process of verifying and enforcing constraint of the type system. So it's just checking and making sure that the parts have been connected in a meaningful and constant way. So for instance, we cannot add a string to a list. It's the type, the type system doesn't work like that. With type checking, so it's going to prevent illegal operations. Like I said, adding a list with an integer, for instance. It provides also a memory safety measures. So a good type checking for a good type system is going to reduce the buffer overflows or the out of bind writes that you can do, which would lead to corrupting the running executable or the memory in place. It also helps for logic errors. So to disallow you working with different semantics. So with type checking, we're going to have type safety. Like I said, type safety is basically just enforcing the types in a programming language. It's a requirement for any programming language. And it's also closely linked to the memory safety of your executable. People are often going to compare it to strong typing versus weak typing. So it's only going to be whether it's what is the type safety, if it's memory safe, is it static type checking or is it going to be dynamic type checking? However, the problem is that for many languages, many languages are too big for having human-generated type safety proofs. So they would require thousands of cases. However, there are some languages that have rigorous defined sceptics. For instance, some ML-based languages. And they have been proven to meet certain definition of type safety. Haskell is another language that if you do not use unsafe methods that are mostly IO operation, you can provide a good level of type safety. I also just want to talk quickly about Coq, which is a programming language. Well, mostly it's an iterative term improver which was written in OCamon. It's over 26 years old. It's a dependency type functional language. And so there's a browser that was written in this programming language called Quark. And the web browser has a kernel that has been formally verified by Coq. So it means that it should be almost bug proof or secure as in no buffer refills or no bugs related to the type system. So we have the type safety type checks. But to have type checks, there's two ways of doing it, of course. We're going to have the static type checking, which is going to be done, of course, at compile time. In a static type checking, every variable is going to be bound to a type during the compiler phase. It provides us with some good things for us. So it operates on the program source code. It doesn't have to run the executable. And since it runs on the source code, it helps you to catch bugs earlier in your development cycle. And it's also going to give you a higher level of confidence, in my opinion. And it could be all depending on which type system you use. It could be also a limited form of formal verification as to whether your program does what it's supposed to do. Some quick static type check language we have is of course CC++, Java, Go. And like I said, Askel and OCaml or some of them. There's a lot other languages that are static type checks. The other type checking we have is dynamic. So instead of doing the type checking at compile time, we're going to have it at runtime. So while the program is executed, while it's running. So it's literally the process of everything type safety during runtime. Compared to static type checking, here every variable is going to be bound to an object and not necessarily to a type. Because the type has to be inferred during the execution. This gives you usually it will allow compilers to run more quickly because it removes a phase from the compiler, which is type checking. It allows also interpreter to interpret dynamically new code. So for instance, Python we have eval. It also allows duck typing and easier metaprogramming. The last two are only for dynamic type checking language. Because for instance, with C++ we have templates. With C++ templates, the type checking is done at runtime. Another example of dynamically type check language, of course, we have Python, we have Closure and Lisp, which is a compile language. It's not only a dynamic language, it means a dynamic type check language. There's also a combination of both. Like I said, for C++ templates, there's for Java and C, when you downscale a variable, the type checking is going to be done at runtime. And for C, you can also just downcast everything to void and the compiler will never complain. So that was type systems. But since Python is a dynamically type language, what can static type help us? As I said, it does help us reduce the number of bugs. And by reducing the number of bugs, it will help us to identify more quickly bugs during our development process. With a static type language, people will often say that since my language is staticly typed, I do not need to run in test. Because my test are handled by my type system. Which is, I don't think it's true. But also people are going to say that a bug is merely a poorly typed, poorly checked type. And this brings also the, I don't know if you remember Harbid, which was over a year ago. When it happened, people complained that if OpenSSL had been written in a language with a better type system, the bug that, the bug caused by Harbid would have never happened. It's up to discussion. Maybe yes, maybe no. We'll see. All right. So let's get back to some Python. Let's say I have this method, Fibonacci method. If I want to run it, Fibonacci 42 is going to give me the 42nd Fibonacci number. If I wanted to test that method, one quick way of doing that would be to assert, for instance, Fibonacci 012. And maybe another bigger number just to make sure that the method does what it's supposed to do. So by only testing those five numbers, let's see what else I can do. So let's say I have my entire set of possibilities of various types, various objects I can have. And I have my set of integers which are here. The test that I have only tested the five integers. So it's like almost nothing in the entire set of possibilities that I have. Which means that if I have floats, for instance, Fibonacci 0.0 or 1.5, in the case of my method, it works. But it doesn't really give me what I wanted. And if I do a Fibonacci 14.32, it's going to explode right there. This means that with my entire set of possibilities, by using static type checks, I would have been able to, here I have my set of floats, but I would have been able to just remove them from my set of possibilities since floats or since Fibonacci cannot be calculated with floats. And same thing with strings, yeah, with lists. I don't need those. So this reduces the set of possibilities that I have to test against. What is the current state of static checks in Python? So Python is a dynamic language, but we do have some static type checks that are happening. One of those that I looked upon is JetBrains Python, so the IDE. So they added type hints and type checker for five years, yeah. And the way it works is they use doc strings for Python 2 and for Python 3, they use the function notation that I'm going to talk a bit after. And this provides the IDE with information on what types, either methods or variables are supposed to be. So it gives the IDE user some basic code completion. So as I said, it works with the parameter pass to a function, return values also, and local variables. Some example that I took directly from Python's documentation. Let's say I have this method, and for Python to do some type hinting, I could do, I could add doc strings with their specific syntax. So here, for example, A, B, C would be integer, and this would help Python to give me either auto completion based on those choices or to warn me if I give a string, for instance. PyTems went from a simple class, simple class types to a really more complete type checker with, like you see, topical types, generic types, function types. And I must say that they did a pretty good job with the community's feedback, so they just gathered feedback from the user's experience to help have a better system. Another one we have is PyLint. PyLint is a source code analyzer, so it's a command line tool, which looks for programming errors and helps you enforce a better coding standard. So the static checks that it's going to do is they're going to do basic Python, PEP8, sorry, style guide, and they're going to do some various error detection. So, for instance, they're going to tell you when you have variable that are undeclared, if you have modules that are not important, if you have a news variable, they can also tell you if you have a return statement and you have code after, it's going to say that the control fold will never reach you, so you don't need this code. It's fully customizable, it's extendable, it's a good piece of library, in my opinion. Last time I checked, there's over 180 different error codes it can produce, and even the current core maintenance gave an amazing talk Wednesday, which was really nice. It integrates nicely with IDs, so with VM, EMAX, Eclipse, PyCharm, and many more, everything is on their website. It's nicely documented. Another one that I want to bring up on is PyFlakes, so just as PyLint, it's a command line tool that will check your Python source code. It says that compared to PyLint, it's going to be faster, because it only parses the syntax tree of the files, and it will never complain about your coding style, and also it will try very hard to never emit false positives, so if there's a warning, it wants to have a real meaningful warning to you, and that may be a problem with the parser. So this is PyFlakes. As I said earlier, there's a functional addition, this is PEP3107, so this is a syntax that was added to Python, just in time for Python 3 back in 2006, and what it does, it allows you to have arbitrary metadata to your method signatures, like you can see here, and with those arguments, we're going to be able to get them from the underscore-underscore annotations super method, so this means that the other libraries like PyLint and PyFlakes are able to use those syntax augmentation to help. The next one is MyPy, it's an experimental, optional, static type checker that has been around for 2012. It was heavily inspired by a Python-based, Python-inspired language, which included an optional static type system. So it's an optional static type checker, it's a command line tool that you can use to run against your files, and it will use, first of all, the PEP484 type hintings. It has a powerful type system and compile type checking. And the thing is, the author wants you to use the tool after writing your program. So start by writing your code, and then add your type hints and write after, and then run MyPy to be able to maybe catch bugs or to enforce a certain type checks on those. So once more, it uses the function annotation, and it's going to use also the PEP484 type hints, and in that case, my Fibonacci method, since it takes an integer, if I pass a string, MyPy will produce an error. So there's also PEP484, which is, of course, optional type hints. It appears since when type annotation started over in 2006, a lot of third-party libraries and applications started using those, but it sprung lots of different ways of using it. So this PEP wants to be just a standard, bring a standard way of doing type hinting. And I believe it's going to be compared to what Whiskey introduced with Web frameworks of having a standard way and a baseline for tools to work with that. And of course, as the author states, Python will remain a dynamically typed language, and the authors have no desire of making type hints mandatory. I won't go too much in those, because even Guido gave a talk. There was a second talk about type hints, so I'm going to just tell that this PEP aims to unify and ease static type checks. Let's go back to our circle of possibilities we have. So if we include static type checking in our program, this means that the sets of possibilities that we have to test against are really lowered, but we still have this huge integer set that we are not sure what we should do with that. We could either do some formal proof with our method, but I think this goes against the principle of what we want to do. So one way of doing it is using a library called Hypothesis. Once again, a lot of people talked about this this week, so I'm just going to go quickly on what it does. So Hypothesis is a property-based testing library. It's based on Haskell's QuickCheck library, which itself is a combinator library, which was written back in 1999. It's designed to assist you in testing your software. So it's going to generate data, random data, and it's going to try to falsify your assertions of your unit tests. And once it finds a feeder, it will try to give you the simplified, at most, the failure. So on the normal unit test, what you're going to have is you're going to set up some static data. You're going to perform some operations, which is the method, for instance, you want to test, and you're going to assert that the result that you get by this operation is what you think it is and is what you expect it to be. The difference with property-based testing is instead of setting static data, it's going to try to test data that is going to be matching a specification you're going to be giving it. And also, you quickly, the way it does. So Hypothesis will generate random data matching the specification. If it finds a failure, like I said, it will try to give you to simplify the filling data. So let's say that you have a big list of integers that fills your test, it's going to try to reduce it to maybe having an empty list that fills or if it's having negative integers that make your test filling, it's going to try to reduce it completely. And of course, the data is going to be saved locally for a faster test after, since it generates random data and it does lots of tests with it. All right, so I have this huge method that we don't really want to care about. It's LZW data compression. I took this from the website. So I have a compressed method and a decompressed method. If I wanted to test it regularly with a unit test, this is what I would do. So I would try to compress my text and then decompress and I would try to make sure that it stays the same, since it's a lossless compression algorithm. But with Hypothesis, what I would do is I would give it a specification which is with the given decorator and I would tell it that it's text. And with this, it's going to try to generate random data that are going to be text-based and try to make this assertion fail. And if I run it, it's going to give me one failure that the method has. And it's, of course, that if I have an empty string, it fails. And if we go back to the test, the empty string will fail somewhere there. I think I forgot to put the stack trace. So this would be a, so having Hypothesis on top of some static analysis would help us to literally test the entire set of possibilities we have. So in conclusion, in the 20 minutes we have, type systems are inherently complicated. But even though they're complicated, it's interesting to know how they work. Both dynamic and static type checking have their cons and their pros, but I think that having both of them living in a type system could really help devs. PEP484 is going to unify type hinting and it's going to give more power to develop more type checkers in Python. And lastly, of course, Hypothesis and other fuzzing library can help you to reduce and find bugs early in your development cycle. So this is it. Thank you. If you have any questions, feel free. We have some time. Just catch me. So you were saying that PILand also does static type checking to some degree? What I want to say is PILand wants to use PEP484 to add type hints in its checks because it wants to use the ASC3 of the Python code you have to give you some hints of fear you could have. So from what I heard, it's something in the work. It will be done later. Okay. Thanks. Any more questions? Okay. Thank you. Thanks.
|
Jean-Philippe Caissy - Static type-checking is dead, long live static type-checking in Python! A few months ago, Guido unfolded PEP 484, which was highlighted at PyCon 2015 as a keynote presentation. This proposal would introduce type hints for Python 3.5. While the debate is still roaring and without taking a side, I believe that there is much to learn from static type-checking systems. The purpose of this talk is to introduce ways that could be used to fully take over the amazing power that comes with static types, inside a dynamic type language such as Python. The talk will go over what exactly a static type system is, and what kind of problem it tries to solve. We will also review Guido's proposal of type hinting, and what it could mean to you. Finally, I will present a few libraries that are available, such as Hypothesis or various QuickCheck-inspired library that tries to build more robust tests, how they achieve it and their limitations. Throughout the talk, a lot of examples will used to fully illustrate the ideas being explained. At the end of this talk, you should have a better understanding of the wonderful world of type systems, and what it really means to you. It should help you decide wether using type hints will be helpful to you and also if an external library trying to fuzz your tests has its place inside your project
|
10.5446/20135 (DOI)
|
Hello. Do I need to do anything about the mic? Probably not. Okay, so thanks for being interested in authentication. Here's the situation. You have a application that you've developed and a large organization is interested in it. They would like to either buy it or deploy it and you are excited about it. The thing is you've probably built it around Django Contrip Auth or something like that and maybe you've extended it a little bit and maybe you've added a nice management interface. But the organization wants all employees, all associates to have access to that application. And perhaps not just the employees of that application, but they have partners and suppliers and maybe customers and all of their people should have access to it. The workflow, the ideal workflow is like this. A new person joins the company. HR is a good HR and they put the person into their central identity management system, active directory possibly. They put the person into groups that somehow match what the person is supposed to do or groups to which the person belongs. And when the person then finally gets their password and logs in, they should be able to log in to any service that is somehow related to the work that they do. So if you hire a new finance person, that person should be able to log in to their finance accounting or whatever application right away. If you hire a new network administrator, again, they should get their laptop set up, log into a domain, perhaps it's a Windows laptop, and then access their network management system, network administration and be allowed to log in. And vice versa, the finance people should not have access to the networking management and the network administrators should probably not be able to access the tax records. The problem is that no one will enter those users into application in large organizations. So when a new person is hired, that application somehow needs to learn about that new person and about their access rights right away. No one is going, none of those administrators are going to type that person's details into multiple of their applications. They only want to do it once. Moreover, different organizations have different ideas and different requirements about what authentication protocols are to be used. Sometimes it's Kerberos, typically if you have active directory. Sometimes it's some chip cards. Sometimes it's SAML, especially when you want to integrate with other companies. Sometimes, often, the organization has a verified audited method of setting up those authentication mechanisms using some front end HTTP servers. And they won't change it for your application. They have standard that they won't use. So I'll probably do it the reverse way. I typically do it. I'll do a demo first so that you know what I'm talking about and if you think that it's too easy, you can leave and won't suffer here. So I have free IP server which is basically something like Active Directory just on Linux. And I have Bob account created here. And I have a very simple jungle application. The application looks like this. It just shows who is logged in and it shows the last login attempt. So only we are currently not logged in. Only admin has ever logged in. That's the situation. And I also have identity provider which is basically which provides me with SAML assertions. And it's connected to that free IP. So I have three machines. Bob currently only exists in that free IP server. Now, I can do it two ways. I can either click login here in that application or I can click login here in that central single sign on solution. Which way do you prefer it? Okay. So please log in. I'm Bob. I know the password. I get logged in. Welcome Bob. Now, when I come to that application, I'm currently not logged in. I click login. I got redirected to that identity provider. I get redirected back. The log now shows that at 1235, it's GMT time. So it somehow matches the reality. Bob has logged in. And not just that, Bob has acquired some privileges in jungle. Now, well, either say wow or leave because it's pretty cool. At the end, I'll show you the same with Kerberos and at the end, you will be able to pick your own login so that you know that I'm not faking it. So how do we do it? I will assume that we have Apache, that jungle is running under Apache and that we want it to consume that typical standard remote user authentication result. So how do you do it? Well, since jungle 1.1, you have remote user middleware. So that's easy. You set it up. Well, there are at least two problems. First, remote user middleware really expects all URLs or accesses to be authenticated. So even if as the result of that access and what I did not show you and I probably should, if I look, I have two different Firefoxs running. So if I refresh the view, you can see that Bob account got created. And actually, we not just know that it's Bob, we also know his first and last name. So there's more than just to the login name. But remote user middleware wants that login name to be present in that remote user variable or whatever you call it upon every request. So you would even need to maintain some sessions in Apache, which is kind of duplicating what Django does anyway because Django created a session for us. Or you would need to re-authenticate upon every access. That might have been fine when HT password files and basic authentication was used. It's not fine if you use Kerberos because you don't want to renegotiate upon every request. You don't want to renegotiate upon every renegotiate sum. Second problem is if you use the standard odd views login, it does not really understand when the user has already authenticated via some middleware. It still shows that login page, even if the Django authentication session has already been created. What's the solution? Well, how do we want it to work? We want extra authentication, the one that the organization requires, be it Kerberos or GSS API in general, be it SSL being SAML, I'll show some example with Smoot. We want it to be only created or only enabled on one login URL. The URL that we were clicking here where? Let me log out from single sign or otherwise I will not be able to log out here. This is the login I'm talking about. For some reason it doesn't show the URL, but it's slash logging. The solution to the first problem is coming in Django 1.9. We have new authentication middleware called persistent remote user middleware. It basically does what it says. It will only require that remote user extra authentication to be present when the Django authenticated session is created or when the Django session is marked as authenticated. It will basically preserve it until you log out in Django. For the second problem, I've come with a solution which basically I'm seeking some comments about. If we check that user is authenticated, but unfortunately we need to write our own view or login. Of course, it can inherit from the standard of use, but if we check that the user is authenticated, we just redirect to whatever the landing page of that login page is. If we click the login page and the user happened to be authenticated before the handler got chance to be involved because some middleware kicked in and found that remote user populated, we'll just say, well, okay, you are authenticated fine. Django upstream is not currently very happy about this being the default. At the end of the presentation, there will be more links about how we want to solve the problem. Maybe even is that a problem at all. Now, if you have a modern application, the username, the login is not enough. Applications want to send notifications to their users. So they need some reasonable email address. Applications want to send notifications to their users. They want to show welcome David when David logs in and welcome Bob when Bob logs in. So they want some additional attributes about the user. So here's the proposal. Since we started with the remote user for the login, let's use remote user and underscore attribute for attributes of that user. We've actually done that in other non Django, non Python projects and it seems to work pretty well. If you are using SSSD based installations, you can use more lookup identity Apache module to configure basically mapping of LDAP or other attributes to those environment variables. If you use more for some, you can do the same. So it's possible with if you are depending on external, maybe Apache, maybe ngenx in the future, maybe some other front end server, but if you are depending on external authentication, we should also somehow expect that external environment, that front end HTTP server is able to populate some other environment variables or headers, whatever you call it, than just remote user. How do we consume those attributes in Django? Well, I have remote user after middleware, which basically checks that the user SVC authenticated in Django is the one matching in the header so that there's no mismatch. And then it just uses those attributes, sets the attribute, uses those environment variables or meta values, sets attributes for user and saves. I would have pointed with a pointer if I had a pointer. So what this means is that upon every request, not just remote user middleware creates the user, which is what you would, what would happen normally, because you need that user record created in your application database, otherwise your foreign keys won't match or won't have anything to point to. But we will refresh that user record whenever authenticating session is started. So we have fresh data about the user. By the way, fearfully to raise your hand and ask at any time when you don't like something or when you're confused or when you want to add something. Now, that's nice, but I said that we also want user membership to somehow relate to permissions that the person has in that application. So the networking people, network admins will be put by the HR to some groups, which will make them have more privileges in the networking application than the normal people. And it's not a bullying thing. It's not like you either get in or you're completely out because, for example, help desk people, you often want them to have access to everything but read only so that they can check status of things. If it's, for example, my account is system and network administration. So if I have IT help desk, they need to be able to see into things. They should not be able to modify anything. Only a few things, maybe. So based on their group membership in Active Directory or somewhere else, they should be given application-specific permissions. The similar way, if I log in back as Bob, the similar way Bob got his three permissions here. So here's the proposal. Since we don't want to tweak the Django schema too much, we could have added special model, special database to somehow hold the mapping. But let's start small proposal. If a group in Django starts with X-con, whenever an externally authenticated user logs in, populate his or her membership in those X-prefixed groups with whatever groups you find coming with that user. And vice versa. If the person is in that X-prefixed group and no longer you see them being member of that group when they are authenticated, remove them. So that when the person changes departments or is no longer working on the project, you have one central place to do that setting. And it gets propagated to any application which gets hooked to the central server while similar mechanism. At the same time, if the administrator admins need to say that, yes, this person, even if they are not network administrator person, they will need to have certain permissions. They can use any other non-X-prefixed groups and manage their permissions anywhere they like. I'm just proposing that those X-prefixed groups are somehow special. And yes. You're assuming that you only have a single external authentication mechanism. I think you want to main space this in order to be able to have multiple ones. That's actually the point. One way, I'll show, I'll show in a minute, I'll show a code that does it. So one possibility is to have separate middle-end for each. But you probably don't want that. But one possibility is to also get users affiliation or what authentication mechanism or domain, perhaps, that they came from, have it part of that reference. But yes, you spotted a problem. I did not want to get that deep in the stock. So you are ahead. So we said that we would have, that we could have remote user email for email. So why not have remote user group for groups? And the way you can set it up in modlookup identity which does lookup using SSSD, which is an OS level authentication and identity solution. And the way it is possible in mod.autbalon in the coming version that I hope will be released soon, is that it populates group underscore n so that you know how many groups are coming and then you have those individual variables with separate names. And another possibility would be to have it just call and separate in one value. That's also a possible way to do it. And then we have the middleware and yes, the prefix is somehow hard coded here. Because I thought that we need to start small and start somehow. And then we look at the groups that we get. We look at the groups that we currently have and we basically update what we have and then call this and save the user and update the user. So I showed you how it can work with SAML. Let me do it with Kerberos. Let me log out so that we have somehow a one-on-one state. What user would you like me to create? Come on, say something. You be very silent. Let me give her a password. Let me try to make this smaller so that you can see more. Yes, it works. Cool. And I will not just create the account. Oh, I've been speaking too long. So I need to re-authenticate as administrator, which would be much easier if I saw the bottom of the page. Okay, so we have Penny created. And let's put Penny into Network Admin's group because she's joined the IT department. So we just created Penny. Let's change email address. Why do we have email addresses here? So that you can trust me that it's not just faking it. So we have Penny created. Sorry. We have Penny created. Let's pin it as Penny and Kerberos slides, realms. By default, VIP creates the user with the password expires. I have a ticket granting ticket for those of you who are fluent in Kerberos, which basically means it's a good thing. So let's verify what we have in the Django application. We still only have Admin and Bob. And now, well, now I need to change the configuration of that application because I have Y which is the client for some of here. So let's remove that. I could have used both, but I don't want to complicate things. And Django, let's put it here. Confid. Restart Apache. And so I'm blocked in single sign on. This is what you would see ATGT if you blocked in on your Windows machine as well. Let's click login. Now, we can see that Penny has logged in, even if she did not have any account there. If I refresh now, the account is there. I can see that her email address is there all right. And I can check that Penny is a member of X network admins group. So her group membership that I created that I set up in free IPA got propagated. And if I look at that network admins group that I had created, you can see that it has listed these four random just for the purpose of this presentation permissions. And if I look at what the application says to Penny, it lists just exactly these four permissions. Conclusion. It is possible to support multiple authentication mechanism without writing Python code. Now, this is a Python conference, so you might not find that a good thing. But first, it might be required that a front end authentication should be used because it's what the company or government body uses. And that implementation might actually be much harder than you can imagine if you start from scratch. In Django 1.9, it is possible to easily use authentication enabled, that external authentication enabled in Apache or in another front end server, just for the login URL. And it will survive. It's persistent, remote user mid-war. You need to be careful, especially if you want that login page to still be available when the external authentication fails. Because you might want users to use Kerberos, but fall back if they don't have a ticket. And by the way, just so that you trust me that Kerberos happened, you can see the HTTP service ticket was created. So it really was Kerberos that authenticated Penny. Now, yes, traditionally, remote user was what authenticated users. And in Django, we had the functionality to create that user automatically upon the first login for some time. But these days, you probably want more than that. Attributes of that user, and yes, you need to find a way to somehow match what you have in your corporate identity management to what attributes each and every application adds to the model. And more importantly, group membership. Because you can link permissions to groups and have those permissions pre-created and predefined. And when the user gets created, they get propagated. And you can use them. Again, we were not writing a lot of Python code. We were not writing some more Kerberos in Python. We are just consuming external authentication. Depending on your view, it's either a good thing or a bad thing. So I welcome your questions and comments. And here are some links to go with this presentation. And now really ask me some questions. Yes, please. What about federated log out? Actually, sorry, what about federated log out? The identity, first it's hard. Second, the epsilon identity provider does support it and someone does support it. But depending on the protocol use, it's either possible or not. So if you are thinking about a kill switch, if the person gets fired, so they're really logged out, I don't think that you'll find a solution, like a standard solution. So I was not focusing on that. Would you be using the same method if you're using OAuth or something like that for login? Possibly yes. It's still on support set. I'm not sure how groups would work. How do you use it for authorization, depending on the group. Yes. The question is how often you will find it used for end users within our organizations. To me, it's more for public use. My question is actually pretty similar. I thought about more OpenID Connect. When I already have a single sign on based, for example, OpenID Connect or anything else, do I depend on the implementation on Django that it's possible to connect it? Or how does it work? Because you were talking a lot about Kibiris or SAML. So what is really necessary that I'm able to connect my single sign on? My answer would be what does your Apache support? Because I'm really trying to not do it in Django. I'm trying to build a framework as a bad word. But I'm trying to find a approach in Django which would make it possible to use what maybe other languages and frameworks use. So if it's implemented for Apache or for some other front-end web server, then you should be able to consume it this way. Without having direct support for it or direct knowledge about that protocol in Django itself. Yes, but at the end of the day, you get authenticated user and some attributes. So what I'm saying is don't try to or maybe you don't want to try to address it in Django. Maybe you want to find an existing solution for that and just consume the result. And it does not really matter how many round trips the protocol mandates to get the user authenticated. Eventually, you will get the remote user or the indication that the user has authenticated. And that's what we use here. Hi. Hello. We actually used to work together. I'm wondering if you have ever made a similar Django app but with Uwizgi or something else other than Apache. With what? With Uwizgi or some other server other than Apache. Well, we actually had a person working on the modules for Nginx. So we're trying to somehow expand the approach to non-Apache. But we are still only focusing on Apaches, our main thing. And again, it would be about implementing those protocols in those additional servers. And then consuming the result in hopefully a standard way. Okay. So thank you. Everyone, please thank you.
|
Jan Pazdziora - External authentication for Django projects When applications get deployed in enterprise environment or in large organizations, they need to support user accounts and groups that are managed externally, in existing directory services like FreeIPA or Active Directory, or federated via protocols like SAML. While it is possible to add support for these individual setups and protocols directly to application code or to Web frameworks or libraries, often it is better to delegate the authentication and identity operations to a frontend server and just assume that the application has to be able to consume results of the external authentication and identity lookups. In this talk, we will look at Django Web framework and how with few small changes to the framework and to the application we can extend the functionality of existing RemoteUserMiddleware and RemoteUserBackend to consume users coming from enterprise identity management systems. We will focus on using proven OS-level components such as SSSD for Web applications, but will also show setup using federation.
|
10.5446/20134 (DOI)
|
Dyma'r gweithio. Dyma'r gweithio, fel mae'n gweithio, ac rwy'n gweithio, ac yn ystod, i'r ddweud y dyfodol y dyfodol, yn y dyfodol, sy'n gweithio. Felly, mae'n James, mae'r ddechrau, rwy'n gweithio ar gyfer ymlaen, ond rwy'n gweithio ar y Pai-Con, a rwy'n gweithio ar y Raspai Weather Station. Rwy'n gweithio ar y Pai-Foundasio Raspai, ychydig ymlaen i'r cyfrannu o'r cyfrannu cyfrannu i ddweud i'r cyfrannu cyfrannu. Felly, rwy'n gweithio ar y Weather Station, ond rwy'n gweithio ar y Pai-Foundasio, i ddweud i'r cyfrannu cerddoriaeth, ystod o'r ffais o'r ddweud o gyfrannu gyfrannu cyfrannu i ddweud i ddweud a'r cyfrannu i ddweud. Felly, mae'r ddweud o'r gweithio a'r ddechrau'n gweithio a'r ddweud i ddweud a'r ddweud o'r ddweud o'r ddweud o'r ddweud Felly, y projekty o'r ffordd yn fwy o'r ffordd. Yn ymgyrch ar ddweud ymgyrch ar y cyfnod, ac rwy'n meddwl am astro pi, wrth gwrs, mae'n ddweud y cyfnod ar y ddechrau. Astro pi yn y fwy o'r ffordd, yn y ffordd yn y fyrdd, mae'n ddweud ymgyrch ar y ddweud, mae'n cyfnod ar gyfer cyfnod ymgyrch ar gyfer cyfnod o'r ddweud o'r ddweud o'r ddweud. Felly, mae'n cyfnod ar gyfer cyfnod ond cyflym Llyrgr거든요, mae Llyrgrinied졌e chir effig co bath o Down Filipino wedyn bach ent have ham honno, y b mieuxm ychynתill fydd amdi, yn hitio fairESut Ymhyrch, sy'n d Defnod ymhellio Edituro ac mae'r bobl saneouslyf am de N i femwys maen nhw'n eich dawn mwy, mae'n defnyddio fy ng Kunst привydig i sefy古f yn unionsolio cardeniau is this picture, the Raspberry Pi team took last week with a Raspberry Pi on a helium balloon at yeah I don't know and what altitude this was at it reached an altitude of about 32 kms, and this picture is actually one of my favorite ones and it's actually picture 314 which is quite apt, and so we're then doing this ous with some kids and teachers as well and showing kids how we can use programming science engineering technology ryngwneud â phryg pwysg Doesn. Rwy'n addysgu'r et, ryngwot yn fain'r Starthaias gy Roedol. A nawr fe wnaeth chi i nhw yn edryadya dib Всё, hynddd eich cy avŵr da rhywun wrth gwitaenlau Gwmionotality, Ben Govern M, yang lŵr sy'n gilydd o mallem C whipping. Bydd arweinydd sy'n'lldu wir rhaglen constraints y wir o lityg maen Will subscriber sy'n y gallu hyffredigau L hagwymau eventuallyad Iüne Blue Lle파 wrthboneg taetw i Gweithg icexuラ wdd y cy Bruce brwynt yn cyffra Primers Y ad plast haf bangannau fy mod regularly?Classachal een aid astag a creu weithchartnag hit norio sydd os ble dicellon Freedom of ID yn y pe다면 bowьяodd tra par datnwch anega sydd wedi'w i phirwyrr optic haf an abroad y ychydig ma 회f benchmarks a f 컬러� gyda Leven Os yn y tyl Bahas. Aethwn ar munawdd Maemor futures. A newydd yr aelum o beth, i'r stations makeriaid. A ac roedd enw MD B1 yn dweud yr iawn ha-ddai mas. A rydych chi'n bwyng trwy bob amser Ym Fatreesi hefyd. **** Fy oedd gennych'r aww ti i gade trwy gweith Ng diseis. Ac mae'n br bragwm siarot yw'rhorro embohi. Nid yw hyn sa chi gweld yn belyniad o'r r scor airport ac rwy'n ni.完成 o cas churches ar y replacement. The main parts of the weather station but we connect up to the weather station with a series of sensors so we've got a this is all just standard kit.. We were storing in the UK called mafflin this is the same kind of kit they stock in there.. So we've got a simple rain gauge which operates with a little seesaw, that little seesaw triggers a little tiny digital switch and it just counts the number of the tips. So we've got nawr ydyn ni amichieleri a sevigoch. Rwy'n gondi'r cyни Poduoan neu mili ciwers i fewni ochr fanyio, mae holl nifer o dyn taran tren ni ac ma onsio. Mae holl Owl arweindien ac this one is a weather vane hen we'n merkie'r diworithgo ASMR. Rwy ydychTs gôl profa a dyau dal dimio yn gweld. Dywقet Traffies挂o'r tanthlu graf panitten.ecdid y fizwellyno arall. Felly byddwn na d來了. Felly oedd wedyn Reading你就 yn 5 gyda'n dydw Ynored Llogin Ond oedd weddynWill Cmwy健on bod noi polli Aber yma a fydd yr Rhyf yn blaid. eraill ei gweld byw gol, a'i稱 demasi指ill ddiddar ac mae'n gweld re immediately. felly gallu пой, mae'n daer un't puedo legiaeth follow a guide, set it up, deploy it and have it just logging data. That's a great big tick. You've got a lovely weather station set up because it can interrogate data. What's even better is if you can show the kids how it works. I'll come back to the project. Let's talk about resources. What we're doing is we're building up a selection of resources which are going to show kids how to go through building their own weather station from scratch. To the first one, reading digital sensors. Roedd y gallu ar gyfer, sef fyoedd, a'Mn nidbur ent mae'r plefadau кан cyfasib ni, Vader yn gwahanir hefyd. Sef du a'rmy innu bydd. Os ydych chi'n neud ar cos sy'n cyffan. Yn Ico ch束 pam chi'r hunain yn diwethaf i difrights. Bydd y rallon hyn sy'n ser grateful Super Roedd pethau poo dip которой. acベntio atleidwch yn alignadwy fruitau, neu nid gフfwrdd gatw son. A'r hollfaen iawn i gwbench fedr spillau fy lab calculator yw'r du methu... fel F yn dda dim yn cas Dyonian Felly mae'r compet yma en nad farmwyr a lle mae deimlo feir mae'r mwynntiaffwyr eu popeth a así fe fyrfl rôl o gefnawyd. Cardiwch yma i fod o gyespace rhagd am y cy puree ar y gwyrgyn amddweith. Felly mae'r fod yn lle amddafol sy'n ffordd ac erioed fawr yma nad yw RudAPL yn fyged yn eich c兴ци yw Chesru. Mae hyn i crus greu rhagd amdano o'r gwelodau gredge. Ac mae hefyd neu roeddbarth eich c� patients o'r chael cy sentiments o sponsored cyfnod o wrth iddo. children reading some of the analog sensors as well. Presenting live data. We've got a simple Py Game GUI that we just not up for a demonstration. So that you can see what those sensors are reading right now. Talking how to do data logging. We now have our sensors collecting all this information. How are we going to log that? Are we going to use text files? Was lots of inherent issues with that. ond iawn hynny yn eich gallu ei gennych. Iethwch i ni'n фydd.arlodd ddechau, ychydig i ni'n edrych i blétaru'u core tycofynd â cydi i gyrf titleโ shef投au. Gweinir hynny bobl iökull y cyflyb g sjf- regionalu y ni. Acfallod os wedi wneud hyn, a ch sy'n lleol representation. Fel Daith Llyrwyr maeAr Llyrwyr y Daith Llyrwyrballs yn cynaf mwyaf yn gwybod o'i achlyni dysgu gennydiิnog arlau cyisi. Ond i wnaeth o loedd y bydd o bobl reciped, ond Codexen y twfio ar y parodydd gŵr oedd ynhefweithio trajectory 59ominous drum S Psycho Gwydig. ond i성wyd nhw yn Gallorllaneol, Oetnetwch chi'n trwy pwysig ar unrhyw iawn.....ad yuralwyr yn scansul yn agorffent y daint. Fally皮 spoestriani Ysbyty Pwysig isb iguald saligarc-皇hwg pajr fa pointingo, ac yn bobi ar y cyd-lellmaewyd yma.....icerma mor ddylu ar y gyfer oedd�56ydd amlaedd ichi, analistanethau brofiwn lle i rhaid i wneud yw dd Erin judgement..... Beverly Queens. Oedden a lledweddwe tackle hoeith y hynny... so mae'r ègar notio ein ffordd y Blend,ade yw'n mynd o. Ry<|is|><|transcribe|> ych Gkarwch Sardd� rebellion. Rwy'n uns gweld arna o'r arweitum iworki ond tro ma'r rai arelo? Fi'n gwneud y pryd yn holl forthist Tys�答ling maen nhw y celfr eighthrhu sydd ni Yong idd rigid! Rwy'n rwy'n fy Talking Among Huw pen y mesta yn gyfiwr o gyflu am Johnson and ac mae'n gweithio i'r ddweud yma'r hwn o'r sgol, dwi'n gweithio i'r sgol yw'r 1-Wethaf, a'r 20 o'r clas yw'r 20 o'r cyffredin. Mae'n gweithio i'r 1-Wethaf i'r ddweud yma'r sgol yn ymgylch. Felly mae'n ddweud yma'r problemau. Felly yma'r 1-Wethaf, mae'n ddweud yma'r 1-Wethaf i'r ddweud yma'r sgol yw'r 1-Wethaf, i ddweud yma'r 1-Wethaf i ddweud yma iddyn nhw. Roed ar anem seating o digwydd, dyna huo'r zod govern. y cysur a ychydig wrth cyfwyrraeth fel y by dolphin tof, clonwch chi di darnини insulted y peth, a") at Oedda llent, wedi ar y llunedd Llywedeg. Ac hanibal cofer tyd, a hefyd yn credu diadжu cael gyfan. Y film gallais bod y tanhael hfمن gallai yn gw semndここbiol, yn gyfarerydd y merchledd wedi gweld bap yn yn baen yn hyn ar gyfer husiaeth. Efallai fairly y eligibility darnt y stygrif sydd wedi am le'r golyfre. Some Llywodraeth H lyraiad wnaeth i ddwy voy nynt loc ar bu amser ac yn gy Firstly 3. I ddaf i'r bwysig. Felly that's kind of all I'm going to talk about about the weather station kit itself. So what I think I'm going to do now is I'm going to try and do a live demo which could be interesting. So we shall see. So what I'm going to do is well maybe, if it's okay is I might at this stage whilst I'm doing some setting up, open the floor to any questions that you might have if that's okay, and then I can be answering those whilst I'm also setting up. So I'm going to give that a go. Ihta alright? Canyonhame peel, siw ddifus o aethau'r Cyfrydyn, sy'n ei ddis友 ar gyfer ysgol, oes yn ôl ystafels i. Ie priests её! F rausodio. Wyd Bug Sall yn cael ei Richard y~)ill, niff ei hyd maeth'r Matchol Collider. The 1000 boards that we are producing, they are almost produced, I think, as far as I'm aware. Currently I'm involved in the resource generation side of the project, that's my primary focus. I'm also working more generally with the project. I'm aware of the conversations. Our aim is to get these into schools early to late autumn this year. Mae'n yn remaining ond bodnd dim o'n cymryd mewnial iawn er mwynhau i'w meddwl. Fe fyddai'n gofyn, os mae'r gewneudau, fin a'r borbydd tenth rwy'n carcadwy! Fe yw accum i ha Siadodd Ac Niu Dy place w でul yn na repe, dwi'n bod ni提 edrydf wedi bod mi rhaid o'n bod nhw siaradol ar taeth talo allan sy modd, felly mae'r cyfan hyn yn brosPAur. mai allan gweithreoli gôl unimanic wedi gael cael botsiaeth mewn am Bachelor降ersonoding pan fel ymylf yn ddigonwattur math o'r achat haes. Roedd o'r ryn store'r data gyda cynedd F quickly Y Llywodraeth, i weithio lyżad certyn lleolaeth hynny. Wath yn fwy ang alley 있지 ac ym mely molecules rêngoedd cymryd rydyn. Wel, yn cheletwch chi ti'n fasgr y n degraden, ar gyfed y celfendraeth sydd mewn y cyflwmp effeith gallwch eich cyfl substitute yw'r lleoli. mae'r peth nodwch er Casgrwp rhywbeth hwn wedi'i seymλλu sydd yn engog penigol wedi'i risod ei unionig gyda'r pobl mewn. Mae fimlu a'i destŷ arnaadau.\llug Y rhaidd fyddwn o ballach – mé'i으면iaid pwyll фильма – fel hynny mae'n b 맛faen, allwn gyflydd mACEPASR, Pwy yuddwch â isau, tynnaid byw na'r bwyолоで Bran Om ar Ieith i digwydd ym Ch Gedanken dysgu, a nag iddieb er gwrs i hynny, oherwydd yn canolflu mewn cael mewn arm am gall Orca newydd yno arall.. Dyna eich cwrnm ходm cyntaf ar gyfer law alw, barking ar enbyrgwyd<|af|><|transcribe|> yw eich cyfais cyrgu 있다는ut yng ngy Afryr Wars Wir ac mae byn ar hostio wnawn On gadaeth o astrwbaeth ar gyfer y pwn. D lack le wrth iddistodd i dnosio'r Pro faux gan gynnyddio'r pair. R Gesetz wedi gael neud, edrych yw'r prwyll sydd hoff yn seimlo. Ie, nexh, pressa mi, gave it something away. No, so things we're currently doing at the moment. So the picture, the space-lawns picture that I showed you earlier on the balloon we did last week, we are working with some teachers later this month to get them trained up on how to launch their own weather balloon launch with a pie attached so that they can, so they mae cyn wedi shoulders a gas siarad o hurwm ni'swyr dyfur og dwi'n meddyliad ychwaneg. Felly awr hyper gyd siarad peth yn wahanol piolaeth. Wrthmi, gael eu rhani, dae'n gwneud sefydliadau ez golygu'r l groomingant. Yn mynd i Hydwch Sian careful ac mae'n cael eu bod cramwyr hosiad cyfyfriedd yn 애fonu am y stand. Yn ymhyng Snaf, bellachen getón ar Fyffeddwyr spiderf ac mae hi ddan seloghef productorwyr mur. 24 oed yn ymdweud o'r cyfnodd. Rwy'n meddwl. Rwy'n meddwl a'n meddwl i'ch gael ymddangos i'r oed yn ymddiol. A'r ydych chi'n gweithio'r cyfnodd i chi, ac yn ymddiol i'ch gael y Pai Kond UK, ysgol ymddiol yn ymddiol, y dyfodol yn ymddiol, yn ymddiol i'ch cyfnodd. Felly, rydych chi'n gweithio i'ch gael eu chyddo i'r cyfnodd i'ch gweithio i'ch gael ymddiol i'ch gael ymddiol, i'ch denodyddio'n modd i mi ddoch chi'n gwneud,<|cy|><|transcribe|> ran y gall nos o fartig yn rhoi, sad yn 50 dysgu os gallwch. On circlerion y sparkle profi, Fel eu wfreun gwfer y barn, angenwysio Crood ynれ achos i mi bethe'i callu ε un o'r cosfau. genuine i lleio'r arddangos gynnig sees i ar gynyddiad Australian I카 potion Won D Oil i phrwng Torun o'r cynghreibu o fel y same, sy'n oed dod i ymd Mommy. cymryd, ole'r amser yn ddwy i'r python나�w'r'm fel pobl... Mae fe dweud, над yna'n gobarod i'u finestu. Nid yw'r mater, ydych chi an recite all yr eich busbfyniengen. Ac Beulla Egypte Llwg. A bwith ar y mae'r mawr particular llych nifer twistedf plane'riencegen yma... rhywbeth hyn ar y p beginnen. Ych chi'n bywfo'r couwch trwy'r gri, doedd alcohol ac gre� KP16. ryn ni'n credu conducts, ei wneud yn bryw спhefódraeth, felly'r mod iffred tri Ryn ni fel ym regardedur Cddohesive a'r gawellio. Po blyniad de oed. Felly, yö iDBicio peisi adroddi票 yn hytrus o'r periyf ond â'r gwneud cyffredin i wneud gwirioned rests. Cydw i i doub widerav, sy'n rhoi wedi gweld dewis-du. Ond gw liningirAS ydweithi i hyn gyflynydiad. Felly byddwn ni gorhwymp mewn Ohvohon Yn gweithio, mae'n dweud yn ymddangos. Mae'n dweud. Mae'r cronjog yn y ddechrau, ac mae'n dweud yn dweud. Mae'n dweud yn dweud. Mae'n dweud yn dweud. Yn ymddangos, mae'n dweud yn dweud. Mae'n dweud yn dweud. Mae'r dweud yn yrwe defnyddio dyn nhw gweithio y cyddeth anur chi Hallu ans llawn. Mae'n dweud hijcelu ein� fy ngosedur i Gweith salty. Mae'r cymay adrodd ei dweud o brofi wneud. Hy fathernydd yn gwneud yn pygame Wortw. Wehe합니다 ac mae hy ferwyd dros yw hyd yn y gyffaint, useryn yma. Yn sínyd arrivingaio ar y stordd fydd ynghylch felly wefyrdd. Dyma, Mae iniwch credu sakeu am greu data ac ymgyrch Cadw yna'r wathgiad ar gyfer pawr. Yn hyfryd yma hyn erbyn eu ces estas, mae yna dull darmarfydd dda. Fyd gallol rhwynt yn tynnu byddwn i wefyrddens y cyr Tacg wider dron gylaeng. Ac gall Fan les modell byddwn o hyvwyno. If that plays...... I took for one degree mum She thought I had killed them The lines at its end It made it feel more alive My mother was stillción Because there is something that brings me more of an incentive from studying who in a way who is late because now I don't have children My archấu is that I fath, three o'clock, was a dip in air quality and I realised that it was because it was on my classroom wall directly next to the car park. And so at three o'clock when all the teachers just were like right on off, got in their cars, all their exhaust fumes just kind of got caught by the air close centre and everyday there was a spike in air quality, there we go. So just bear with me, I think it's demo and then if it's still running in the background we should get… Yep so we've got a little raspberry pie, we're the demo. assumptions on a phnodd, ac bod yn five minutes selected. Mae hynna enterwg mai wouldo sicrhau fearsристu neu cynglynlu cydryg itr关ci coffer a wedi cynehaf iais. Felly yn dechrau bod thysgu wereth gael, cael Fernando trof takeseradau, mae hwns ar mynd yn equ Pois bobl yn Med Skic! I was wondering about the screen-planned monitoring project. I was reading a little bit about it and I don't think setting up where I do some voluntary work. I think there have been some projects in the UK testing it and can you share any time frame? The plant monitoring is a small scale project which is happening. The moment those greenhouses are in the exhibit, it's part of an art. We've produced about 100 for this art project going on. The creator of those is mainly looking at how she can engage adults and children together on the same project. The teachers have come together to the workshop and have learned how to use it and taken it home. Lessons learned from that may inform some of the schemes of work. I think this water station project may be a little bit complex for children. What age do you think a child could be happily involved in this kind of project? I had the different schemes of work that we are proposing. The first scheme of work that I have written is with Kiesa's three children. In the UK that is 11 to 14 years old. They have done some very basic Python. Their Python programmes consist of 12 lines of Python because all they are doing is they are importing some stuff, setting up some pins and they just sit there and count in a loop how many events there are. Every few seconds they present what the current wind speed is. That is certainly very accessible. The graphical user interface side of things, we are looking at up a secondary there, 16 onwards. We have another idea which we have proposed. There was a talk earlier on, we were talking about physical computing. It is in theory possible to interact with this weather station using other languages such as Scratch. One idea we have got is to develop a way of displaying data using Scratch. For your young children they could be reading data and displaying it in a more graphical, engaging way. With really young children they could have a rain cloud appear. You could do some really simple stuff. Once you have got all this data that has been collected that might be the point at which the primary schools can get involved. You have got this data, we can start talking about what is weather, how do we log it, what does it mean, how do we predict and they can start using that.
|
James Robinson - Raspberry Pi Weather Station The Raspberry Pi weather station project introduces young people to using python programming to solve real and technical problems. The weather station consists of a range of sensors including: Anemometer Rain gauge Wind Vane Temperature Probe Barometer Air Quality Sensor Hygrometer 1000 kits are being given away to schools to take part in the project by following our schemes of work which will involve. Programming basic interrupt based sensors Advanced Sensors using ADC chips Create a pygame based UI Logging data to MySQL and Oracle Apex Presenting data to a web app Deploying the weather station Integrating Apex database We would love feedback on the project from Python Developers and support in updating some libraries from python 2 to 3.
|
10.5446/20133 (DOI)
|
James. I'm an ex teacher, and I now work for the Raspberry Pi Foundation. And I want to talk to you a little bit today about Pycon UK and in particular the education track that I attended last year. Just before I get started, I'm imagining that the majority of the audience are developers, is that the case? Are there any teachers or educators in the room? Awesome. So we've got some educators. Awesome. So I want to talk a bit about Pycon UK and to do that, I'm going to just step back I will be re-emphasising a few points that Carrie-Anne made in her keynote this morning. But I want to just step back to how I began as a learner. So sometime in the 80s, and we'll keep that vague, I was embarking upon my early education and my mum particularly was very good at supporting me with just exploring the world and finding out things. She'd often come home and I'd have a theory for about how gravity worked. It was wrong, she'd come home and I'd greet her with theory. But as kids are I was keen to investigate, I liked playing with things. The things that informed my early education were things like LEGO, which I played with for hours on end, followed the manuals. I don't think I had that set, I think I coveted that set. I think I had the smaller version. About the age of eight, I think we got our first family computer. We got the cheaper version of the Commodore 64, this is the Atari 65XE. Technically it was a family computer, but my siblings knew that really it was mine. I made that very abundantly clear. I began after playing some games, I started using basic and writing some very simple programs. I was also really into craft, anything, glue, scissors, cardboard, I'd make anything. I'll come back to this idea in just a second. I also progressed from basic LEGO, I progressed to Technic. Technic I loved and still love. I got Technic this year for my birthday and that was the best present that I got, it was awesome. Recently our house has been packed up as we're doing an extension. My LEGO came out and I was a mouth, very brilliant, there's my crane thing. For me the point with LEGO was I didn't just learn about how to connect things, it taught me about mechanical systems. The first kit I got that had a differential gear in, I was like this is amazing, now I understand how a car works, brilliant. I learnt so many things from these, again that's the set I coveted and bought when I was 25. LEGO really informed my early years as well. Knitting was something that my mum taught me quite early on. I lost the skill and I've recently almost picked it back up again, I was on a recent trip and I took some knitting with me. As well as starting out with some basic programming, I picked up some of the Osborne books which were fantastic and I specifically remember this book and a battleship style game that I built. The point I want to make about these educational experiences, these were before school or on top of the stuff I was doing at school, is that all of them were playful. All of them I approached in a playful manner. Every single one of them I could put down at any moment, I could make something in LEGO and go, ah, I'll throw it away. This comes back when we were talking about ideas during the keynote. One of the things that I really wanted to bring up at that point was that when you've got to use an idea and you've got to make a project, you're committing to making a project. Whereas if you've got this nice boilerplate scratch pad style area, you're playing. You can play as kids do and then just chuck it away. The other thing about these learning experiences was there was purpose. There was something that I wanted to get out, something that meant something to me. It wasn't someone telling me to do something. I wanted to build that airport. I wanted to knit something. I wanted to build the technical thing and I wanted my game to work and was frustrated when it didn't. The final thing was that there was progression. There was always somewhere for me to go next. So I started out with Print Hello World and I exhausted all the examples in the Atari manual that came with it. But then I found somewhere else to go. Lego, there's just so many places you can take that and there's so many crafts out there. There's always somewhere new to go. So those three things, playfulness, progression and purpose, are something that I think is really important in educational experiences. There's a middle section to my life where I went to university, did computer science, lots of interesting anecdotes involving me mainly losing stuff. That doesn't happen anymore. I'm a reformed character. But then this was me as an educator. So I moved on from being a learner and in 2004 I started a career in teaching. So I started my time as a math teacher and as someone who could use a computer and there was an outgoing head of ICT, picked up this role and really enjoyed doing it. It was a great experience. But I did get quite quickly frustrated with ICT and this again echoes some of the points that Carrie Anne was making earlier. A lot of ICT was being driven by office app based programming, was limited to Excel, macros and formula. A lot of the engaging and difficult, what I would describe as engaging and what a lot of teachers might consider difficult, concepts or activities, were left to the end of the year. A lot of primary school teachers that I was talking to around the time was, oh yeah, yeah, we do Lego robotics and we program them in the summer term. Sometimes we don't get the full six weeks, sometimes we do a week, because those things that leave to the end are the fun activity to end the year and then at the end of the year, oh, it fell off the end of the year, never mind, we'll do it next year. And those kind of experiences were always being left to the end. So as I started I was really trying to change the way that I was doing things in school. I managed to convince the school to buy some Lego robotics. We started a Lego club, we entered the first Lego league. Games factory, as anyone ever used games factory, it doesn't really exist anymore. But that's where I started. I remember hanging out in a shop in my local town called Microfuture, where they had this system called Click and Play, and it was a drag-and-drop games engine. You put a sprite in, you pressed a button and it pinged around and bounced off walls. That was your starting point. They developed it into an education product and I used that in school. It was really good. Scratch, again, lots of scratch stuff, simulating games, those kind of things. And we also tried to do some HTML. There was some HTML before. Sorry, I keep turning and the microphone keeps losing me. I'm a fidgeter. So HTML was something that the school had previously done through like a front page, or a dreamweaver or something like that. And I think you start using those big heavy ideas and you start losing the basics of how that works. That's not how I learned HTML. I learned HTML with a notepad in one window, a browser in another window. Save, refresh, save, refresh, save, refresh. So we did some basic HTML stuff. I moved on. I went to a secondary school, but I began collaborating. This is the point where things changed a lot for me. Because those frustrations that I had in school, being the only teacher who kind of, in my mind, got it, was really difficult. So starting to connect with other teachers was really important. I joined CAS, which is a group in the UK called Computing at Schools. I started working with local primaries, helping them deliver robotics in their schools. And then, all of a sudden, in, it was a few years ago, it was possible to teach the GCSE in computer science. So there was a pilot phase and there was a phase where teachers could pick it up. And I was super excited by this, because this was finally, you know, all these things that I've been banging on about were going to happen. And I could pick this up. And I was really excited. And then I suddenly realised, well, actually, how am I going to deliver this? What's my route to deliver this with kids? And the first question was what language? And that, sort of, again, like we mentioned earlier on in the keynote, this took me back to my childhood experiences using basic, and Python was really the clear choice for me. And then this holds up, oh, how do I teach this? How do I go about sharing this with kids? You know, I know how to write a loop or write a function or do this and that. But to me, it seems self-evident. How do I break that down and take my understanding? And put it, get it across to the kids? And how do I challenge kids? I had a kid in my first GCSE class who had been programming since the age of five, and it never stopped. Unlike me, when I was a kid, we didn't have the internet. We were very late giving the internet. Eventually I convinced my dad to buy us a three-month trial of this internet thing. We got on there and suddenly, you know, my horizons were expanded. But he'd been learning Python for years and C in Java. And at the time he came to my classroom, he was a far better programmer than I think I will ever be. So how do I challenge kids like that that have this experience? Because there is, unlike traditional subjects, there is this difference of experience. Some kids would have been doing this for years and years and years and have a wealth of experience, and it would just click. And other kids might be aware of it or might have been put off by it in the past. So where as another subject you might have a fairly level playing field, you've got a whole different range of experiences. So the collaboration with other teachers was imperative, was really important, and it was around that time that I gave a talk at Raspberry Jam. I met Cary Ann shortly afterwards, went to Pycademy, joined Twitter, went to more Cows conferences and things. And that was the point at which I became aware of Pycon UK. And submitted my application hurriedly because there was a deadline, and if you quit, you'll get a free place and they'll fund your cover costs. So great, yeah, I'll do that. So that's how I got to Pycon UK. So Pycon UK was a fantastic experience for me. It was last September, and particularly the education track, which is where I spent my day. Unfortunately, I had just moved house, and so I had a wall to go and knock down or something at home, so I could only stay for the Friday. But there was a two-day track. The first day, the teacher's day, was brilliant, and I'll talk a bit more about how these worked in a moment. The second day was a kids day, and developers were welcome to come to both of those two days, encouraged to come to both of those two days. In fact, there's a picture somewhere of Nick in a minute. I think it's the first picture I've got on here. Yeah, there's Nick saying, go to the education track, you should be there. So it was fantastic. It was filled with workshops, demos, training, discussions, quadcopters towards the end of the day, and cake, which is going to get anyone there, really. The cake's here. If they don't wet your appetite, nothing. They're amazing. We were joined by members of the Rangipai team. So Dave is in the bottom right-hand corner. He's not with us today. He's the guy that Ben was previously talking about, who is doing loads of space stuff. We've got Ben, we've got Karianne, and we've got Alex, who's down here with us as well. And there's the cake. So the cake was amazing. So one of our teacher that I hadn't met at this point, that I was aware of from Twitter called Cat, she'd been on a following, the pycademy after me, and we'd sort of been chatting a little bit on Twitter, and she was like, I'm going to bring cake. So she turned up with this cake, and it just vanished. I think I saw it in the beginning of the day, and then by halfway through the day, there's that cake gone, and there was one cake left. So it was a fantastic day. The link at the bottom, I've nicked a few things from Nick Tolway's excellent introduction to Pycon UK. If you want to have more detail over the two days, you can follow that link and it will take you there. And we were also joined by members of the community. I forgot to thank two people at the beginning of the presentation. This is Alan Adonahy, who, along with Nick, provided last minute a lot of the images that I've used in this presentation, because I had some camera failure issues. So, yes, loads of community people that had given up their time to come along to talk about education, how we educate, and the projects that we should be focusing on. So this was sort of how the teachers' day was broken down. So first of all, we had a session on Minecraft Pi, and we can see all the teachers worshipping Minecraft there, or Martin, I'm not entirely sure. So this guy you can see standing at the front there, this is Martin. He's got an excellent website, Stuff About Code, and that's how he usually starts his workshop with a, okay, please pay your respect, worship, that kind of thing. And he gave everyone a little demonstration of how to use the Python library to interface with Minecraft and create amazing things. And if you think back to the first point I made about my early experiences with Lego, Minecraft is often described as digital Lego, okay? This playground, this area where kids can play. I don't see Minecraft. When we talk to teachers, I don't try and put Minecraft across as, oh, here's a game. Minecraft for me in education is a medium. It's a way of expressing things. The idea we talked about earlier on, the story about, oh, we want a game where we've got, you know, it's set on a planet and there's aliens and you shoot them, well, with things like 2D graphics, Pygame, to get kids to make something graphical, I never bothered with my GCSE kids because it was just, it was more of an A-level topic anyway, but for the amount of time they had to put in to get something simple, graphical on the screen, it just, you know, Pygame just wasn't worth playing with at that time. Minecraft, we can build all sorts of things with Minecraft. So we set my kids in school to challenge. We said, go away and I want you to build space invader. And they went away and they did that, they built space invaders, but then they turned those into flashing space invaders, space invaders that chased you, space invaders that dropped things on your head, space invaders that chased you. They hid space invaders around the world and just went and find them. So Minecraft is a medium and this workshop was fantastic. And there were loads of people, it was their first experience of Minecraft, first experience of programming and they loved that. We then, there was a session we had all about physical computing. So some of the things that Ben's just been talking about in his talk, there was a session flashing LEDs, using motors, we had some spinning flowers and bees, is that right? Yeah, okay. I wasn't in that session, there was a bit of a split at that point, but there was some physical computing. We also had a session with an introduction for secondary teachers, particularly to object orientation. For lots of those teachers, object orientation was something they might encounter with libraries that they're using, but it wasn't something they were overly familiar with, it wasn't something that they had used for ever or for a long time. So it was a great introduction or refresher. We also, we had a team brought over some, oh, that's gone too far, there's a robot there, it's going to disappear. So the robot there is a little, it's got a now robot and a team brought those over, they danced, they sang, they did all kinds of clever things, you could take them for walks and they talked about their programming interface behind that and how schools should be doing more robotics. And then the afternoon session, we had an unconference style session, so we started off, we got all the teachers and developers together, we pitched some ideas and then we all broke off into groups and we kind of worked together as teachers and developers to either do some resource development. So I want a resource that helps me teach this concept. Let's go and make it together, brilliant. I want some coaching, so there was a lady who wanted some help with, I think she was doing some web stuff and she wanted some help with Django, the rings of eight bell. And so she sat down with a developer who helped understand how Django works and write the beginning or the framework for a scheme of work to take it back with her kids. And there was people that were sort of talking about how to make some libraries or some tools. So this screen here, there's a thing for one of the GPs they have to do a project on a little man computer. Are you guys aware of little man computer? It's kind of, I don't teach that spec partly because I don't like little man computer. But I think it's basically, it's a simulator to help you teach kids machine code type, like very low level code stuff. So very basic operations and instructions. And the complaint from one of the teachers was, at the moment this doesn't run on the Raspberry Pi and we're using them. So they sat down with some developers and they ported it to the Raspberry Pi. And that was like an afternoon's work for a developer. So this bit, the second part of the day was really, really valuable because it gave the teachers the chance to direct the help, the support they needed. And I think this collaboration is really important. So if we think about the first group of people that are there, the teachers. The teachers, they're really great at delivery. They've got years of experience of doing that. They know how to take a concept, to break it down into parts, to work out how to deliver that, to explain that, to get that idea across. They're also great at that sort of idea of progression. So this is your starting point. We want to teach you this concept. What's the pathway that we're going to take to get that? They're great at assessing, knowing where kids are up to. Sort of getting a really good measure of what kids understand, what they haven't grasped and what their next steps are. And they're also good at engaging pupils. It's their job, it's what they do. But what I need help with, and this is going back to my experience as a classroom teacher, what we often need help with is our background knowledge. I programmed as a kid, I did bits as I was growing up. I did a computer science degree where I did bits of Java, but it was more systems kind of base. I haven't done programming for a long time. And so my knowledge is rusty, so having that support. And some people don't have that background at all. They need more support. They may need some help with exploring possibilities, knowing what is out there that can help them. So they want to solve a particular problem, or they wonder if, oh, can I do this with code? You guys, the developers, you're the people to ask. Relevance. A lot of the teaching materials out there, first of all, for teaching computer science was, well, here we're going to do a little maths game where you guess the number, and we're going to use some selection to see if you're right or wrong. Or here's a teacher tool to, we're going to write a piece of code, which if you put in a score, it's going to tell you what your grade was. Which students are going to care about that? They're not. It's not particularly relevant to them. But developers, you're out there writing software, doing all kinds of cool things on things that kids might be aware of, whether it's robotics or web interfaces or whatever. You've got relevant experience that can show kids cool things. And also challenging people. There's 16-year-old that I had that was a far better programme than I. Talking to developers means that I can find ways of extending them, pushing them forward, and finding new ways to challenge them. And enabling learners. So this comes down to the barriers point that we talked about a little bit earlier on. As a teacher, there's lots of little frustrations that we have, either to do with network administration, the interface we're using, the libraries. Those problems, it's good for us to be able to air those somewhere and have those conversations. And developers. So you guys are great at creating solutions. It's what you do. Here's a problem, create a solution to solve that. You've got the really in-depth, detailed knowledge about how the libraries work, how they interconnect, how we can use them. You're great at writing libraries and tools that teachers can take into their classroom. As I mentioned, you've also got that relevance, that experience which is current, which is relevant, which is sort of real-world stuff. But perhaps what we can help with as teachers is finding new ways to engage learners. So that idea that you have about how, you know, it's a cool thing that we can do, we can maybe work out how to make that more engaging for learners. And making Python more accessible. The reason I chose Python for its simplicity and accessibility, I think that Python should be almost the de facto language, the text-based language that kids are using. And in the UK it kind of is. You know, lots of kids in the UK, lots of teachers in the UK have chosen Python based on other teachers' recommendations. Whereas in the States we went over there recently, and it's, you know, Python is used a little bit in education, but not as much as C and Java and processing and JavaScript, which I find weird because they're just so syntax heavy. So that's what I think those two groups bring to it. So the teachers, we kind of helped sort of sort of that progression, the bit I was talking about from the learning experience early on, the progression of how we get from A to B. Developers, you're great at bringing the purpose, the projects, those kind of things. But the bit that's missing is the playfulness. And that's where day two comes in. So day two was the bit that I missed. And this was, I saw pictures the next day of all the stuff that was going on, and I really wished I could have been there and knocking down a wall or whatever. So I just, I nicked some pictures from Nick's presentation, but you just kind of chose the journey that kids go through. Okay, so here's some kids sitting down on a computer. They're doing some coding. I've got no idea what it is they're doing, but they're doing some Python. It might be Minecraft, it might be GPO, whatever. And you can see there's a slight look of anxiety, maybe in a couple of other, they're a bit unsure. There's one hand on the keyboard. There's a few little things we can notice there. A few minutes later, everyone's trying to get hold of a keyboard, the mouse, they're all having a go. And this comes back to the point that Carriam mentioned, her keynote, that kids are inherently sort of investigative and playful and want to learn. Okay? And they probably, I mean, there is a whole series of, there's like 10 photos in this series, and they go through a range of emotions, like every human emotion you can almost imagine. You know, there's despair when it doesn't work and all kinds of things going on. But then we get this moment at the end where they've solved that problem. They've got through all those little barriers, they've worked out what the problem is, they've solved it, and the look of satisfaction on their face is fantastic. Okay? And that's why we should be getting together as educationists and developers to support these guys. Okay? To make sure that they're having this experience with coding. That they very quickly, they don't hit that point, the first picture, and stop, that they're able to get to this point where they see the value, they get the gratification from coding. I've got no idea how I'm doing for time, so I'm probably running a little bit short, so I'll move on better. So what's in it for the different groups? So first of all, it's fun. Okay? Whether you're a teacher, a developer, or a kid coming to a Python education event, it's tremendous fun, huge fun. We've got a picture up here, and the first picture, you can't really see it, one of the teachers that was on the course, Sway, this is one of her tweets shortly afterwards, she tweeted, I've just written something in Minecraft, I'm having so much fun, yay. Okay? Here are some developers who are helping us out on the day. Again, smiling developers, having fun, enjoying themselves, doing something playful and creative. And here we have the kids, just in awe, probably as Ben, I'm not sure what was going on there. Were you talking at that point, I'm not sure. So the teachers, what they get is they get coaching, support, confidence building, and that's really important. So many teachers did not have the background that I had, playing with code at a young age, and having a little bit of experience of it. And that confidence is really important for them. They get ideas for lessons, they collaborate with other teachers, they collaborate with you guys, they collaborate with the kids, which in fairness is actually where some of these ideas should be coming from. Okay? Because I don't want to teach kids how to do something that I want to do. Because that's my passion, not theirs. I want them to be leading things. So actually getting together with the kids, they come up with a good idea, you're like, awesome, let's make that happen. So we might get some new tools or contacts out of that. Building on network. Developers, you get to adopt a teacher, right? Which is awesome. So I had this experience, I sat down with some developers, and my project wasn't particularly to do with a scheme of work or a learning issue, it was more to do with enabling my use of GitHub in the classroom. So I sat down with some developers and I said, what I want is I want to be able to push things to GitHub, but it'd be private. I want a local GitHub, we found GitLab, which we played around with a little bit. And then what I want is I want an automatic backup mechanism from my pies. So it pushes there, and I've just got a repository that I can comment on and collaborate with the kids for. And we started excellent discussions, we played around with some ideas, they adopted me for the day. And unfortunately, shortly after Pycon, that kind of contact, those conversations very soon disappeared. Now, it's perfectly understandable. Developers have day jobs, I'm a teacher, I'm back in the classroom and doing things. So it's difficult, but if you are going to do this, adopting a teacher, it's really fantastic if you can maintain that relationship. Because it is really beneficial to both, and I'll explain why both in just a second. So you get to engage teachers and kids with the Python community, which is a fantastic community, and we need to get the teachers and kids engaged within that. You get that warm fuzzy feeling, which is great. But also at this point, it reinforces, and I'm paraphrasing Nick Tolwell here, it reinforces a deeper understanding of your own clarity of thought. So just to unpick that a little bit, when I was a maths teacher, and I said to a kid, right, well, I don't know, here's this mathematical concept, and the kid would say, well, why? Why is that the case? And I go, well, it is. It is, I don't know how to break that down any further. And that was me at the beginning of my teaching career, not being able to understand that some things aren't self-evident. At that point, my knowledge of mathematics was not deep enough that I could explain that concept to a child. So being able to explain things to people that are non-experts, being able to unpick why things are to you maybe self-evident, actually gives you a deeper understanding, and gives you a better clarity of thought. And the kids, what the kids get is they get a safe place to play and learn. There's no measuring their progress, there's no sort of feeding back on what they're doing apart from in a positive way. Oh, I didn't work, brilliant, try this, have a go. So they're supported, they're encouraged, and so on. They get to collaborate with others, and that's teachers, that's kids. Lots of kids don't get a chance to collaborate with adults. Lots of kids don't get a chance to collaborate with other groups of kids. So bringing them all together on this kind of day means they get that opportunity. And they get that sense of pride and achievement. I found a photo last night that I couldn't put it back up here. There's a picture of Nick standing at the front with a kid and a robot, and the kid is just beaming because he's sharing his robot, robotic creation with an audience. So they get that sense of, you know, I have achieved something, I have done something, I am part of something, and that's, again, really powerful to the kids. And did I mention the cake? Because that was great, okay? So I'm going to just take a brief pause. This is another perspective. This was a teacher that I'd met on Twitter. I met for the first time at Pycom, sat down, caught up, chatted with her. She sent me this because this was her point of view. I'm going to shut up for a second and just give you a chance to read what her thoughts were. I think there's a couple of points in there. Sorry if you're still really on par with me, I'll just move on a little bit. A couple of points in here. She regards herself as a relatively knowledgeable computing teacher, but really needs that support, or values that support from developers. What her kids need is something exciting which will inspire them. Teachers don't have a huge amount of time. The working week is very long. Anything that developers can do, or other people within the community, to say, here's this really cool thing, by the way. I went to a conference and I was shown, here's a library that you can use, which if you just run a command, it would access the tube times. Which is not super exciting, but it's way more exciting than crunching a few numbers arbitrarily in a fairly trite example. This is really powerful. There's a link there. Sway wrote a whole blog post all about her experiences of Pycon. I recommend that if you're interested, you check that out. How can you help? I will be reiterating a few points from Caryham's talk earlier on. Firstly, the adopter teacher. That doesn't have to be through a Pycon event, or any kind of physical face-to-face meet-up. In your local area, wherever you are, find out what your teachers are doing in terms of computing, programming, that kind of side of the curriculum. Find out if you can support them. Can you go in and talk about computing in your local school? Go and talk about the fun bits about your day job. Go and explain the cool things that you do with code and how it empowers you and what you love about it. Help develop tools and libraries that support learners. You've got the homework that Caryham set and said, we should all be working on that. Run a workshop or a talk for educators. That could be at an event like Pycon. It could be at a local networking meeting. It could be that if you run a local Python interest group thing, invite some teachers, get them to come along and find out what Python is all about. Have they got questions about Python and how to use it in their classroom? Code club. I don't know how many people are aware of code club or how you access them, but code club is a great resource. It enables a partnering scheme for getting experts to go into schools, work with kids, run workshops. That's great because it means that the teacher doesn't have to. They can attend and support you, but they don't have to be providing all the material. You get a chance to network. It gives you an in. It's difficult for me to get in with a school. Going by something like code club might be the way to do that. There's probably a whole ton of things that we've not thought of, amazing things that you could do to help engage kids and teachers, A with the Python community and with the language itself. If you've got any other cool ideas, come and find anyone from Raspberry Pi Education team, come and chat with us. If you've got an idea about something you'd like to do, then we're happy to have that conversation. I've left some time for questions. There we go. Thank you. APPLAUSE Thanks for your talk. It was very interesting. I liked the kids' day thing. Do you know how old the kids were? I'm going to have to defer to Nick, who was organising. What was the range of the kids there, Nick? The youngest was about five years old, and they were at 280. I think that's important as well, that it's not limited on age, because it means that the older kids can mentor the younger kids. The younger kids can mentor the older kids. There's lots of great opportunities for those collaborations. Nick, Dawn, you can have a look. Sorry, I've got my hand up. I feel like I'm a child at a primary school. Those photos of the children that we saw programming, in the series of... The longest series of photos, actually the kid in the middle, at the last one, he came along and mentored the other two. They had a problem. He came along and there was a picture of him pointing at the screen, and then something goes wrong. Then altogether they put it right. The other important thing about this slide is that there are no adults involved in this learning as well. hen Verdwn across Mars Tfceld Pel Pacadegreid replied to socio Union Llyfrannidll camera Time to pass that call. The weather felt us and S容易��choniki I took two Schemes Rhoad yn ysgolwch ddylch, fawr, r��, drwyf.
|
James Robinson - Pycon - A teacher's perspective A perspective of the impact of the PyconUK education track from the point of view of teachers and educators. Having attended the education track at Pycon UK 2014 as a teacher, my talk will share both my experiences and those of other teachers attending. The education track bought educators and developers together in a way that allowed the teachers to get support and advice whilst developers get to support teachers in developing exciting & real applications for teaching computing. The talk will focus on two aspects of the education track. The workshops delivered for teachers by python developers and how this helps build teachers confidence. But also the breakout sessions where educators and developers with common interests can work together to develop something. This might be a program / library or a teaching resource, some developers gave a hands on and bespoke training session to a group of teachers. If we are to get more young people programming or at least having a positive experience of programming then we need to minimize obstacles to that experience. By having educators and developers working together we can identify those obstacles and eliminate them!
|
10.5446/20131 (DOI)
|
Thank you, Yurki. I'm delighted to be here today. It's interesting. One always thinks that we know that the Python community is what makes it what it is today, really, the whole ecosystem. But when you hear, you sort of see up close that it's all the individuals that had their little bit of goodwill to make it. So it's really nice to be able to see that. Before I start a couple of words about myself, I come from far away, South Africa, where I've spent my entire career either programming or working very close to it. Most of that happened in the financial industry. And about 15 years ago, I met Python for the first time. Not long after that, I was part of a team who developed quite a complicated financial system user interface using the first Zope. I don't know if anyone worked with that. At any rate, that sparked my interest in web frameworks. And I started wondering about how things ought to be done and had very idealistic ideas. I started investigating these things. A little while later, I was lucky enough to get a paper published in an academic journal about where I surveyed 80 different web frameworks and tried to see how one can decide what the differences are and the similarities, etc. A little bit while after that, a colleague joined me, Craig. And since then, we've been working on making those idealistic ideals a reality. And that's what I'm going to present to you today. The last version was just over 26,000 lines of code of which 40% are tests, if that's interesting to anyone. Of course, anything like this needs to be funded because you don't really make money from open source software, right? How I do that from my side is I'm lucky enough in the financial industry, what you see is these huge big systems that live for many, many years and grow many thousands of lines of code and become very complicated for people to deal with. So I basically consult on a part-time basis back to some of these companies who allow me to then sit down with teams of programmers and help them to try and improve some of these designs so that these guys can actually deal with them. It's a tricky job, but it's given me kind of a feel for small design issues that over time can become a big problem. And I really like that. Of course, the knowledge that I get there, I try to work into real. And similarly, the knowledge I get in real, where we have complete freedom to do what we want to do, I work back into those environments. Quick overview. I'm going to try and tell you what real is, how it works and why you should care. Those are the important things. Of course, I can't get into too much technical detail. But I'm going to try my best. Please ask me. I think there's a mixed audience here. So if there's something that you don't know, come and talk to me. I am one of those introverted people. So if I'm at a conference like this and I stand outside at a coffee break, I have to work really hard at meeting people. So you will be doing me a favor. I also want to talk a little bit about strategy and our status. And I hope I'm going to do this quickly enough to give us some time for questions, as long as they're simple ones. So what would you expect from a web framework these days? You probably would know that an HTTP request will come in and that the framework must do something to map this to some of your code that's going to execute, where you'll do something like read from a database and eventually use a template language to produce HTML that you send back. On top of this, there's probably all kinds of odds and ends that get added to be able to reuse bits of templates, for example, to be able to deal with CSS. And there's now a big ecosystem in the JavaScript and CSS world, more tools and tools to help you deal with that. What strikes me about this is that it has very much a technology focus. It's almost as if the tools target particular layers of technology that we use. And I don't want that. That's what I would like to change. What if you could take all of these layers of technology and push it down, and indeed, a layer of Python that allows you to actually just think in terms of and focus on what you're actually building and not on all the different bits of technology that's getting you there? How we think that one ought to do this is, firstly, you need to be able to focus on the different views that your user will have in this application. Also, on the particular user interface elements that are on these pages, and how the user gets to move in between the different views. And something like the button that I've got in here is we will take care of that by implementing behind that Python clause a vertical slice of all these technologies just for that button. I hope to explain that a little bit better. Firstly, so there's no template language in real at all. So what is a page then? Well, a page is a widget. And how you will build it is you will compose your own page, your own widgets, by adding other widgets to it as children, just like they do in GUI frameworks most often. There's another trick to this that I won't have time to talk about though, and that's layout, because you also want to add them in particular places and be responsive and all of that. We do care about that. But here's a simple, very, very simple example of how you can compose widgets or a more complicated widget from a simple one. Our most simplest widgets are actually one-to-one corresponding to HTML elements. And in this example, we create a div, and we add a child to it, which is a paragraph, and you see what it will result in what you would probably expect in a browser, right? But simple examples like that can confuse you, because widgets are very, very much not HTML. There are a lot more. This example shows what we call a sliding panel. It's actually three different divs. And it only shows one at a time. We've zoomed in quite a lot, so the controls are quite big. What happens, obviously, in JavaScript, it switches between different panels, and if you click on those controls, you can control when it switches and in what direction, et cetera. Just take a moment to think what it would take you to build something like this. You'd probably build some HTML. You'd probably get a JavaScript plug-in. There are a couple of different things that you're going to have to think of. Here's what it will look like in real. So you create a sliding panel class, because that's what you want on your screen, right? And you add a div to it as a panel. All our add methods actually return what they added, so it just makes it easy to write one line is like that, assign it to a variable, and then you add a div to it. It's a variable, and then we can add other widgets to that thing, and that's all there is to it. The rest can just happen for you. There's actually more to it than that, though, because look at this example. It's still the same thing, but what we've done here is we have switched JavaScript off completely, and it's still working. Obviously, it's degraded a bit. Nothing happens automatically, but you can still click on those controls to make it move. To make something like this work, you can probably guess that, oh, you just add a query string to the URLs, the H raves that are behind those controls. But if the user now clicks on one of these things, it's obviously going to run to the server, and the server must now have logic that will allow it to render the HTML differently with a different div showing. So there's some server-side logic necessary for this to happen. That's what we can do. On top of the technologies, we've also added ways for us behind a Python class, like sliding panel that you can see there, to do other things, like add URLs to your app. Doesn't matter on which view you're actually using these widgets. To add query string parameters. To add server-side logic that happens. Of course, building a complicated widget that uses those things is something other people could also use, but we haven't documented things up to that level yet. So for now, we just like hiding it. Something that you that's probably might throw you, if you don't know about this, has to do with how long widgets live. I just want to show you one example. The red area there, we've decided when we built this particular example that that is a widget, the heading with all the addresses in it. And we called it address book panel. I want to show you what that code looks like. So an address book panel is a div. And in its init method, we add children to it. We add a heading, we query the database, and we add an address box for each address that we find in the database. This should get you thinking, what if the database changes? How does this widget actually change? And the answer is it doesn't, because it doesn't live long enough for it to matter. Widgets are created when a user requests a page in the beginning of the request. They do their job, whether it's rendering or doing server-side logic, and they kill at the end of the page again. That's really all there is to it. The other two things that I said we want to focus upon, firstly, is what views there are in your application. And secondly, how a user moves around between these things. Let's look at how you define a user interface. A user interface is an application in real. In this case, you'll inherit from user interface. You'll have a special method called assemble, in which you can define all the views and execute all kinds of code. In this case, we have two views. You'll see the first one gets defined on slash with the title, and we set a particular widget as its page. There are other ways to do this, but I wanted to show you only the simplest examples. Homepage is a widget like any other. Funny thing, though, we don't construct the widget here. We create a factory for it, because we want the framework only to create all these widgets, depending on which view you're going to be visiting. So the creation is a little bit delayed. The other thing is the last bit of code there is to say how the user gets transitioned between these different views. Well, what it basically says is that if you are on the add view and the user clicks on a button that triggers the save event on the server, then the user will be moved to the Addresses view again. So that's what we wanted to make clear in our code. I'm going to change gears quickly and spend a little bit of time trying to convince you that you should care about this. I can talk for days about this, right? So these are the things that I thought are important. First of all, if you can forget the technical details, then you can actually focus on the menus and the layouts and the stuff that you really care about in your application. And I think this makes a difference in the quality of what you can build. It's also a lot of stuff that you don't need to know. And I'm lazy, so I want to know as little as possible. And that's why I try to build this. The other thing is Python classes. All those words you see up there are Python classes. They're a wonderful unit of re-use. It's probably much, much better than anything you can do with template languages and includes and macros and things. We try to leverage that as much as we possibly can. And because we can do that, we can actually deal with other subtle problems that sometimes crop up. For example, and we haven't done this one yet, but we will, there's this small little issue of someone who double clicks on a button instead of clicking on it. And then your server receives two post requests. Have you ever dealt with something like that? How do you sort that out? Do you really need to worry about that yourself? Or can you just create a button and have it be happening? There are also a couple of complicated requirements that are difficult to do in a reusable way otherwise. What you see here is a table. It's got a list of all the speakers at EuroPython. There are, I think, just over 200 of them. And in this particular example, we wanted to display only three of them at a time in this table. Obviously, you don't want to actually get all the speakers, info, all 200 of them, to be able to do everything that you see here in JavaScript. The only way for you to be able to only get three at a time, for example, would again be to have some server-side logic. And here, we can package it in one thing, a Python class that has the server-side logic and it has the JavaScript and even the styling. In this particular example, we can actually sort as well. And when we sort this table, it's sorted server-side. So it's the entire table that gets sorted, not just what you see there. You guys have to help me create a new buzzword today. I've searched Google and I don't find the word no HTML. And there's a word no SQL. I thought, please do something, tweet, create a word for us here. There are quite a couple of frameworks that are playing with this idea that they want to get away from HTML. They've got different focuses, different ways of doing it. The only thing that I saw that's sort of similar between all of them is that they all have funny names. But when I talk about strategy, I need to think about how to differentiate real from all these others. So I'm going to tell you a couple of things about strategy with that hat on. Real also, if you sort of get into the details of how it works, you'll see that it actually works quite a lot more than like the mainstream frameworks that you're used to, as opposed to these other guys. They generate a lot of code. We don't. Things like that. So firstly, we like to maintain the web semantics. We're writing a web framework. We're not writing a GUI framework. And we like certain things about web interfaces, like you have tab browsing, bookmarks, things like that. We also want to support the ideals of the web, things like the fact that if someone switches off the JavaScript in the browser, that your framework will still continue working. You might think this is not a big deal, but it helps a lot with bots and crawlers, for example. But there are other usages for it as well. Things like device independence, responsive designs, accessibility, stuff like that. There's a whole lot of knowledge out there about these things. The other thing that we are quite adamant about is if you look at the web world, the platform that the web gives you, it's a lot different to what GUIs look like. We've got multiple servers to balance load. We've got multiple clients that serve at the same time. We've got distributed execution, some of it happening on the service, some of it happening on the clients, different devices. So it's quite a different ballgame from, say, your traditional graphical user interface platform. And we think it's not a good idea to take an existing graphical user interface framework API and pretend to do exactly that on the web. We rather want to grow from the bottom up in this environment and provide you with an abstraction that's very high level, but grown in this particular environment so that we can learn from it. We also aim for higher level issues, like this particular one, which already is implemented. You can actually mark a particular method on a class and give it some code by which it can decide whether the current user is allowed to execute this method or not. And the framework will automatically figure out, if I'm not able to execute that method, that the button it's attached to should be grayed out or not visible, things like that. Because we don't generate code, we can actually use all the methods that have grown over time with current web frameworks. And this is a nice example of that. On the left side, you see what we currently have, an example that we've got there. And you see that we've cobbled together some styling there all by ourselves. On the right is the version that will come out next, which makes use of Bootstrap. And you can see the difference. It's different when you have professionals who focus on something and you can make use of their tools and you can make use of their knowledge about what sort of widgets should exist in the world and how things should be displayed. We try to use that as much as possible. A quick word on status. You know, many years ago, we had this idea, this dream. And a dream is a funny thing. It's sort of you see it from afar. It's like a place you want to get to far away on a mountain. It's difficult to describe to other people. You don't see it clearly. You don't know how you're going to get there. You don't know what obstacles await. It's not as if you can just create a project plan and do it and get there, right? There's a lot of risk involved. And I think what we have accomplished up to now is we've built a foundation that shows you that this is not idealistic. It's not impossible. We've implemented all the important features. The risk has gone down very, very much. And we've got something concrete that you can play with so that you can get an idea for what this dream was about and what we're heading towards. Of course, the road is not quite finished yet. We've got a lot still to do. There are a couple of things you won't be able to do with real that you could currently do with other web frameworks. We need to work on it to get those things done as well. We need to add more widgets with bells and whistles, things like that. But we also need to get people interested and build a community because this thing is too big for us. Or let me say, our dream for how big it could get is much bigger than the resources we have available. So that's why I'm here today to get some of you involved. So this is a dream and you're invited. There are a couple of things you can do. First of all, what I've given you here today, you probably need a lot more details and meet. Go look at the examples. They're on the website. Join our mailing lists. We have one that's on which we will only announce new releases. So if you don't want to be bothered and only want to know when something new happens, please go and join that. It's a little bit more effort, I suppose, to install something. But please go for it. Play with it. If you've installed it, you can follow our tutorial. We will be very, very glad to help anyone who struggles with whatever. And in that tutorial, we really explain a lot more detail and a lot of the flexibilities and more complicated use cases. And of course, we need people to help. And there are many different ways people can help. We need marketing people. We need programmers. We need all sorts, because we're going to have to grow up from where we are now into the future. That's all. Thank you. Couple of contact details there for anyone. Any questions? Hi, thanks for the talk. Just this morning, I saw the keynote about education. And that's probably an aspect that you didn't consider at all. It somehow jumped to me just a minute ago, because focusing on only one technology, on only Python, might make it especially easy for beginners to grasp also web technologies. So do you think that maybe on the education level, this could have a go? I'm not quite sure. Do you think that are you asking whether it would be easier for beginners to write websites? Or do you feel that because we're hiding so many things, they won't get a chance to be exposed to what we're hiding? Just repeat, please. Well, I think the first thing that you said was a good thing, because there is a lot hidden. The HTML probably is produced at some point, and the CSS is produced. And because of hiding that, I do feel that it is easier to start up. I can imagine that maybe for, I don't know, 15-year-old using just one technology from one tutorial would be easier than going through a Django tutorial and having to write HTML and CSS as well. OK, I understand. Yes, it's not something we really considered, but we want to make things easy for ourselves as well. Thank you. Hi, thanks. How easy it is to plug in like CSS or something to style the website that you're generating? At the moment, what we're working on, the next version that we'll release will make use of Bootstrap. And what we want to do is more and more allow you to customize the Bootstrap, that whole JavaScript environment with the tools that have evolved there, like less and sass and grunt and all of that. So that's how we want to make it customizable. So is it easy, are there a lot of examples on how you can do this? At the moment, we don't have that. At the moment, if you want to customize CSS, we basically just say, here's how you add a link to your own CSS, and then you write your own CSS. But that's what we want to do with the next release, to make it based on Bootstrap. And I don't know if you know Bootstrap at all. There, it's pretty easy to change things because you just set a lot of variables, and it looks different. OK, thanks. I'm just going to stop on the way here. Hi, you mentioned some things you couldn't do with really yet, which can be done with other web frameworks. Can you give an example for that? The most important one for me, there's a book. It's been on the internet for a while called Web User Interface Design Patterns. I don't know if you know that. We want to try and follow those design patterns. One of them is called Responsive Disclosure in this book. It's when you have a form, and you select something, and based on that selection, more of the form appear, and so on. At the moment, you won't be able as a user to do that. That's actually the very first one we want to target, because we feel that's a very important thing. So we sort of have an idea how we're going to get there. But currently, as a user who don't know much about the framework, you won't be able. So I'm not really aware of much else. Obviously, we don't have zillions of plugins and things like that, but we're taking it one step at a time. OK, my question would be that there's been some discussion about the performance of different template rendering engines that's been going on. And you're not really doing the same thing, but have you looked at performance? How long does it take to speed out the actual HTML from your objects? You know? My stance on performance might shock you a little bit, because I feel that you have to first optimize, and then you can optimize. First, not optimize, profile, and then you can optimize. And I think you can do that best when the actual model of how this thing works is better fleshed out. So we want to concentrate on that. There are some things that we have done for performance, and they usually mean that things that are dynamically generated get cached. But you must understand, think of something like assembly language. It's really, really fast. OK, but you do prefer Python. So there's a trade-off. Sure. Any more questions? Does all computation are done on the server side? Because you show examples, for instance, with the sliding window. Each time the view and the reposal change, is there a request to the server? When we do Ajax, usually there is a request to the server, because we would like for the HTML to only be generated on the server using the same stuff. If we do Ajax, for example, we won't reload the entire page, but we'll just render one widget server side and come back. OK, now I have another question. Do you plan to be able to execute things on the client by compiling a Python code to GES? You know, there's a lot of exciting stuff happening there at the moment with Asm.js and all sorts of things. I think that's too far in our future to worry about now. It would be nice. Like regular expressions in Python are different to regular expressions in JavaScript. And it's difficult to deal with that sometimes. But it's certainly something one needs to think about, but not right now. Do we have any more questions? No? All right. Thank you, Ivan. Thank you.
|
Iwan Vosloo - Reahl: The Python-only web framework Reahl is a full-featured web framework with a twist: with Reahl you write a web application purely in Python. HTML, JavaScript, CSS and all those cumbersome web technologies (and a few other lower level concerns) are hidden away from you. As far as web frameworks go this is truly a paradigm shift: away from the cobwebs of all the different web technologies, template languages and low-level details -- towards being able to focus on the goals at hand instead, using a single language. In this talk I will give you a brief idea of what Reahl is all about: why it is worthwhile doing, how it works, where we are and what still needs to be done. I hope to convince you that this is an important direction for web frameworks, and of how unique Reahl is. Developing such an abstract framework is an ambitious goal. I'd like to convey the message that what we have achieved so far, and the strategy lessons learnt along the way demonstrate this goal to be realistic and practical.
|
10.5446/20129 (DOI)
|
I live in London but I'm from Spain, from the north also, quite close more or less to here. All right, hello everyone and welcome to our last session before lunch. Please join me and welcome to the stage talking about how everyone can do data science in Python. Hi everyone, thanks for being here. My name is Ignacio Lola and I'm going to talk a bit about how to do data science in Python and what data science is for me. So a quick overview, I'm going to talk quite a bit about what I do, why I'm here actually talking about this. A bit of an overview of what data science means for me, what is the flavor of data science that I'm going to be talking about. And then we will do a quick overview of the data science cycle with some examples in Python, data acquisition, cleaning, processing and also using that data to predict some stuff. So that's me with a bit less of facial hair. And who I am, I'm not a software developer by training, I study physics actually so I came from the maths background or point of view. I've done some research in systems biology, complex systems, always interesting in how things work between each other and things like that. That drives my attention to big and small data not so long time ago and I started coding in Python around three years ago. You need to have in mind that my previous coding experience was doing Fortran 77 during university and I'm not kidding. It was not so long ago. Probably they are still teaching Fortran 77 in physics, I'm sure. And yes, 77, not even Fortran 90. I become obviously in love with Python very easily and I become also engaged in the start-up world doing a lot of data science and those kind of things. I'm also a huge advocate of pragmatism and simplicity and you will see that in everything that I'm talking about today. That's why this talk is also pretty much a beginner's talk into data science because I believe that with very little tools you can do a lot actually. You cannot solve everything, that's for sure. There are still problems and things that will need very clever people to work on there for a lot of time but most of the stuff actually can be solved quite quickly by most of us. Now, on contrary to saying that I'm a big advocate of pragmatism, I've done for the very first time all these slides in Python notebook because I thought, you know, it's a Python science, I should give it a go and do all my slides in Python. It makes sense. It took me forever but I'm actually, so it was not very pragmatic but I'm actually quite proud of the result even if it doesn't look as good as if, you know, I will try to use PowerPoint or whatever. I'm also one more thing. I'm also the man between you and the food from lunch so I will try to be a bit fast and do this a bit fun because, yeah, I'm looking forward to both after all the introduction early today about it. I also work at import.io. This is relevant because of some of the stuff that I will be talking about and also because of the vision of the data that I have and the kind of data science that I do. And what is import.io? It's a platform that has two different things. It has on one hand a set of tools, free tools for people to use and get data from the web. So to do web scraping without having to code, it's just half a UI and you can interact with it with not really a lot of technical knowledge and get data from the web, even doing crawlers or things like that. And it's also on the other hand an enterprise platform for just getting data. So we use our own tool and other things and we just generate very big data sets that we sell. I've been working in import.io for a couple of years as a data scientist and more recently as the head of data operations. So heading basically the data services that put those data sets together and deliver those to customers. Now let's go into the topic. What we talk when we talk about data science. There's a lot of hype around data science which obviously came with good things and bad things. When you have hype, there's some good things about it. There's a lot of jobs around it. So it's easy to find a data science job. You can get very well paid to do it, but also there's some bad connotations on it. So usually a lot of roles are ill-defined. So you can find unless you can find with the same tack things that are really, really different. And expectations sometimes can be actually quite not fair to what it is. To define what I mean with data science, I'm going actually to just talk about it. To just talk about what is the cycle of data science for me as it could be the cycle of development. And we will just see it on the go what I mean with data science. And I'm going to start that introduction cycling around this nice picture. This is called the Here's Journey, which I took from Wikipedia probably. I'm not even sure if the context of this image was like talking about movies or books or whatever. But it's a very nice metaphor for, I think, for most agile development cycles. And a very, very good one for data science. That thing that is called the call to adventure in that diagram is what I call the problem to solve for the business questions. Everything needs to start with that. All pieces of work that we do in data science need to start with a business question, with a problem that you need to solve. Otherwise you're just doing things for the sake of it. And I will come back to this thing probably two or three times over the presentation because it kind of upsets me because I see a lot of times the opposite. So, yeah, here is where myself, the pragmatic is coming. That's always the starting point. Then that threshold between the known and the unknown is when we start actually collecting data or uncleaning data to try to solve that problem, all those questions. We need then to do exploratory data analysis which is usually what drive us to some kind of revelation where we can actually start to have some insights and knowing what we can do, what we cannot do, and so on in the framework of the business that we are working on. Then it comes the algorithm and machine so trying to use that stuff to make some predictions. And the last things but not the less important is at the end we need to answer those questions that we try to solve or to do a kind of MVP. And we need to remember this is a cycle. When you usually arrive to your first model, it's just the first step into making it better. It's just the first step into actually solving that issue. You might then realize that you have learned something but you have learned that model is not the kind of the correct model that you need to use or that you need to change the kind of data that you were doing. As far as you have learned something from the first iteration of the cycle, you are going in the right direction. I also want to mention that when we talk about data science, especially in tech talks like this, most of the time we just focus on the machine learning and the algorithms, which is fine because it's a lot of fun. And if you are talking with people that came from mathematical backgrounds or from programming, they will get really deep into this kind of stuff because we find it fun to be playing with Google's deep dream code or to do stuff like that. Now, actually, most of the time that we do data science or something similar, we are not playing with those kind of stuff and we are doing many other things. Like data cleaning or exploratory data analysis usually takes much longer than playing with algorithms or tweaking them. Not everybody talks about those kind of stuff. And usually a lot of the pitfalls are there. So I'm not going to read all of these things, but I think it's a very nice list of sentences that I agree with most of them. And I will just highlight a few things that data is never clean. Yeah, most of the tasks will not require deep learning or things like that. Most of the tasks actually could be done with very easy tricks and we will see that. Yeah, this is basically a lot of the things I believe. I didn't write this. I quote the person who wrote this. But it's very pragmatic. I like it a lot. I think there's a lot of truths about data science there. So let's go inside that cycle and see some examples and let's try to do some stuff and see how that goes. This is a cycle which basically is you get data, you process data, you use it. And that's like a mantra. We need to be careful with that mantra because if you go deep into it, you can just, you know, you can try, you can be biased by yourself, biased by the data that you have. And then because I have this kind of data, I'm going to predict these kind of things because that's what I can do or bias by, oh, I really like to do a neural network right now. So I'm going to do that. Those kind of things happen and happen all the time. And actually what you should be biased through is through the business to say, okay, I'm trying to solve this issue. I'm trying to predict this thing. So what data do I need for that? What is the kind of algorithm or model that I need to make that prediction? And that's the right approach. But sometimes you might end up using, yeah, the data that you have and doing that cool neural network, other times you might be doing a very simple regression or just drafting some KPIs, but that's fine. The goal always is actually to have an action after what you have done. Your goal is that when you have finished your work, something is going to change. Something is going to change in your business or something is going to change in your business. So people use your product or in how you see your product or whatever. But there needs to be an action. If it's just like knowledge for the sake of it, something is going wrong. And you need to fix it. So let's go into getting data. This is a very important part. I'm not going to stop a lot on it, but it's a very important part because we can also be biased in getting data. Not a lot of people talk about this, but we can get data from, you know, our internal data store, which could be my SQL database. Getting data then means doing a SQL command or a series of SQL commands and putting that into maybe your Python code or a file that you are then going to process and make predictions on. Now, this is very important because usually when then you are going into the machine learning and doing cool stuff with the data, you don't think again about how did you get the data. And if you have done a mistake or if there is some kind of bias in how you get the data, you will have to be, you will be conditioned for the whole rest of the cycle. This is the very first step of the funnel. So you need to be sure that you are doing it right or that if you are doing something that is where you have questions, you at least have written down those question marks. So you know where to go in the future if you need to review this. As I was telling, we can get data from what can be internal sources, like, yeah, the database where you have data around your web page or around your customers or something like that. Or you can get also external sources, which for me, and obviously I am biased here because I work on this can be things like, like, with data, data you get from crawling or things like that. The next step is to process the data. And what I am talking about for processing data, I am meaning digest data. Digest data, so we get from that data that you got from a SQL query, let's say, or whatever that is, into the actual NDA array that you are going to use in Python to make a prediction or to make a plot. That is when the data is ready. And there are steps in between where things can go wrong or where things just can take time to make. So we are going to do a very simple example. This is a web page called a speaker pedia, which I find by pure coincidence some time ago. And it is basically like a Wikipedia or a list of speakers around the world or kind of topics you can find, I don't know, Obama there or things like that. And how much they cost if you want to put them in your conference. Basically, this was for me a surprise because I didn't know people charged to speak in places. But apparently some people do that. So I crawl the whole site and I make a database of this stuff just to make some analysis and some quick fun stuff or insights into how that strange world of people who receive money for speaking work. I've done that within Portillo, but I'm not going to go into how I crawl the whole site. It's pretty easy. And if someone is interested, I can show it to you. Probably take like 10 minutes or so to set it up. And I'm using the pandas to see the data and also to clean it a little bit. As you can see here, I'm just consuming the CSB that was let's say my in the output of my crawling. It was actually. And we got around more than 70,000 speakers and we got a lot of information. I'm just plotting here some of the ones, sorry, showing here some of the ones that we have. We have the speaker name, the fee, we have the location, tags, stuff. There's a lot of things to clean here, which is very common in getting data from the web. And in some cases, you can just do the cleaning while you extract data. It's the same when you are calling a database or when you are crawling. If I use the right regex, let's say, I could have turned ahead those fees into a number that will be read as a float here and not as a string because I have the case. But I've done it very plain and naive just to showcase how these kind of things happen and we need to deal with them. Same thing happened for Twitter's where we have that inside the list or many other things. I'm actually putting only a few columns here, but I have many others. So I'm showing here how can we clean, for example, the fee data because if we are going to do something simple, the very first thing that I would like to see is, you know, how much people charge for speaking and how many people actually charge and things like that. So I can replace very easily the case for zeros in a string and then reload the, say, that column of the data frame as a float. And we have then this ready to be used, to be consumed. That's what I'm calling basically process the data, getting it ready for that, getting it ready for using it. And there's a lot of things to do in using data before going into making predictions with it. And a good example is a dataset that we just saw. One thing that is called, that thing that is called exploratory data analysis is basically knowing, okay, I have that dataset, I thought that was cool. We need to make now something out of it. We need to know where we can start. And I'm breaking my rule here. I know I have no business context or question in this problem. Okay? This is just for fun. I just don't know that thing. And I'm for fun. I don't really have an objective in so far as this. We will see other examples later where I have that objective and are more like more real-worth examples. This is not the case. But exploratory data analysis point is very similar. You need to see what your data look like. And if I see what, if I want to see what my data look like in the previous example, well, I can print the average, the median and the most of the fees of that dataset. And we see very easily here, well, we have an average fee of more than, what is that, $20,000, sorry, $12,000. But the median and the most is zero, which is already telling us, okay, a lot of people actually charge zero. So that average is probably meaningless on that sense. If we do a box plot, we actually see that. We actually see that. We see that, but we see something else. The box plot is not even a box. It's just a line. Because there are so many things close to zero. And we see that that's also because we have like three outliers here. Three outliers that are, I don't know, like a really crazy number. So crazy that I can think probably maybe it's not true. Maybe it's, you know, I don't know how a speaker Pedia works, but we can go back to the source and think again, and this is why we need to think about this kind of stuff. Well, maybe if a speaker Pedia is actually like a Wikipedia and people can edit things, that might be not true. That might be something putting something crazy because that's what, 10 million or whatever, you know, that might be, or even if it's true, it's changing a lot. Anything that I do in my dataset, I have 70,000 people here and just those three guys is going to change all my numbers. So I might want to exclude those liars in any further analysis. And one more thing to comment here. I really love box plots. I think they are like one of the most important things, important plots that you can think about. And probably if I can choose only a few plots to work for the rest of my life, it will be like only three or four, and I think I can do it with those. Probably a scatter plot, a box plot, line plot and an Instagram and who needs something else. I don't know, journalists to plot pie charts, but really not people who is doing like actual stuff. Now after saying this probably tomorrow, I'm going to use something else and see that it's super important, but that's what I think. We can go deeper into this and say, okay, let's actually see the Instagram, but avoiding those crazy guys to see how actually this is distributed. The distribution is something that we will expect ahead, and if we again do the same thing of calculating the median, the mean and the mode, we see that the average is much lower, but we still see the same because there's a lot of people charging zero. There's a lot of people who are not charging. They are just there because it's a list where you see people by location and people by topics and things like that. So what makes even more sense to do is something like this, where I'm seeing how many people do not charge anything and how many people is charging and what is the average for those people, which is around $20,000 for a talk. But we see that only one between four people in a speaker pedia do that. This is getting me back to my previous point of knowing always which are your data sources and how you are biased from the very beginning, because the right conclusion here is 25% of the speakers in a speaker pedia charge an average of $20,000. It's not that 25% of the speakers charge at all because most of the speakers at all don't charge. It's just that you are not in a speaker pedia. I'm not. And that's a very important point. It's kind of obvious in this case and maybe it's not so obvious when you are working with your database on Hadoop, but it's actually the same and you need to have it clear. Other things that we can do here and we are not going to do, but we could do stuff like repeat this kind of analysis for a speaker topic and see how different topics have charged different maybe or have a different ratio between people who charge and people who don't charge. That's something very easy. We have a column already for the topic. We can do, I don't know, we can do location versus fee. How fee correlates with the location of the speaker. All those kind of crazy stuff. Very interesting. Basically when we do splattery data analysis we always want to do that kind of thing of knowing what is our median, what is our mean, what is our mode, what are the percentiles, plotting the data to see how it actually looks like, which of the layers we have and also which variables correlate or can correlate with others. I'm not going to speak a lot about correlation, but I'm going to give you at least one comic about it, which I think is kind of important. This we can do a whole talk just about this, but I think the comic probably makes the point even better. So, okay, we were using data. This is an example of a very quick and dirty exploratory data analysis. Another thing before we go into predictions is KPIs, K-performing indicators. What are the metrics of the thing that you are trying to solve or the thing that you are measuring? Because sometimes just monitoring the right metrics can save your business. And very simple things can have a huge impact. So we shouldn't be afraid of going sometimes for simple tools to do simple jobs. Every tool has is right for one job. And we shouldn't be afraid of things like Excel. The fact that we can consume data in pandas and do really cool stuff, that doesn't mean that sometimes, I don't know, Excel is not the right tool. I'm saying this because actually it's how most of the people consume data. CSB is how most of the people consume data and how most of the people is also going to read your data. So a lot of times the output of an analysis or the output of a report or whatever is going to be on the end of CSB. And it's important that we know also how to not how to work with those tools. It's not so difficult, but how to make both use of them. There is even a whole book written by John Forman called I think Simple Data, which is just about how to do data science only in Excel. And it's a lot of stuff about modeling and machine learning only in Excel. When I'm talking about Excel here, I'm talking just about something that can give you a graphic interface for viewing and editing at CSB. Not really about Microsoft Excel, even if I choose that picture because I think it's kind of amazing. Okay. Let's go now into making actually predictions, to doing some machine learning and modeling. I'm going to do super simple stuff here, but going to use different examples and some different whole bunch of different algorithms. First of all, when we go to this step is when we separate the data into what is called a train set and a test set. This means whole world. This means everything into data science because this is the basics of what you will be able to, in theory, prove why your predictions are correct. This means that all the data that we were preparing before, we are going to split it into two pieces. And one piece is going to be used to train our algorithm, train our machine learning model. And the other one is the one that we will use only to test the results. So it's the one that we are going to test in the model and then see, oh, if we were right or not. Because we know the answers for that one. So we can see what is the answer for algorithm and if that matches. And we can have some kind of accuracy into our predictions. It's very easy to get biased by this. It's very easy also to your data set not being specific enough. You have a sample set that is actually not good enough for the problem that you try to solve. But then you divide it, you train your model, you test it with your test set and you say, wow, I have 90% accuracy. And when you suddenly go into a real data set outside your very big data set, the accuracy is completely wrong. That happened a lot of time. It's a very big problem. So we need to be doing this all the time. It's, well, the train set and test set is what is going to tell us how good our algorithm is, but it's not like a magic thing. It's still biased by how was your first data set and where did you get it and how did you get it. After doing that, we have basically only one question to answer from my very simplistic approach, which is, do I want to predict a category or do I want to predict a number? If I want to predict a category, I'm in a classification problem. If I want to predict a number, this is just a regression. So there are only basically two things to do. I'm being simplistic and taking edge cases aside. But we can put almost everything in those two buckets and are very, very differentiated and they depend on what is the output. It's going to be a number or it's going to be a category. Let's start with the regressions because I think it's what is, what everybody has done. Everybody in high school has used least squares. And least squares is a machine learning algorithm that will make predictions with some data for where some, yeah, it will predict other points for the data using some trains data that we have. There are others, things like lasso or things like support vector regression, for example. We will see an example, but least square feet is basically a machine learning algorithm. Any other regressions that we do are basically going to be the same or the same in the theory. The only thing that will change most of the times is how we are defining the distance between the dots and our perfect line or curve to those dots. How do you define this distance if it's this thing or that thing or any other crazy thing is what will change between having a very simple algorithm here or having a more complex one. But on the end, we are basically doing this. Maybe we are doing this for 20 dimensions and not for two and, you know, we have maybe a lot of a whole bunch of other problems. But on the end is this, what we are doing. And I'm going to do another example here. The data that I'm going to use now here is more business oriented. It's hard drives prices that I also scraped from the Internet. So I have a whole CSV with like features for hard drives and prices. And I can basically do very easily a linear regression which is I think is least squares, this thing that I'm doing here after dividing into test and train set my data. I can see basically more or less what is the variance score for that linear regression, see how it looks like. And we can very easily using skill learn doing more complex regressions. Support vector machine is just two lines. It's just two lines to train, two lines to print a score and probably again 20 lines to make a plot. But on the end is very easy to do. And we can get some results. We see that the results here are not much better than the results that were from least squares or just like 5% improvement or something like that. Which might mean a whole world in business context but it's actually not a lot. Very quickly, some classification issues. Let's try to do as an example, let's try to put our heads into how people is using a platform for example. And here again I'm doing a real world problem. I'm trying to get to know better the users of import IO, the free tool, the free platform, plotting and dividing how they use our product. So I'm going to be looking into how much people is using the platform, how many volume do they do of queries and how often do they do that usage or that volume of queries. And I can try to divide that in clusters. That can just tell me something that I didn't know about that dataset and hopefully make me do better decisions in the future. We again load some stuff from SQL learn. We load the data with pandas. We do a quick model with using MinSeed with this one way to do clusters. One algorithm to do clusters. We plot it. I don't like how it looks like because we have bands of stuff. So basically the only clustering that it has done is in one of the access which kind of not sounds right. So I say let's do Cummins. If you Google for clusters, most of the people do Cummins. So let's try it. We find basically the same thing. And the issue here, which is very obvious for anybody who has done some clustering before or even some material before but not for the real beginner is that you cannot be doing this. This is absolutely wrong. You cannot be working with an ax that go from zero to, I don't know what, and one that go from zero to one. That's it's never going to work especially in clustering. So we need to clean the data. I'm not going to do it, but we just basically need to normalize the two variables that we were trying to plot. And then we just repeat the same thing. We have now two accesses that go from zero to one. And we actually have some kind of clustering that makes more sense visually. But also when I go to the data, because if I now use this stuff to see, okay, which user is this? And I see it with real examples, I see that it makes a lot of sense. And one of these users can be, I don't know, the user who use Python and has connected an application with our API and is doing millions of queries versus the guy who is using the UI to do crawling without knowing even what crawling is. And making that prediction might be very valuable because you can implement that into your, I don't know, your help desk system and the customer support guy that you have working in your company can know in a heath when a ticket support coming if that guy is actually a very technical guy or is a less technical guy or is doing this kind of user or that kind of usage. And that will improve the experience for the user and the support that they get and also the life or your friend at the support desk. Last thing that I'm talking about very briefly, we're running out of time, is a web page classificator, I think, decision three, which is another way to classify things. In this case, the context is I'm trying to do, I'm trying to basically know which kind of website is a website just by looking at very simple attributes of that website and which type of website I mean classifying the content. So trying to know, okay, this is an e-commerce website or this is a map or this is a job application board or this is events data, things like that. For that, very easy again with SQL learning, just two, three lines to make a decision tree and also to plot it, we plot this thing here and again I'm doing a very nice mistake here, which is when you see something like this, a decision tree is supposed to be simple to read and simple to interpret, simple to know what it's telling you. When you see something as big as this, it's because you're doing something very wrong, you're overfitting your whole dataset into a lot of very small conditions that will drop into this huge list of categories and decisions to then make the classification of categories. We can very easily change that just by doing a lot of things actually, but the most simple one you can just say no, the maximum number of leaf nodes that I want is this and then you've got a much simpler decision tree, which you can read and try to see if it makes sense, which you can make a prediction very easily also in only one line with your test data and see actually how it works out. And that's it, the recap, always know what problem are we trying to solve, clean your data and get it ready to use, be aware of very common problems like overfitting, I try to make an example of that or normalization of your data, I try also to make an example of that and always try to have an output which is something actionable, something that you say okay, we finished this analysis and now we need to change this in our business, now we need to change this in how we do support with our people or in how we are doing this in our product or in how we are dealing with this data. If there is no that kind of action, basically the whole thing has failed and you need to learn from that cycle and go again into the loop and make it better. So that was it, just telling you that we are hiring a lot of Teamport IOS, so there's a lot of different positions, DevOps, Frontend, QA, and Python with a lot of data connotations in the role, so anyone that want to talk about that or about data science or about Python or about web scraping, I will be here for the next few days and I will be very happy to engage in any conversation. Thanks for your attention. Do we have any questions? I've just seen that you jumped over the abyss in the adventurous cycle, the abyss like the death and rebirth, is there something in data science too like that in the hero cycle in the beginning? Oh, in the cycle, sorry, what was the question around the cycle? I cannot hear you very well. You didn't reference the abyss, the rebirth at all. The what? The rebirth and the abyss like the death of a friend. Yeah. You're referring about this. The very bottom. Oh, sorry, I know now what you mean. I didn't refer about that, but I think that's precisely the moment. I have actually words for all the things there, so I have the metaphor very well in my head. And the abyss basically is that moment of realization where you know what kind of problem are you really trying to solve from a mathematical point of view, so what algorithm is going to work? Because when we are doing just exploratory data analysis or when we are doing the data cleaning, we might not even know in that moment for a complex problem, we might not know in that problem and we are going to do a regression of a classification. We might not and even less what kind of algorithm is better for that classification problem or for that regression. That's the point of the revelation basically when you think you have an idea of how to solve that and then you just need to apply it, which is much easier. What is your experience with SK Learn when you were a beginner? Do I have to know, trying around with different parameters until I get a result or don't have to know the internals of the algorithms? It's very easy to use SK Learn. Basically, in the documentation page, there is even a tutorial of how to approach it from the sense like depending what kind of brand do you have, what algorithm do you need to use, which is like great map into how to do machine learning with it. Once you know what algorithm you are going to use, which is usually just a few lines of code to put in there, knowing which right parameters are going you need to use, if we are objective, it's a very hard problem and it's basically the whole thing around this is how do you fit those parameters. But from a simplistic point of view, it's not so much. You can just use basically some defaults or something almost random. You can basically do a loop and iterate through different parameters and see how it looks like. You always need to have an output from your model, which is either a plot or a prediction or even better the two of them. So you can see, okay, I put these parameters, this is my output, do I like it or not? Let's change the parameters till we fit something that we think it makes sense. That will be a simplistic approach into how to change parameters and fitting the right things using SQL learn. Thanks. All right. Do we have one last question? No. Thank you, Nace for a good talk. Let's all head out for the Apprentice Fabulous launch. Thank you very much. Thank you.
|
Ignacio Elola - Everyone can do Data Science in Python Data Science is a hot topic, and most data scientist use either Python or R to do their jobs as main scripting language. Being import.io data scientist for the last 2 years, all of them using Python, I've come across many different problems and needs on how to wrangle data, clean data, report on it and make predictions. In this talk I will cover all main analytics and data science needs of a start-up using Python, numpy, pandas, and sklearn. For every use case I will show snippets of code using IPython notebooks and run some of them as live demos.
|
10.5446/20128 (DOI)
|
Why hello everybody. I'm from the Internet. You may or may not have heard of me. I do quite a bit of writing and coding and whatnot. But I have little time and I'm not that interesting. What's more interesting is the way I work. I work for a very small other hosting company and no mind regs are. And the reason why it's interesting to you is because we are big enough such that we need proper metrics and logging systems in place to be able to even function. But on the other hand, we are small enough that we don't have a team who does that for us. We have to do it on the side. And it's just one part of our work. And I think that makes it kind of relatable to you. At least I don't think that Google sent their logging team over here to learn something from me. So to make it more convenient for you, I made a page with all the links, all the concepts, all everything I'm going to mention here. So just relax and listen. And the agenda is three things, basically. I'm going to talk about errors and how to get notified about them. I'm going to talk about metrics and how to know what the hell is going on on your service. I'm going to talk about logging and how to centralize it and whether you even need it. So one question I have, who is happy with their logging and metrics infrastructure? Liar. I'm not promising you happiness. It's computers after all. But maybe we can make it, maybe I can provide you with functional unhappiness, which is nice sometimes. So errors. I'm going to start with them right away because they happen. You have to deal with them. And they are the quickest wins to make. So I'm starting with them while everybody is still fully awake. And again, I have three expectations from my error log in, from my error notification system. And this is really timely notifies. I want to know right away when something happens. I want to be notified only once because this happens to people who use an exception to email logger. So I'm Hinnick and I had once 500 emails from such a thing. So I would also like to have some useful context of my errors because monitoring may tell you that something is broken, but this is not really helpful to have any idea what is broken, what's going on. So obviously there's a huge market to solutions. I'm going to talk about only one of them, which is Sentry. Sentry has a lot of things going on for it. Most importantly, it's owner and famous expert to DBA. David Cramer bought me a Burrito once. So consider it as my full disclosure. But it's also open source software. And it's written in Python using Django. So if you're deploying Python services, you may already know how to deploy it. And if you don't want to do that, there's a paid solution. The plans are pretty affordable, I think. And there's also a free trial and a free plan. So you can be up and running within seconds. So if you don't have any error notifications, you should try it out. So what do you get? You get instant useful notifications by email, but also by Slack, or whatever you want. There are plug-ins for that. Or it contains a trace back and some metadata. And the most interesting button, of course, is the view on Sentry one. A nice touch I find is that those emails have a reply to header set to your whole team. So maybe you're on a train and you see something that exploded. You can just hit reply and give them some hints how to fix it. So the web interface offers much, much more. And my favorite button is this button, because this button is telling me I've saved you 100 emails in your mailbox. So once you've fixed this exception, it gets marked as resolved. And if it happens again, it gets marked as regression and you get your notification again. So basically, it does exactly what you want it to do. As you can see, there's a lot more going on. There's a lot of metadata. It's much of it is collected automatically. So you can think of it like the Django stack trace view that many people still are serving to the customers, but it's just for you. So how do you get your data in there? So the short answer is JSON over HTTP. So you can use it with any language, any framework, even assembly if you want to go through scale. So there are nicer clients for various languages. They have usually the name Raven in it. And the Python one supports both multiple transports, which is how the others are delivered. So using GUM, the new I think I think IO, Twisted, request, and so on and so on. But also multiple integrations, which is basically how data is collected automatically without you doing it explicitly. So for example, logging, you install a logging handler and every exception that arrives there is forward to sentry. You don't basically have to change anything in your code if you're logging errors. For Django, there's great support. There's general whiskey support and nine more. Maybe it's even more since then, since I did the slide. So let's start simple. How do you do it when they love? You instantiate a client using a URL you get from sentry's web interface and then you capture it. You're done. This is how you capture errors and report them into a nice interface. What I personally like is this. For ad hoc tools, which you may or may not have a lot of in your operations, every exception is caught here. That happens in this function, get forward to sentry. So you don't even have to change your functions. You can just add a decorator to it and your errors are caught and forward it. Integrations make it even easier. I've already mentioned that it is built on top of Django. The author's thing or two about Django. The support is the best as far as I could tell. You add a single line, you get all of 500 reported and you can import a client from anywhere. That's that. We are already done. Deploy sentry or give David a few bucks so it can be via Miya and the burrito. Install Raven and add a few lines to your project. If you don't have error notifications, and I really have to stress that you are missing errors. Your customers are seeing those errors. You are not. You are losing customers. Get something done. To make it even easier, David was nice enough to issue a nice promo code, which is, I think, 100 bucks. I'm not getting anything out of it, but if you want to try it, there you go. There we go. There we go to metrics. What are metrics? Metrics are numbers in a database. That makes them a time series data because they are associated with a time stamp. They are basically the difference between guessing and knowing. If you want to make decisions, you need facts. I think it's an accepted wisdom. Otherwise, you spend weeks and months building something that's useless or even harmful. Metrics are those facts. We will give them a quick role. I would just think between system and application metrics. System metrics are something you observe on the server, like the load or how much traffic is going through. Very important should be collected using something like collecting, but not really part of my talk. What I'm talking about is app metrics, which is something you measure within your app. In the simplest metric, you can have our counters. Something happens and you increase an integer, which is pretty fast, even in Python. Then timers. Maybe you want to know how long your database queries take. Maybe you want to know how long your requests take on average. And finally, there's gauges, which I find undervalued because they are really useful if you want to debug something. They are just numbers, which you want to keep track of. So it can be the number of customers online or the number of connections in a connection pool. Think like that. I find them super helpful. There are much more, but these three are, in my opinion, the most important one. So what can you do with metrics? So we said there are time series of data so you can plot them. And such a plot gives you a lot of information that bare numbers don't. So for example, you see development over time. So you can tell that you are running at 99% capacity every day at 12 p.m. And if you don't do anything, it might fall over next week when you get one more customer. And you also see trends. So you can tell if you need to buy the, if you will have to scale out today, tomorrow, next week, or maybe never because you are using customers because you don't have proper error handling. So when you have graphs, you can correlate them so you can see, like, requests per second versus latency. How much requests per second can you handle? And since they are just numbers, you can do math on them. So, for example, if you have a graph of a counter, it's just a raising line. It's not really interesting. But you take the derivation of that and you have requests per second. If you have timers, taking the average is not very useful, but person types are very interesting. So for example, what is the average request time for the slowest 0.01% of your customers? Because what if every 1,000 requests takes one minute? You wouldn't know from the average because it gets smoothed out by the other 999. But this customer gets regularly, for some reason, one minute request, he may leave you to. Thing is, math is hard. The average human has one ovary and one testicle, which is true, but it's not very useful information. And you can do the same mistake with your system or your app metrics. So unless you know what exponentially decaying reservoirs are, use tools by people who do know what it is. So one I think you can also do monitoring on top of metrics, of course, because you can set a hard limit for acceptable latency. If the threshold is exceeded, ring the bell. Zero rates. If you have a busy application, you usually have always some kind of errors. If they go out of whack, something is going on. And it's actually true for any kind of anomaly. If, for example, benign errors like 404 or 401s go out of whack, there's something going on you should investigate. And there's actually a whole stack called Kail by Etsy that is just made for finding anomalies like that. So we've said they're living in a database but probably not SQLite. So what we're looking for are so-called time series databases, which have various features like special querying and everything. One of the most important ones is that you have a roll-up of your data, which means you have various resolutions of your data for the past, because you probably don't have enough storage to store a second resolution of all your metrics for the past year that might get expensive really fast, even if you're big hard disks. So you usually smooth it out somehow. So you have like, you want to know what the average load was a year ago per day, but you want to know it very precise for the past hour. So I'm going to introduce it to three. The first one is paid and hosted. And it's really, really nice. You can get started immediately by using curl and uptime and you have a curve of your system. I've done that. We started it like that too. The graphs are beautiful. There's a lot of goodies. It's a lot of fun to work with it. If you want to host them yourself, the current Inherfond gorilla is still graphite, which has been popularized by Etsy too, and it's written in Python. The front is in Django. The back end is in Twisted called Carbon. It's finally in trusty. You don't have to build it yourself. And you can say that it's a widely supported standard nowadays. So the network protocol of Carbon is supported by other applications too, just for compatibility. So the thing is, it is a little bit long on a tooth. So the storage configuration I just talked about, the roll-ups and limits, it's a bit finicky. And it might be not the most pretty interface you've seen today. It's XJS. If you don't have a pleasure to work with it yet, this is what happens when programmers build interfaces. I mean, it's open source of there, so I'm not complaining, but it's clearly kind of a problem. But that one is solved by Grafana, which is something that's really just to build pretty dashboards for Graphite. And once you install it, you will probably lose a few hours to it because it's so much fun to play with it, and it looks so good. And Grafana also supports InflexDB, which is the next generation time series database written in Go, because that's what you do nowadays. It has a company behind it that sells hosting, so let's hope they don't pull off FoundationDB. And it is used by Heroku, so it's not an obscure toy for nerds, but it is in production. It looks better. It's easier to manage its storage. You can tag values, which anyone appreciate would ever put the server names into their metrics names that you've seen in the slide before. Now you don't have to put any tag on a value and have clean names. It offers a SQL-like query language to those metrics, and a Graphite front-end, which means if you're running Graphite right now, you can point your tools just to InflexDB, and it should work, but it's computers, so I'm not sure. If you start on today, I would recommend to look into it first. If you run Graphite and are functionally unhappy only, then I would not abandon ship so quickly. It's not that big of a deal. So, collecting, how do we get the data into these databases? And there are basically two approaches. The one is that you aggregate externally. So something happens, and you send out a UDP packet to StatsD, or protocol buffers to RIMAN. StatsD is older, comes also from the Etsy ecosystem. Simple to use, simple to set up. RIMAN is by a super smart person, and it's configured in Clojure, so you probably have to be also super smart to use it. The good thing is it has no state. It's super simple to set up and to use. The bad thing is you have no direct introspection. So you need at least one more service to even see what metrics are coming out of your system. In the case of StatsD, you need even two, because StatsD does only aggregation, and then forwards it to graphite. With RIMAN, you get at least a kind of dashboard. So the second approach is that you aggregate your metrics within your application, and then deliver it to your metrics database. This approach has been polarized by CodaHale and his talk, metrics, metrics everywhere, which you totally should watch if you want to get into metrics. It's super interesting. It's super funny to watch. And this one gives you immediate insight into your application. You get some kind of dashboard out of your application, and this is useful both in development and in production just as well. Of course, you've got state. State is bad. State means bugs. But I personally prefer the second approach, because it's more practical. So the question is how will you do it in Python? So for StatsD, there's a gazillion Python clients, pick one. They work all the same. You instantiate a client with a URL, and you should packets around and don't look at return values, because they are not, it's UDP. Everything's going to be okay or not, because if your system is burning, UDP might not be the best way to message your state. So the only known working solution in AppMetrics to me is Scales. So it comes with a plethora of stats, but you have to set it up. These are the two I use most. The meter stat is for something that happens per second. So basically a derived counter, and a PMF stat is a timer. Nothing else. So how do you use it? For metering, you just call mark on it, and for timing, it is a context manager where you do something inside of it and you're done. Now, by doing this alone, you get a nice web dashboard out of the app. This is the metering. You've got already the average for the past minute, five minutes and 15 minutes. Even nicer is the thing you get out of your timing, because you get your person tiles for free, plus some more nice statistics. And all this data you also get as JSON, so you can collect it from, you can collect D or whatever. I personally use the graphite periodic pusher that comes with Scales. You just define the period, how often you should just send out the metrics, and you're done. We are done. You know how to collect metrics and how to store them. Now, we come to logging. In an ideal world, we wouldn't be logging, because you want to know about errors, which we now have error-centric, and you want to know the state of your system, which are metrics. So there are people like Armand Roenacker who just refuse to log anything. I personally cannot get away with that, simply because we needed some kind of bookkeeping, because the customer calls us, they always lied to us. They always stated they did not log into the server. They did not change that file. And we have, we need a way to double check what they are telling us. And that's usually not me. That's someone from support. And those people usually don't have the SSH keys to our servers. So this data should be somewhere searchable in a central place. So we are talking about centralized logging. And I can talk about centralized logging and not mention Splunk. And please, my, see there are more moneybags next to the name than on the other slides. It is for a reason, because this is enterprise software. And it's not just one web interface. It's a versatile platform. They have literally an app store. It works both on premise and in the cloud. It's great if you can afford it, but it is enterprise software. So the home page is full of PDF white papers. There's a lot of webinars for you to attend if you know that kind of things. So more down to earth, there's PaperTrail and Logly, which I have heard good and bad things about both. So it's a matter of taste. I'm sure you're going to be reasonably happy with any of them if you choose. And if you want to save your log files on foreign servers, which I personally don't. And that's why we are running Elk. Probably heard about it. It's currently the most popular stack. And it consists of Elasticsearch, LogStash, and Kibana. Let me just quickly show you how it works together. So we have servers that are generating log files. Those log files somehow get into LogStash, which parses them, adds meaning to it, and saves it into Elasticsearch, which is a database that is easily searchable and easily clusterable. And now the data is in there, you can view it using Kibana, which is a web interface to all these things. Yeah. And that's all. That's the Elk. There's a similar solution called GreyLog. It also uses Elasticsearch for storage and search. But Kibana is only a view on Elasticsearch, but GreyLog does more. Because, and I'm quoting here, Elasticsearch is not a log management system. So overall, it's a bit more integrated. They do more. But I'm personally not particularly fond of having a Mac vendor in my infrastructure. So you have to decide yourself. I haven't found a compelling reason to switch from Elk. But I'm sure there's someone. If you have any questions about Elk, Honza Kral is somewhere, probably in some pub here. And he works for Elastic, the company behind Elk, so he will be happy to answer all your questions. And he's also the maintainer of the client to Elasticsearch. So one more thing, Kibana is much more than just a web grab. They have a lot of nice things going on, like geo stuff and everything. So there's a lot of things to find. Now let's come to the finicky part. How do you get your data input? How do you produce it? So I'm going to say this should be the goal for you. A timestamp and something machine readable with as much useful context as possible. Because that makes configuration really simple. You literally tell LogStash there's a timestamp and JSON. LogStash will figure it out. Of course it's just one line, but I thought you might find it more readable in that size. So how do we get there? It's a matter of context and format. So you want to look out everything important and you want to format it in a machine readable way. And if you try to achieve that with the standard tools, you might find like I did, it's rather tedious. So I wrote something on my own called StratLock. Does anyone know StratLock? Okay, let's change this. So StratLock is not a logging system. It's not a replacement for LogBook. It's not a replacement for StratLock. Instead it gives you a bound logger that wraps your logger. So if you're going to ask me, does StratLock work with X, the answer is yes. Now it also gives you a context which you can bind key value pairs to it. And once you decide to log this event out, this context you saved before is combined with the new key value pairs to one event dictionary. And this event dictionary is run through a chain of processors which are just callables. A function that gets a dictionary in returns a dictionary. Nothing else. The last processor, the return value of it is passed into the original logger. So if you're using the standard logging from the standard library, you would return a string. For example, a JSON string or whatever format you want and return XML for all I care. StratLock comes with JSON and key value pairs for a matter. So the thing about the processor is it's really cool because it's really just callables. You can do whatever you want. You can plug data out of it. You can collect metrics from your log entries. You can report errors to sentry from them and enraged with the context you've collected. This is really nice. So this handles both context and format. And let me give you a few examples because it's a bit abstract. So simple case. You get a logger which is everything is pretty much configurable. And now you can log using key value pairs. You can stop writing pros. And if you're any like me, I hated writing pros before, but what's even worse is parsing pros. So this output is completely configurable. This is the default which is just key value pairs which is human readable in development. So I find this is already a huge progress over Stunner library, but you can do more. So this is incremental data binding. So again, you get your logger. And now you can just start binding key value pairs to the logger. And this lock object is a new object every single time. This is immutable data. And we have no mutable site at all. Ask your Haskell friends. It's a great property to have. So in the end, everything you bound to the logger gets logged along with the event. Again, output is configurable. And please notice that you don't care at all how the data is represented within your business code. That's something that you care about somewhere else in a processor or in your logging module, but not in your business code. You just bind key values and just log them out. So now maybe even more practical, how do you use it in practice? So this is a pyramid view, a very simple one, but it would probably work the same with any other. So at the beginning, you bind a request object to your logger, and then you lock something out. And how do you do something useful with that object? So you write a processor that extracts the data. So you try to remove the request from the event. If there's one, so you've removed it, but now you add some data from the request, like the IP address of your client or the user ID of the user. And you return the new dictionary. And this is what you get out of it. In case you have a JSON format installed. Again, you did not care about what you want to log out in your view. That's something that you decide elsewhere. That's all I'm going to say to StratLog. If you have any questions, just talk to me. I'm pretty proud of that one. Now to something slightly sadder. Let's talk about standard library slogging. I'm going to say this is all you should do and ignore all the rest. Just log to standard out and handle the logs outside. Because UNIX had over 40 years to develop solid logging tools and there's absolutely no need for us Python people to reinvent the wheel like date stamping or leuprotation. We are doing it worse. Stop doing it. Just go to standard out. Also, I've heard that it's not that much fun to use. But you be the judge. So now we have structured data on standard out. What do you do next? Send it into a file. Or send it to syslog or any other queue like Kafka. Pipe it into a logging agent like log-forwarder. You can do whatever you want. It's just a pipe. I'm personally a bit paranoid. Because I don't log a lot. But what I log is important to me. So I don't want to lose any log entry. No network in this world is as reliable as X4. I save everything in a file. This file is rotated for 48 hours and those entries are deleted. I ship it from there, from this file. So while I do not want to have to use grep, I still want to retain the reliability of greping through files on a file system. So let me put it all together. So I use struct log to bind data to log things out. Struct log makes it a JSON string which goes into logging and log in and sends it to standard out. Now, I use run it to run my processes. It doesn't really matter what you're using. But run it comes with a demand that takes standard out, add a timestamp to it and write it to a file. Now my log entry is safe. This file is watched by log-forwarder formerly known as lumberjack. And it sends to log-parsed. Log-parsed it, sends it into elastic search. Logging is solved. So, yeah, we are done here. Let's get some pinchos. We are not. We have three nice components. We forgot about the pragmatic part. How do you put those three things together without making it gross? Because this is gross. You can barely see the logic hidden in the jungle of reporting, measuring, counting and whatnot. So I want it to look like this, which is much nicer. Something happens, I tell the log system about it and I'm done. Of course, that's not always possible. But I would really try to hard. I would really like to try hard to get somewhere there. So with errors, it's pretty easy, I dare say. So either use some handler that comes with a sentry, either just logging or if you're using running Django, using the Django app there, using, or just track log. That's what I do when I use a pyramid. I just plug my errors out of the logging stream and I can drop entries if something is not interesting. And in that app, there's usually also a way to define error views. And this is really, really cool because, again, pyramid, you get the exception and the request object. Now, you serve back the error ID, which is served from sentry. So now, when your customer calls you complaining about errors, they can exactly tell you the error ID and you can look the error ID up and you have the exception that the customer saw. And this is so great that we've seen something that's even rarer than a white rhino, which is a happy armored runnaker. So although I have to say, since I made a slide, he joined sentry, so take it with a grain of salt, but still. So onto metrics. Most metrics can be observed from the outside and outside can mean outside of your views, outside of your app, outside of your server even. So maybe let's have a look at whiskey containers. The two major ones have both knobs that will help you with that. So G unicorn offers stat C integration right there. So you add one command line option and you have average request times in your stat C and in your graphite. You don't have to change your code at all. Micro whiskey, as usual, goes far, far, far further. They, of course, have stat C too. They have direct carbon, aka graphite support. They have a whole metric subsystem, including nightmare inducing things like SNMP. So you get your stuff done with that. And with this, you get a big picture of the state of your application without even touching your apps. So go for it. Then you can write middleware. Middleware is no dark magic. Again, pyramid. This is a tween, which is a very awkward contraction of between. And this is called on every request that comes in. So you have the request object. In this case, we just measure the time. But you can, of course, look at the data within the request object and start splitting up your data depending on the view or some argument that you're passing into your view. Probably don't have to because there are things like pyramid stat C that already do that for you, but you always have the possibility to do things from within your app but outside of your actual logic. Then, of course, you can extract data from logs. Because if you lock something out, you shouldn't have to also count it or measure it. So LogStash will do that for you. It supports all major metrics backends. The drawback is that you have to change the configuration of LogStash, which may or may not be a problem for you. It's not a really problem to me, but it adds friction, which I do not like, so I do not do that. I don't want to annoy those people who are responsible for that to fix it for me because I've added a new metric. Of course, you can do the extract log. It's what I do. You can just count events by their names and you already have something useful. Okay. Finally, you can also leverage monitoring, which is even for the outside. Any monitoring system has some support for metrics numbers. In the worst case, you just measure the time it takes to execute a check and save it. So you get a really external view of your data, external view of the behavior of your apps, which is not very precise, of course, but sometimes it's useful to see how your system feels from outside boundary and not from within your availability zone or your computing center. Okay. So what's left? What do you have to yourself? So if you want to measure code path, you probably have to add some code to your business logic. For example, database queries. Or if you have certain major use cases, like a view that sometimes uses only cache data and sometimes hits a database, it's not very useful to average those two numbers. Not to say it's completely useless. So you may want to split that up. And of course, gauges, if you want to expose numbers from within your application, you will probably have to touch your application in some way. And now we are really done. So what did you learn? Proper error logging is important. Sentry is awesome. Metrics are important. Influx TV is probably the future. Graphite is the present. Use whatever you want from those two. Centralized logging saves you a lot of pain. And maybe you need it. Elk will have your back. Strucklock will help you to get your data there. And now you know how to use all of them with Python without the gross code duplication. So I hope everyone learned something. So go forth and measure. Study the talk page. Follow me on Twitter and tell your German-speaking friends to get their domains from Vario Media. Thank you. And I'm sorry. I'm not taking any questions because whenever I did, I completely misunderstood the question and said something very stupid. So if you have any questions, I will be outside. I'm here through Sunday. I will be at the conference. I will be at lunch. Just check me out. I'm happy to answer any questions. Thank you.
|
Hynek Schlawack - Beyond grep: Practical Logging and Metrics Knowing that your application is up and running is great. However in order to make informed decisions about the future, you also need to know in what state your application currently is and how its state is developing over time. This talk combines two topics that are usually discussed separately. However I do believe that they have a lot of overlap and ultimately a similar goal: giving you vital insights about your system in production. We'll have a look at their commonalities, differences, popular tools, and how to apply everything in your own systems while avoiding some common pitfalls.
|
10.5446/20127 (DOI)
|
So, hello everyone and good morning. So I'm here to talk about what's beyond the basics with Elasticsearch. I work for Elastic, the company behind it, so we've seen a lot of use cases and some of them actually surprised us and definitely surprised many people that are familiar with Elasticsearch as sort of the full-text search solution. But before we get beyond the basics, we first need to know what the basics are. So super quickly, for us, that's where we come from. We are a search product. It's an open-source search product. And search is not a new thing. Like it's been around for a while, for a long while. And the basic theory, the really down-to-earth basics haven't changed that much since those times. We still use the same data structures. We still use the same data structure that you find in any book at the end, the index, especially the inverted index, which looks something like this. It looks the same in a book as it does in a computer. It is a list of words that actually exist somewhere in our dataset. Notice that they're sorted. And for each of these words, we have, again, sorted, the list of documents or files or pages when it's a book, where actually these words exist. And we have some additional information stored there, too. For example, how many files does actually contain the word Python? Or how many times is it present in file one? And on what positions? And stuff like that. Those information, those statistics will be very important for us as we go on through the talks. So this is the data structure that we use. So how does search work, then? Well, it's super simple. If we're looking for Python and Django, it's the same search that you would do if you were looking for those things in a book. You locate the line mentioning Django and the line mentioning Python. You can do that effectively, both as a computer and as a person, because, again, it's sorted. And then you just walk the list, and if you find a file or a document that is actually present in both lists, that's your result. Naturally, if you want to do an OR search instead of AND, you just take everything from both lists. And that's not enough, because this will give you the information what matches, but it doesn't give you the most important thing for us. And that is the information how well does it match. What is the difference between the Django book that talks specifically about Python and Django? And the biography of Django Reinhardt when it mentions in one passage that he had an encounter with Python, the snake. Obviously there is a big difference between those two books. And the difference is in relevancy. It is a numerical value, a score, essentially, saying you how well does a given document match a given query? And a lot of research has gone into how best to calculate the score. And again, it hasn't changed that much since the beginning. At the core of it, there is still the TFIDF formula. Those are fancy words, fancy shortcuts. It's a term frequency and inverse document frequency. It essentially represents how rare a word we are looking for, and how many times have we found it in the document. So this essentially represents that if you found the word the in a document, that doesn't really mean much. Like every document in the world, if we are talking in English, will have the word the. That's not a good information. Because the IDF, the inverse document frequency, that's the part. It will tell you that this is not a specific word. It's almost in every document. If you, however, find the word framework or something like that, that is fairly specific. So that's the IDF part. And the TF part is just how many times did you find it there? If it's only mentioned once in a book, doesn't mean much. But if it's there 100 times, that probably means more. And we can keep building on top of that. So for Lucine, for example, adds another factor to it, which is a normalization for the length of the field. That's essentially the equivalent of saying that, yeah, there is a fish somewhere in the ocean. Probably true. Not really that relevant or surprising. But if you have a bucket with water and you say there is a fish in it, that is much more actionable information. So that's the second part of it, the normalization for the field length. If you find something in a super big field, okay, if you find it in a much shorter field, for example, title compared to body, that probably means much more. So already we have a formula that's baked into Lucine, it's baked into Elasticsearch that does very well for text and for search. But sometimes even that is not enough. For example, you're not dealing with text, but with numerical information. Or you have some additional information that Elasticsearch is not aware of. For example, you have the quality of a document. You have some user-contributed value, or even somebody paid you to promote this piece of content or something. Or you want to penalize or favor things based on a distance, let's say, from a geolocation or distance from some numerical range. So how do you do that? We have a few ways of expressing that. And the best way to show it is on an example. So this is a standard query for Elasticsearch. And it's using the function score query type. The function score query type takes the regular query. So normally we are looking for a hotel and we are looking for a hotel that's called the Grand Hotel. So far so good. And then we want that hotel to have a balcony. We want our balcony in our room. But we don't want to just filter just the hotels that have balconies. Because then we would be robbing ourselves of the opportunity to discover something else. But if a hotel has a balcony, we want to favorize it. We will just add two to the score. So all the hotels with balconies will be towards the top. Then we want the hotel to be in central London within one kilometer of the center. If it's within one kilometer, it's a perfect match. The further away from it that it gets, the score decreases. It will still match. But the score will be smaller. Again that means that the hotel that perfectly matches our criteria will be at the top. But if we have a super good match outside, it will still show up. And then we also have the popularity. How have people been happy with the hotel? And let's take that into account. So we have a special thing called field value factor, which is essentially just telling us at Search. There is a numerical value in there that determines the quality. Put it into the score. And finally, we add some random numbers. And this is actually taken from a real life example, because people use this to mix things up a little bit, to give users a chance to discover something new, something they wouldn't otherwise see. So all of these things together will make sure that you find your perfect hotel. We're not limiting your choices. We are not just because you say that you want a balcony. We will still show you the hotel that is almost perfect for you, except for the balcony part. We are also not just sorting by popularity, so that's something that's really not that good a match, but is really popular would be at the top. We're just taking all these factors and combining them together. So this is one of the main ways what we can do with the score and how we can use it in a more advanced way. Just take all the factors that go into the perfect result and just combine them. You're not limited to just picking up one and sorting by it. You can combine them all together, and then it's just a matter of figuring out what these numbers are supposed to be to one and what will actually give your application the best results. Some people actually use machine learning techniques to figure out the best ones. They have a training set and everything. It's not that hard because you have only limited number of options, and typically those are fed numerical values. So if you know what a good match would be, you can actually train the perfect query for you. So this is if you're doing search when you already know what you're looking for. But sometimes it's the other way around. Sometimes it is you don't have the document, but you have the query and you want to find the document. So imagine that you want to do something like alerting or classification. For example, you're indexing documents, you're indexing stock prices, and you want to be alerted whenever a stock price rises above a certain value. Sure you could keep running a query in a continuous loop and see if there is something new. But what we can do instead with the percolator feature of Elastic Surge is to actually index that query into Elastic Surge. And then we just show it a document, and it will tell us all the queries that matched. And that is very powerful, especially because it can use all the features of Elastic Surge. So that's the alerting use case, sort of the stored search functionality. If you supply your users with a search functionality and you want them to be able to store the search and then be alerted whenever there is a new piece of content that actually matches their search. With percolator you get it essentially for free. You just index their query and whenever there is a new piece of content you just run it by the percolator and it will tell you, hey, you should probably send an email to that user that was here the other day. He was really interested in that. That's the sort of stored search. You can also use it to do a live search. So if you've ever been on a website, you did some searching, you were looking through the results and suddenly there was a pop-up that there are five new documents that match your query since you've been looking at it. Again, easy. Once you execute a query, you also store it as a percolator. And then whenever there is a new piece of content during that time, you can just push it to the browser to say, hey, there are new results, more recent. So again, something that's otherwise fairly hard to do or would require some busy loop or something and you can do it this way. But we'll go a little bit further than that. We'll look at the classification use case. That is essentially if you use the percolation to enrich the data in your document. So imagine that you're trying to index events and all you have as location goes is a search set of coordinates and you want to find the address. This is something that's easy to do the other way around. If you have the address and you want to find all the events in that location, you just do a geoshape filter that you're looking something that falls within this shape, within a shape of the city of Warsaw. And that's a super simple search. Well with percolator, we can make it into a super simple reverse search. Let's say we get our hands on a data set with all the cities in Europe or in the world. It's not so much. We index the cities into an index so we don't have to construct the polygon every single time. We store it in the index called shapes under the type city. And then we create a query for each city. We register it with a name and then when a document comes along and its coordinates, the field location fall within that shape, we will know that it is actually happening in Warsaw, Poland. So something that is super simple to do one way but difficult to do the other. We can do with a percolation just essentially using brute force but in a smart way and outsourcing the brute force to elastic search. We can do it very effectively and in a distributed fashion. So that's geoclassification. Another thing that's easy to search for but not that easy to do the other way around usually is language classification. Generally any language has a few words that are super specific to that language. They don't exist in any other. These are some of the examples. This is essentially just a test how many Polish people there in the audience. And the assumption here is that if we look for these specific words and we find at least four, because four is always a good number, because 42 would be too high, then the assumption is that this is actually a document that contains Polish language. And sure, it's a simplification, it's a heuristic but it actually works fairly well. It just depends on the quality of your words. And are super good for Polish that is. So again, and if you have a set of words for each language, you can just start a collection of queries like this. And then when a document comes along with a description of an event, with a geolocation and a description, you can immediately get back the classifiers. You can get back the location in actually a human readable format that it's actually Warsaw, note that it's 473 minus 74.1, which is by the way not Warsaw, but whatever. You also get the language back that it's in Polish. You can use a similar classifier to determine the topic. Like if you have within keywords something like programming and Python and Django, it's fairly accurate assessment to say that the conference probably is something about Python. So this is how we can use percolation to enrich our data and sort of to determine something that otherwise would be hard to do. Another use case for this is imagine that you have a blog, a CMS, and you have a category defined as a search. That's super easy to do one way, but then if you have a blog post and you want to see in which categories this blog post is, that's the harder part. Again with percolation with something like this, it's super easy to do when you can actually tag the blog post with the categories as they come in. Then you can do obviously a little bit more with a percolator. You can attach metadata to the percolators and you can filter on the metadata. You can aggregate them. So as the response, you will not only get the percolators that matched, but also let's say their distribution across categories. You can even use them to highlight something. So you can search for some words in your documents and then just highlight the fragments that actually contain those and store them separately in the document for easy presentation, et cetera, et cetera. You can get the top 10 hottest categories for this piece of content or something like that. But those are if we're working with individual documents. We can also look at more documents at the same time. So this is the traditional search interface. You're just looking something and you get back the top 10 links. What we also have here is something that's called faceted search. This part, the search part is really good when you know what you're looking for. This part shows you what is actually in your data. So you can immediately see the distribution. You can see that if you're looking for something related to Django, the most results are in Python and some in JavaScript. So it allows you to discover data. Some people have taken it even further and we have allowed that with aggregations with multi-dimensional aggregations that you can aggregate over multiple dimensions at the same time. But that is still boring. That is still just counting things. And that's not really interesting. Any database can do that. What we need is we need to use the data that we have, the statistics. So to do that, let's look how we would do recommendations using Elasticsearch. This is our data set. We have a document for users and then for each user, we have a list of artists, of musicians that they like. And we want to do recommendation. Something that I like these things, like what should I listen to next? So we have two, in this case, we have two users, they have artists being common and there are three other artists. So the naive way to do it is to just aggregate. Just ask for the most common thing that they have in common. So give me all the users that like the same things that I do and then give me the most popular artists in that group without the ones that I already know. That way I will get the most popular artists but not necessarily the relevant. It's like asking you, like, what is the most common website that you go to? Probably Google. Not interesting. Because everybody goes to Google. But if I ask the people in this room and I think about it, what is the more specific part for this group compared to if I asked somewhere on the street, it will be something like GitHub. You probably all go to GitHub. Nobody in the outside world goes there. Nobody even knows that it exists. That is relevant. That would be a good recommendation. And we can do that with Elasticsearch. We have all the information. We have the statistics about how rare a word that is and what is the distribution across the populace. So we can ask for the significant terms. It will use all the score, compare it to the background, and then the results will look something similar. This part is important. Because what I would expect is all those dots to be on the diagonal line, because that's what would happen if I had a random sample. It moves away from the central line. The more specific it is. And that is how we can do relevant recommendations. Because we see that this dot here, it is obviously much more common in this group than in the general populace that would be here. So it has moved greatly. And because we have all the information, because we have analyzed the data, because we are the search people, we understand the text, we understand the frequencies, and we can use it, we can actually produce something like that. There are obviously some caveats. For example, if I like a very popular band, like One Direction, then it will skew my results, because everybody likes One Direction, right? So I need a way to combat this, because otherwise I would just get completely irrelevant recommendations. And again, we are the search guys. We understand data. We understand documents. So we can find and sample just the users that are most similar to me. And we have all the tools already at our disposal. Remember TF, IDF, normalization and everything. TF, the people who like the more things that I like, the better they match me. IDF, the people who like the rarer things that I like, put them to the top. And then just take 500 of those best results and only drive the recommendations based on that group. It will make it both faster and more relevant. It will allow you to discard all the irrelevant connections that you might find and only focus on the meaningful connections, on the things that are relevant for your group, in this case the group of people who like the same things that you like. It will provide you with a recommendation. So just by applying the concepts that we have learned from search into other things like aggregations and everything, we can get much more out of it. Another example would be if you have Wikipedia articles when the labels and links are the words and you apply the same concept, you get a meaningful connection between different concepts. If you try to do it based on popularity, it would always be linked through something like yes, that person and that person, yeah, they're both people. Okay. Not exciting. But if you apply this principle, you get something more out of it. So if you combine aggregation and relevancy, all the statistics that we can do, that is actually how we as humans look at the world. If I ask you what is the most common website that you go to, you'll probably not say Google because you know that's not interesting. We as humans have been trained from the very beginning to recognize patterns and to spot anomalies at the very same time. And this concept can be used for other things as well. For example, if you use the same principle, the significant terms aggregation, and per time period, so you split your data into time period and you ask what is significant for that period, how do you call that feature? Well, it's a very common feature that we now see. It's what's trending. That's just it because it's more specific. It's not more popular than in any other area, not necessarily. But it is more specific for this one time period, for the current time period, let's say, compared to yesterday, compared to the general background. So again, once you're doing these aggregations, there's again one single caveat that can happen is that you can have too many options, too many buckets, too many things to calculate. And if that happens, so imagine that you're looking for a combination of actors that star together very often. So I'm looking for the top ten actors, and then for each of those, I'm looking for a set of top ten actors that act with them, that they appear together. If I just ran this, what will happen in the background is I will essentially get a matrix of all the actors and all the actors, and it would be huge. It wouldn't fit into memory. It would probably blow up my cluster. Actually, LSE search would probably refuse to run this query because it would say, hey, I would need too much memory. This is just not going to fly. So what you can do is you can just say, just do it breadth first. Just first get the list of the top ten actors and greatly limit the matrix that you will need to calculate. And then go ahead. So it will be a little slower. It will have to run through the data essentially twice, but it will actually finish, and it will still finish in quite a reasonable time. So that's just how to find the common caveat that people get into when they start exploring the aggregations, especially with the multidimensional. So just to wrap things up, because we are approaching the end in questions, the lesson here is that information is power. We have a lot of information about your data. We have all the statistics, all the distributions of the individual words. And if you understand this and if you can map your data to this problem, you can get a lot more out of Elasticsearch than just finding a good hotel in London or the conference events in Warsaw. So that's it for me, and if you have any questions, I'm here to answer them. There's a question in the back. Questions? That's a long question. Thanks. You show the example how to search people like the 500 more like you. Can you do that more like people that have 90% of being like me instead of a fixed number? Because fixed numbers you have to find and tune in. Of course, you can do that by a simple query. Because aggregations are always run on the results of a query. So we can very easily remember the example that I gave with the language classification when I was looking for at least four words. I could do the same. I could say give me only the users that have at least 70% or 90% or nine. I can use both relative and absolute numbers of the same artist that I like and use those as the basis for the aggregation. So yes, absolutely. And it would actually be much simpler. You wouldn't even need the sampler aggregation. Thanks. Any other questions? Is anyone still awake? Okay, I'll take that as is. So a question? A question going once? Going twice? Sold? Are there any performance implications of running, say, hundreds of percolators? Of course. But you can scale way beyond hundreds. I've seen people doing millions of percolations and it still works. It scales very well with the distributed nature of Elasticsearch. Essentially the only resource that the percolation consumes is CPU. So add more CPU, either to a box or add more boxes and it will scale fairly linearly. So and also just the more boxes and more CPU you will have, the faster it will get. You don't need anything else. You don't need much memory. You don't need faster disks. You only need the CPU. So it's very easy and fairly cheap to scale. To give you an idea, I think that if you want to run hundreds and thousands or millions of percolations, you will need like five reasonable boxes or something like that and you will get responses within milliseconds. So it actually does scale very well. Another question? Could you give us some examples of the customers you mentioned that you had like cases that were really impressive for you and you didn't expect those use cases? Could you give us some examples of the use cases from the customers that you mentioned that you didn't expect them? So some of what we didn't expect was the percolator example. There are some people running big clusters of Elasticsearch and they don't store any data in it. They have a cluster of 15, 20 machines without storing any data. That is a weird experience for essentially a data store. So that's definitely one of them. We also always run into these issues where we have a feature. We recommend people to use it and then people listen to our advice and we find out that we might have underestimated the people in the wild. For example, we introduced the idea of index aliases. That you can have an alias for index essentially like a simling or something. So you can sort of decouple the design of your indices from what the application sees. So you could have like an alias per user but all the users can live together in one big index and the alias will just point to that index and a filter. And that works very well unless until we encountered a user that had millions of users and suddenly we had millions of aliases and we didn't thought that that would ever happen. So as with anything else with computer engineering like assumptions, assumptions, assumptions. So we encountered something like that. We had to go back and fix it and rework the aliases. So these are the two most notable examples where we got really surprised by how our users used our product that we really didn't foresee. And it's good because we always learn something new and it allows us to sort of reorient ourselves better to what the users actually need. Okay? Any last questions? So hello. I have a question regarding reverse queries for language classification. So basically elastic search supports the Ngram indices. So could you use those actually for classification of languages? So Ngrams have the problem that they have a very wide spread. So they might give you some correlation with the language but they will definitely not be precise. So just to explain Ngrams essentially if I split a word into all the tuples of letters, for example with things I would have T-H-A, H-A-N, A-N-K and then I would essentially query for these triplets. And it will obviously have a correlation but it will by no way be decisive enough. Especially for something like language classification where you're really interested in the probability. Ngrams are very good for as an addition to something else because of their nature because they always match something. That's why you typically don't want to use them alone. But they're fine if you have some more optimistic methods like exact matching and then the regular like fuzzy matching and everything and then you just throw Ngrams into the mix to sort of boost the signal if it matches and sort of to catch some things if nothing else matches. So I definitely wouldn't use Ngrams for language classification and I typically only use them with a combination of other query types and other analysis process. Make sense? Okay. So I think that we're running out of time. So thank you very much. If you have more questions, I'll be outside.
|
Honza Král - Beyond the basics with Elasticsearch Elasticsearch has many use cases, some of them fairly obvious and widely used, like plain searching through documents or analytics. In this talk I would like to go through some of the more advanced scenarios we have seen in the wild. Some examples of what we will cover: Trend detection - how you can use the aggregation framework to go beyond simple "counting" and make use of the full-text properties of Elasticsearch. Percolator - percolator is reversed search and many people use it as such to drive alerts or "stored search" functionality for their website, let's look at how we can use it to detect languages, geo locations or drive live search. If we end up with some time to spare we can explore some other ideas about how we can utilize the features of a search engine to drive non- trivial data analysis including Geo-enabled search with relevancy.
|
10.5446/20125 (DOI)
|
Yeah, thank you. Thank you, everybody. So basically, this talk is about multi processing, multi trading, concurrency and palism. So this talk is basically for trading and processing. So basically, while doing a programming using a threads and processing, we do mistakes. I had to use rates, I had to use processes in different different programming language, different different compilers, different different conditions and methods to execute the thread processes. So in this talk, I will start from basics, like how to start a thread, how to process and basic thing from that, we will slowly go to advanced topics and we will cover some internals of C Python as well. So that bit about tell me about myself, bit about myself. So basically, I started my career with a swing or there I was a software engineer and currently I'm working as a senior platform software engineer at substance. I have experience of development of large scale scale, large scale fault or else and mission critical systems and some material as well. And I'm passionate about a web backend and infrastructure and I'm a Pythonist and I'm a go for. So basically, what is parallelism and what is concurrency? So usually people do a mistake in these terms. So many people think that parallelism and concurrency is same but reality, they both are different. They both are not same. So basically parallelism is like to add more processor, more code, CPU codes to make computation faster. It is to add workers to add more codes to your task. So suppose you are executing something, I mean executing some task or doing some computation or something, then you are adding more cores, more processors, more workers to make it faster. Another term, it's a concurrency. Concurrency is related with parallelism. Concurrency is to permit a multiple task to proceed without waiting for each other. So for example, I have four processor like four CPU codes or four processor with me. So how do you utilize them perfectly? How to, I mean, parallelize the task in between? It's like two. So for one example, I have four CPU codes and I want to divide all the workload with all the CPUs. So it's about concurrency. So concurrency deals with that things. So that's a parallelism. One example, there is eight boxes on left hand side. And there is a two ways. So there are two guys, two workers whose task is to take one box from left hand side and put it on right hand side. So here two ways are there and two workers are there. So they will do this work. I mean, one by one. I mean, first of all, it will, in a parallel manner. So first of all, that both guy will go or take that box. They will go on wave one, both the ways and they will put it on right. Now parallelism is to add one more, I mean, one more worker on both sides. So like I have added two workers. So total worker is four workers. So it's a parallelism. So now here in this example, I have two ways. So like even though I have four workers, then also I cannot get speed ups here. The reason is that there are two ways. So at a time, one worker will work. He can, I mean, four worker cannot work at a time. So here concurrency comes in. Concurrency is like to create more ways to execute it. Yeah. So that's the term palism and concurrency is about. Another term is multi-trading and multi-processing. So it's a very simple term. It's like to see, I mean, your operating systems, ability to, to pal to, to run your multiple tasks in parallel manner using a threads or processes. It's a multi-trading and multi-processing. So very simple term. It's, it's to, and the threads are nothing, but they are part of processes. Some of you can call it a lightweight processes and processes are, it describes a program which we are executing. So multi-trading and multi-processing is like that. There are too many puppies and they are eating everything. They are, they are dealing with their, their whole stuff. So it's a parallelism. They're parallel, they are eating the stuff. Okay. So let's have a demo how to start a thread. In a Python, it is very easy to start a new thread. Their API is really, really easy. So basically, thread, trading is a module which deals with the threads in the, in a Python and thread is basically a class which deals with the normal, I mean threads. So first of all, let's create one function which will do a simple, very simple stuff. It will do a print hello world. Oh, sorry. So it will do very simple stuff. Now what I will do is I will creating a new thread. Check. So in a target, we will specify a function name and in arguments, we will specify argument. It's basically blank because there is no argument. T dot start, it will start a thread, T dot join. So it's a, it's a method which will wait until your thread execution stops. Now I'm going to run this. So, so it's thread started. It printed it. And that's it. So that was a very basic example of threads. So in a Python, it's very easy. So basically, this is a Python trading. It's a module. It's a high level module under the, under, I mean, it's a base model is a thread module, which we should not use it. Basically, it's we should always use a trading module because the thread is inherited in a trading module. In Python 3, a trade module is renamed to underscore thread. So that's why it's like, I mean, we should not use it. I mean, another thing, another model added into Python 3 is dummy trading module, which provides, so whenever, so if there is no three underscore trade module available, then it will raise the import error. That's why in Python 3, there is a dummy trading module, which we can use at the depth. Yeah. So basically Python, Python threads are system threads. So whenever we start, we have, we start the thread at the time it will open, it will request the operating system to start a thread. And operating system will manage everything. So for example, I just called a start method. So it will call the system, I mean, operating systems API to start a new thread. And in a Linux machine, it's a POSIX thread. So basically POSIX is nothing but in older days, there was like many, there was many hardware, hardware manufacturers, they used to provide their own APIs and everything and it was creating some trouble for developers because they had to write different, different APIs for different, different vendors. That's why one new standard defined for that and it's a POSIX. So in a Python, there is a P thread libraries there. So it is being used by Python threads. In a Windows, I mean, API is provided by Windows, which is implemented in Python. And all the scheduling is, will be managed by operating system. So how to switch a thread, how to how to schedule it, everything will be managed by operating system. Python will not deal with it. Yeah. Okay. So, so, so that's the one thing. Let's execute. So let me show you some other example of threads. So basically I'm going to use a P thread. So basically I'm like looping until n is not greater than zero. So now I will try to. So now I specify a target function name to Cal and RGS2. Okay. And I'm defining another thread because I'm going to start two threads at a time. Okay. Now I will start a thread two, a thread one first. So after this method, this, I mean, it will start starting thread two. I will wait for thread one to complete, wait for thread two to complete. Okay. Okay. Okay. Let's, let's keep this for now. So, so basically it's like whenever I will execute two threads at a time. For example, one method, it's taking two seconds to complete it. And what I'm doing is I'm executing two, I mean, two threads using a, using a same method. And actually it should, I mean, running parallel mode. But in a Python, when we are starting a more than one thread, they will not run in a parallel. So basically in a, in a Python parallel, parallel running of threads is forbidden. I mean, of number of, any number of threads you are opening, you cannot run in, I mean, parallel. So it's a, it's a lock at the processor level. So basically if you have a two processor, then it will be parallel. But if you have a single processor with a multiple, multiple cores, then it will, I mean, then one thread will run at a time. So that's a GIL. It's a global interpreted lock. So it's a, it's a global interpreted lock. It's implemented in Python. So whenever one thread is running at the time, it will take a lock and it will not allow any other thread to execute until that thread execution completes. So that's the, that's the one thing. It's a GIL. It is suitable for IO bound operations. The reason is that whenever there is IO operation or something running onto it, that the time it will release its lock and it will, I mean, it will give its control to another thread and that will start, I mean, that will, another thread will continue, or continue the execution. So that's why it's a bit better for IO bound application because it, it also releases GIL, global interpreted lock on read, write, send, receive methods. Like, yeah. Then it is bad for CPU bound applications because the reason is that like it may possible one thread takes too much CPU and it will never give a chance to another, I mean, another application. But there is a handling for CPU bound applications as well, but although it is not suitable to use it. So basically global interpreter log is runs like this. It will run whenever there is IO at the time GIL will release its lock, then another thread will run. And this is how the global interpreter works in a Python. So one more thing is there in Python, so to, for, for a simple case, if there is IO bound or something IO application is there at the time, it will release a lock. But what for application or function or calculation which is taking too much time, it is holding a CPU at the time. Python has handling for this kind of application. So what Python does is like Python, Python has a tick event for CPU for most all the threads. So what it will do every 10ms, it will, it will send, it will, okay, it will, what it will do, it will, I mean, release a lock, it will send operating system signal to unrelease that, I mean, thread and so and again acquire. So operating system will reschedule it by itself. So it's like to send a signal to opening system to, to release a lock of that specific thread even though it is running. And so, so, so, so operating system will reschedule it to, to this thread, give a acquire or and you can change it using a sys.set, I mean, set interval. Thread pools are to whenever you want to restrict a number of threads you want to open and you have too many, I mean, too many tasks to do and you don't want to open more than, I mean, more than allowed threads. So at that time you can use a thread. Threads are nothing but it's kind of a queue where you are adding your task and it will assign it to, assign it to, I mean, open threads. So, so basically thread pool is like when you are starting you have to tell it that I want to start with just force if you and if you will then it will queue that all the inserted arguments and everything and it will put it, assign them to all the, all the started threads. So it's like pretty good and there are some methods. So I cannot, I'm sorry, I can, because of time constraint I cannot process demo but it is similar to processes, threads. It is very similar to threads. It is a, it is a module to interact with the processes to start, stop and to do various operations on the processes. So Python will create a system level processes. So whenever we'll start a process at the time it will create a child process under a Python process. Yeah, and good news is that like it, it will bypass a global interpreted log which is there in a threads. So if you are starting a module that one processes then it will, that's your calculation, your program will run in a parallel in a threads. It will not happen. If you are starting a fourth thread then at a time one thread will run, but in the processes it will run, I mean, when it works on both Linux and Windows. Yeah, like a thread pool you can start a process pool as well. If you want to restrict the number of processes like at a time that, that, that should be four processes running then you can do it and task will be distributed across them. Yeah, there can be this kind of situation as well, as well like many, many, many processes can interface to variables and different, different memory accesses to different processes. So how to deal with that? So it's a basically deadlock kind of situation where you have a resource to more than one resources. Resources can be anything or network resource or files or any kind of resource can be, can be. So like a thread a bond to bond resource one and, and thread B wants resource one and they both allocated, I mean, resource accordingly. Also there can be one more situation where like this can occur. It's, it's like thread one wants object one and three two also wants object, same objects at the time situation can occur. There are same offer, same offer for this kind of situation and it was invented by a Dutch computer scientist. Sema four can be of three types. It's a binary semaphore counter semaphore and mutic semaphore in a Python binary and mutic semaphore are same and in a counter is also provided. So let me show you. Okay. So basically it's a lock and a, a, a trant lock. So locks and the trant lock are two, two, if you are executing something at the time, if you want to restrict some code that if one process is executing it, then no, but no, no, no, I mean, no, no other process should execute it in a parallel at the time. You can use a locks and a lock. A lock and lock is different in the terms of execution. So a lock can be used in a recursion. Whenever you are doing a recursion at the time, you should use a lock because it is for that and locks are for, I mean, normal code. So this is a code. So there is a part one and part two. I can acquire the lock and I can release the lock whenever I want. So that, that a code which relies in between a lock acquire and rock release will run. I mean, so whenever that part is running at the time, no other process will run. It will wait until that, that code execution stops in some process. So, so it is for synchronization between processes. Yeah. So there is a library for semaphore. Semaphore is another way to deal with these kind of situations. Semaphore is to, so, so, so semaphore is we can define like a number, a number like what are the maximum count? Like if I will define some number like nine, then release and acquire should be according to it. So, so it is like if I am acquiring a nine locks and so, so, so whenever I will acquire a nine lock, so there will be a current value of semaphore. It will be a zero because the reason is that you have to release that lock. So whenever someone come, I mean, some process will come to execute it at the time. It have to wait until some release will, I mean, lock release will come. So it is for, it is for semaphore. This kind of semaphore is suitable when you want to have some limit, some network limit or something for example, you want to send, I mean, some very less number of activity requests. You want to send 10 requests at a time. At the time, you can set number accordingly and you can use it. Okay. Bounded semaphore is like whenever we reach to zero at the time, instead of waiting for waiting, it will raise a value error. So it is basically suited for same application, but it is of different type. Yeah. There is one more thing is that events. So events are basically it will, so basically there are many conditions occurs in programming where we want to wait until some condition. So for an example, I want to say that whenever I will get some flag from network or activity request response, then only all the process should start or it should start its execution or start, I mean, its processing until that it should pose or something. So in this kind of situation, what we can do is like, I mean, whenever we can, we can write event.wait in different, different where we want to wait, where we want to stop, then we can event.set. So whenever event is set, at the time, wait will not work, but when we will do event.clear at that time, all the wait will, I mean, so, so whenever that program will come to that point, it will wait until that until we will call event.set. So it will be a blocking until we will call event.clear. Timer is also there. Timer is to execute a function after some interval. So if you want to execute some like function after seconds, seconds or something then we can use it that function. And yeah, delay can be there. For example, I have said that I want to execute this after 30 seconds and it may possible it will execute after 31 seconds, 32 seconds. The reason is that it is using thread internally. So because of global interpreted log or something, it may possible, I mean, delay can be there in this function, in this pipes. Pipes are basically a data channels that can be used for inter-process communication. So it's a channel. So it basically returns two file descriptors, one for a write and one for read. So whatever you write to write, it will be catch by kernel and you can read that using read object. So basically, Python provides two types of pipes. It's OS.pipe and multiprocessing.pipe. OS.pipe is interface on top of Linux kernel. So whenever we request to open a pipe, it will open a pipe on top of, I mean, it will ask operating system to open a new pipe and it's just interface to start and stop like pipe and to deal with it. Pipe is one restriction. It's in a Linux. It has 64 kb of limit and it uses encoding and decoding while sending and receiving the data. And in a Linux, it is implemented using a like of 4.6, 4.6 standard. In Windows, it is implemented using a create pipe method API which is being provided and multiprocessing.pipe is socket implementation. Sockets are files of memory mapped in-memory file sockets and it is a full-double duplex. So it will also give you the read and write objects. But I mean, both has, I mean, on both way, you can do a communication. It uses a pickle to send the data. Pickle is nothing but it's a kind of comparison, not a comparison, but on an object sending, you can send object by a, I mean, pickling your data. Q is also there. Q implementation is threads. Python supports three kind of Qs, first in first out, last in first out, and priority Qs. And it is process and thread safe. Set state can be used. So if you, like, if you want to set some variable directly in between processes, then we can use a set state. Set state is basically, I mean, set state are to set some variable in between or some data, Python data structure in between. So, yeah. So in pipes, we can set some textual data or that kind of data only. It's a simple file like object. You can write anything and it will be received on the other end. But if you want to use a data structure in between inter processes, then you can use set memory. So, yeah, Python provides as them as well. And they, all the structures are thread and process safe. So basically, the values and arrays are multiprocess modules, I mean, classes. So you can use them. Yeah. So that's the example of set memory process, value and array. So it's an array is basically nothing but it's a Python array implementation with some more trading and support for multiple inter process sharing. And there is another thing which we can use is the manager. Manager objects, Fox and new process, whenever we start it. So manager gives a one benefit that we can use, we can start, we can start, we can take a dictionary object, we can take a list object. So it's very good for if you want dictionary and list, but it is lower than set state that I mentioned array and values. But yeah, we can use it and manager starts a new process, whenever we will create a check of it, it will give it gives a proxy object, which supports names, spaces, blocks, return log, semaphore, pound log condition, event, queue, value and array. Yeah. Manager can be used to set state on different computer processes as well. So for one example, I'm on computer one and I want to share that data with another computer on another computer's process, then I can use a manager to do it, but it's lower than set memory. And yeah, in the end, don't do it. Or it's set state as much as possible. The reason is that it will decrease your speed in the end because it's a trading and I mean they are they all are third safe and everything. So that's why if you are writing something into it, so until you write into it, it will block you until that time. Yeah. So also one more thing should be taken care of whatever object we are using in, I mean, multiprocessing argument, I mean, that's a state. It should be P cable. It should be, I mean, P cable in a pipes. We should also care about this thing, whatever object we are using, it should be P cable. I mean, P cable in the science like we should be able to convert it into pickle format. Zombie processes. So we should take care about zombie zombie processes are nothing but like even though my program stops, I'm executing something. I started multiprocessing under a master process. I have stopped that master process, even though that child processes are running. So there are the zombie processes, zombie processes. So how to deal with that processes. So whenever we stop some program, we do controversy at the time some signal Linux and Windows sends some specific signal to Python and we can handle that signal and we should send that all the signal handle that signal and we should terminate that process is or we can send the same signal to that processes as well. Yeah, our terminating process is. So instead of terminate, try to close it or try some events or something so that or conditions or something so that what we so that so terminate will gradually stop. It's like you are, I mean, taking a plug off your computer. So it's it will give you very bad results. So for an example, if you are fighting some file or I'm doing some stuff, then it may possible. I mean, you will get corrupted data or something. And yeah, you I mean, in global variable, whenever you are using global variable, then it may possible in the child process and it is possible that in the child processes, you will not get the same values as you are getting in a global variables. Yeah, that's it. Okay. Thank you very much. It's all. Unfortunately, we don't have time for questions. Sorry, there's no time for questions today. But just crash it to afterwards if you have something to say and thank you very much. I look forward to seeing you the rest of the week.
|
Hitul Mistry - Python Multithreading and Multiprocessing: Concurrency and Parallelism In this talk, people will get introduced to python threading and multiprocessing packages. This talk will cover multiprocessing/threaded development best practices, problems occurs in development, things to know before multiprocessing/multi-threading. After this talk attendees will be able to develop multiprocessing/threaded applications. This talk will cover threads, global interpreter lock, thread pool, processes, process pool, synchronization locks - Lock & RLock , semaphores, events, condition, timer, pipes, queue, shared memory. This talk will also cover best practices and problems in multiprocessing and threaded application development.
|
10.5446/20123 (DOI)
|
Hello, everyone. Thanks very much for coming. Hooray. All right. Thanks for coming to this talk. I appreciate that I kind of mistitled it, and it's a very boring title. Maybe you think I'm going to talk about how to output data to a spreadsheet with Python. That's not what I'm going to talk about. I'm going to talk about how to build a spreadsheet application with Python, how to build an alternative to Excel. That's me. My name is Harry. My Twitter handle is the HJWP. Like all Twitter addicts, I'm more followers is crack to me, so be followers. My website is abaythetestinggoat.com, where I talk about testing. But here, I'm going to talk about magical Python spreadsheet. So that's what we're going to do. Does anybody here know how a spreadsheet works? How a spreadsheet calculates the functions you put into it? Yes, sir. You do? Anybody else? Good. All right. So you at the end can tell me whether I gave a good description. And everybody else, my plan is to try and demonstrate that making a spreadsheet is a little bit easier than you think. I'm going to try and build one up step by step. So first, I'd like to take you back to, let's say, 2005. It was a simpler time. There was no, well, Facebook had just started out and just got out of universities in America. MySpace was still all the rage. I had a MySpace page. Britney Spears had just released her seminal comeback album, Toxic. And of course, it was a great, great year for the beginning of the North American ragged jungle renaissance, if you don't like Britney Spears. So in this time, a series of programmers got together and they had this idea that everyone loves using spreadsheets, but actually working with them sort of sucks because you have to use VBA. And wouldn't it be great if instead of having to use VBA when you want to script your spreadsheet, you can use Python? Wouldn't it be great if there was a Python spreadsheet? So these guys got together and they had this crazy idea and they went off and they built a Python spreadsheet. It was a GUI app. And then four or five years later, I joined the company as we were just re-implementing this app in a web-based form. Come in. You're not late. The conference is late. It's lovely to see you. Hi, Paul. Hi, David. So we're going to build then a Python spreadsheet with our colleagues here back in 2010. And it's going to be a web-based tool. And what I want to show you is take you step by step through how can we build a working spreadsheet with Python starting from scratch. And it's going to look a little bit like this. So first of all, we'll assume that the GUI is a solved problem and that we can make a two-dimensional grid like this with all clever JavaScript that's going to allow the user to interact with it. And we're just going to build the back end, the engine, for recalculating the spreadsheet when the user does stuff. So here's a spreadsheet. I want to be able to do things like type into it. Yes, it's a talk with live demos. It's going to go wrong. I'm going to type in things like this. Sorry, ignore that. So I'm going to type in two. And I'm going to type in and three. Fine. OK. So what have I got so far is I've got a sort of two-dimensional grid. And if I want to store this data, I'm going to be able to say, OK, well, a1 is 1, b3 is the string, and 3. So we'll start off nice and easily. That might look, I'm going to propose to you something like this. We're going to have a dictionary. It's going to be indexed on a tuple of row number and column number. And that's going to contain a cell object. And that cell object is going to say, oh, I'm 1, I'm 2, I'm 3. So far, so good. Hooray. A round of applause. Yeah. Yeah. Yeah. You guys, you guys, you guys, you can't just applaud when the speaker asks for it. That's like super cheap. But I'm the cheap one. I'm not you guys. All right. So that's a pretty useless spreadsheet. Let's see if we can't make it do something better than that. Like over here, what if I want to go equals 2 plus 2? A spreadsheet should allow us to do some maths. OK. So who do you think is going to happen when I press Enter here? No, it's going to work. Hooray, 2 plus 2. So it works. That's not too bad. And notice what I've introduced here is that there's a difference between the formula. The formula is equals 2 plus 2. And then the result, or the value of the cell, is 4. So now I've introduced a cell. It's not just a bit of text. It's also there's a distinction between formula and value. And that might look a little bit like this in code. Let's say I'm going to have a cell class. It's going to have a formula and a value, which we initialize to a sort of magical undefined special variable. And then when we want to calculate the worksheet, we just go and find all of the cells in the worksheet. Don't look back at your slides. That's in my speakers tips. Are we going to go through all the cells in the worksheet and we're going to see how? Does this start with a little equals? In that case, it's a formula and I have to do something special. Otherwise, the value is just what the user entered. So what special thing can I do to get 2 plus 2 to turn into 4? I can basically just do eval, cell.formula. Room answer, just call that. And so 2 plus 2 is going to turn into 4 because I'm going to call eval on it. Hooray, eval statements. The best thing about Python. They never go wrong, do they? OK. Now is anyone particularly evil that would like me to change the formula that I've put into this cell? Any suggestions? Yes? OK, for the liberal people's time. I could do, yes you can. And you can put it in a string. How about if I do this? Everyone wants to see this, right? Error. Oh, hang on a minute. I thought of that too. Somehow we have to handle this. We have to notice when the user does something stupid. We have to catch some errors, maybe give them a nice little trace back and show them a little warning. There's division by 0. So how can we do that? OK. Well, we're going to put try accept, classic. We're going to call eval cell.formula. And then when we catch an exception, we're going to go and populate that error on the cell instead of calculating its new value. Fair enough. OK, well, so hooray. Well done. I've got a spreadsheet that's basically a two-dimensional calculator. None of the cells can talk to each other. This is still not a spreadsheet. It is still no better than a calculator. So what we'd really want to do is maybe be able to refer to other cells in the spreadsheets, right? In my things, I want to be able to go something like a1 plus a2. Everyone think this is going to work? Yes! Hooray! a1 plus a2. OK, so what have we done there? We've taken something that looks like a1 plus a2. And we need to somehow turn it into something Python's going to understand. So in a way, we're taking a1 and a2. And we've got our worksheet object, which is a dictionary, and it contains all of the cell objects. And basically, we want to translate the sort of string a1 and the string a2 to become some valid Python. So if I manage to turn a1 into worksheet11.value and worksheet12.value, then I could call eval on that. And so the way we're going to do this is we need to transform things that look like Excel formula, things that include cell references, into things that look like that Python can understand. And we've already got our worksheet object to refer to cells. Does that make sense so far? OK. All right, so what are we going to do that? OK, fine. We're going to have a little formula, a little setter for our formula. We're going to say, hey, if it starts with an equal, then we're going to go and parse the user's input and other one. And even, I'm going to show you that, you might get a formula error. So if there's a syntax error in their formula, you can pop that in there. And meanwhile, you can transform a user's formula into a Python one. Does that make sense? Now, would you like to see some of the magic of parse to Python formula? Enthusiastic yes. Yeah. Yeah. So this is the first little bit of recursive fun. And I thought I'd present it using some tests. So we want to be able to say equals1 should turn into 1, equals1 plus 2 should turn into 1 plus 2. Equals a1 should become worksheet1.value. And then you're going to have crazy formulas in your things. You can have x times a1 for x in range 5. That is a valid formula that you can enter into this Pythonic spreadsheet. And that's going to turn into this. So all of these sorts of things will happen. I'm not going to go into the details of the parser. Parsers are parsers. This is a special one that knows how to understand Excel and turn it into Python. But you're going to have some recursive fun. You're going to look at a node. You're going to say, hey, if it's a cell range, we rewrite that. If it's a cell reference, we rewrite that. And then we're going to call the parser the rewrite function on each of the nodes inside it. So if you've got a cell reference or a cell node, you've got some children. So that's the first bit of recursive fun. You parse things, and you look at your a1s, and you've transformed some a1, a2s into valid Python. OK. So that makes sense. But we're not really finished with that job. OK. So we've got equals a1 and a2. What if we have something like this? I'm just going to go and browse. Sorry, I should have opened this before I started. What if we'll out? I'll just show you on this spreadsheet instead. OK. So I've got equals a1 plus a2, and that's fine. Now what if over here I have another one, which says equals b3? Well, now I can't just evaluate these cells in any old order, because before I can calculate this one, I need to know that it depends on this one. And then this one depends on these two. So what I've introduced is a kind of dependency graph. That's the fundamental structure that underlies a spreadsheet. You have a series of cell references, cells point to each other, and those are the dependencies. And in order to actually do my calculation of my spreadsheet, I'm going to need to know what that graph looks like so that I can calculate the things that have no dependencies first, and I calculate things that depend on them, and then I calculate things that depend on them, and I do it in a sane order. And so we had a little bit of recursion in our parser, and now the real recursive fun begins. Recursion, it's the only truly fun thing about programming. We're going to do something a little bit like that. We can parse the dependencies out. We've used our parser to recognize a1s anyway, so we can also say, oh, OK, that I can find for any cell formula. What other cells does it depend on? I can have a little function to calculate a cell. I'm going to go, OK, evalves cells Python formula. And now I'm going to be able to build a dependency graph. And this is going to say, what order am I going to be able to re-evaluate my cells in? If I build a dependency graph, find out everything that depends on everything else. I can find all the leaves in that graph. I start with the leaves. I pop them off this sort of queue. And each time I calculate a particular cell, I can then remove it from its parents. Does that make sense? We're going to say that when you've calculated a cell, the things that depend on it don't depend on it anymore, so you remove yourself from the parents. That means that we did have a graph that was a one-way graph. We knew what every cell depended on. But now we need to know who and so we know the cell's children. And now we also need to know the cell's parents. So we need an algorithm for parsing a one-way graph and turning it into a two-way graph. And let me see if I can explain that to you. So build a dependency graph. We generate a cell subgraph for each of the cells. And we're going to parse in these arguments. We look at the worksheet. We keep track of the current graph, which is the two-way graph instead of the one-way graph. We keep track of the current location, and we keep track of what things we've completed already. So like any recursive algorithm, the very first thing you need to think about is what's the exit condition. The exit condition is a thing that you come across and know that you've already done, in which case you can exit out. Fantastic. Next, we look at our cell. We know who its children are, so we're going to add them into the graph. And then for each of those children, we're going to recursively call the same function to go and find, you know, do the same thing. So we do the children, and then we recurse down into each of the children. And once we've done that, we've completed that particular cell. Adding the children involves creating a node for the parent, and for each of the children, we create a node for them, and we say that you are in the parent-child relationship with it. Does that make sense so far? Hooray! Who now? That's quite a lot of hard work. You have to wrap your head around it, and then pretty soon you'll be thinking, hang on a minute. What if I have A1 depends on A2, and A2 depends on A3, but A3 depends back on A1, and I've got a circular dependency, so I can't quite do it like that. I'm going to have to track the current path that I've taken through the graph as I'm recursing down into it. And if I spot that the current location I'm at is already in the path that I came to in the current stack of the recursion, I'm going to raise a cycle error. And that means that I need to catch the cycle errors when I make the recursive call down into the algorithm that's looking at each of the subcells. So that is going to make sure that I can also catch cycle errors. Still OK, everyone? Yep. OK, fantastic. Who knows a better way to do this? It's rated tool? Not sure about that. You can do a thing that's I think called NetworkX, which is a network analysis package. And you can give NetworkX a one-way graph, and it will just give you back a two-way graph in two lines of code. So a lot of wasted effort and hard work on the part of the dirigible spreadsheet developers. But it was character building. We enjoy recursion. It made us joyful, so fantastic. All right, so far so good. All right, great. I've now got a little thing that can go, OK, b3, a1, a2, a3. Now, what if I wanted to go into my spreadsheet and actually start having some proper Pythonic fun? So if I wanted to define some custom functions and now start using more Python in my actual spreadsheet. So if I wanted to go def foo of, like, say, x, and we're going to return x plus 42. All right, great. It'd be lovely to be able to use this function foo inside of our, what should we call it, spreadsheet, right? Am I going to do foo of 3? Do we think that could work? Yes! All right, how are we going to get that work to work? We've now got two sets of user inputs. We've got all the values and all the formulas they've put into the spreadsheet. And we've got also some custom user code that they've put onto the right-hand side. So what we're going to need to do then is we've got two sets of evaluations. First, we need to evaluate the user code. And second, we need to evaluate each of the cells. And the way we're going to do that is we're going to start isolating our eval calls from the global context, which is probably something we should have done ages ago. And we're going to call eval the user code. And that's going to populate the foo function into this context. So a context is just a dictionary. All of Python's namespaces are just dictionaries. Namespaces aren't they great? And then I'm going to pass that same context to the eval context of each one of my cells. Hooray, I've now got custom functions. But I am sure I hear you ask, what if we want to write a function that actually can access things that are already in the spreadsheet? So supposing rather than having a function foo, I'm going to have a function that say, have I got it over here, is going to say sum everything in row A. And I'm going to want it to say, OK, well, look at things that are already in the spreadsheet. So I can't evaluate my user code before, say, I've already done a cycle of loading some of the constants in the spreadsheet. So I can always load the constants if I know, if something is a constant, I don't need to evaluate it. So I could evaluate, before I evaluate each of the cells, I can evaluate some of these functions. And then I'm going to have something like this. All right, so I'm going to say, load all the constants, then eval my user code. So the user code has access to the constant. So I can have a function that looks at what's already constant in the spreadsheet. And then evaluate the formula. But as I'm sure you're wondering, what if I wanted to write a custom function that can access the results of evaluating the cells? Well, then I'm going to have to say, maybe I'm going to let the user input two types of user code. One to be run after we load the constants, but before we evaluate the formulae. And one to be run after we evaluate the formulae. And sure enough, that is what one can do. And you're going to do something a little bit like this. So you've now got a load constants function. You've got an eval of the user code pre-formular evaluation. You evaluate the formulae. And you eval the user code post-formular evaluation. Who is still with me? Hands up. OK, almost everyone. So that's fair enough. So here's the realness. So that would mean then I would have to put two little code panels over onto the right-hand side. And hopefully, this is the bit of the presentation where something slightly magical happens. If we have a look at this bit of code here, where we're doing, OK, let's build a context. We add the worksheet to it. We call this load constants function. We call the evaluate formulae function. And then we've got this user code that's kind of before and afterwards. What if instead you had a thing? Oh, incidentally, you can now, if you put things into the pre-formular evaluation functions, you can actually put formulae into the spreadsheet from the user code panel before the formulae evaluated. This is a certain. I mean, this amazing Python spreadsheet is a surefire set of guns pointed at both feet at the same time. It's brilliant. OK. So what about, this is what we're doing right now. And then what the user actually sees when they log in to a dirajor spreadsheet, and then I'll start a brand new one for you, is they see this. The user code panel is pre-populated with a function called load constants, which you're noticing is exactly the same name as the function that I used in the actual evaluation of the spreadsheet. And it's got a function called evaluate formulae, which you might remember was the same name as the function we're going to use in the real evaluation. And so here what I'm going to do is I'm basically going to turn the whole spreadsheet on its head and say that the spreadsheet is the user code. And I have my load constants evaluate formulae in my calculate function. I take my context. I put the worksheet into it. I put the load constants function into it. I put the evaluate formulae in context, which takes the worksheet and carries it into my normal evaluate formulae function. And then all I do is exec the user code. And it's just your user code panel that loads the constants and evaluates the formulae and all of that. And that means that you can do totally crazy things in your user code, like you can evaluate formulae multiple times. Or you can put nested recursive calls to the spreadsheet itself. You can make spreadsheets that call other spreadsheets. You can populate formulas done programmatically. All of that sort of fun. The spreadsheet is the user code. The user code is the spreadsheet. This is the most pythonic spreadsheet available. Am I telling you that you should use this spreadsheet? Absolutely not. If you want to work with spreadsheet type data in Python, just use an ipython notebook and pandas. This is just a little bit of fun that I thought you might be interested in. The dirigible source code is all available on our GitHub at github.pythonanywhere. forwardslash dirigible spreadsheet. If you fancy taking a look at it, I've now given you a tour about it. And hands up, please, if you found that easier to understand than you had thought it was going to be to understand how a spreadsheet works. Wait, no. Hands up if you now understand how a spreadsheet works. That is like a 90%, I'd say. Hooray. And who thought that was easier than they thought it was going to be before they arrived into the room that they thought that they thought that the spreadsheet would be evaluated for what it was going to be? Irish just held up a sign saying, are you kidding? So hands up if you thought that was easier and simpler than ever. One person. All right, that's good enough. No, no, no. Please, hands up if you genuinely thought it. And if you thought that was really confusing, there you go. Hands up if you think I should never do this talk again. All right. So by implication, that's two people putting their hands up incidentally, any program working groups that are considering talk submissions for future conferences, two people put their hand up, which means about 200 people in here think I should do the talk again. Thank you very much, everyone. Good night. We have four minutes for questions. Hooray. Hi. Did you think about security? Yes, we did. So obviously that was one of the major things we were giving this out to random strangers on the internet saying, hey, will you eval your Python code on our servers for fun? So yes, we have a sort of sandboxing model that allows make sure that each of the users can only access a restricted part of the file system and basically the whole containerization story that you've heard to death at this conference already, no doubt. That's the story. I have stripped all that code out of the dirigible that's published online, although you can find it in the history if you're interested, because it's easier to demo and easier to understand without the security stuff. But yes, you can. Good question. OK, maybe one more quick one? Thank you for the talk. It's more common than a question. Well, to be serious, when I look at my colleagues work, and they work a lot of time with Excel, I think it could be some kind of killer application or serious, because we can work with pundas and do it at this level. But during our work, often people really want to change single cells, among 1,000 cells, because of some special thing they just heard or whatever. And if this is easily trackable and expandable with Python functions, I think it has potential for really serious or useful tools. No, you're right. I mean, the spreadsheet is a wonderful tool. I wrote a thesis about it for my masters. And that whole instinctive two-dimensional, I can see the numbers, I can change them, I can visually represent the relationships. It's all wonderful stuff. But it turns out nobody wants to use a spreadsheet other than Excel. And so as soon as you try and build a competitor Excel, everyone's like, oh, I want every single shortcut key that Excel has to also work in yours. OK? And that's quite a lot of work, catching up with 20 years in Microsoft. So however, any iPython developers, iPython notebook developers, who want to integrate some sort of spreadsheet component, would be happy to point you around the Dirigible Cobase and see if there's anything in there that you can use. There you go. What time do we have to finish, Iris? Two minutes. Two minutes. Hooray! That's time for another question. Here you go. Again, hooray! That's someone really engaged. Everyone else is like, how long before we can get? Maybe a weird question for a Python conference, but did you think about implementing this in JavaScript? Uh... So the straight answer to that is no. But I'm thinking about it now. Yeah, so one of the interesting things about it is that you run things on the server. So remember when we executed that, we looped around all the leaves in the spreadsheet and we had a sort of queue of them. It's very easy to start parallelizing that. And then you can do things like recalculate your spreadsheet across a cluster of machines. Because once you've got a leaf node, that's totally independent from all the others. So we were thinking maybe massive parallelization would be an interesting market spot to be in. You can't really do that with JavaScript. Is that the right sort of answer, Charles? Yeah, yeah. And no JS didn't exist back then. No, but you were thinking maybe you'd have the whole thing on the browser. Well, mostly to run it all locally. Yeah, all in the browser. I'm sure you could rerun it. All Python to JavaScript is pretty easy to rewrite. Knock yourself out. Exactly zero minutes. Let's rush to the lacking talk to see more Harry's performance. Yeah, I do have another job. Thanks very much for coming, guys. There you go.
|
Harry Percival - How to build a spreadsheet with Python Do you know how a spreadsheet works? Can you imagine building one, from scratch, in Python? This talk will be a whirlwind overview of how to do just that. Based on the source code of Dirigible, a short- lived experiment in building a cloud-based Pythonic spreadsheet (now [open-sourced], for the curious). We'll start from scratch, with a simple data representation for a two- by-two grid, and then gradually build up the functionality of our spreadsheet: - Cell objects, and the formula/value distinction - Evaluating cells, from simple arithmetic up to an Excel-like dialect - Building up the dependency graph, and the ensuing fun times with recursion (arg!) - Integrating custom functions and user-defined code. Showing and explaining code examples, and alternating with live demos (don't worry, I've done this before!) And it's all in Python! You'll be surprised at how easy it turns out to be, when you go step-by-step, each building on the last... And I promise you'll be at least a couple of moderately mind-blowing moments :)
|
10.5446/20121 (DOI)
|
So I'm going to try to break a personal record in how fast I can go through these slides. So type hints. The last time I talked about this subject a few months ago, it was a PEP that was a proposal. Now it's being accepted. So that's good news. But I still want to start with really ancient history. In 2015 years ago, there was a type sik that was already discussing some way of adding type hints or static, at the time we called it optional static typing to Python as a completely optional way to sort of communicate either between developers or between the developer and the compiler about the types of arguments of functions. And we actually sort of one of the proposals which I at the time already favored looks exactly like the annotation proposal that eventually got accepted. But at the time it was too controversial. I picked it up again in 2004, 2005. It was a series of again incredibly controversial blog posts that sort of the sky was falling, people hated it. But at the same time, my own thinking about the top jack sort of slowly continued and I started introducing things like generic functions and generic types. Not too much, not too long after that, I realized that the whole topic of defining types for arguments was too controversial to actually introduce in Python 3 as such. But as a compromise, we got PEP 3107 which was accepted, which introduced the function annotation syntax with little or no semantics. The annotations would be introspectable but otherwise they would be entirely ignored. And this was very much a compromise approach intended so that eventually experiments could be carried out like what has happened more recently. Because the recent history is that a few years ago at PyCon in Santa Clara, I met an enterprise young student who was writing, at the time he was writing, I think, a dialect of Python that he wanted to have gradual typing. And I convinced him that if he created the dialect of Python, his language would be great and it would have one user. On the other hand, I said if you actually tweak your syntax and add some compromises and sort of mess around a little bit so that it fits in with the existing PEP 3107 syntax, then maybe your work will not just be to earn you a doctorate but will also be useful for the Python community. And he actually took that to heart and started experimenting with notations like list, square brackets of T. I also ended up working with him at Dropbox. However, my pie still didn't seem to be going anywhere until last summer at EuroPython Bob Ippolito gave a talk, what Python can learn from Haskell. He had three very specific proposals, two of which to me seemed completely inactionable and the third one of which was we should adopt my pie. So that sort of inspired me and a few other people including Lukas Lange who drafted the first version of a PEP for type hints. And this again was an incredibly controversial explosive discussion on Python ideas and afterwards on Python Dev and on IRC and everywhere else where I didn't want to look. But finally sort of in my head at least it began to gel what this was good for and I met someone who had been thinking about this kind of stuff for a while, Jeremy Sieg who I'll be mentioning a little later. So the architecture that we eventually agreed on and I think my pie was very instrumental here, static type checking is not a function of the Python interpreter. And this is the sort of big light bulb that went on in my head and in other people's head even for Yucca I took a while for this to sort of gel. The first version of my pie that supported Python would actually just run your program and just before running it it would type check it. And then he added a command line option to only type check it. And then eventually we removed the command line and removed the option to actually run the code. We said to run it use the C Python interpreter or my pie or whatever. On the other hand if you want your code to be type checked that's a separate thing just like pylint is a separate thing it doesn't slow you down at execution time at all. And this architecture suddenly a lot of things started making sense and this helped sort of decide how to design all little details of the static language. The second part of the architecture is sort of obvious we're using function annotations for these type hints. You put them in your code only the type checker cares about them. The third thing that was really important that was sort of also a somewhat smaller light bulb is you have to have another way of placing type hints separately so that the type hints are separated from the code that they annotate. We call these stop files you could think of them as header files but they're not really the same thing because they're not actually ever used at execution time they're only used by the type checker. Before I go into more detail about all these things why do you actually want the static type checker and this sort of there have been many reasons why people have proposed optional static typing for Python and some of those reasons were very runtime oriented people were hoping that at runtime they could catch functions being called with the wrong argument or people have hoped that at runtime a just in time compiler could generate better code or maybe sort of module import time type annotations could be used to generate more efficient code. This is an idea that for example, Scython actually uses albeit with a slightly different notation. However, the real reason why static typing is an important thing is that it is not that it makes your code run faster because that's an incredibly complicated thing it is that it helps you find bugs sooner and the larger your project the more you need things like this and in fact people who have really large old code bases maintained by dozens or hundreds of or thousands of engineers are already within their organization running various things that are in some sense static type checkers. There is an additional thing that especially inline type hints help you help when you have a large team working on a large code base which is that new engineers are really helped by seeing the type hints and it helps them understand the code. And it's in part it's just a communication mechanism from programmer to programmer which in general is always one of the criteria I use for designing parts of Python. Let's see. So the type hints in particular help a type checker. Python is such an incredibly dynamic language. There are so many clever hacks where you introspect a dictionary or module or a class or use a dynamic attribute getter that very quickly if you do traditional sort of program symbolic execution of a program trying to figure out what the types of an argument are so that you can then check that that argument is used consistently with the argument type. Well, you can't even find where the call sites are because everything is dynamic and there might be four different functions named keys and you can't actually tell which one is being called very easily. Type hints help a static type checker sort of get over those humps. There's a little statistic that the authors of PyCharm told me. PyCharm is an IDE that has its own sort of partial type inferencing for Python programs so that it can show you not just when you're making syntax errors but also when you're calling things that don't exist or with the wrong number of arguments and it can make decent suggestions about what methods starting with K might occur at a particular point. So they told me that they can correctly infer the type of maybe 50 or 60% of all expressions in a Python program which means that almost half the time they don't know the type of an expression which makes it impossible for them to then give any useful hints or do any checking. In the case of an IDE, of course, what they have to do in that case is be silent or use some other fallback heuristic to give suggestions, not say your program is wrong. But nevertheless, if there were type hints in a program, they could often produce more accurate predictions and so on. I did mention the additional documentation. When you find coding conventions at companies that say in the doc string every argument must be described and the type of every argument must be indicated. Well if the type of the argument is already part of the syntax, you save a little space in the doc string. Also if you don't have a doc string at all, a document generator can still use the annotations to generate better documentation. So why do we need these stub files? Why do we need to be able to put the annotations elsewhere? Well the first use case that you think of very quickly is C extensions. When you start thinking about static typing anything in Python, you realize that there is a huge number of built-in functions and built-in modules for which you also need to have type information. And you can't easily scan the C code and then figure out what the types of all those functions and classes are. So you need to have some dummy Python code that declares the types for your corresponding built-ins and built-in modules. So this is the first use case for stub files. The second use case, and there's a series of use cases that have to do with Python code that you might want to annotate, but there are reasons not to put the annotations in the code. And so it could be that this is just third-party code and you can stick annotations in third-party code, but now you have made a local mod and every time you upgrade that third-party package you have to do that again. Or that's a lot of work. You can't always push those changes to the third-party because they might not care, they might not be a maintainer, you might be using an old release that doesn't get maintained anymore, maybe they want to be source compatible with Python 2 and the annotation syntax only exists in Python 3 and so on and so forth. There are too many things to sort of try and annotate everything. So stub files are a lighter-weight approach to annotating code that for some reason you don't want to annotate in place. So when I present all these ideas I still get a lot of very sort of critical negative looks. A lot of people really like the fact that Python is dynamic and they don't see any reason why they would pollute their code with stuff that in their mind is associated with troglodyte languages like Java or C++. Well, and nevertheless, the people who are maintaining very large code bases often have some form of static analysis. They have things that look in the doc strings and use some convention for storing types in doc strings and use that in their analysis. Or they have some kind of static analysis but they don't have annotations at all, not in doc strings nor anywhere else. And their type checker just isn't very effective. PyLint can only catch so much. So in some sense what this whole proposal is actually introducing is more or less just a standard notation that you can use in case you already want this. It's very much optional. In Python 3.5, the first version where it's available, it's also provisional which is a technical term for new standard library modules and new peps in general where we say, well, we introduced this in the standard library but we're reserving the right to sort of change the API for one full Python release. So in Python 3.6, the typing module may look a little different, perhaps it's unlikely but it could even look quite different than it looks in 3.5. And this is something that sort of falls outside the normal guarantees of backwards compatibility. You can read up on this in PEP 4.11 which sort of explains and defines the concept. The key thing is that in 3.5, nobody's code will break. And my plan is that beyond that we won't break your code either. But at the same time, I do want to sort of take a position. I don't want to say, well, we have the annotation syntax without semantics, let people just do whatever they want to do. They can use Mypy. If they want to, they can use their own doc string based convention. They can put type annotations in decorators. Let a billion flowers bloom. I think that we've had enough experiments and sort of attempts at doing this that it's better to get everyone behind one proposal. And I was very pleased to see that Google and PyCharm, for example, were both very supportive of this proposal even though they're not planning to adopt Mypy itself. But they are planning to adopt this new syntax. Some people said, well, okay, maybe you're right. Maybe we need a syntax, but you can't sort of force it down our throat. It's unripe, immature, needs to be thought about more. Let's wait until 3.6. But really, that's not going to help anybody. If you want a notation that uses angular brackets instead of square brackets, introducing that is just as hard in 3.6 as it's going to be in 3.5. So I sort of, I mean, I started this with what I thought was plenty of lead time. We had a large number of very productive discussion threads. And I just pushed on everything to sort of reach a compromise and get something working. And so if you were hoping to use this for code generation or if you still believe that type annotations mostly are useful to make your code faster, sorry, that's not actually very high on my list of use cases. PyPy is doing fine without type hints. We'll see what Siphon says. Siphon, I believe, can already optionally use annotation syntax instead of the traditional Siphon notation. Maybe they'll prove me wrong. But Cpython certainly is not going to suddenly run your code faster if you put annotations in. And that is not at all part of the plan. So there's one more thing. Step 3107 is now, hmm, it's not quite 10 years old. Maybe it's eight years old. There are definitely people who have used annotations creatively and done something completely different with them. Here's an example of something. I made this up, but I saw something similar where someone had written little language for marking up functions that would be invocable from some command line where the annotations specified say the option name used. That's cute. That's not going to break in Python 3.5. However, if you run code like that with mypy in order to type check it, mypy is going to choke on that particular notation because mypy expects the annotations to be something else. Of course, you may not need to run mypy. You may not care at all. Or if other parts of your code you actually do want to benefit from type checks and you think you want to run mypy, but you still want to use this particular notation in some part of your code. There's actually a decorator defined in the pep that you can use to shut it up. It basically tells mypy this function or you can also use it as a class decorator. This class ignored the annotations because they're meant for someone else. So that was mostly an apology, a history, sort of the motivational part of the talk. Now I'm going to try and outline a bit how this actually works. How do you think about type hints? If you really want to know, you should probably start with 4.8.3, which is sort of a simplified theory behind this stuff. But let me go over a few of the basics. Here's a very simple function named greeting. It has an argument with a type and it returns a type. It happens to be both our strings. Then there's a function greet that calls the function greeting. Greet does not use annotations. Greet is not type checked. The basic idea of gradual typing is that both functions can occur in the same program, even in the same module. And a type checker is required to accept that code. If inside the greeting function there was some use of the name argument in a way that is incompatible with it being a string, the type checker will complain. However, in the greet function where there are no annotations to be seen, if you invoke greeting, it's not going to, perhaps the biggest thing to sort of understand is if I could only get a mouse. Okay, well, you can see dev greet of name. Clearly name could be anything. Print greeting of name. The greeting function only accepts a string. However, we're not going to get complaints from the type checker that we don't know for sure that name is a string in this greet function. And that is sort of that in case of doubt don't complain, that is one of the basics of gradual typing. And that's sort of different from, for example, if we were to assume that name given that it has no annotation has the type object, then we would actually have a type violation in this code. Because greeting doesn't take all objects. An object could be a list and a list is definitely not acceptable for greeting. At least it's not a string. So instead of being sort of picky, a good type checker using type hints, sort of checks, thoroughly checks code that has annotations and backs away from code that doesn't. And lets the two be combined in a useful way. Also, if the annotated code calls something that is unannotated, it will always just assume that the best possible thing will happen there. So this is sort of the principle, I think I'm repeating myself here, which is unfortunate because that means less time for questions. Code without annotations is always okay to the type checker. There are some hand wavy things here because there are some subtle differences. But basically there is this magical type named any, which is different from the also somewhat magical type named object. And the absence of annotations in first approximation can be seen as annotate everything with the type any. And any has a bunch of magic properties and I'll get to that here. So any is confusingly both at the top and the bottom of the class hierarchy or the type hierarchy really. On the one hand, if you ask for any object x, is it an instance of any? And this is of course a question that the type checker asks itself. It's not a question that you ask at runtime, although I use a runtime notation here to express it. It's always true. Everything is an instance of any. Also everything that's a class is a subclass of any, which really means it's a subtype. Apologies to Mark. On the other hand, and this is the weird part, any is also a subclass of every other class. And I'm going, this is, you can see I should not try to draw squirrels, but I can draw a very simple diagram with boxes and lines with the help of PowerPoint. This is a very simple class hierarchy. It has object, which is the built-in object. It has number and sequence, which happen to be abstract based classes. It has none type, which is the type of the variable none. Now, let's add any. So any is sort of a superclass of object. It's even higher up in the type hierarchy, but it's also at the very bottom. And if you were to think of this in terms of a classic subclass relationship, everything becomes a mess. Because now you can prove that every class in this hierarchy is a subclass of every other class in this hierarchy, which completely collapses everything to a big muddy ball of everything. So we don't want that. We want this version. And there is a separate relationship, which is formally called, is consistent with, that is just like the subclass relationship, but special cases any on either the T1 or the T2 position. And you either got this at this point, or I'm going to ask you to look it up later. Actually, Jeremy Seek has a very good blog post. What is gradual typing? So what do we have in our typing module? Typing.py is a single pure Python module. It's the only thing that the PEP actually adds to the standard library. Very easy to ignore. This is where you import things like any. So again, there's no new syntax. Syntactically, we are constrained by the stuff that Python 3.4 or 3.2 even can already do. And with a little clever operator overloading, that's actually not such a terrible constraint. We're not actually adding any type annotations to other parts of the standard library. So if you're looking for examples of type hints, you're going to have to look elsewhere. Also this typing.py itself can also be installed in Python 3.2 or 3.3 or 3.4 using pip install. What does the typing.py module do? It defines a whole bunch of magic objects like any and union and dict and list with capital D and L that are used for expressing types. So here is a little example class. It's kind of messy. There's a chart class and it has a function set label. And you can see that it's being annotated with some argument types. I don't give the function bodies. Now there are also some plain functions, make label and get labels are not part of the class. They're plain functions. And I just include them to show that you can use a class as a type annotation in some other part of your code. I'm also showing here that you can use the built-in list type as the type at the bottom. You have the argument points which is a list and the function get labels returns list. However that is incomplete information. Because we would like to be able to express to tell the type checker about these lists what is the type of the element of these lists. And so there is a new notation using a capital list and a capital tuple which are just some more magic classes that you can import from the typing module. And now we can say let's look at the return type first. The return type is list of string. So it is written as capital list square bracket stir square bracket close. You can also combine more complicated types. We can have a tuple of a float and a float which is a tuple of length two each of which item has type float. And you can use that as the argument of a list type. So now we know exactly what the type of that points argument is and we know exactly what the return value is. You can go one step further. Instead of list you can write ABCs. The typing module exports modified versions of the standard collection ABCs like iterable and you can actually say the argument can be any iterable of tuples of float and float. However we still keep the return type. This is pretty idiomatic type hinting. The return type is a concrete list because we actually sort of declare that it returns a list and not some other sequence. So what exactly happened there? Typing.iterable is almost just an alias for collections of ABC.iterable. However, it has a little bit of magic behavior added to it. But it is still usable as a standard ABC. It's usable in all the contexts where collections.abc.iterable is usable but it is also a type. The typing.list type shadows the built-in lower case list. And tuple has some resemblance to the built-ins. Tuple, however, it's not an immutable sequence. It's more like a structure. I have been incredibly imprecise in my terminology. Technically, we should talk about types when we talk about things that the type checker cares about and classes when we talk about things that happen at runtime. The reason that most of the time things work out fine if you are fuzzy about the distinction is that all classes are usable as types. When you define a class, that class is always also usable as a type. However, there are a few magic things that are considered types like any end union that aren't classes. So in the very little time I've got left, if we want to have any Q&A, a complete enumeration of things that can be used as type hints. So anything that's a class can be used as a type hint. There are these generic types list of int. There are the magic things that I haven't all explained yet, although I've given enough of an explanation of any. You could also define your own generic types. The first thing that I haven't mentioned yet, which is pretty standard in type theory, is a union type. You could easily have a function that takes either strings or numbers as argument. And you might use a union like that. A very common special case of unions is an argument that is either a certain type or it's none, and we can express that using optional. Optional doesn't necessarily save you any characters to type, but it certainly gives a very clear intention to the human reader. The type checker actually just expands it to union of int and none. So tuple, I already sort of tried to explain how tuple works. It really is a structure with a fixed number of fields, each with their given type. It's sometimes called a Cartesian product if you read academic papers. For those people who use tuples as immutable sequences, you can say tuple of some type and then dot, dot, dot, three literal dots ellipses. That's actually immutable sequence of floats of arbitrary length. Callable sometimes you want to say an argument is a function that takes such and such arguments. We have a notation for that. It's not a very elegant notation, but given all our constraints, it's the best we can do. If you have a really complicated argument signature, you can just put an ellipsis there and then it will take anything, and then at least you can at least, you can still talk about the return type. Generic classes. I'm going to cut this short, but you define these by deriving from a special thing named generic using a type variable. Type variables have to be defined explicitly using the type var helper function. The collection ABCs like sequence themselves are all generic and can be used in this way automatically. You can also define generic functions. Again you introduce a type variable. Type variables can be used. If you only ever need one type variable in a particular module, you can just use t everywhere. You don't have to define a new type variable for each function. This is something I'm going to skip in favor of more question time. There is a built-in type variable that can express something that is either a string or bytes, which is a very important idea in Python 3, mostly for Python 2 backwards compatibility, but there we have it. Oh, yeah. Now we get into the sort of slightly ugly stuff. Sometimes you have to have an annotation that contains a forward reference. It needs to, there's an argument, but the class that is used as the argument type hasn't been defined yet. One common example is recursive types. You can put the whole annotation in string quotes and then the type checker will sort of evaluate that while Cpy from just sees it as a string. There are also some cases where you want to annotate variables, especially class variables that are used as instance variable defaults. This is very useful. We have a type comment for that. And there's also a cast function. If you somehow need to tell the type checker, everything's okay. Don't worry, little guy. So stub files have a PYI extension. The bodies in the stub file contain literally three dots. In stub files, you can define overloading, which is also something I'm going to skip explaining. You can disable your type checks in probably too many different ways, but this is to sort of make the people who don't like type hints or have other uses for the annotations function as happy as possible. And then finally, here's a list of alternative syntaxes that have been proposed at various times for and what we ended up on the left. I'll actually skip this. I do notice that nobody actually proposed return type, Perenn arguments for a callable. The reason that we ended up with the somewhat clunky syntax that's actually in the PEP is that it needs to be easy to parse. We don't want to introduce any new syntax because we want to be able to backport typing the PY to previous Python versions, at least 3.2 and up. And we really don't want to have to change other standard library modules. So if you're a type theoretical academic, you're probably very unhappy with this proposal, but we can iterate over the next few years and at least we have the first iteration in our hands rather than in the air. The PEP has been accepted. Thank you, Mark Shannon again. The status is provisional. The code is in 3.5 beta 1. And I'm very happy that much of the discussion is behind me. So let's start some more discussion. So we don't really have time for questions, but we can make time for questions. If the next speaker can come up on the stage now. Thank you. So first question. Thanks. I really like the idea of type hints. I'm sure that will help us write better or more high quality codes, but I'm not so sure I like the idea of having two options for specifying these type hints. So in a stub file or inside the source code itself, that somehow doesn't seem very pythonic, that there's two options to do one thing. And I'm thinking, I have also heard some comments from other people that say argument lists will become very long, so the code will become harder to read. Would you perhaps recommend always using stub files as I can see that IDEs could perhaps inline these in the source file as you're working on it? Can I ask you to wait for the question to be finished if you want to leave so we can hear the answers? That was a long question. My position is that there are really quite a few downsides to stub files. It's sort of difficult to switch back and forth between the stub and the main code. And so when you're reading the code, on the one hand, the argument lists become longer, but if you put all the annotations in the doc string, your doc string becomes longer and people are OK with that. In many cases, the annotations aren't actually so verbose. Some of the examples I gave, for example, are impractical. In practice, you would always use a type alias, which I forgot to mention. You can just say A is some type expression, and then after that A is usable as a type alias. And so using type aliases, you can make your annotations shorter and also more meaningful. So I think that the case for inline annotations is still pretty strong. At the same time, there are absolutely cases where stubs are the only acceptable solution. So I think that we have to have both. Over here. I'm raising my hand. Hi. Sorry, where's the speaker? Yeah. Give me. So to add what I'm, I don't know what the proper term would be, but there are effectively arguments to things like list or callable. Parameters. Sorry? The parameters. Parameters. Parameters. In Python, we use parentheses to specify parameters to things. Why did we make square brackets for these instead? Because usually the thing before the square bracket is a class, and calling a class has already the meaning of instantiating the class to an instance. Also the square brackets sort of make you wonder, whoa, what's going on here? Something interesting must be going on, and sort of parameterizing types are something quite different than calling a function or instantiating a class. So sort of the square brackets came out because they're notationally sort of, they stand out a little bit, and yet they are actually already part of existing Python syntax because you can just use, we actually implement the square brackets by overloading get item on the meta class. That would be the last question of AI. Sorry. I have two questions actually. So the first question would be, is there any way to express variance and contravariance? Yes. Great. I didn't get to this, but it is in the PEP. You can have invariant, covariant, and contravariant type variables. The default is invariant. One quick question. How do numeric types work? Like floats, ints, can I pass an int to a float? That is currently done by a little bit of special casing in the type checker so that if the specification says float and the actual value is int, that's actually considered a subtype and acceptable. Well, thank you again, Guido. My pleasure. My apologies. Thank you.
|
Guido van Rossum - Type Hints for Python 3.5 PEP 484, "Type Hints", was accepted in time for inclusion in Python 3.5 beta 1. This introduces an optional standard for specifying types in function signatures. This concept was previously discussed as "optional static typing" and I similar to the way TypeScript adds optional type declarations to JavaScript. In this talk I will discuss the motivation for this work and show the key elements of the DSL for describing types (which, by the way is backward compatible with Python 3.2, 3.3 and 3.4). Note: *Python will remain a dynamically typed language, and I have no desire to ever make type hints mandatory, even by convention!*
|
10.5446/20119 (DOI)
|
Mežic...sinob in sedajje pravam leurally o pra offerom. Katera se cel igne, staj dr년om喂ovnikom Sa vsap dej matki regionsga? I zaten da sem bahel je na 3vej pske, a, izgledes napicne,... na Team P&M? Pa za sem mu victory iz izglede? Daj srp, ko je allih ročili? Plaza se tak், MHän Staj se paprimah tudi, Oglunajte sanje devet, pa neXi du kompletnost minim, boதim shouted성 v Xuan! Poletice Prežil, da z tem ras, ker pri extended vorherje bil s bolj poznat. Ch Belgija je modiv izaj pala vrovna, Dakle, ki je l Vel忘 je polo. P Moni je vol bursting, priživajo zelo v ljudle, pasa čekaj je na vsakov. Minds na in vse čekaj tak je visitorsi, to ne očal Brahm, strateg地je sanja od explodiraj, boš so podnovil, odotoč si ljudi陋. R April 15 tesad ho Audi pravli....o se zelo. Then these huge areas, you have to check for land mines step by step,...like spot by spot, very slow, it's a very slow process, you are always in danger,...is slow. When you are close by to what you think is a mine, we approach it very slowly,...try to clean the area around it. So, autorizemaker sp nevertheless, gredno do<|pl|><|transcribe|> celu odpalupu nimi posiada. İsz flavorowuj Michael stelloba Sloboj oporiestom za na Race z nami oby mladji u okorleben kolej conductorov ven, da se nada brandedne v rečje. Toms je claro nekaj na reminiscent$, te bude zvarani tutorira, v tabletu, ko videli, scooter!!! Se začugen? Nene agujte večne reč Varim favorin in echoesči mobilce, kaj bil TriUU, Tako od normace jednje razni lango poval jobsa možete k��j SEGRLENje obeđ pancakes, boljli izgled boneh 1953 pa smo od vr demeve in di capabilities je jedne ze digravijo nekaj hepko, z strs сначалаje reagrecije z držb, ki alm te bljenje, izroadっと s predaj izrod84편 dužizaj Latino proti trajsne in iz relations driftopnoądno poč Laureal nemamo iz ližnic za nedi od katahutnih zdovrfnovoj z ni no PVPL. svoj si jazem horses, in imamjutih vsojem kontro, prozumote sajeli prepared, što mislim ko ali dub cam je ne�aj što ne si voda napravila. Vi艿 Prof, s televisijima sredne fружiteče... je dodaril bridj 不是 in ako mate že vlaibtyi v Show Изbitno njevala šelet겨i. Sa lidem Specisetibernji os watersh.. sem плат na veseljei bars gre neko monetaryya izgleda razglet. Obtihit проjez al tartarje po objectives, bi smo vzato ledsili o teh vse Lasimi in pogledajooris ni vsak elektrolam iz sitov de in kislin- uporodingar, thinga zdajse publish pike in koloče. In prav stitchelse v 아니嘲iec minumi id charm nas na ò이었eque. Per pom rottenu posodle hr za Plettn baking, zelo po好好iksi v ridicu to ne razreč這麼, Prez battlefield, z liquidikoti tudi vokon quizaj factičiti tudi, druža misi hrabba, to je v unknown, ki pošli skup bit, li se z v kom Disney v In izmenje herbice do о performac boss throughout map. directions, disappeared Node. If you build a new site of you remove atreys... if you use a road, you can realize here sactivity and you can see from satellite. Tako, tudi je zelo aktivitivače klasifovate v zelošnjih zelošnjih aktivitiv. V kajšli, v zelošnjih zelošnjih zelošnjih, kaj je aktivitivače. V zelošnjih zelošnjih je, da ne so vsebe, ali nekaj zelošnjih zelošnjih zelošnjih, kaj je nekaj aktivitiv. To je več vsočen na lendom, ker, kako je aktivitivače na zelošnjih, zelošnjih je, da je zelošnja, zelošnja je, da je vsočen. To je več vsočen. Neljavno, da je več vsočen, je to izgleda v Bosniji, je vsočen 5 km x 4 km, zelošnjih je, da je 20 km2, da je vsočen na vsebe, da je vsočen na mene. In, da smo nekaj zelošnjih, zelošnjih je, da je aktivitivače na zelošnjih. Zelošnjih je, da nekaj zelošnjih, da se nekaj nekaj zelošnjih, ker je zelošnja, oče všajano ki je na viotshipt,� zo nekaj nekaj zelošnjih, hen ki se spinacho, Convention recently, u voldoveratku dohajala naB. To je na, leto zelošnjih je, da je ovolj. Dajbi ovolj, Tukaj jeection na 10 hè Z kartudja nebar 7 Joza V 2014 je ten dola kare Kako vova da smo poživili za vsejoikošor sodatelvanie 11 am. Protoč να tukaj hvaj gции drha, za stre bouncing, je Og stuck softly hotart out of pockets in dolo stranji, pone ležice are trivno m. Quント M honourenja pa morajence好的 nosete. pred defenem na 3. del어서vej od 2014. Na druge rane SurvIV20 selection pode pokazati doteženice. Samo phases battlefield, kve implementite inform serverusni nabeist, ker predan delajo kaj naj norm국 však naj letánsko izpotip, vač ga pozbine prokcessivne iz Graceljo. Je lvisko, je brojbalite očgodne vné po sp spicicice44.ронą več, z kindvoje, je tripod, pravidno ta štika lahko može vačno pri mejoreso. z bowlami wispoji k musclej. Imam tako zape, do dgega, kako jaz s lemami povoljaj, mislimo pa, da jaz nalacil informacije, je dolor na poisonedog vstranodish, toty močат kot o potvrbi. Vo domittenje šte есть, ki podjeli recommendation ko zaатakajo zgriniti戴ina rotacije grooming snakesh in da bom ima soemie pojočen Maryva v probe zpratlegt. Igore neighbors, iz plat Nico no pralobil, nekaj prej, Okej, k statues bila se oč nekaj ca zastanjem in oč Franki temあぁo, takoba mam del dičnem glasba in qtom samom koš remembrance bo to jest na zonjaấu nekega, graber je čas fighterстik, lahko je za Capta, pa in se premača maja. In da je prejudatuje tend award, in vzupračite mike zaamine pos groupo a movement, v čarbe spasipping�je obredito. Musimo da smo nič obredino res v Harris-R shutter, pa vš падirih, imamo drip, vividite ma app kon JUDGE, uga. Matej spasibo preso YOUH image, bits on Inuary znamo, da bile več iznik želo se, da jo sne zan propsEt je zbičen novičuča bodo heata kur na睛 achievements vo del地ško ta izveč naji boards lep underway do radi vs, iz, tem p moi SEG So je izveč indoesi preizvestovati in posoleni on je tako različpatna. Zdaj Perfect jo sneč guarding. In Tab phulator vues transmitted das Krist chill pen Friday Musite Oloudl about us to be and then compute the Stretching in this new dynamic just to show the fx of the step this is the data input and steps so before adding a stretched or any filter and this is the same images once it's been stretched and filtered so every feature in svoje ego Smar VR areớite bija. Da vid talked about the bolts. Pa smo Profes镜ki vedeli, kje labo tudi strap starimi voj march PO운ih za odskutiti crush weakj or ta tr 티z 아주ейst craf pound, z仮iju vsobjetite aj na owo so in pain fark か Makah so edge pospojen approximately Masko tak�, tak le cheat, tak je to, pri veliko verja. Tzemo je tkozda, ki nebo odmišleraj se v pondaj t风a imazi. 근非常的 sv porta, so ima na se svoj counts gov prijem, z učem drej tva Ivaka ljud. Njih uštrat po demobili. landssega, letak, Šikut. Nicko pro 기�ous laga. Nord один Grant de Sen. Texty so prese, kot jega mu pod Laboj 17 and etc. Servi ma namne od Hayd v drtiljejegodne, Ko še文je meni v RF, jaz risen Dåare, lont Brezinal in vinja qared stewih zakejاخovama. Kako jaz dol t Cherry bo vse kuzt v Krzy, ne za nothing z masnu burfiči sevunovati. uploaded v ter rumpiči im Hence 자� s da tvoj postavlj punishment ne transformerite prezentrem tvoj mandan lepa v stahuccessno nazimno uga ob quietly p Laranju kar delova skup on blocks. Nezavonega igrača. Je to enakšnimi strojetiti o strane in ŽЛEW encourage.FR To je instopov tem lagige in vznoarén.之 na t Titani In igrač nezavon, z rosni trampečine. Mi je啊rčit thusilovila aktivit שנive z Dusk не zobrano. So z Nestorjeri tudi zelo czemo dakle in domale razložimo. Raž convicted preri déžret tem, ko je pro athletes leash Cub felony early fa poddanje al di pro FOza i de<|pl|><|transcribe|> Σenca granicaz legislative Kw!* have to add geographical information, and Jidal library does brilliantly this job. It's very simple code. So, this is the step to pass from this, that is just a matrix with a value and a color associated to it, to this. So, a layer with geographical information that can be used by professionals, can be used on Google Earth, or every GIS software is in commerce, like ArcGIS, Kujis. So, just for concluding my thought, I will just find some final remarks. The activity map, map so is a second level product that is supposed to help to support the mine clearance professional, trying to reduce time and so money invested in this process, and to reduce the danger of doing such a job. This activity map is a product that was developed in the frame of an European Space Agency project, called Space Assets for Enhancing the Mining. And this project is going to be tested in real field next September in Sarajevo. I'm down. Thank you very much. So, my question is less about Python and how widely available are these remote data sources. Did you have to make special requests to get coverage of that area, or is the data from every pass, every day, somewhere available? Well, you mean the satellite images. Well, this data source, you have to get in contact with the company that produces this data. So, you just ask for them, you place an order, you say, and it's fixed. So, the satellite will pass over the area of your interest just in that date, and then it will keep passing every 24 hours in the same point. Just because the satellite keeps orbiting around the Earth, the Earth is under it that is rotating. So, after 24 hours, there will be the same configuration again. You need the same configuration to compute the coherent images, because you need the phase. So, you need that the satellite has to be in the same position to get the same geometry of the images. You could have, at minimum, at least four images a day on the same area, but they will not be on the same geometry. I mean, what I'm saying is that the satellite is looking under Earth from a point of view. Then after 90 minutes, the orbit is complete, and the satellite is again in the same position, but the Earth is slightly rotated. So, it will just see, it sees again the same area, but just with another angle. So, if you want just to see the area, you can do it at least, I said, four times a day. It depends on the latitude. On the north you go from the equator, the more passes you have, but you just can do the phases analysis using images that have been taken with the same geometry. So, that's why you have just to wait for 24 hours for the second and third images. Ne more questions? OK, so, thank you, Giuseppe.
|
Giuseppe Cammarota - Activity Map from space: supporting mine clearance with Python Removing UneXploded Ordnance (UXO) from minefields at the end of a conflict is a very time-consuming and expensive operation. Advanced satellite image processing can detect changes and activities on the ground and represent them on a map that can be used by operators to classify more dangerous zones and safer areas, potentially reducing the time spent on field surveys. We exploit space-borne radar Earth images together with thematic data for mapping activities on the ground using numpy, scipy and gdal. The Activity Map generation process to be shown will be implemented using IPython Notebook.
|
10.5446/20117 (DOI)
|
.............. Prydych chi, yna,'r fanod idde, yn unig arnos y bofal, o wildlyniad grafio a bosddai hwn. Mae ari shots a chyfnodwch hyd o ein ahyddoch er llejoad yn effigsoch i chi sydd wedi eu dr 잘못io'r Children's y safon.. Fy fydd yn tryn y hy 연� y bwerdd. A'w gofio yn fawse peth o bach eisiau bydd hynny felly roielinegardd me'r fawr yn fy iawn i fyni speeder y gymhenn Nad hayro ignorance Nu'n rhoi'r bilyodus Felly gallewch'i weld tan meddwl hefyd afternoon, Alun gearbox, south bydd0 o goedwm. Ac i liciunai ddech chi, Donkey Ac we also, how many people here have worked with Pandas? Okay, almost everybody. If you haven't worked with Pandas, it's fine. It's a lot like Excel, but way better in Python. And it's not a lot like Excel, it will change your life. But you can, the core of it is you can pass it a dictionary and you get out a table of data. And there's lots of other ways you can read from Excel, from JSON, from CSV, and so on. If you haven't used pandas before, I would totally recommend Brandon Rhodes' pandas tutorial from Pycon 2015. It really boils down like how to really just understand it but keeps it at a nice simple level. So you can make your column data source from either a dictionary or from a pandas data frame. And your data ends up represented like that. And you'll notice the one difference between when I pass in my dictionary and when I pass in my data frame is that I've got this extra key value here. The index has come from the pandas data frame. It's this left index here that it's named index for us. So that is the column data source. The next thing that's at the heart of bokei is the plot. I'm skirting over some details here because I don't think they're that important. There are models, plotting, charts, these three different ways to use bokei. But they have a lot in common and you should find it pretty easy to jump between the three levels. But I think it can be a bit confusing knowing exactly what to do from the get-go, which is why I'm trying to show you all three. So at the lowest level we have the plot object, the next level up we have the figure, and finally we have the charts. They're all different things. They come from different modules within bokei, but you're going to see they all have the same attributes attached to them, like a toolbar location, background fill, and so on. So once you're working with one, you can reasonably assume that the other one's going to behave similar. And they all have three really important methods on them, add tools, add layout, and add glyph. And if you can remember those three methods, you're going to get a really long way into building all of your plots. Glyphs, I hear you say. No one said glyphs. OK, glyphs, to me the word means nothing. I was so confused for a long time. It just means shapes. And I don't know why it's called glyphs. I'm sure there's some very, very logical reason. But glyphs are just shapes. And so this is actually from the examples. And these are all of the different kinds of shapes that you can spell within bokei. And in some combination of all of these, there's a whole nother set here, in some kind of combination of all of these things, you can make pretty much anything that you could think of. If you couldn't, I'm sure we would add another one, but we really have got most things covered here. So glyphs are just shapes. So we're building a dashboard. We've talked about data. Let's talk about charts. But I keep going on about this models plotting charts thing. And I want to just lay it out a little bit more. So there's these three modules. Bokei.models, Bokei.plotting, and Bokei.charts. Bokei.models is the lowest level of bokei, but it doesn't mean it's a level you should be afraid of. It's an awesome level. I love to live in the models level. Everything above it is built on top of models. So it offers you the most control, but it doesn't mean you have to do all the work yourselves. On the complete opposite end, you have bokei.charts. These are one line functions that let you spit out a bar chart, spit out a line chart, spit out a horizon chart. It just takes your data frame and gives you something awesome. So it's very, very quick to use, but it does all the magic for you. Bokei.plotting lives in the middle, and you have to organise your data for it, but it then tries to pick sensible defaults so you don't have to add the axes and add the grid and so on. So it tries to give you something you want, but it gives you a bit more control. Where you want to work is really a matter of personal taste. Personally, coming from a web world, I started with charts, and charts for me were a gateway to models, but I think a lot of people coming from the data science world find plotting more intuitive. You'll see them all, and you can choose what you will. So let's start with charts. This is one little pocket of the dashboard that I was demoing, and this chart here, this bar chart here, has come out of the charts interface. So one of the things about the charts interface is that it kind of does some magic. It just takes your data and spits out a chart. So you kind of need to know what it wants the data to look like, otherwise it can't do its magic. So often when I'm getting started with a chart, especially if I haven't used it before, I'll just set up some dummy data to make sure it's roughly doing what I expected to do before I put in my real data, because then who knows whether it's OK or my data that's being weird. So we just have a simple thing here. We've spent some time with ponies and unicorns and other magic things. Now we get to importing our bar chart. So we've imported it from the charts function. We import bar, and now we just call it with our data frame, and we show the bar chart. One line, and here we have our chart. We have zoom, we have reset, we can save it, we can do that, we can resize it. And this is a web object. You could share this HTML. It's all going to work beautifully for you. So that's working how we expected to. So now we're going to try with some of our real data. All of the code that gets all the real data is in that repo that I shared with you earlier. I'm not going to cover how I made it today. This is the raw data, and then this is the sort of process data that I did. So once again, I spit out my data. I spit out my bar chart, which is almost how I want it. So now I'm going to add a couple more options. I'm going to stack my bar chart, and I'm going to set a palette. And now it's starting to look a lot like what I showed you earlier. Styling is later in the talk, so we'll get to that in a bit. So there we go. One line, bar chart, it's interactive. You can add hover, you have zoom, and there you are. What I forgot to put in was a list of all of the different kinds of charts that you can import. But at the moment, it's all the ones you might expect, line and time series and horizon and scatter and so on. So next up is IO. What you might have seen me do at a couple of the top of those notebooks is from bokeh.io import output notebook and show. And when I call output notebook in an IPython notebook, I see this bokeh.js successfully loaded. So that's how we work with bokeh in a notebook. And then there's also this output file, which is the equivalent. And so then at the top of my notebook, I would call output file and I would give it an HTML file name. And then when I hit show, instead of showing in the notebook, it opens a new browser with my static HTML has been saved and I could share that with somebody. I like to use this extra argument mode equal CDN, which instead of dumping all of the JavaScript and CSS that power bokeh inline into the HTML, it just puts in a reference to the bokeh CDN, which takes your HTML from 1.6 megabytes down to about 6K for a small plot. So it's a decent bandwidth saver. So yeah, that's how you get started with spitting your stuff out. So plotting. We've done the charts interface and now we want to look at the plotting interface. This is a plot that I couldn't build with the charts interface. There isn't a chart that's sort of pre baked for me that does this. I have this categorical axis on the left side here, splitting up the different kinds of activities that I get up to. I have a time series axis on the bottom here and then I have these rectangles that I've plotted across. So once again, I'm going to try with some test data first. This time I'm going to build my own column data source. I didn't have to do that with charts. I just threw in a data frame and it just spat out a chart. But this time I'm going to have to build my own column data source and I'm going to do it by passing in a dictionary. And then I'm going to instantiate my figure. Now figure doesn't do anything on its own. In fact, if I just do show P here, what I get is a very helpful error message. It's just come into bokeh saying you have no glyph renderers. I'm not going to be able to plot anything interesting. We're starting to this validation and error helping is new and so it's a bit spotty. So there are some bits that do it and some bits that don't. So if you ever see a place where you're like, why didn't it give me an error message, just throw up an issue on GitHub because we keep adding these in. Once we have our empty figure, now we have a whole bunch of methods on it. Remember all of those different glyphs that I showed you over and over and over again? For every one of those glyphs, there's a method. So there's P.quad to plot a quad. There's P.rect to plot a rec. There's P.circle to plot a circle and so on and so forth. For our one, we're going to use the quad method and spit out a plot. This was just dummy data so I just have linear axes on the left and bottom here. Now let's think about how we might do this for our plot. Here's our data. We have the start and the end in time of each of our blocks of time and then we have our categorical name for our label for each of those time blocks. So we're going to try plotting this. It's not looking great. It hasn't given us anything. If I open the console, what I'm going to see is although I tried to pass it some data and I really did try, if I open the console, it's going to say it could not set initial ranges. If you're thinking opening the console, really, this is where these requests for more validation can come in. This is totally something that we could have put one of those error messages in for we haven't done it yet. We should do that and I'm sure it will come. Going to the console is always a good place to look if things didn't work how you expected to. What happened was, if I had linear data, like I had up here, just a series of numbers, when I instantiated my figure, it was clever enough to figure out that I can figure out the ranges. This is no big deal. When I passed it, my data with time and categorical axes is like, I don't know what you want me to do here. I'm going to try to give me some help. Now, when we're instantiating our figure, we are going to specify the range, the time series, and we're going to specify the categories on the left side. When we do this, we get out something that looks pretty good, but not quite right. What you can see is our little blocks of time are teeny tiny. I don't know if you can even see that in the back, but they're there, but they're just really small. What happened is, how do you specify height on a categorical axis? This is a bit of a tangent, but it's a good one to know. When you have time, or 0 to 10, or whatever, it's pretty obvious how you specify height or width. When you have a categorical axis, what does that mean? The way we do it in bokeh is, from the left to the right of your categorical axes, or the bottom to the top, is 0 to 1. You append a colon to your label, and you add 0.1 for the bottom to 0.9 for the top, or the left, or the right, or whatever. Once we've done that, if we try again, now it's come the height we expected to. If I change this to be 0.4 to 0.6, they'd be much narrower, and so on. You can play with that as you wish. One of the things you're seeing here is, think about that original data frame. The chart that I showed you was essentially a switched-around version of this. The chart function did all of that magic for you. It magically figured out how wide I'm going to make things. It processed all the data, set that all up for you, so that you didn't have to think about it, which is awesome until it doesn't quite do what you need it to do, and then you have to do it yourself. To be honest, one of the ways I learned bokei was I started using the charts, and then I went and read the code for the charts to figure out how it was doing it all, and then I was able to build my own. That was plotting. A little bit more verbose, but a lot more control. You can do lots of different things there. Last but not least, we get on to models. This is where we use the lowest level of bokei to build up our charts. In the docs, I'm pretty sure I read somewhere that it's like most people wouldn't need to do this, but I like living down here, but each to their own. Once again, I've pre-processed all my data, and I've made a dictionary of data sources that I want to plot, which looks something like this, which is nice. This should look somewhat familiar. This is our first attempt at making a plot, and this should look a lot like what we just saw in plotting. We are specifying the x range and the y range, because if we don't bokei or free cow. Instead of calling a method just yet, we're specifying our line here. When we specify our line, we've imported line from bokei.models, and we're saying I want the x to be powered by this column called timestamp. I want the y to be powered by the column called cumesum hours. Then what I'm going to do is loop through and add lines for each of my data sources that I want to add. It's allowing me to add lots of different lines. To do that, I use that adgliff method that I talked about before, and I pass in a column data source, and I pass in the specification for my line. When I do that, this is what I get, which is what I asked for, but it's not beautiful. That's because down at the models level, we really do have to spell everything out. The first thing we might want to spell out is some color. Bokei has a whole bunch of built-in pallets for you that I would recommend using as defaults. You can of course spell your own pallets. Pallets are just a list of either hex or IGBA tuples or a whole different bunch of things, and you can see the reference at the pallets documentation. Let's try again with a little bit more style. Again, we just have our ranges specified. Now we're going to try a bit of background fill, a bit of border fill, and mess around with this. Now we're going to add some layout. This is where we're adding our linear axis on the left and our datetime axis on the bottom. Now our line, the specification, has got a little bit more complicated. We're not only specifying the X and the Y, we're specifying a color. We're specifying some line width and so on. Once again, we're just using this addGlyff method, use the source, use the source, and add the line. Once we do that, things start to look pretty promising. We can really start messing with this. If I just turn this into a method so I can make it nice and repeatable. Now what I've done is I wanted to add... This has no tools on it. The charts and the plotting came with all those tools just poof out the box. Now we have to manually add the tools ourselves. To do this, it's very easy. You just use addTools and you specify the tool you want to add. Now we have pan added, and that's awesome, except I don't really want it to pan up and down, because that doesn't make any sense for my data. I'm going to try constraining it. Constraining equals width. That didn't work. That's weird. What happened? Oh, thank you, Bo-K. It has told me that constrain is an unexpected attribute, and the things that it's expecting are dimensions named plot, session, or tags. This part of the error reporting of Bo-K I find incredibly useful. If I can't remember what it wants, I'll often just put in garbage so that it tells me what it's expecting, which is really handy. It said it wanted dimensions. Oh, I Python fail. I think I just deleted my cell. Hang on. No, where did it go? Oh, it's there. Thank you. No, I changed it to markdown. There we go. It said it wanted dimensions. I think it said that. It did say that, but now it's saying it got an unexpected error. It's looking for a list, and the list has to either contain width, height, x, or y, but it just got width. It didn't get a list. Again, there's the properties reporting, and then there's this value reporting, which is part of what we use to help us be able to serialise from Python to JavaScript, but it's also very helpful for debugging your code as you go. Now, has our. It pans left and right, and I can't go up and down however much I might want to try. We can keep going and going with that, but I will spare you for now. Quick mention of the examples in the Bo-K repo. There's the gallery on the website, but if you download the repo and find the examples directory, there are just a bajillion examples in there, and in particular, for getting started with models, they're incredibly helpful, so I would recommend downloading them and running them. Styling. We started to see a little bit of that just now on the models, but styling is possible for every single level, for plotting charts, models. It just depends on how you want to do it, really. I'm going to show you for charts and models and take a pic. In the interest of time, I'm going to skip the little prototyping, but this is us making a bar chart, just like we did earlier, and once we've got that, we can then just start to set all of the attributes, the toolbar location, the background fill, the outline, and so on. We go and we grab the glyphs and we can specify their fill colour and line colour, and we set the attributes, and we set the attributes, and we set the attributes, and then finally we show it, and it looks gorgeous. When you're in the IPython notebook, one of the nice things about it is the auto-complete, and the fact that you can really use IPython Notebook to help you see what the attributes that you might have available to you are, and so that's one of the ways that, especially early on, I used to play around with bokeh a lot. Now, with models, we spell it all first, right? We don't create something and then fix it. We spell it all as we go. What I like to do is, and the other thing we like to do if we're making lots of plots and we all want them to look the same, is we can start to build these dictionaries of properties, like our line colours and so on, and we can just splat them into all the axes, and we can splat properties into all of the plots. I typically build up this style sheet, and on top of the style sheet I'll have my colours and padding and things like that that will mirror my CSS if I'm building a website, and so that I can have those common colours and things like that across. Then I have this plot properties dictionary, axes properties dictionary, and so on, and I make my two axes, and I make my two rectangles. I make my two rectangles, and I make my two rectangles. This is why I like working with models, is because I'm doing a lot of heavy customisation, and I find this fairly clean as opposed to that big, quite verbose version that you end up doing on a website, and I find that fairly clean, as opposed to that big, quite verbose version that you end up doing where you fix all the attributes on a chart. But each to their own, anything is possible. Layouts. If you're just working within Bokeh, you're not interested in putting stuff on the web, there's plenty that's in the box to help you make something nice from the get-go. In particular, there's Hplot and Vplot that are in Bokeh.io. What they do is they just make rows and columns, and it's pretty self-explanatory. We've got here a Vplot, and we've nested inside it an Hplot and two charts. Our Hplot across here, we've lined up our three things. This is not my dashboard, this is just giving you a demo of what you could do. Then I've stacked it on top of some other plots. You could spit that into an HTML using the output file method, and you're going to get something that's pretty presentable, and all of these plots can interact with each other once they're in the same document together. If you spell it all in a nested thing like this, all of those things will be able to interact together. Because I typically sit in the web, if that's the very high level, I use the very low-level embed function. Again, to use this, I use the components method, and I take a dictionary of all of my plots, and what the components method gives me back is my HTML script and a dictionary of all of the divs that I have to stick into my HTML. I'll take a template, this is a very simple one, where I've just imported bokeh and maybe some CSS at the top, and then I've started to template in the divs that bokeh has given me back, and I template in very important my script. Rendering that out with the context of the script and the divs, or this works in Django, Djingo, whatever your templating language is, it's all the same. Bokeh comes with Djingo built-in, so that's convenient. You can start to build, this is an HTML page right ahead, H1 building a dashboard, my time selector, and this is the actual HTML page that I could either be rendering through a web application or as a static HTML. That's layout, and I'm pretty sure Fabio is going to talk a bit more about embedding and stuff in more detail later, so if that was too fast, there'll be more. So, with all of that under your belt, I promise that adding to it the pandas data frame method to HTML, which turns your data frame into a nice HTML table, you can make everything that you see here. These tables here are pandas tables, this is our plotting plot, these are our bar charts, these are our line charts, and we've got it all. What's different about it is I've used a nice template, I've used some nice CSS, and it all comes out very nicely. Last but not least, interactive. There's two parts of interactive. The first part has been built into Bokeh for quite a while. We have our tools like HoverSelect Zoom Pan, which we've been talking about a little bit, and the shared selection and panning. I'm going to quickly run through how to do that. Just getting started, I have a simple set of data. The first thing I'm going to do is share some ranges. What I've done here is I've got three figures that I've sat side by side. In the second figure, I've used P1's x-range. In the third figure, I've used P1's x-range and its y-range. When I do that, it means that if I pan left to right, it's going to pan out on all of the plots. When I pan up and down, because I only shared the y-range with one other plot, only one is changing. There's lots of different times when you might be wanting to see different dimensions and how they compare, so you can plot different dimensions on different plots, but start to move around in sync, which is very handy. The other thing that we've done here is we've shared a source across them. Instead of making different sources for different plots, I use one source with all of my data in it, and I shared it across my three different plots, but I plotted different things. My left plot has x and y, my middle plot, and so on. Once I've shared a source, when I start using selections, they'll share across all the different selections. I'm running a bit low on time, I'm very sorry about that. This is the spelling for how to do hovers, which I'm not going to talk about. I'm going to whip through interactions to show you. This is a quick example of something very nice I did with linking selections. When I click on a country, the name of the country updates, the value updates, when I click over on this other tab, I've also shared it across tabs. Marley has updated here. When I click on Ghana, if I go back to my first tab, Ghana will still be selected. That's a bit more sophisticated example of what you can do with that linking. Last but not least, I have four minutes and 14 seconds to whip you through callbacks, and then Fabio is going to really take this and run with it this afternoon. Callbacks have just been added to Bokey, and they're super awesome. They allow you to write a tiny piece of JavaScript, or a giant piece of JavaScript if you want to, but you write some JavaScript in your Python code, and you can start to do really sophisticated interactions. They're currently available on a weird and somewhat ad hoc subset of all of the components of Bokey, and that's the list there. As a web developer, I really love this because what I really want is that rich experience. Nothing crazy fancy, not doing computation on the server, but giving people the chance to play with their infographic a little bit. This has really allowed me to do this. I built recently, they've got this slider at the bottom, this is a Bokey slider. What I'm doing is very simple. Whenever I slide, the value of this number at the back here changes, and the data source that powers these circles switches out. This is something that is a completely standalone piece of HTML, all done in Bokey, using this new callback mechanism. If you hate JavaScript more than life itself, we're starting to, in the same way that we made charts, as these canned one-liners that will do all the magic for you, we're starting to build up these actions. A common thing might be you tap on, select a glyff, and a URL opens. The first action we've made is open URL, but there'll be more pre-built things, so you never have to write JavaScript coming. Here is an example of how to use the callback. Excuse that type of where it just says code equals blank there. We make a callback, and we have some code, like change the selected value to a new thing. Now, instead of just adding a hover tool, we add it with our callback. What's exciting here, and is a little hard to see, but when you get used to it, you realise how magic it is, is I've passed in these arguments, and what I've done is I've given a name, my column data source, and I've passed in the Python object. Now, in my JavaScript code, I can use my column data source as a JavaScript object, and I can start playing with it, and it's like I was playing with my Python object, so you don't have to do any magic, you can be playing with your Python objects in JavaScript. If you use that components method, where you have them all in one script, that's how you can start interacting across all of your different plots. Fabio is going to show you all about that later. With one minute left, are there any questions? Yes? You can shout. I can repeat it. Yes? You mean like a Google map or an open street map or whatever. The question was, can you embed a Google map or an open street map or something like that? There is actually a GMAP plot, like bar chart or something like that, like a plot that you can instantiate, which works okay, but we wanted to make it better, and one of the things I meant to be doing this summer is actually working on our mapping support in general, which is not as hot as we would like it, but it's really high on my to-do list, because it's something I want to see done. Sara told us in the afternoon we have the possibility to learn a little bit more about bokeh, so I want to thank you for your talk. I'll see you in a minute, and I'll see you in a minute, because now...
|
Sarah Bird - Getting started with Bokeh / Let's build an interactive data visualization for the web..in Python! As a web developer, I find myself being asked to make increasing numbers of data visualizations, interactive infographics, and more. d3.js is great, as are many other javascript toolkits that are out there. But if I can write more Python and less JavaScript... well, that makes me happy! Bokeh is a new Python library for interactive visualization. Its origins are in the data science community, but it has a lot to offer web developers. In this mini-tutorial, I'll run through how to build a data visualization in Bokeh and how to hook it into your web application. This will be a real-world example, that was previously built in d3.js. Along the way, I'll provide tips and tricks that I've discovered in my experience including how Bokeh works wonderfully with the iPython notebook which I use to prototype my visualizations, and many data science people use as their native way to explore data. For those of you who already know a little Bokeh, I'll be covering the new "actions framework" that lets you write JS callbacks in your python code so you can do lots of interactions all on the client side.
|
10.5446/20113 (DOI)
|
Hello, so I'm Flores, I've been a PIDA test contributor for a couple of years now. Recently PIDA tested split off its plug-in system that it's been using and I'm going to talk a little bit more about how that plug-in system works, how you write plug-ins with that. PIDA test itself is a testing tool that hopefully all of you know and hopefully use as well. It's been around for a long time and it actually has a lot of plug-ins itself. I think there's over 150 plug-ins, that's very unscientific, that's just the number I've quoted from someone else. There's a lot of plug-ins and they've been around for a long time and its plug-in system seems to work quite well. Recently it's been split off because it's interesting to try and it was nice to be able to use that in other projects as well, the plug-in system. That's being called plug-in. It's on GitHub actually, under Holger Cracol's username, it's the repository. It's a standalone version of the PIDA test plug-in system, it's a little bit different than what it was originally. It's very similar, it's just a couple of details and PIDA test specific things that have been moved out of it. One note of caution though is it's still not version, we haven't been releasing it as version 1.0 yet. It's using semantic version so in theory you could break the API at any point. As I said, the API is basically taken from PIDA test so hopefully that shouldn't change too much. A little overview of basically how I'll go through the talk. Basically I'll kind of with a simple example introduce how plug-ins work in this sort of world. I'll talk a little bit about what advantages that brings and then towards the end of the talk I'll talk about this. It's basically designing your entire application consisting out of plug-ins which is what PIDA test itself does as well actually. I think it's an interesting way of looking at an application. You probably all know what a plug-ins system is and why you'd want plug-ins in your application. Basically it gives you certain points in your application where you just allow other codes that you don't know to execute with some extra information etc. The benefit is that people can extend it to the things that you never thought of being useful or basically make it send email or something. Everything sends email in the end. The plug-in approach to it is one of hook functions. There are several ways to write in the plug-ins I guess. The idea is basically that you find your extension point as a hook that the application then calls. At the point you want to extend the application you call that hook with some arguments. Any of the plug-ins that have been loaded are free to implement that hook. They all get called. It's a one-to-n kind of call mapping. Plug-in does support one-to-one as well in a way. I won't really go into much detail about that. It supports that sort of functionality as well. I'll start with a very simple example of writing this. This is basically just the application that I'm going to use to demonstrate it. It's very simple. It just gets a URL and prints it to standard out. I'm using the requests a little bit funny here. I'm breaking it out so this is basically request.get. It's spread out over four steps to a slightly lower level API you can use in there. That's because it will allow me to extend the application. This is very straightforward. It doesn't do very much. The first extension point I'm going to create is basically allow the plug-ins to modify the headers that I send in the request. In this case, that means that after a request that prepares, that object has a dictionary of headers. I'm going to allow the plug-ins to modify that dictionary before actually sending off the request. To do that in plug-in, you define your hook points, your extension points with hook specifications. Generally, we use hook.test.hookspec.py to write this in. It's a simple module. All you need to do is basically create this mark for your application first. The hookspec that you can use as a decorator at hookspec decorator there. All that you really care about, or all the plug you really care about here, is the signature that you define. That is actually important. The name of your hook and the arguments it takes are important. Because this is your API, it's a good idea to write a nice dockswring about it and explain what it does. The name itself, again, by convention, prefix it with your application name or something, that is not strictly necessary, but that's fine. The last thing is that there's absolutely no code about it. This is purely about the function signature. Having defined the hook specification, let's just skip ahead and look at what the plug-in implementation is. That is really straightforward. This is a different module again, plug-in.py in this case. I'm using for the example. Again, to write in a hook implementation, it's called you need a decorator to decorate your implementation. When it loads your plug-in, it can scan through your plug-in and find your hook implementation using that decorator. Generally, the application exposes that. As an API to the plug-in writers, I'm importing the application itself, curl in this case. The rest of the implementation is really simple. The argument is just a dictionary and I modify it. Do you notice that the implementation that I used only uses the one header's argument? If you go back to the hook specification, it actually takes another session argument as well. The plug-in doesn't force you. It basically allows the hooks to only accept the arguments that they need, and plug-in will look at the signature of your hook implementation, and from that, will figure out basically which arguments you need. That is a good feature when your hook's evolving and you're extending them, because that gives you easier backwards compatibility. To actually look at what you need to do to change application, this is still fairly similar. I'll give you a couple of more imports. The first thing we do here is the hook marker to create that decorator. That was just the public API that we decided to use. The main application is the first four lines are concerned about creating this plug-in manager. It creates a plug-in manager. Again, give it a name. I imported the hook specification here. The hook specification is just a module object in this case. You register all the hooks and using that decorator in the hook specification, plug-in will scan for the hooks and find them, et cetera. In this case, I'm using just a simple import lip to dynamically load my plug-in. It's a kind of hard code here. Then you just need to register the plug-in. The following is basically the same. It's like the earlier request stuff, so great session, great request, et cetera. Then actually calling the hook. I got the plug-in manager object. The plug-in manager object has this hook attribute, which is a hook relay. After you've added hook specification, for each hook that you have defined, it will create a callable in there, which allows you to call your hooks. This is where you're actually calling the hooks from in the plug-ins. In this case, the thing to notice also is that when you're calling the hooks, you have to provide obviously all the arguments because you don't know which arguments you're plug-in using. You have to specify them or give them a keyword argument as well, because plug-in looks at the names of your arguments to know which arguments that your hook implementations need. They need to be pasting as keyword arguments. If you forget an argument or get it wrong, unfortunately, as I discovered while I was in the slides, you don't get a terribly useful error message at the moment, but hopefully that can be improved. The rest of the application is just the same. That is essentially everything that you need to do to create plug-ins and start using plug-ins. One more thing, basically, because the hook that I wrote didn't actually return any value. You can have multiple plug-ins all implementing this hook, and they would all have passed into them, and they'd all be modifying the same dictionary. When your hooks want to return a value, that's also possible. Basically, it returns a list of all the return values of each hook, and then your application has to decide what it's going to do with the hook. As a quick example for this, I'm going to add another hook specification. This one is not much special. It's not the best example. I couldn't really think of anything much better. The idea of this extra hook is basically that it will return to a false, brilliant, whether you plug-in things that you should make or request, so plug-in can deny a request or something. Hook specification service is pretty straightforward. Plug-in is a plug-in as well, and it's really straightforward to implement. I don't really care. I just don't want to filter anything, so yes, I just return true. In fact, I didn't even have to take the arguments in this implementation, because I'm not using it. I could skip that as well. This is actually the application. It's getting a bit big, but not much has changed really. Everything up-to-the-first hook is being called the prepare headers. The hook being called is exactly the same. Now I'm doing after this. This is exactly the same. Basically, call the new hook. It returns a list. I just built in all plug-in here. I'll function in Python to just see if any of the hooks doesn't like it. That's it. I just send the response, send request, bring response, all that stuff. That's using return values. That was the very short introduction to how to write plug-ins. There's a lot more features, actually. Essentially, the application defines certain hook points. The thing that matters is that they are function signatures, so the function signature matters. The implementations of plug-ins, the arguments are optional, so if you don't need arguments, you don't have to use them. You get one call in the application resulting in lots of calls in the plug-ins. There's a bunch of more advanced features that I haven't really gone into. It's just like the entry points set up to the entry points integration, so you don't have to write that all from scratch if you want to use set up tools. You can also influence the hook ordering a little bit, so sometimes that might matter, so you might want to run a plug-in. You might care that it's hooked because you've been run early or late or something like that. Hook wrapping is something similar. It's just there. Your plug-in actually gets called before all the other hooks get called, and then it gets the result back as well. It gets to run code at the beginning and at the end of all the other plug-ins in a way. Doing that gives you access to, allows you plug-in to actually see the results that the other plug-ins have produced and you could even modify them or return something else or something like that if you want to. The last interesting feature that you can use is basically plug-ins. Writing plug-ins for your plug-in. If you just pass the plug-in manager object to your plug-in, there is nothing stopping from adding new hook specifications to the plug-in manager, which then can be called from other plug-ins, et cetera. It's a slightly unique way of writing plug-ins, I think, in the Python world, compared to having static classes and your functions and signatures being strict. It allows you more flexibility because it doesn't force you. It's just a function. It can implement it as a module as it just did. You can implement your plug-in as a class as well. The feature where basically the hook implementation doesn't have to be request or doesn't have to use all the arguments, it also allows you hooks to evolve a lot easier because you can change the API by adding new arguments and all the plug-ins will keep working. That seems to have worked quite well for Python tests, really. The other thing is that because there's no class or anything involved, it doesn't force any behaviour or workflow or state that you have to keep or anything like that. If you want to, plug-ins can keep states, they can implement themselves as a class as well, so you can keep state, but it also leaves if you have a very simple plug-in, it is very simple, it doesn't force anything extra on you. Next, I'll talk a bit about how you can actually use that to design your entire application. This is an interesting way of thinking about things. Basically, at SMBs, you saw creating and setting up your plug-in manager is very small. All you need to have is a very small bootstrap module that will import some core built-in plug-ins, and those plug-ins can then be responsible for running your entire application, and they can in fact be responsible actually doing more setup work, so you can just use a couple of hard-coded plug-ins, and then those plug-ins will be responsible for looking at setup tools and points, and learning more plug-ins, or using namespace packages, or whatever kind of mechanism you'd like to use for that. This sort of approach has been used, and the test has been pioneering that, as far as I know, and that's sort of for command line tools, short tunes. I've also used it for long-running demons as well, so it scales very well for different types of applications, basically, and the interesting part of that is obviously that your entire application exists after a few plug-ins, and if you're using your plug-in API, it ensures that you make a useful plug-in API. It makes for very flexible and extendable applications as well. Basically, if I keep modifying my little toy application here to start and do that, is what I'll be doing in the next few slides. In here, basically, I'm replacing the main function here, taking a more classical argument in this case. In this case, all it's doing is basically the same, creating a plug-in manager. As you can see, I've actually created a list of core plug-ins, and I've just implemented this core plug-in, because I can hard-code a small set of plug-ins. I just import that already, and when registering the hook, I just iterate over my core plug-ins, because it's just one register. Once that's done, I can just call this a new hook that I've specified, and that's the end of the application in a way. There should only be one plug-in that really implements that hook, although all the plug-ins could implement it to override it, but it's slightly more advanced. Then just use that return value to exit the application, and that's all the rest of it. This is the new plug-in arrow, core.py in this case, implementing this sort of workflow. This is actually now the one that drives this curl main hook, is the one that drives the rest of the application. You'll notice I've done a little bit more here than strictly necessary. When writing your entire application, this way around, inside out, as I sometimes think of it, it's useful. You'll notice after creating a plug-in manager, I create this configuration object, and later on I create a session object. It's a nice abstraction, and it actually works quite well. Generally, I make the configuration object, and I make that responsible for doing things like passing the command line, reading config files, et cetera. That provides some application states, which are essentially your static configuration. The session, because of the request session that we were using earlier, I had to call it CLI session, that session object allows you to keep runtime state about your application if you have any. It's a nice pattern. Once you're doing those two, as you can see, I call the hooks Configure and Session Start, and Session Finish and Configure. I'm not actually going to use these hooks in this example, because it's quite short. It's a nice pattern, and it allows you to hook in on those points, and do extra static configuration, maybe enrich a static configuration, things like that. After that's done, you can see I've moved the argument passing into the config object. I'm just passing my RV. Basically, everything indicates the rest of the call of the application I decided to implement in just one hook. I could have split that out more, but it makes the call to look at a bit more. This one call make requests hook will be responsible for executing a code that we saw earlier. Because it didn't all fit on the same slide, this is still the same module. Just showing basically that config object, so I'm not doing anything clever at the moment, I just still know that I just passed in my URL. Basically, creating a little bit of stage in the CLI session one. It seemed like a reasonable idea to put the request session in there. If my application grew some features where I would request multiple URLs or something, they could share the HATB session for that. So they all get their cookies, et cetera. The other new hook I created was this make requests hook. The implementation there is basically very similar to what we saw before. It's exactly the same as before. The only difference here is basically that now the URL is taken from the config object and the HATB session is now taken from the CLI session object. That's kind of the only difference. One thing like, I think is a nice example in this way of writing your application, is sort of the handling of configuration in a bit more detail. Because it's basically, you can even like creating your argument path to generate minus help, you can let plugins interact with that. It's nice because in this case I'm implementing a plugin that's basically responsible for making this request. The URL argument on the command line, I'm actually making a concern of the plugin that actually wants to use this. It's a nice way of separating concerns. There's no need to have one central argument passing thing that needs to know about what everyone wants to do. Again, pretty straightforward to implement. All I'm doing here is this config object that I already had. I was already responsible for parsing the argument lines. I'm actually doing it properly this time around and creating my own pass parcel. The small trick basically to get your arguments passed by plugins is just define another hook basically before you pass your arguments. You pass the parser around to all the plugins that want to add any arguments. In here I've just in the same plugin directly implemented that hook as well, adding my URL argument in here. That's it. The make request changes a little bit here. Again, not really anything significant just because the URL is now in a slightly different location on my config object. Everything else is very straightforward, very similar. This is a way, as I was saying, that the pilot test itself uses this pattern as well. This is a very simplified look of the hooks that pilot test defines and how it calls them. It doesn't cover everything because it's a little too short. Also because it's very complicated. Oh well, reasonably complicated. You can spot the same pattern. It starts with creating its own version of the plugin manager. Then it creates this config object that helps in parsing of the command line and just the next few things. You've got the pilot test add option which adds the option, then command line parsing is implemented as a hook as well. Then it basically spawns off this command line main one. That is then going to be the hook that drives the rest of the application. In there you basically see the same pattern as I was saying earlier. You create a session, then you give all the plugins a chance to hook into those configuration and session setups. With the pilot test configure, the pilot session starts, then finish and then configure at the end as well. The main work that pilot test does is then split up in these two main steps, which is like a collection of all the tests and then running of all the tests. Then inside those, there is again parsers of the work to even more and more hooks. There is more hooks than are actually shown as well. It shows that in large applications, if you think about how you structure all your hooks, etc. It's an interesting way of thinking about designing applications and it works quite nicely in some situations. To summarise a little bit, it's an interesting plugin system. It allows you to evolve your plugin API quite nicely. It forces very little on your new plugin writer, so you can keep simple things simple. Likewise, if you do need to store state and all that sort of thing, you can do this as well. Despite the fact that I've shown writing your whole application as a plugin, there is no obligation at all to do this. The first example I started from a static existing application, it's very low overhead. All you need to do is create a plugin manager, so it's actually not that hard to have. If you have an existing application that you want to grow in a plugin system, there is no need to go, you know, this whole everything needs to be a plugin design. You can just go very gentle, traditional kind of adding a few hooks and there's not much overhead there. I do find it an interesting way, if you're starting from scratch, it might be interesting to consider. It's kind of a fun way to try and design your application and it works quite well in the right situations. So, that was what I wanted to say about the plugin. Any questions? Hi, thank you for the great talk. I also find this very interesting. But when I think about it, usually I would do that with objects. One object overloads another and that way extends the functionality. This is much more dynamic in a way and it gives me the feeling that you can really plug them in at a certain time. But I was missing that. How do you now configure your application together from plugins? Is that now intended in this design or not that the user, when the user has the application, he can have a variable set of plugins, right? But how is that configured? How does it activate or deactivate plugins? How is that configured? That's where I was hand-waving referring to set-up tools integration. It leaves it fairly open to the application itself. Typical things are set up to all entry points. If you've got any distribution, it's the right word apparently. If you've got any distribution installed at the entry point, they'll be registered. Your application can decide if that doesn't give you enough control. You can do command line switches, you can use configuration files. It leaves it up to the application to select which plugins are used. You'll have a set of corp plugins that will provide the basic functionality. There's a new functionality in the new version of Python, which is called functions.singleDispatch. It's a generic function mechanism. It can be quite useful for some types of powering infoctionality when you want to dispatch based on some type. Is there a way to integrate that with a plug-in or a best practice to do that? Or you just use them side by side? So generic single dispatch is not the one you're talking about. I haven't really used that much myself. That's where it looks at the type of the first argument and then decides which. Can you register new implementations on the fly? I'm not really sure how to answer the question. I haven't really thought about how that interacts. A plug-in doesn't really do anything like that. You just get your arguments in the implementation. You say, if it's this type, I don't care, and just return on, then it doesn't influence anything. Hi, so if you have multiple plugins, it's up to the application to take the results of a hook call from each different plug-in and then somehow integrate them. So can you talk a bit about how... So that's then something that the application author needs to think about, I suppose, if two different plugins want to, for example, modify the headers. And also, if you're passing dictionaries to your hooks, then the order really matters, right? Because they could be modified in place. So the order can be relevant sometimes, as you say, and that's where plug-in has... So when you're implementing your hook, you can say... So the decorator, I didn't go into this, but the decorator takes... You can basically call it like a function and then give it keyword arguments. And then you can basically... You can influence the ordering a little bit so you can say, I want to be called at the beginning, I want to be called at the end. So obviously, modifying... Yeah, when you're modifying an object, you don't get a very strict ordering guarantee. The wrapping method was another one, basically, where you will be called at the beginning and at the end. And the other part of the question was... Well, I guess if you're the application author, you need to think about what kind of plugins people might write, I guess. Yeah, the way you return... If multiple plugins return a value, you get the list with them. Yeah, your application... I mean, there is also one of the things that you can do on the hook specification, as I said. I refer to you, you can do one-to-one calls as well. Basically, on the hook specification, you can say something like... If this list of return values is not very useful. I skipped over this detail when I was showing the inside out architecture in a way where the main function... So if multiple plugins were to implement main, something funny would be happening. The way to solve that is by specifying that hook, you say, I want to create a one-to-one call, really, which you do by adding the first result, I think, keyword, or use first result or something like that. And then, at that point, you get a plugin ordering going as well. So at that point, if your plugin could go... Actually, I want to be called before the normal main one, for example. And then basically look at, do I want to... I can decide, do I want to return a value and actually become main, or do I just return none? And if I return none, then the next hook will get a chance to be main. So, yeah, that's kind of... Thanks. Just another one about return values. Is there any way to map what return value has come from which plugin? Strictly speaking, not really, unless you happen to know the order of the plugins. So the plugins are actually... So if you tightly control the order, so when you're the plugin manager, I was calling.register.thePluginObject, that order is actually respected. So if you know exactly the order that things are registered in there, and if you know that all your plugins implement the hook, because mostly only the plugins that implement the hook will have a return value, then you can sort of... Yeah, not easily, no. Because if you're dropping none as well, then you're going to... Is there anything to consider when testing plugins? Sorry, say that again? If you want to test a plugin, is there anything special you have to do to test it? I... Not really, because the way the decorators work on your hooks makes it quite nice because it doesn't actually affect the function at all. It just marks up the function, which means it's still the same function object as you see, so in your test you can just go and call it directly, and it actually makes testing easier in some way, yes. Sorry, so the question is, plugins can be disabled from the outside? So the question is, can you disable plugins? You can essentially, so you got in your plugin manager where you do register, you can also do unregister, and that's how you disable a plugin. So, without any point, you can iterate over, you can see what plugins have been registered in the plugin manager, so if you don't want to go, you can iterate over it and look like, oh, I don't want that, I want to be loaded, and you can go and unregister it, and... Yeah. I have a little concern in the benefits slide. You mentioned that you can basically pass any argument you want, but that's more or less the behavior you would expect from JavaScript and from Python, because this could lead to some nasty bosses. So you can't pass any argument you want. It has to be, so when you... So in hook specification, when you actually... those are the arguments that you have to implement, and you can't call it with anything else if you don't... The only thing that you can do is, on the next slide into implementation, you can leave out one, and that is basically just implementing a signature inspection. So it is... it's maybe a little bit unconventional in Python, but this is what Python allows you to do. It provides you all the runtime inspection things. If you used to Python test and it's fictious, it will seem very natural to you as well. Yeah, it's okay. That's a little bit older first, but it doesn't really violate anything terrible, I think. It doesn't allow you to pass in on existing things either. It's still quite strict on that. It's for the recording and for the audience to be nicer. Hi. One thing is I'm thinking about because plugins, normally they shouldn't be part of your application, and normally on applications you start to do something, on application exits you do something else. If everything is a plugin, how do you manage what application is starting? Because there is no application if everything is plug-able. You know what I mean? Because if you can not disable a plugin, it's not a plugin. So I'm a bit having a problem understanding the plugin approach. Right. Yeah, a plugin could completely screw up your application if you really want to. But that's coding in Python. You can screw up everything anyway if you really want to. What's the term? Contenting gentlemen or something like that? In practice I think it's not really an issue. As you said, it's been in use for a long time in Python tests. You could completely replace the main loop and go and do something else. But at the end of the day you can't help anyone either if you do that. All right. We are out of time.
|
Floris Bruynooghe - The hook-based plugin architecture of py.test The hook-based plugin system used by py.test and being made available as a stand alone package allows easy extensibility based on defined extension points which can be implemented using hook functions in the plugins. Plugins can themselves call these hooks as well as define future extension points allowing for a very flexible design. py.test itself uses this plugin system from the ground up with the entire application being implemented by built-in plugins. This architecture has proven powerful and flexible over the years, on both command line tools as well as long running daemons. This talks will describe how the plugin system works and how it deals with passing arguments and return values 1:N hook calls. It will also describe how to design an application consisting entirely of plugins. While not specifically talking about py.test it will also give a solid understanding on how plugins work in py.test.
|
10.5446/20111 (DOI)
|
Welcome, welcome, welcome, welcome, welcome, welcome, welcome, welcome, welcome, welcome, welcome, welcome, welcome. I will speak about the other structures, discipline with Python. I'm a professor, computer science professor at Fatec. Fatec is a public university at Sao Paulo, Brazil. I love teaching. Data structures as a difficult discipline to teach is also a difficult discipline to students. It's also a difficult discipline for my university. My thesis is C language has a problem with syntax. I like C language very much, but sometimes C have a bit of dangerous problems to students. I love data structures because algorithms have a lot of independence of operating systems and languages. For example, binary search, the number of steps or measures are independent of operating systems and languages. The number of steps is the same in the windows, Linux, or in C, or in Python. Data structures at my university, Fatec. 2008 is a bad year for me because we have this number of students retain it. But with Python, in this year, only 10% we have improvements with the change of the C language to the Python language. In this talk, I will show the improvements with some codes. The summary, 2008, retain it for students. When I submitted the talk, we have improvements of 12%. Now, when I finish my semester, we have two points of improvement. In Brazil, we have a national contest, a national 125th grade. My university received the maximum. We have received the contest programming first last year. It's a good before the change to Python and 2009. In 2009, we moved all the classes to laboratory classes. We have four classes for week. No chalkboard classes, only laboratory classes. We have four lab projects for semesters. All lab are in Python. We introduce a strange name, Big Brother. Some students, best students, could help the other students as coaches. We maintain C language only for proposal to compare the algorithms. All the algorithms are in two languages. C to show the details. And Python to show the essence. Both the students, the exercise are submitted in Python. The projects are submitted in Python. The C language only to show the details only. The majority of the university in USA are using Python to introduce courses of programming. Because usability is a problem for introduction to programming. I think usability is a problem for data structure courses also. Donald Kinnut in an interview of people of ACM said the most common fault in computer class is to emphasize the rules of specific programming language instead of to emphasize the algorithms that are being expressed in those languages. Sometimes the teachers are struggling in syntax problems of C languages or Java languages instead of teaching the algorithms. That's the most important. Donald Kinnut showed me the code. I will show a lot of the code of my course. As in Python variables are just names. All the variables are pointers. A42, A Python are pointers. If you have a list, if you have a segment 42 to A index 0, A, B are pointers to the same list. A0, B changes to because A and B are pointers to the same object. If you need a new object, you need to explicitly create a new object. It's good to a professor, to a teacher, to explicitly create a new object is good to teaching. Python is cool. There are big integers. In Python, Python 3, there are natural division in C language. I think that's strange. 1 divided by 2 is 0. For beginners, it's a bit strange. There are cool things like multiple assignments. There are no need to attach variable A, B to swap for B. All manipulate year, month, day like this. All the middle, like this list. All phones. And put these phones. With Python. Identation. The programming activity shows the ability as a process of creating work of literature, and it's a very important thing to be read. The C language has some indentation problems like this. If you put this, you print only one text. This is a problem also. Is a false indentation. This indentation confuses the students. It's not a problem only in education, but also in corporations like Apple. This is a code in Apple. Some guys go to fail. Go to fail. This is a crazy. Bracket, bracket, go to fail, go to fail. Who? Go to fail. C language has some problems that is very dangerous. Not only for students, but for corporations. Recursion. To understand the recursion, the most first understand recursion. People not have a problem. People not recursive. We solve this using dictionary, for example, with Python, have LLU cache. We can also show this high level to students. This creates some solutions to that to being problems. Because the dynamic aspect of the types. Linked lists. The implementation of linked lists in C language has some details. For example, we need to have a head in order to avoid the special case to analyze the empty list. This head, we need to put the reverse order 3 to 1. The linked list. We don't need to do this in Python. Kills. We have this matrix, this node 3. We need to calculate the distance of the node 3 to the other nodes. Node 3 to node 4 have one point of the minimum distance. We calculate this in C language by these algorithms. C language have some issues, some problems, because the int star that returned in this function. Int star in C language have two meanings. It's a pointer for one integer, on the other hand, is a vector of a dynamic number of elements. Sometimes it's a bit confusing to students. In Python, it's a more clear algorithm. It's more direct and readable. Stacks. We need to see the well-formed expression. The equivalent of this C language algorithm in Python, it's much more clear than the C language. It's a direct way to see the essence. One sort, the manipulation of the index to see the minimum of the vector and swap for the first is a bit more complex than this algorithm equivalent in Python. Of course, we need to see the complexity. But if the complexity or the number of the steps is the same, it's better to use the more readable algorithm to students. Quick sort. The quick sort is a very difficult algorithm to teach if you use, if you see the C language, you have a pivot, you need to put all the smaller elements to one side of the higher elements to the other sides. In Python language, we have a list comprehension. It's much more clear to the students the concept of the smaller of the higher. We have a pivot. The smaller elements is a list, the one side list, the higher elements is the other side. We solve the problem of the in recursive manner with the smaller and the higher. In C language, it's not clear the essence of the algorithm. This is the most important. Some exercises like the word count, we have a text. We need to count how many occurrence of the word have in one text. In C language, if you have to count the number of occurrence of Alice in Wonderland, it's SMS, we need to tokenize to memory allocation to have many pointers. It's a crazy to code this program in C language. In Python, for a teacher, Python is good because all the code is put in one slide. The code count program in Python is an exercise in ten minutes. We open the book Alice in Wonderland. We read, remove the special characters, and we use a dictionary to count the words. That's enough. Simple, direct, and displaced. Some projects. The first project that my students work is a comparison with a magic sword, quick sword, selection sword, and a native sword, a Python, native sword, is a good practical project. The students like it very much because the time of the native sword is so good. The team sword is a modification of magic swords. The second practical project is a simplification of art or mailing games. We have some girls, some guys. We need to arrange a marriage of the girls to guys. We have kynites. We need to put the kynites in a table. The code to have made the numerations of the guides is a code to generate the sub-sequences. Lexicosequence is used, the yield operator to improve it. It's good to students to know the difference of the return of the yield. The third project is to detect the binary regions. This region is connected. This run is connected. It's a problem in simple to announce, but difficult to resolve. It's a cool project. The last is a graph theory project. It's a risk to solve the minimum degree, greedy to maximum independent set problem. One, two, four, and six is the maximum independent set of this graph. The implementation of this problem in Python is not so hard. I will show is 20 lines of code. In C language is 500 lines of codes. Conclusions, there is a trade-off to choose a language to teach. C is good for optimization. To see the details, the low level. But Python is also good to show the essence of the algorithms, the high level. But if the algorithm is the same, the number of steps, the complexity, parameter optimization is also in teaching data structures. Thank you. Thank you very much, Fernando. Do you have any questions? No, no questions? Okay, then. Thank you all very much. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you.
|
Fernando Masanori Ashikaga - Data Structures with Python Data Structures is traditionally a “bogeyman” discipline in Computer Science courses and has a high degree of failure. In FATEC São José dos Campos we are adopting a hybrid approach, with C and Python languages. The failure rate decreased from 85% (2008) to 12% (2014). The talk will be extensively illustrated with code in C and Python, addressing the various concepts taught in this course: recursion, linked lists, queues, stacks, sorting algorithms.
|
10.5446/20110 (DOI)
|
What is the driven? That's the important bit, because that's what actually changes when you actually do it like that. So it's a software development process that is based on a very, very short cycle that you repeat very often. You start by writing a test, and then you make it run. And it fails. And this is called the red phase. And then you write the minimum amount of code to make the test pass. And then you write the test, and the test passes. And this is called the green phase. And then because you have written the minimum amount of code, maybe also a rough code, you want to refactor. You can refactor both the code and the tests. I suggest not at the same time, because it's not wise. And when you start with it, it's really useful if you take very small steps. So you write a little bit of test code, and then a little bit of code to make the test pass. You refactor, you go back. You add another little bit. You make that pass, you refactor, and so on and so forth. And the more you do this, the more in your brain, basically you learn how to work with this kind of back and forth between tests and code, which means that at some point, you're going to be able to take longer steps. And then you go like, I'm taking longer steps. And you take them really long. And at some point, you're in trouble. And so you go back to short steps, you fix whatever you have to fix, and then you try increase a little bit more until you get some comfortable size for you to work with. But where do we start? It all starts with a business requirement. So that's very important. You have to understand the business requirement. And then when you are clear on the business requirement, you enter the TDD cycle. You start writing a test. You make it pass your refactor. Test, pass, refactor, and so on and so forth until that business requirement is fulfilled. And then you get to the next business requirement. The frog at this point thought the training was completed. But the masters disagreed. And they kept giving examples. They said, this is just the mechanics. If you know how to use a steering wheel, the gear switch, accelerators, and brain, but you never driven a car, would you call yourself a pilot? Of course not. And I've talked to some people, some coders, they say, I know all about TDD. I read an article. But have you ever tried it? No, it's not really for me. Well, if you don't try it, you don't know. So you need to try. What changes in your brain when you use TDD? Without TDD, you go from the business requirement to the code. And when you think about the code, you have to take care of two things. What the code needs to do? We need to deliver this functionality. And how it needs to do that? We need to cycle using a for loop. We need to call this function. So there's two very different things, the what and how. And we take care of both of them at the same time. So our brain is fully concentrating on two things. On the other hand, when you use TDD, you basically mostly take care of the what when you're writing your tests. Because the tests are testing what should happen in the code. And then you take care of the how when you're writing the code itself. So you have your brain concentrating on each of those two different times, which means it's like having two brains. It's much more powerful. Your code changes immediately. There are some common aspects to working with TDD. The key is principle, to keep it simple stupid. Which means basically by forcing yourself to have a test that represents something that your code needs to do. And to, again, force yourself to write the minimum amount of code in order to deliver that. The code basically remains simple. You cannot go and take care of super architectural things or whatever. The code automatically stays simple, which is simple, is really, really important. And then the Yavni principle, you ain't going to need it. If you're focusing on understanding the business requirement, writing a test, and making it pass, it's not likely that you're going to over engineer your code. Because you don't have the freedom of a code that says, I need to do this. And then while you're coding, there's nothing telling you what that code should do. It's just in your mind. And you go, oh, I'm also going to add this. Because maybe tomorrow I need it, which is something that with TDD you don't do. Three strikes in refactor. I've taken this off the test-driven development with Python from Percival, which is a very nice book. I just read it a few months ago. That basically says, when you're in the factor phase, and you find that you've got one functionality and basically the same functionality, just wait for the third occurrence of something really similar. Because if you group these and factor out the mix-in or whatever too soon, when the third occurrence comes out, you may realize it's not as easy to make that work with the mix-in that you just done. So you may also have to do a lot of refactoring work. On the other hand, going for four or five, that's not good. So three strikes in refactor is a nice balance to achieve. You can do all the architecture design when you refactor the code. And the beauty of it is that because you have the tests, you can refactor with confidence. And then you do triangulation. Triangulation was puzzling for the frog. So let's see an example. Let's say that you're writing a test, and you're just making sure that your square function that takes minus 2 is producing number 4. If you think about it, at this point, all we have in the code base is to fulfill that requirement, which basically says, I just want to know that your function returns 4. So you can do this. You can cheat. It's actually suggested by TDD authors. And it's called fake it till you make it. But of course, we don't want to have that code in our code base, because that's not correct. So we do triangulation, which means we pinpoint to the same function from two different angles, which would cause us to have to fake it in two different ways, which is not compatible. At that point, we need to write the real logic. So you write the actual logic when you have triangulation like that. And this is very nice, not just because it's like a theoretical thing, but it's because you get something done, you get a test that basically is there. And then when you triangulate, you have a test there. It builds on your test base. So the main benefits are you can factor with confidence, because you have a set of tests that when you touch the code and you change something, like in the first example, we had the boundary that was jiggling a bit. You have a test that fails. Readability, because it's much easier to see code that was designed with tests first. Because basically, even though it's like a given, you take care of the design part when you're writing your tests, because when you're writing something, unless you're writing integration tests, but if you're unit testing, you have to test a unit of code. So you have to think, how do I structure this code? You can't just write something. You have to give it a thought, which means you're thinking twice about the code. When you test and when you actually write it, which means it comes out much better, much more readable. It's more loose-coupled. It's easier to test and maintain. It's easier to test, of course, because it's coming from the test. And it's easier to maintain, because it's well structured. And when you do test first, you also have a better understanding of the business requirement. Because in order to start writing your tests, then you have to have very, very clear in your mind what is the business requirement that will drive the design of those tests. If you're not clear, you will find yourself blocked at the test level, which basically will prompt you to go back and try and understand better the business requirements. And then by testing in small units, it will be much easier to debug. And also, you will have the perks of having the tests act as documentation. Because by having small tests, easier to read, you just go through it very, very quickly. And you say, oh, OK, my code does this, which is sometimes even easier than read an English sentence that describes that functionality. Because English can be misleading, especially if I write it. So having a test, which is Python code, that you know what it does, it's very useful. And higher speed. It takes less to write tests and code, than to write code, and then having to debug it. I can tell you this by my personal experience. I was working for a company called TBG. We were competing with other companies to get the preview access to Twitter's advertisement API. And we had to deliver a proof of concept in about six weeks. We succeeded. It was a monolithic Django application. And the order from above was no tests. We didn't have time to do them. So no tests. We do the coding. We do overtime. We go on Saturdays and stuff. And the last two weeks were spent just to fix two bugs that they drove us crazy. And it was just one small Django website. But it grows so complicated, touched in such a short amount of time by six, seven people, that basically we were going crazy debugging. Because we were changing something here. It was breaking something there. You go fix it. You break something there. And then you've got a ripple effect that flows through your code. On the other hand, had we done tests first, that time that we spent debugging, we wouldn't have needed that. The main shortcomings of this technique is that the whole company needs to believe in it. Otherwise, you're going to fight all the time with your boss, which is something that me and some of my colleagues know very well. We need to write the test. There is no time to do the test. But then we'll have to debug. It will be a problem that we will have later. So we go and write just the code. And then, ah, it doesn't work. It's your fault. So you don't want to, you have to really, really convince everybody that TDD is the way to go. And because it's really hard to see what happens in the long term. All we, myself included, most of the time, we just see what happens tomorrow. We need to deliver tomorrow. We need to make the client happy tomorrow. And we tend to forget about the long term. Blind spots, if you misread or if you don't understand, or if they are incomplete, the business requirements, it means this will reflect in the type of tests that you write. And this will also reflect in the code. And without the business requirement part, take a look at the tests. Perfect. Take a look at the code, makes the test pass. We're done. So the code and the tests, they would share the same blind spot, the same thing that you missed from the business requirement. So in this case, it can be harder to spot something. Pairing, for example, helps on this. Because you have to discuss what you're doing, which means you bring up a discussion about the business requirement. And then you realize, oh, I understand it this way. My colleague understands it that way. You go ask whoever it is that you have to ask for clarification. And then this does not happen. And also badly written tests are hard to maintain. For example, tests with a lot of mocks, they're very hard to maintain. Because when you change a mock, you're changing an object that basically is just a puppy. You can do whatever you want with it. And you're not really sure if you're breaking something. Because you just change a mock. And you make it do something else. And then you make your code pass. But if that mock is representing a real object, as it should, and that object is no longer in sync with the mock, then you have bugs. So when you have mocks, you have to take extra care in order to reflect to your tests and your code. So these are the shortcomings. I'm going to give you a few real-life examples. Because one of the things that I'm told many times is like, yeah, OK, it's all good and nice when it's from a theoretical point of view. But then in real life, what happens? So for example, you hide by a company. You tell them, I really want to do tests. They say good. They don't say you are the first. And they have a legacy code that is not tested. So you have to cope with that. You have to change something. So what do you do? You read the code, you understand how it works, and you write a test for it. And this is wrong. Because if you read the code, understand how it works, and write the test for it, what happens? You're inverting the cycle. Basically, you're going from the code to the tests. What we want to do is to go from the business requirement to the tests to the code. So a better way of approaching this would be, read the code, write in reverse engineer what are the business requirements that were behind that code, and then write the tests for it, concentrating on the what part, not on the how. If you do this, it's very likely that the tests will concentrate on the how part. Changing a horribly long view. Let's say we have a Django view. And we need to insert pagination, filtering, and sorting in it. We do a bunch of things. We get the data. The data is from a search. So it could be empty. It could be 10 things. It could be a million things. Then we do another bunch of things. And then we render a template with some context. So we need to add pagination, filtering, and sorting. And it's not possible to do it at the API level, because it's too complicated, and it's not tested. So something that I did after discussing it with my colleagues was let's write filter data, sort data, and paginate data, insert that into the function, into the view, without changing anything that was happening before we get the data or after we got the data. But those three functions, let's write empty dd. So at least we're changing the code, yes. But that bit of code will be rock solid. And this is a very good way to go about it, because in the end, the function was not caring before about what data was coming from the search. So it's not going to care now if you have paginated, sorted it, or filtered it in any way. Introducing a new functionality into an existing code, which is not test. So that's a nasty piece of code that does a lot of things, like this function. Uncle Bob would cry if he saw it. So how do we change? Because of course, you will not have the test, the time to go through every if clause to check if that for loop is correct or not. So what do you do? Well, one possible solution would be to come up with one test for the new functionality that you're trying to insert. And then you change the function, which was not tested before, it's not tested now. But at least that new functionality is done. The next time you go back to this function to refactor, what happens? You have one test. You come up with another test. And then you go back again. And at some point, you will have a bit of tests behind this function, which means either the function was well written, and therefore you just keep adding tests. Or at some point, you realize the function has a bug. Because you start having tests for it. So the frog was in Zen mode after all this. He went back to the princess and passed the exam, they married. And when the minister said, you can kiss the bride, nothing changed. It was just a talking fog after all. What's the moral of this story? The princess should have tested first. So should you. Thank you very much. Thank you very much. Great. We have a few minutes for questions. Do we have any? So first of all, thanks for the talk. It was really enjoyable. And my question is, how do you put your test in directories structure? Because I have some tests of all kinds. But I have trouble to checking if the code I intended to make already have some. So how do you put them? So for example, when you start the Django project, you've got tests in the application folder. First thing that I want to do is to put the test in the folder. First thing that I do is to remove that. And I reproduce the code structure in the test folder preparing everything with test underscore. There's just two main advantages. The tests, basically, are easier to find. Because if you know the tree of your files, you just go to the same tree with the tests. And also, when you deploy, you just delete that folder. Because sometimes when you deploy and occasionally someone tests on production, you get your production database that goes, and that's not nice. So I just reproduced basically the tree of my code. Probably there's also other approaches. But this works. It's good. They're out of the way. And then for functional tests or integration tests, you may want to have a separate folder as well. Or maybe even a separate repository. For example, if you are doing integration tests, it's likely that they won't be testing what's just in one repository. Maybe it's more repository, more services, or more application. So you want to have a whole project dedicated to your tests. We have a question here somewhere. Hi. Great talk. As humans, we are, I believe sometimes there can be mistakes on the test data, either by typos or bad hand calculations. In your experience, how often do those mistakes happen, and how do you cope with them? They happen more when you have to deliver by yesterday. When you have the luxury of delivering by tomorrow, they happen less. What you want to do is to take good care of your fixtures. And especially when you change the tests or you have migrations, take good care of those. Because basically those are the things that you test against. So they need some love. Me, for example, I tend to write unit tests more like an interface style, which means I have some inputs and I check on the outputs, rather than mocking everything out. I tend to mock the least possible, because it's dangerous to mock. And what was the second part of your questions? In your experience, how often, for example, in the examples you printed before, one of the numbers could be wrong by a typo and you don't notice that while you are writing tests or the assertion, how often does that happen and how do you detect it? The way we detect it is to have pull requests. So we use a branching system. We have story branches, and then we have a staging branch and before master branch. So we have at least two pull requests for people other than the guy who wrote the code to take a look at that code. And we read it. So most of the times the stuff is detected on either of these two passes. That's very helpful. Thank you. My pleasure. Hi. I was recently involved in discussions about did the test and how the code was connected to the code. And I think that some guys were defending that if you focus too much on tests using TDD and write it a test first and et cetera, you may leave for later thinking about the architecture of the code, the design of the code, how different parts of the code is coupled between them, the performance, et cetera. What would you say to them? I think that the type of methodology that fix everything and solves every problem, you still have to take care of the architecture, take care of the design. And you can do it in the refactor phase. What people who do this and find themselves neglecting the rest of the code is because probably they believe TDD can do too much. TDD gives you a very good way of writing your code and a very good solid test code base that basically shields you a bit more when you're refactor. But it's not that you can go refactoring like a monk because you've got tests. You still have to give it your best shot and take good care of the code you write. But with TDD, you've got this extra thing, extra like a guardrail that basically keeps you on the right, on the right path. It's just something that you have more but will not solve all the other things that you have to do. Great. Do we have one last question? No? All right. Now I don't really want. Thank you.
|
Fabrizio Romano - TDD is not about tests! TDD is not about tests! Well, actually, it’s not a about writing tests, or writing them before the code. This talk will show you how to use tests to really drive development by transforming business requirements into tests, and allowing your code to come as their natural consequence. Too often this key aspect is neglected and the result is that tests and code are somehow “disconnected”. The code is not as short and efficient as it could be, and the tests are not as effective. Refactoring is not always easy, and over time all sorts of issues start to come out of the surface. However, we will show that when TDD is done properly, tests and code merge beautifully into an organic whole that fulfills the business requirements, and provides all sorts of advantages: your code is minimal, easy to amend and extend, readable, clean. Your tests will be effective, short and focused, and allow for light-hearted refactoring and excellent coverage. We will provide enough information and examples to spark the curiosity of the novice, and satisfy the need of a deeper insight for the intermediate, and help you immediately benefit from this transformative technique that is still often underestimated and misunderstood.
|
10.5446/20109 (DOI)
|
and less beauty together. So hello, everyone. We'd like to welcome you to the EuroPython 2016 session. And we're going to do the talk together. I'm going to do the first few slides, and then later on, Mr. Pavie is going to join in. And of course, if you have questions, just feel free to ask during the presentation. We don't want to put everything at the end, because then we just lose the context. First of all, a question. How many of you know what the EPS is? How many don't know? OK. So the EPS was founded in 2004 when we found that organizing EuroPython. EuroPython started in 2002. And by 2004, we wanted to have an organization that basically holds the IP rights of the conference and makes sure that it's a continuous process, that the selection process works for each new location, that the transfer of knowledge works from one location to the next. And the EuroPython Society was supposed to back up the local organizers at the time. So the way that it worked was that the EuroPython Society had the trademarks, the logo, and it was supposed to have the social accounts for the conference. The EPS used to select, just select, the conference location. It did not actually run the conference itself. That was being done by local organizers. But every now and then, when we switched, we had a CFP process to then determine a new location and then work with the local organizers to actually make it happen. In 2015, we realized that the old model of doing this would no longer work out. Because we had serious problems finding new local organizers doing the conference because it had grown this big. This year, for example, we have 1,100 attendees. Last year, we had 1,200 attendees. So that's a size where you cannot easily just do the conference organization like that on the side. It's actually a full-time job for quite a few organizers. And so we thought that we'd use a workgroup model to make it possible to scale this up in a better way. And the workgroups are meant to allow work to be done remotely. So there are lots of things that you can do remotely. For example, you can manage the website remotely. You can do communications remotely. You can do marketing and these things. You can do marketing and these things remotely. And this is what we started this year for the first time. And it has not yet really worked out all the way. But I think we're doing fine in developing this kind of model. And we're definitely going to continue using it. One of the most important things there is that we wanted to reduce the loss of institutional knowledge from location to location. Because we always had in the past, we always had an issue having the knowledge that was gained at one location being transferred to the next location. Usually what happened is that we got a complete new website software set up. So EuroPython has, I think, about six or seven different website systems right by now. All the little details that you find when organizing a conference, those details were usually not transferred to the next year. So the same mistakes happened over and over again. And of course, there's a financial risk associated with this. And we wanted to reduce that risk and have the EPS sign contracts instead of having local organizers sign the contracts. Now, this year, we tried that. It didn't work out because we couldn't manage the VAT taxes with the EPS. We tried to get a VAT ID for this, but the Swedish government wouldn't give us one. So we have to deal with that next year. So under the new structure, the local team is meant to just do the onsite work. And we try to do most of the other things remotely. Now, in practice, for this year, it hasn't really worked out that well. So what happened is that even though we had the workgroups, most of the active members in those workgroups were actually members from the onsite team, plus a few members from the EPS board. But we just started this year, so we have good hopes for the next years. Right, so just a few dates maybe to show you how the whole thing developed. In July last year, we had the election of the new board. Then in October, we did the CFP for your person rather late. By December, we announced the winner. The ACPYSS is a Python, Sans, Sebastian user group. And in January, mid January, actually, we started the work with the ACPYSS to run the conference. So the conference itself, the whole conference, was set up in a very, very short time, as you can see. In January 2015, we also, because we started with the workgroups, we set up all the infrastructure for the workgroups. So we're using a meeting list for that. We're using Lumio for voting. We're using Trello for a few workgroups to manage various tasks. We set up a pre-launch website because people were starting to ask where the website was. And we set up this pre-launch website because it took a bit longer to actually launch the final version of the website. Then in March, we managed to get the final website running. The website was using the code base that was used in Florence because that was very complete conference software. And as you can see this year, again, there are lots and lots of cool features in that software. And once we managed to get it working, it really worked very nicely. And so we didn't have to do anything much to make it work like what you see now. So then in March, we also launched ticket sales. And then we ran the call for proposals, again rather late. Then in April, we had this podcast issue, which was actually the first bigger COC issue that we had. In May, we had the FinAID program set up and announced. And the talk voting was started. Talk voting was not done last year. It was done in Florence. And it was also invented in Florence. It was a very good kind of, it turned out to be a very good way to figure out which talk proposals to accept. It works kind of like crowdsourcing effort. So basically all the people that have bought tickets can then go to the website and vote on the different talk proposals. And then the program work group would then take those results and based on those results would then go ahead and actually do the selection. So the selection was mainly done based on the results from the talk voting. But the program work group, of course, also took care to, for example, increase the diversity of the number of the speakers, of the number of talks that were done by women. It also took care that each speaker only, well, we tried to have each speaker only do one talk so that we can get more and more speakers instead of having one speaker do five talks and the other just maybe one. So I think I'm one of the notable exceptions here because I did one talk for your Python in the Python context. And I'm now together with Fabio doing two other talks, just the EPS stuff. But this is more on the side of the conference. It's not really the normal case. Right, and then in June we had the schedule online. Took a bit to get that set up, again, because we had a few issues with the website. But we finally figured out how to actually publish the website, the schedule on the website. And then in July we followed up with the guidebook application. And now we have EuroPython and today we have the General Assembly. So overall, it took just six months to organize this conference, which I find really impressive. And we really have to thank the local team a lot for this because without them this wouldn't have happened. So I think we should give them a good applause. So thank you for that. Right, and just to give you an impression of how the development was with EuroPython, it started very small and with 240 people in Charlery and Belgium. And then every year the attendee size increased a bit. When we went to Vilnius in Lithuania, it went down again a bit because it's a bit far away for many people. But it was still very interesting in Vilnius. Then in Birmingham it picked up again. And last year we had with 1200 we had the peak so far. In Bilbao it decreased by about 100, which is not really much. And given that, Berlin is right at the center of everything and it's very easy to get there. And there are very good train connections and everything. So it's not surprising that Bilbao has decreased a bit. We hope to increase that again a bit by maybe a few hundred next year. We'll have to see how that happens. I think that when people go home again and they know how well organized this conference was and how beautiful the location is and beautiful Bilbao is, I think they will definitely come back and get us more attendees next year, which is good because more attendees of course also means more work. But it also means more sponsors. And sponsorship is very important for the EPS because in the long run what we would like to be is something like a European kind of version of the Python Software Foundation so that we try to use the conference to make money and then to redistribute that money to the European Python community. So what we ideally in the long run would like to have is a grants program where you can go to the EPS and ask for grants for running smaller conferences or running projects in Python. And so this is our long term goal for this. And for that we need to increase the sponsorship from what we currently have. At the moment for this conference, for example, the sponsorship revenue is about 175, 180,000 euros, which is really not much. I mean, if you look at PyConUS, PyConUS has way over a million dollars in sponsorship money every year. So there's quite a big. They have about, well, between 2000, 3000, I think it's about 2,500 at the moment. So compared to Europe Python, Europe Python has about 1,000. But I mean, it's just quite double the size. And the sponsorship money there is, well, I mean, in the US it's much easier to get sponsored because they're very open to sponsoring things. And in Europe it's not that easy. So we have to try to talk to all the big sponsors that you have at PyConUS, for example, whether they would like to do the same kind of thing in Europe. You can see that with Google, for example, it has worked out great. Thanks to Fabio's good connections to the Google sponsors contact. They really invested a lot of money this year. We're trying to do the same with Intel. We're trying to do the same with Microsoft and with Facebook. So we need to get better connections to them, build better relationships with them, and then make everything bigger. So now I'm going to pass over to Fabio. He's going to talk about the work groups. Fabio, thank you. Last question. Are you guys going to get to what's coming in the future? Will it be in the future? We have a slide. We have a slide for that, yes. I think it's on the general assembly, right? Yeah, maybe it's not on this one. Let's see. One thing, a couple of things I wanted to mention. Sorry, I know Bill Bowie is in the center of the world, but Berlin is more connected. That was the statement. And the other thing was I wanted to mention is that I had many people asking about how around, regarding the website and how they could, that it was hard to help and how we chose, how can we chose the Italian version instead of the last year version. And the background is that we contacted both the main developers of the Italian and the German versions and based on the support, because we didn't have much resources, both were very, they said to be very honored for the request, but the Italian team turned out to be ready to do more work. So it's to say that both software were great software as well. One, so the work groups, the work groups concept we had in mind is that we had one group, work group chair person or two, depending on the work group size, working members that would focus on specific topics, specific things of the conference, that could work remotely, and we could organize those work groups with voting members and a few other working no voting members. The teams were built looking to set the tackle for the conference administration, so having to run contracts with the venue, tickets, support, one finance group to control our budgets, do accounting, spellings, check out, connect with other work groups that needed something from the budget. Sponsors, so for instance, sponsors and finance were very connected work groups. Sponsors of course were in charge to contact new sponsors and all sponsors logistics, explore sponsors needs and everything. Communications, do press release community relations, work on the diversity and outreach, code of conduct, follow up with meeting lists. The support, contact attendees, help with business, financial aid, handle the grant selection, set up grants, organize aid, put up information, marketing, do their jobs, the program that Mark has already told about, which is repeated. The web work group to handle basically the website and everything related, the media team to do video recording and translations. And of course, the most important team, the on-site team, to handle as a glue for those work groups and work locally on all the providers, the venue and all the needs regarding catering, printing, logistics and everything. We thought since the beginning this was quite important to have guidelines to get communications right, avoid putting people in a place where they couldn't actually understand where they could help. The guidelines are necessary for those groups also because to welcome new people and address how they could help to keep the knowledge and hopefully act as a code base to improve the new editions. The purpose is that the guidelines evolve pretty much in line with the conference editions. So every year we learn something new, every year we have new members, new ideas and we really think that those guidelines should just be written by the volunteers and the people doing the conference and hopefully from the feedback from the people enjoying the conference as well. We did a lot of, we had a lot of contents to Wiki and a new Wiki with the content, we set up a few procedures and documents to handle all the communications, the conference workflows and needs. So at the end of the day this year was pretty much a dense year, hard year, but overall I think this year was really a big step forward on building something that can stand for the future editions. We built a lot of things. Many things may not be great on the workflow side of things, but we really, I think this is a real proof that we can handle this. If we could, especially the team from Ubao, could organize a conference like this in six months with a lot of attention, a lot of small time and many unplanned surprises like for instance starting with a number of volunteers and ending up with a fraction of that number pretty soon. That proves that we really can do incredible stuff if we have time we can do even better. So I guess that's that. Those are the UB society contacts. For information you can check out the website. We have a blog and every request, there are many channels you can contact the UB society even through the board or the mailing list or other channels. We invite everyone that wants to help to step ahead, help propose ideas and keep helping us doing what we do. That's all. Thank you. Thank you. Hello. Thank you very much for organizing it. It's an amazing conference. Just two things I wanted to raise. Talk voting, it was amazing compared to some centralized commission which was picking up a few years ago. The only thing is I think that everyone should have a limited amount of votes that you can give because otherwise you have a lot of talks that had very catchy titles and we're not fulfilling the promise and I think that the catchy title itself because of the freedom of voting for as many talks as you want I think that it biased those catchy title stores actually being selected soon. Perhaps Alex wants to say something about that because he is one of the... Well, one of the two Alexes anyway. My only thing is whenever you are going with the model of crowdfunding it's like you always seem to introduce the concept of some kind of currency otherwise if it's unlimited then you introduce a lot of biases and it comes from the fact that everyone has a limited time here. So basically it makes sense that everyone can vote as many talks as you can attend basically a 20 or something and that's it. So basically I can answer that directly from the voted program workgroup. Basically I have some similar results because on talkwalking we spend a lot of time on the algorithm comparing and it was not just like a number of votes. It was one more complicated thing. Also the algorithm is more like preferences. So it didn't really matter if you voted for many talks because we didn't not only sum up the points. We did comparison, averages and said that we finally used the algorithm with also Italian broad to us. We changed it a bit so there was no possibility of down voting. But there are some things I think we can improve on the voting process because for me it's a bit limited only to people having a ticket and I want to suggest that we include people probably having attended one Euro Python because basically we want to build a great conference for attendees but we don't want spamming with people making multiple accounts. So basically, yeah. What was good last year is that there was some kind of feedback you could give, you could add questions to the people creating the talk. So this year there was... So my only issue here is it's actually amazing that you decided to go this path rather than actually making the whole voting closed. But I'm saying if we really want to achieve the transparency of the selection either we go with the crowd voting and then crowd voting that is fully transparent or some hybrid system where you say you can vote but actually you don't vote because after that there will be some algorithm that would do something with whatever you did. It's all available on the website and it's just nothing really hidden. The way that it worked was that we basically used the system looked at all the voting results and then the voting results were then used by the program work group to then make the actual selection. So what for example the program group did was instead of having 10 talks about maybe I don't know, async.io, which is very popular this year, they just selected a few of those and then instead gave the preference to other subjects which were a bit underrepresented in the talk voting results. So the voting results were used as basis for the selection but I think it's already a large step forward because you don't have this kind of committee decision process for the talks. We sometimes get biases from those committees into the talk selection. So that happens sometimes with conferences. That's really good but if everyone is voting and it's like a decision of everyone to have certain kind of talks and I think that's a community decision. The community decided they want to have 10 async. I just wanted to make one comment that takes account also like why the talk voting was introduced in 2011 and there was a progression. On one side, imagine handling 300 talks that you need to review, look at it and it turns out that public voting, if you see the people voting, I don't think anyone vote all the talks. I can't remember anyone that actually voted every talk. You need to spend the whole day clicking. It's a very difficult thing and it's a very interesting topic. We picked what seemed to be the better compromise. Actually there are many theories about how we should do this. It's a very interesting discussion. Other sources, we could try and maybe put some link regarding the algorithms we have used. If you care about that and you want to improve it, I would really like invite you to make suggestions or try to improve that. We are open to improve next year. We have some ideas too. I'm saying it from the perspective of a person that is willing to contribute to that. No, no, sure, sure. It's not a really completely clean discussion. It's really because it's a hard thing to do. The thing that we chose to not do everything voted is to avoid some sort of bugs in the attendees' preferences. We care a lot about diversity. We care a lot about giving the opportunity to everyone to talk. People are biased by their friends, by people that they already know. It would be hard for someone young that don't have enough experience, that don't have enough visibility to get into talking just because someone that is really known have 10 talks. It's a hard thing to balance. It's really hard. Just to close, this year the program workgroup was fairly limited. You can understand by people nothing. The program workgroup is more or less in the front row. There are more. It's just to say that we had a lot of... You can count on two hands. We made a lot of requests for people to contribute. Really, if you care or other people care about doing this, we are open arms. I just wanted to add one thing, because we have to switch to the General Assembly now. If you want to sign up for one of these workgroups, and you're interested in helping out and improving your price and being organized, then please submit an application request for that. You go to our website. Where's my mouse? You go to the website. You go to Europe Python Workgroups. This lists all the workgroups again. At the very bottom, you... Where is it? There you go. At the very bottom you see the mailing address here, the board mailing address, which is an email there saying, okay, I would like to work in this and that workgroup, and then basically we set you up for it. Does that... continue on your workgroup in this event? Do we have to re-sign up? No, no, you don't have to re-sign up. The idea behind the workgroups is that they persist from year to year, and you can stay as long as you want to. If you don't want to be on the workgroup anymore, you can just tell us. And then we just remove you from the workgroup again. So it's a very kind of open process. We're not making any big restrictions there. The only thing that we sometimes do is we ask the chairs of those workgroups whether they need more help in a specific area, because sometimes people sign up for multiple workgroups, and then we sign them up for those workgroups that need more help rather than workgroups that are very popular. For example, the program workgroup was very popular, so we had lots and lots of members in it. Unfortunately, not all these members were really active. So in the end, even though we had lots of members in the workgroup, the actual work was only done by very few people. We've seen similar things in other workgroups, but for the program workgroup, it was really an exception. So we're trying to change that, and if you have suggestions on how to change that, that would be very welcome, because we don't really know how to address this. One suggestion that was made was, for example, to assign tasks to people and do that specifically, so tell someone to do a certain task and come back maybe in two weeks with that finished task to do it that way, so that people feel addressed and don't just see something happening on the mailing list and then thinking that, what if someone's going to do it anyway, so I don't have to look into this. That may be one way to try to improve this. We'll have to experiment with that a bit. If you have other suggestions, please tell us, because we're all volunteers, we're all working together, and we don't want to put too much pressure on people, but on the other hand, we also don't want to have a few people burn out because they have to do too much work. There was one question back there, yeah? Over here. I wanted to, first of all, all of you have to know that I was involved in the program work group. I saw people corresponding in email at 2 or 3 in the morning, right? They worked very, very hard, okay? All of us, I made a very little contribution, but all of you was something huge, the work that you did here. For my point about the voting that we were talking about, I and I think other people were in some kind of deadlock situation. I wanted to come here. I submitted a proposal when it was the submission of proposals, and my university financed my arrival here in condition that my proposal will be accepted. But I could not vote if I not pay the ticket. But if I don't have the budget, I cannot pay the ticket, but I don't have the budget because my proposal was not accepted. Then as a deadlock, the last day I paid from my pocket for the ticket and then voted, and my proposal was accepted as a poster, and then I received the financial aid of the university. Then we have to find some way that people that presented, submitted a proposal but did not pay the ticket will be able to vote for all the proposals. Then when they have the money, they will be able to pay for the ticket. It's a complicated situation. I know it's not something so simple, but we have to think what is the best solution for this deadlock situation. We will have to experiment with that a bit. We simply took the code that was there and switched it on and let the web code take care of this. I'm not sure. I did one experiment and found that you don't have to buy a ticket to do the talk voting, but maybe that was just because my user account is switched on as a staff account. We definitely have to look into that for the future. I actually didn't want to refer to that, but I'll quickly add to that that other people who ask for financial aid and won't know for a while if they get it and can come here would be cool for them to also be able to vote. I actually wanted to say about the work groups, because I just said a look and the programme work group is indeed huge online, wouldn't it make sense to maybe kick people out again that do not contribute because otherwise the groups will increase in numbers, so number of members, and it doesn't make sense to have a huge list of people to refer to that don't do the work, and it's just a handful of people who actually do the work. That's a very good suggestion. I think we should probably have some kind of policy for this. If a member is not active for a few months, then maybe just remove that member from the work group because otherwise it wouldn't make sense, like you say. Yeah, Jacob? That's the last question because we have to start with you. Last comment about, as we are having about a lot of ideas and new things, actually another thing that we would like to be more, try to be more transparent during the process to also let people know more about how much we are spending or on financial aid, how much we spend, because financial aid is a very difficult topic because it's hard to say how much we'll be able to spend. Many times we did financial aid and helped with a red budget situation, risking a lot of things just because... No, no, no, no, no, no. I wanted to add this to the discussion saying that as much as the community understands the difficulties of doing this, the more maybe people can make suggestions or ideas about how we can tackle those problems. I would like to look a little beyond 2016. I think it's really, really important for a Euro-Python to have a vision about where we want to go. I'm actually somewhat involved in how Euro-Python got to where it is today because I was the founder of the Euro-Python Society and I influenced the decision to go to CERN, to Vilnius, to Britain and to Italy. And we had an idea of why we were doing this each time. We went to Vilnius because we wanted to spread Python in Eastern Europe. We need a vision in what we're doing and it's not enough just to think about what are we going to do next year. We have to think about what direction do we want to take. Something I see as a problem right now is that I find many of the talks rather boring. There are no really new inspirational talks at Euro-Python this year and I think that's something we need to address. Yeah, I agree. I mean, we need to do this as well. But, yes. Sorry? No, no, it's an opinion. I fully agree. But I also think that this vision should be shared with more people. This year was a turnkey year and I think that we need to build something that is more than a group of people in a more engaged community that drives those decisions, drives those visions. It's a hard problem. It's easy to just take for granted that it's going to be a new Euro-Python next year and I would like that it would be not obvious but this vision is much more open and we have a lot of people, vibrant people from Eastern Europe. That means that we would like to have more teams proposing from Eastern Europe. We would like to have more workgroup members from that region of Europe. That's part of the plan that we are trying to bring. And I think we are running out of time. We need to start with a general assessment.
|
Fabio Pliger/Marc-André Lemburg - EuroPython 2016: Help us build the next edition! We need help with organizing and running EuroPython 2016. In this session, we will explain how the EuroPython workgroup model works and where you could help.
|
10.5446/20105 (DOI)
|
Hi. Do I need to address it? Hello. Do you hear me? Again, great. My name is Eugene and I work as a team leader at Scrapping Hub. I was lucky enough over a few past years to work in different areas of IT, like games for social networks or collective blog. Again, like Reddit, but just for Russia. So, major part of my job was to review code and I might say that readability of tests, as we see it now, is not really good enough. But actually tests are very important to our work. If we look at top 100 Mod Start projects at GitHub, we can see that 23% of code is located in test folders. So, this means that if your job is to just read through the code of this project that you would spend two hours a day just reading through tests. And tests become more and more important. If from this 100 projects, 34 projects have revision history as much as five years deep. So we can take a quick look into it. It's a percentage of tests to the code it tests and we see it steadily grown. And if we take a look at some absolute metric, like a number of lines of code, we can see that growth is even more significant. It's from 143,000 of light to nearly half million lines for just tests. So, if we take a look at some general tests, for example, I'm sorry. Yeah. This is how most of tests are written, at least the ones that I see. But when we look at it, we can do better job of describing our intent. That means that when other people would read this code, they wouldn't have such question like what this code does. Is this test actually test what it's supposed to? And if there is an error in the test run, was it because of there is error in the code or because test is written wrong? So what actually is test? How can we understand? All tests consist of three main things. First one is environment. Environment which we have before we test something. It could be no special environment. It could be fixed date and time. It could be if you test in some chat service, we have a profile of user. Another thing is what we actually test. We can test that to multiply by two equals four. We can calculate yesterday eight. We can make these fictional VIP users try to swear in public channel. And after each event, we have some expectations. This means that we expect that two multiplied by two equals four. That yesterday calculated correctly at that. Swerving user would be banned no matter of his status. So nearly a decade ago, father of behavior-driven development then North published an article where he proposed given when then template to describe such behavior. Let's take a look at Wikipedia example. Here first line describe our environment. We have customer who has bought a sweater from us. We have stock with this sweater. We test that when sweater is returned, then number of sweaters at our stock will increase by one. And this is very clearly where we see what we have, what we do, and what we want in result. In behavior-driven development, they are right these tests as a text. So managers without any programming experience could write those. Or even users. But they have additional layer of transforming text to actual code. Benefit from this template, we actually don't need that layer. What we can do is simply extract methods, generate an environment to the methods which start with the word given. And so the same with when and then. And let's have an example. I showed you this test before. What we test here is when we registered a logger for Sentry, it doesn't register same or twice. So same error wouldn't appear twice in the Sentry itself. Let's take a look what is environment here. Creating loggers in environment. So we move it to a separate method. Same for action. We can see that there is two actions. It's a little bit alarming we talk about it later. And what results. So now when we transform code in this way, it's much better to understand what it does. And what we expected of it. Actually, when we have two actions, it's a good alarming thing because we almost never want two actions in our code. Because we want to check that progressing from one state environment with an action, we get to another state. And when second action we do, we want to generate that previous state. And the first action we just moved to a separate test. So I didn't highlight it. I forgot to highlight it. But you can see the second test method self given handler registered. This is where we describe what we had at the beginning of this. Also, when we put it in literal words, we can see when something wrong with what we wanted to test. So here, last expectation of which test, then number of sentry handlers registered. Actually, what we want instead is to check that every sentry handler registered is unique. So we move it to another method. Of course, we change behavior of it. I just put it here. Another thing, when users and developers firstly approach this template, they tend to put everything. And I mean everything in the code itself, so it's read correctly. It's kind of again common sense. We don't need this. We have setup. So move common, don't be afraid. You still check up setup when you read this. So this is a good way to structure single test. But single test is not only our problem in test. We also want to organize better tests with its multiple tests we have better organized. So let's take a look at another example. If you take a look at every aquarium, it has some equipment. It has environment in it like salt water, fresh water. This is a Dutch aquarium where majority of space is taken by plants. So this is kind of like classes. So we can take this example. So here is an example of these aquariums. And we see that they inherited from each other. And also we want to check single method. A lot of methods could be... But this method on the top of class and it's inherited by each, every one of them. So of course with aquarium we have a lot of fish. And usually with such many test we have test like this. It's a lot of repetition and I mean a lot. And there could be hundreds of cases. For example if you test in parcel for its removing malicious code from it. There actually could be hundreds of cases. So we don't want such long tests. Usually it's transformed in a loop. So we have aquarium and for each data we loop through it. We check, we do some action, we test. Same but it's just a little bit better. I can't stress enough how bad loops are for testing. For this particular case, if there is an error with Goopy, a fish. You don't know whether errors just with Goopy, always every other fish, always some particular fish. But this information is very important because you can spend half a day looking into a place where you shouldn't look for this particular problem. Because problem is somewhere else and you would see this if you immediately have results on which test, pass, which don't. Here for example we see that tests pass for Goopy and for Goldfish. And don't pass for Raspberry and for Leopoldy. Your train of sorts could be, how are those connected? Because you sort of something else at the moment. And thinking about it you would be able to find exact spot which unites those particular problem. So how do we want to translate data to separate tests? I very like NOS parameterized model which looks like this. And by the way, if you don't know, this month NOS parameterized updates and actually applies not just for NOS but Pi test and unit test. So you transform your tests. So test methods now have parameters. You can see after self there is a fish parameter. And you write your code, your data in a list decorating this method. Unfortunately, particular for NOS parameterized there is a problem with inheritance. So if we inherit from freshwater aquarium, and noticed the different data. So we changed Drasbora for Grammys and we don't test for Leopoldy. Our expectations could be either we run tests in Dutch aquarium just for this three particular data. Or it unites all of them and Drasbora and Leopoldy added to those three. Unfortunately, what's actually happened is this. We have four tests for each aquarium. And what happened here, so you would understand. We have four tests from freshwater aquarium. We replace three for those we have data. And the last test is just a leftover which is inherited from previous one. So it's fails and it's not something. It should be considered. So what do, how can we apply inheritance to our tests? So our goals would be to have tests for afterized. Like NOS parameterization like NOS is good enough, but it need to be able to deal with inheritance. So inherited test data. Also, we probably don't want to repeat each test each time for the inherited test case. We want it to be defined in parent class and seamlessly used in the child cases. And we don't want to apply, we don't always want to apply all the data from parent cases to the child cases. So we need a way to control this. So these goals could be a region to next environments for parameterized virtualization. We need to apply single test method to different data as much data as we want. For inherited test data and inheritance of test itself, what we need to solve this is just an access to parent class and we can extract data from it. And for control execution what we want is a way to exclude data. So let's take a look what Python tools and approaches could help us to deal with these goals. Decorator is most important part of parameterization. It's used in every approach which I would show. And unfortunately usually it's diminished to just create a new function which puts some code before or later a regional function. But being function over function, the curator is actually transforming a regional function almost to anything. You could see this, for example, skip if decorator of unit test. It transforms a regional function depending on the condition, either to original function or to skip function that raises it. So decorator can be also applied to classes, not only functions. So this is also transformation of original class. So how would we apply decorator to achieve those? Novels parameterized show us very good way to define data. It's clear, it's understandable. So we decorate each test case with data that it requires. But only decoration doesn't create multiple tests. So we need a way to transform this function with assigned data to it to multiple test cases. So here we just assign to the method data in any way we want. And with second decorator we take this class. We have all these methods defining it. And for those methods which define parameterization, we create additional method. With this approach we target test parameterization, but it's not very good for inheritance. It's applicable because you can do this, but for each class which inherited from this particular, from this test case, you need again to reapply this decorator for new test because what we simply did here is instead of some test case, we created some test case with more methods. It's not applying the behavior to its child cases. Another approach would be metaclass, which is a way to configure how you create tests. So typical approach is, and we would use typical approach here, is we have a name, we have basis which is all parent classes of this particular case, and we have a name space which is dictionary of methods, parameters from which we create this class. So we can manipulate with dictionary before we create this class. So we have some tests which have data assigned to it, which are from all of them, and for those who have data, we add some keys and values to this name space. It's seamless and it is working with inheritance. So when you create a child test cases, this behavior is also transformed there. So unfortunately with metaclass, there's also trade-off. It requires that each metaclass is inherited in a direct chain, not a sum-3. So it's not a problem for your single project when you start packaging your things. When there is one metaclass in one package and one metaclass with another package, you're not correlating with each other. Your users could not be able to use both of them. So this approach is useful at some extent, but also... We can also not return some class, but anything we want. Frames is another approach and it's used by NOS parametrized. When you have a tracebook, it's actually a listing of frames and it's kind of name spaces where cursor is... our execution cursor. NOS parametrized take a frame from which its declaration was called and it is inject new method there. And also... so it's... Here name space would be definition of class test freshwater aquarium. And it's transformed like that. It's like we written it this way. I'm a little bit short of time, so I quickly approach last one. My favorite. Before that, here it is very understandable why we can't approach... why we can't get parent from here. Because at that very moment, we don't have class to get parent from. We're just defining the name space. So with frames, we can't target the parent class. So my favorite one is custom loader of unit test. Loaders responsible for gathering tests from your code and creating suits from them. So we don't need actually anything at this point except to mark some tests with data. And we don't have to think about inheritance. What we have here is something like this. We get names, we iterate through all of names, and if it has some data, we extend some... we extend all of this. If it's not, we create usual test case class. Actually, worth mentioning that unit test uses loader to create different instances for each test method it has. So it's not one class with a lot of tests... with a lot of methods with tests, but a lot of instances for this class which tests single method. So it's very straightforward, so you can read through it, and it's very understandable. So here, what we will do, we decorate data, and when we approach actual test... test run, we create additional tests. We don't change anything into class. We just decide if we want this test, if we don't want this test, or if we want to create multiple tests from this one. I didn't mention for previous approaches, control execution, which skipping inherited data. This is because there's two general approaches for them. For those where we have access to parent classes, we can clearly tell what we want to do with the data from parent classes. We can extend it, we can remove some particular data from parent data, we can completely replace it. Another approach which don't need to create different decorators for this, is just to insert in your test body something like skip if, but I like very much JUnit approach where they have assume. So we can have assume here, which would go nicely with given, then, and when. And it would skip test if it's not applicable at the parent, is a child class. So I want quickly to give some mental experiments. I would read out loud a few things and you try to get to yourself feelings about it. So it's not related to test, it's just get and get help. Now SVM, now, first of all, we had before folders with different approach, and changing code live on production. After that, I am very glad that we have get and get help now. And I am very sure that using approaches that I demonstrated today, whatever briefly it was, I'm sorry, would be able for you to create a framework that applied to your particular project that works best for you, that for you and your team, it would be very easy to create tests, it would be no frustration, it would be low maintenance, and you can quickly navigate through it. Thank you. APPLAUSE First part with new functions, it looks like BDD, isn't it? First part with new function, when you change the part of code, with new function, with full name. You mean organizing it in a different given? It is actually very easy, it has some current cases, for example, when you want a width patch, so you can approach it simply like a patch isn't only, can be used with width, but it has start and end, so you can create given function and start patching it, and then you stop all started patches in tests. Yeah, it looks like, do you know about Cacumber, maybe something that this framework, it's Ruby, right? Yes, it's behavior-driven development, but it is another layer which we actually don't need, because some of our projects are small, some are not. Cacumber is great, and we can benefit without using, just clarifying what we want to do from our test. Okay, thanks. Thank you for the question. Last question. Okay, thank you. Hi. So, hi. How you manage, like, if you have a lot of tests, because my previous project ended with having a lot of tests, how you manage to do not create a lot of functions that just either do BDD, but they do something different underneath, and how to maintain a lot of helper code to make the test, because from what you were showing on the slides, you move out something to the different function to have one responsibility, and then test is easy, but, you know, maintaining that. So, you're not only worried for that, but for the length of a single test where you have a lot of... So, from experience, there is a great deal of work that you can do to make sure that you have the best results and that you can do a lot of things, but you can also do a lot of things that are not very good. So, you know, you can do a lot of things that are not very good, but you can do a lot of then-asserts. So, from experience, there is actually no harm to combine those then-function to some single then-function. So, you would have... Then user is banned, and... In this method, you would have... Then it's deactivated. Then it can do this, then it can do this. So, you don't put all of this in the test, but you can do a test, a method that describes it. So, generally, what you want to do is to follow the organization of your code. So, you organize your code somehow, and it's better to follow the same structure in the test. Okay. Okay. Kind of answered my question, but probably I will have more. Okay. Feel free, of course, to talk to me afterwards. Thank you.
|
Eugene Amirov - Sustainable way of testing your code How to write a test so you would remember what it does in a year from now? How to write selective tests with different inputs? What is test? How to subclass tests cases and yet maintain control on which tests would run? How to extend or to filter inputs used in parent classes? Are you a tiny bit intrigued now? :) This is not another talk about how to test, but how to organize your tests so they were maintainable. I will be using nose framework as an example, however main ideas should be applicable to any other framework you choose. Explaining how some parts of code works I would have to briefly touch some advanced python topics, although I will provide need-to-know basics there, so people with any level of python knowledge could enjoy the ride.
|
10.5446/20102 (DOI)
|
Okay, let's start. First I briefly introduce myself. My name is Mity Trophimov. I work for JetBrains. I am team leader and developer of PyCharm IDE. But today I won't be speaking about PyCharm in this talk. If you want to discuss anything about PyCharm, come to our JetBrains booth. PyCharm team will be there during all the conference, ready to answer all your questions. And I'm very excited to be here. I'm a developer of PyCharm in Bishay. Mostly I develop in Java and Python. Java has been my primary development language for many years. And Python for me is more than a language that I use. It's like a subject of constant investigation, like a snake that I examine, which step of scales it has, which teeth and what is inside of it, how to debug it, how to profile it. And a half year ago I started to play with Rust a bit. I knew nothing about it at that moment. Still I don't know much about it now. But I hope that I will be able to introduce it to those who are not aware of it at all. By the way, who have you heard about Rust? That's pretty much it. How many of you have tried it already? Develop some stuff. Oh, that's cool. For those who have tried it, I hope it will be also interesting, because I will show some corner cases. Yes, so what is known about Rust? Yes, and answering one more question. Why did I pick Rust? Why did I start to invest my time and learn it? I found it interesting. And also I wanted to develop a profiler for Python and to make it work fast. Okay, I will tell about it. Rust is a Mozilla project, and they actually are using it already for the new browser engine called Server. The project started in 2010 as a site project of Mozilla employee grade on Hoare. And the version one was released on May 15 this year. So now it is 1.1. And before version one, things were changing at very high pace in Rust, breaking compatibility. And now it's still changing. Standard library is not polished yet, and ecosystem around it is just starting to emerge. But now it has backward compatibility, and this allows to develop production applications in Rust. So what is Rust? I don't know about it when I started to learn it. What exactly captured my eye? They told that it is fast, prevents nearly all sick falls, guarantees threat safety, close to the metal, has zero cost abstractions, pattern matching, and type inference. And that sounded very cool. I thought it would be very interesting to learn it. So I started to listen to some talks on YouTube, nothing became clear for them. Then I found some specification, it turned out that it is already outdated, the language had just changed. And then I found Rust Book. For all of you who is interested, I recommend this book, The Starting Point. It's online and very well written, and I don't think that a talk can help you to teach the language. That's why today I won't explain basics of Rust at all. I don't have a goal to teach you the language, but to give you a feeling of it. Okay, so let's start with a small but real problem. And as the spear is my advantage to Rust, let it be a computational problem, like computing primes. So the problem is to compute prime numbers between 2 and n. And the prime number is the number that has no divisors except itself and 1. So like 2, 3, 5, and so on. And we will solve our problem with the help of an algorithm called Rappatian-Cif. Algorithm is simple, the work is this way. We first take all the numbers from 2 to n, and then iteratively throw away those who have divisors. So something like that we start to take 2, and then we throw away all events, and then we proceed to the number 3. We take as a prime, throw all the multiples of 3, and then it will be 5, and all multiples of 5 are thrown away, and 7, and so on. So here we see a Python implementation of Rappatian-Cif. It's quite beautiful, isn't it? Here we initialize our nonprimes as an empty set, and then we iterate through a range. And if the current number is not in the set, we increment our counter and update all the multiples. We put it into the set as a nonprimes, and then return it from the function. Okay, so let's run it. This is our function, and also we have here main function, which takes a common line argument as n, and then executes our function and prints output. Okay, so it's something like... Okay, so we have 4 between 2 and 10, and for 100 we have something that seems to be correct. Okay, let's measure the speed, the time of this program. We will comment out the output because it always slows down the execution. And okay, and for 1 million... Oops, make something like this one. Yeah, it's more than half a second. It's pretty fast, but what if our task is... What if our task is to implement the algorithm as efficient as possible? One of the obvious solutions is to use C programming language, because everyone knows that C is very fast and efficient, and very well. Very good programs are written in C, for example, C Python and written in C. So let's implement this algorithm in C. So the program now is a bit longer, but it's pretty much the same. So we have our dot-thnce function. Unfortunately, there is no set in the standard library of C, so we use array of primes, and we're... where all items mark as 1, this means that it's prime, and when it's 0, it's not prime. And then we iterate through it, and if it's prime, we increment our counter, and then update of the multiples as 0, and then we put our array into the result structure. And unfortunately, there is no tuples in C, so we need to have this result structure, which holds a counter, and our array. So in here, we execute it, and we print the counter, and then we want to print all the primes. We first put them into another array, which is the length of the counter. So we iterate again, we put them, and then we print them. So it looks quite similar. So we need to compile this, and we run it for 10. Okay, it works correct. Now we run it for 100. Oops. And we have segmentation fault. Anybody knows where the error is? Yeah, I know that it's difficult, but let's see. It managed to print the count of primes. So probably the error is somewhere here. Let's examine those lines again. So we have all primes array, and which is the length of the total primes, and we iterate through our first primes array, and if the number is prime, we increment this counter and put it here. So probably we could think that we go out of range here, but it's impossible, because we start from zero, and we increment it with the same condition where we increment it here when this is true, so it shouldn't go out of the range. And here we just print the data out of the array. What is the problem? And actually that kind of problem that could be called like a beautiful journey in C, because you have no idea what's going on. And I'll tell you what's the problem. The problem is not here. The problem is here, because when we returned our primes array, we thought that we were returning an array, but actually what we do in C is that we return the pointer to this array. So it's the pointer to array. It's not array itself. And array was allocated in the beginning of the function on the stack, and actually it was valid only in the scope of the function. And when we, after we return the pointer to this array, and we go out of the scope, this array, it's just no more. It has C to be. It expired and gone to meet its maker. So, and we still have a pointer to it. And that is the problem very common for C programmers. So that's called a dangling pointer, the pointer that points to nowhere, to some random memory. And that's why we have a segmentation fault here. So our task was to get our primes as fast as possible. And we have a solution was implemented in C, but C is not a solution. I know that there are programs implemented in C, and probably there are people who are convenient with C and who know how to use C efficiently, and probably they don't make such errors. But something still tells me that sometimes they do. And I personally, after years of Java and Python, just can't imagine how to live in the world where you can suddenly become a pointer to a random data memory. So let's carry on to Rust, finally. Let's implement the same in Rust. Let's see if there's one. What we will do now is just to reimplement our C program but in Rust, and we will see how Rust compiler handles this situation. So here we have our C program, and here we have quite the same Rust program. If you don't understand little syntax details, it doesn't matter because I just want you to understand one basic concept. For example, what we do here, we have the same structure, and actually it denotes the same as in C. We have counter, that is integer, and this is the pointer to array. There is no pointers in Rust, but this denotes, like, it's called slice in Rust. So it doesn't hold the data, it just points to some data external for that structure. So it's actually the same as in C. And then here we allocate our vector, initialize it, and we iterate and increment the counter, and then we do the same as in C, and here we return our vector as a slice. Okay, so let's compile that. Oops, I messed up with the typing. Yes, and we have compilation error, and Rust compiler tells us that primes does not leave long enough. So what it tells us is that exactly what we have here, that, hey, man, you cannot compile that, I won't allow you because you just want to return the pointer to the memory that will expire after we leave the context of this function. And actually this seems to be very strict, but what is better to get this error just in time before you run your program, or to debug some mysterious cementation fault just in a week after you deployed your program to your users, for example. I think this is much better. But let's run it, let's make it work. We won't fix this exact copy because it doesn't make sense. Instead of that, we just implement it from scratch in more idiomatic Rust because Rust has a set, and it has tuples like Python, so we have much shorter solution, and it resembles us Python. And, okay, let's run it. Okay, and for one million, something like... So yes, it's 20 milliseconds, and it's like 25 or 30 times faster than Python. And the concept that helped Rust compiler to deduce the error that we had is called lifetimes, if you are interested about it, read Rustbook. So, concluding our comparison, Python is 25 times slower than Rust, and C doesn't work just. So Rust is fast and safe, but that is exactly what they told us in the beginning, nothing new. And returning to our main topic, can Rust make my Python shine? Yes, but... If you search in the Internet about communication between Rust and Python, you'll quickly find some tutorials about foreign function interfaces. You will even find examples like this. These examples are quite clear and simple, and they work if you try, so this allows you to call Rust code from your Python code, but it's not enough. What if I want to access the Python internals from Rust? What if I want to convert Python string objects to Rust string? What if I want to return a complex object from Rust? What if I want to make Rust library importable as a module in Python? And actually, that is what is needed in real applications. For example, a Python profiler. By the way, who have you ever used a profiler for Python? That's cool. But for those of you who hadn't, I'll tell you what profiler is. Profiler is a program that measures frequency and duration of function calls of another program. And normally, the less overhead it has, the better. So let's make a Python profiler in Rust and to see how it goes. Actually, that was my initial idea when I started to experiment with Rust, and I tried to make, for example, a simple tiny Python profiler. There are two major types of profilers, tracing profilers and sampling profilers, also called statistical profilers. And statistical profilers, they periodically capture frames for any program. And normally, it has less overhead than tracing profiler, which traces all calls in the program. Let's see how to implement statistical profiler in Rust. Here, we won't go step-by-step implementing all the program because we don't have so much time. We'll just focus on two important aspects, and maybe we'll learn something all the way. And the aspects are periodically and frames. So how to run tasks periodically? There is no time in Rust and the library yet, but there is a wonderful library called Mio, Metal I.O. Mio is a lightweight library providing effectively different operational system abstractions, like timing. And we just create an event loop, set up a timing event in it, and then we run a new thread, and we pass our event handler. And what is interesting here is an event handler. Our handler will capture frames and save them to statistics map. That is a sampler object that we create. It's called sampler. And as our timer works in a separate thread, that means that our sampler is a resource that is shared between different threads. So it's a shared mutable resource, which is believed to be very dangerous, as everybody knows, that shared mutable state is a root of all evil. But not in Rust. Rust guarantees you a safe shared mutable state, which sounds like a lie, but it's true. What we do is we just put our sampler into mutics, a mutual exclusion primitive, useful for protecting shared data. When you create a mutics, you transfer ownership of the data into the mutics, and it's immediately given up the access to it. And then any access to the data through the mutics will block threads waiting for the log to become available, thus making the data accessible only through the mutics by one thread at a time. And to pass the reference to another thread, we wrap it with the ARC. ARC provides reference counting through atomic operations, and it's also safe between threads. And having done all that, Rust Compiler guarantees us that we won't have any race conditions. Never. It's just impossible. And not having done that, well, you can't access that object from different threads. Rust Compiler won't allow you to do that. It won't allow you even to pass this mutable data to another thread, so it will be single thread usage only. So Compiler guarantees you that your program will work. And to understand this better, read about ownership and Rust. So capturing current frame. For simplicity, we'll capture only current execution line, as we are not interested in call three at the moment. There are three pieces of information, file name, function name, and line number that we will collect at every tick of our timer. In Python, there is a function in module Cs that is called current frame. Under the hood, it uses a function by thread current frames. Looking into the C Python internals, we will find out that the structure that we need is called actually underscore frame. So we have this underscore frame structure that points to some PyCode object that we also need. And what we need now, we need to convert it somehow from C structure to Rust, structure to be able to use it in Rust. And that could be sometimes hard, because some C types are not very obvious how to map to Rust types. There is no strict mapping, no direct mapping, because they are just absent in idiomatic Rust. So there are special Rust types for that to fulfill that gap. For example, C void is analog of void, and asterisk mute is a special type that reflects C pointers. And normally, there is null in Rust at all, but to check this row pointer, so we have special method is null. So knowing that, we write our code, and remember that when one tiny thing, which is important, when you are calling C function or using row pointer, Rust can't guarantee safety anymore. All such expressions should be within unsafe block. And I think that is what they mean when they say that Rust prevents nearly all C-codes, because that is the word nearly means that it doesn't prevent C-codes in unsafe blocks when you work with C-codes. So knowing these mappings, knowing how to use unsafe blocks, we just create our structures in Rust, and since that we are very close, there is everything that we need, but how to convert Python string to a Rust string? Funny, but that was nearly the hardest problem I faced, because actually it was difficult, and at some point I came up with something like this. So it's just to convert the Python string to the Rust string at the last line, and it did work sometimes. It didn't handle some differences between byte strings and unicode strings, and I was already starting to implement that, but then I came across a library called Rust C Python by Daniel Grunwald, and my life became much more easier. That is a beautiful library, and I highly recommend it. Actually it appeared that the most things that I needed to communicate with Rust and Python were there. I only needed to add some details for my specific case, and also it's a very good example of Rust code. For example, a string conversion using this library looks like that, and also it handles all the cases with unicode string representation, and also it provides very important abstractions like special log corresponding to the global interpreter log in Python. Yes, and this is how you can expose your native Rust library to the Python module using Rust Macros. It's amazing. Read the source. It's very, very cool written. So just this line, there are a lot of more under the hood, and they are very interesting. So, enough. It was much code, and very much details, but I'm not finished yet. Now, as we have a couple of minutes, let's see how our profile works. I lost a bit. Let's keep it named. So, Python's... Let's see. So we're profiling this. Erlothance.py that we have written at the first time. Okay, that's not interesting. We need to... Oops. Wow, I thought that it will happen at some point. It's impossible to make a live demo without fails. Python, through or ahead? Yeah, so we comment out our print, Python, snake, matter. Yeah, let's start with million. Okay, and this was fast. Let's make it for 10 million. Okay, it's very simple now. It's very basic, but whatever it tells us, that the 85% of our time went in the light 8, and 14% in light 7, actually nearly 1% was for output of this line. So line 8 is this one. So what we are doing here is we are updating our set in array. And what we are doing here, 15, 14%, is we're incrementing counter inside the array. It looks believable, so it looks very logical. Oh, yes. And if you still care about performance in your Python applications, but you don't want to dig that deep into the native code, I recommend you to listen to the talk of my colleague, Katrina Tuzovar, that will take place on 23rd of July and Thursday. She will show you how to write performance Python code without using any C or REST. Thank you for your attention. Questions? Is it possible to distribute REST code as a Python package which can be installed using people? Yes, I came across a little library that allows you easily to write your setup.py this way that it will compile your REST code if you have REST compiled and installed. Maybe it even downloads it. I don't know, I never tried it. I wanted to try it, but had no time before. And you just type setup.py build to install, and it builds your REST from scratch. Like for C, setup.py allows you to install C extensions. So there is a library that allows it for REST also. More questions? As you said, only one question. Yeah, that was a good one. Thank you again. Thank you.
|
Dmitry Trofimov - Can Rust make Python shine? Rust is a new programming language from Mozilla. It is fast, safe and beautiful. It is also a very good option when needing performance. In this talk we're going to look at Rust and see what it offers and how we can leverage it as Python developers. And we'll do it with a case study: a statistical profiler for Python.
|
10.5446/20098 (DOI)
|
So, hello everybody. My name is Kozmin. I'm one of the co-founders of the RO Python, the biggest Python community in Romania. And I'm doing this with a lot of cool guys which are present in this room right here. Also in my full working time, I'm working for Cloud-based Solutions and I'm a Cloud engineer there. Working on a product. So, today I'm going to talk about Argus, the omission CI. I'm calling like this because he knows everything and is present everywhere when we're doing tests. Just before I talk about Argus, we need to know something about the Clouds and the Cloud Inutilization Services. Also, I'll talk about the components of Argus, the most important thing of it, the configuration file in which the actual tests relies on, and how to use it, and also creating a test. So, what are Clouds by IT meaning? Simply put, the Clouds are some places overhose under a network in which data is processed, stored, and served by the user rather than using your local computer. And such Clouds came across tact as platform as a service, infrastructure as a service, and software as a service. But we will focus on infrastructure as a service. And there are many popular infrastructures like OpenStack, OpenMebula, CloudStack. If you don't know, OpenStack is the biggest project, open source project in the world, mostly written in Python. So, what we are doing with these Clouds? We mainly create instances, VM, virtual machines like you used to create, I don't know, virtual box or VMware, and they are put to life by your local hypervisor installed there. Most important things about Clouds is when we're creating these instances, we also need to initialize them, to configure them. So, these make the metadata providers to be very important and to know them and how particular aspects of the instance which should be configured are served. So, a solution to this are the Cloud initialization services. Cloud init, which is developed by Canonical and OpenSource, is working very fine, but some guy came across with a new idea for creating Cloud Insighted Services for Windows, and many other platforms will be supported in the future. So, he came with Cloud based init, which is mainly used to initialize Windows machines. So, this is how it looks like. It came in the shape of installer, runs as a service, it installs with various options. It's designed for the anti-systems Windows, is open source and written in Python, and supports many popular Clouds. By supporting many popular Clouds, I mean he knows about most of the metadata providers. And it's also independent of the hypervisor, of the method which the instance is virtualized, because the service itself doesn't have nothing to do with infrastructure. It only installs on the instance and configures it, just like that. And now, some talks about merging the Cloud based init project with the Cloud init, and we're talking with the guys from Canonical to do that faster. So, this is how it runs. It's normal services. And after it runs, it detects the metadata provider and using that service, that loaded service, it's using some plugins to configure the instance, because this is the most important thing. Like Clouds name, networking, local scripts, it can execute scripts on that machine. Also, store SSH publicies, and it also relies on a configuration file, which looks like this. You can specify an username, groups, where to store logs, which metadata services you want to use in particular, and also the plugins. If you don't specify them, all of them will get executed. So, enough with that. This talk is about Argus. We need somehow to test it. So, at first, the init testing and other gates like PEP8, Flake8 are not enough, just to test the project itself. We need to do some kind of integration test to actually test how it works and if it does what it really should do. So, we've done some manual testing under VM instances, directly on your host or through Gress, or even through infrastructures like I've said earlier, like OpenEBula, Cloud Stack, and mainly OpenStack. So, we need somehow to automate these things. This is just a horizon interface and on the right is a VNC console, so the testing part was pretty hard. We can also use an RDP. So, we've came with Argus. Argus is a beautiful open source project conceived and designed by Claudio Poppa, which is core maintainer at Pylint, but also the only maintainer. So, through this project, we are creating integration tests, not only for the cloud-based init project. We can create tests for everything we want to test it because it supports any kind of testing concept. It just creates a virtual machine and from there you can do everything you want by writing some code in the project and bundle all the aspects you've been doing there in our configuration file for Argus. So, it's written in Python. It uses Tempest to create the instance itself. Tempest is another project by the OpenStack community which tests the OpenStack infrastructure and how the instances respond and if everything works as it should. So, it's scenario-based and it gives unit tests like reports. Like you're seeing in the right, there are some failures, tracebacks and like unit tests shows you when you run a test, it gives you the number of failures which tests have run and in what time. So, Argus, to understand Argus, we need first to... To understand it, we need to first understand its components. So, it's created of scenarios. Those scenarios are encapsulated, are bundled together. The recipe, tests, introspection module and also the runner which runs everything above. And the most important thing, the configuration file in which you create these kind of scenarios. I will show you in detail what are these. So, this all is looking a scenario. It's just some base code and abstract methods. You don't have to customize these objects. You can use the already created one or if you want to do different things, you can inherit from there and create your custom things. This is how it looks. It all looks a recipe. The introspection module which gathers the details from the instance and the actual test which are run not over the instance but in your local computer. So, how these are working together is the scenario which encapsulates the recipe. The recipe configures the instance after it's created by the scenario. Then, using the introspection module, you gather details from the instance and comparing with the expected ones. So, the tests are getting executed and you get failures for success. The configuration file. Here we can have basic settings like credentials, details about the image, the flavor of it, details in particular for the project you test, cloud-based in it for example. A basic scenario which bundles everything and you can inherit from there and create your own scenarios with your own tests. You can customize all the other things like the recipe use or the introspection module if you don't like the default one. This is all looks. This is just a sneak peek from the configuration file. Above, we define image you must use because to install, to create a virtual machine, that infrastructure also need to provide an image. That image you deploy through glance and open stack. After that, the infrastructure, you can see it's a reference ID and put it there. This is how you tell Argus to use that image when creating the instance. Below is the basic type of scenario in which you gather the things you want to use. From that basic scenario, you can create custom scenarios with your own recipes, introspection, even images and other settings. Also, you can customize the instances you create by providing metadata to the instance which is passed through Tempest. Also, custom user data. Through Argus, you can also create environments. Suppose you want to customize the environment before creating an instance. For example, like modifying a configuration file. A solution to that is to stop the services and configure the configuration file of a component from the open stack ecosystem. For example, Novaconf. After you write your desired details, you restart the services and run Argus as normal. Argus supports these kinds of things through environments which can be specified in that configuration file. Argus will do those automatically. Another cool thing about Argus is that you can mimic different infrastructures. You can make cloud internalization services which run on an instance to think that it runs on a different infrastructure. Like cloud stack or open eBula or everything to support. You can do that by creating custom web servers through Argus and do additional changes to the scenario and the recipe. In a way to make the cloud internalization services when it runs, you can use that particular services. When it uses that service, it just thinks that it runs on a different infrastructure. By seeing that new metadata provider is available other than the open stack. Also, you can attach a drive with a shell on it named context.sh for example. When it finds that file, it thinks that it runs under open eBula even if it runs on open stack. This way you can test different structures which is a cool thing in Argus. This is how we use it. There are a lot of options. We mainly focus on the cloud because most of the time we are testing cloud based in it. But with Argus, you can test everything you want. As unit test has a fail fast option to fail at the first encounter error, you can pass the most important thing for it. The configuration file, you can put a pause before the test will begin. Also, you can test various operating system types or scenarios types. When you define that configuration file, you also specify a type. Another important thing, you can specify an output directory. This is very important because most of the things that are on the instance are also making some logs. Those logs need somehow to be retrieved because after they are retrieved, you check those logs and this way you make sure that everything runs smooth on that instance. You can also test custom builds and custom architectures. Another important thing is that you can modify the actual installer by providing an archive with different binary files. Cloud based in it was using Python on the x86 architecture and I wanted to test that on x64, so I built a custom installer which I passed it through the patch install command. So I test this way, I test the x64 too. With git command, you can test patches which are not integrated because this is the main reason we are using Argus to test cloud based in it. Just to make sure the patch really does what it should do. Below is a real life example. With Argus, I test cloud based in it. With the Argus.conf configuration file, I put a pause on it if I want to manually intervene on that instance. I specify the output, a directory named cblogs and the architecture x64. The most important thing is the git command which specifies the git patch and check out commands retrieved from the review.apponstack.org site. So this way we can test a particular patch. So how you develop a test? Wherever you have a patch or not for cloud based in it and if you want to test something, you can just use those available by default scenarios, RISP test or introspection modules, but you could also create custom ones. You gather all these things together in a scenario underscore and whatever you like, name, group in the configuration file and there you put all these things. So then you run Argus with that configuration file and when it came, it come across that particular group, it just create a new instance using the specified scenario, then it configures the instance with your recipe or the default one. Then it runs the actual test using the introspection module to retrieve details from the instance and comparing the expected ones just to make sure that what you expect is the same for not giving favors. So I think it's perfect. So thanks. This was a quick talk. I'm not usually put on the last slide many other links because people don't click them, but below is the Argus project. So we have Argus GitHub and above are other details about the project we test cloud based in it. And that's all questions. No questions. All right.
|
Cosmin Poieana - Argus - the omniscient CI Bring the continuous integration to a new level, through a platform/project independent framework able to give you unittest-like reports. Argus is a scenario-based application written in Python, driven by custom recipes under configurable environments, that can be used for testing a wide variety of small and big projects, with the ability of querying live data from the in-test application. Until now, it's successfully used with [cloudbase-init] (a robust cloud initialization service for instances) under OpenStack and not only, due to its extensiveness and the ability to mimic different infrastructures. The goals of this talk are to show its generic scalability, how simple is to create such kind of recipes, the relationship between scenarios, introspection and tests and, but not last, the unlimited freedom of creating very custom aspects of these entities which lead to relevant and in-depth ready for analysis logs. There are no major prerequisites to understand it, just to be familiar with Python and optionally have a focus on cloud infrastructures.
|
10.5446/20097 (DOI)
|
Thank you guys. It's really nice to see so many people interested in static analysis here and this year is a special year for Pylint because it's 12th birthday. So, yeah, basically it's the oldest static analysis tool in Python that is still maintained, which is really nice. So in this talk, I'm going to talk a little bit about Pylint's history and we'll go through a detour in static analysis and we'll see where that goes. So, what is this Lint thing I'm talking about? Basically, a Lint is like a tool you use to analyze your code in order to find bugs or potential errors or style checks or stuff like that. But saying this about Pylint is an understatement since it can detect a lot more than a normal Linter Linter. It's also a style checker which enforces Pepet roles. It's a structural analyzer looking at your classes and at your special methods and checking if they are implemented correctly. And it's also a type checker. So, Pylint works by using a technique called static analysis, which is basically the act of analyzing code without actually executing it. And if you don't use static analysis right now, you should use it. You should use any other tool, not just Pylint, because it can really help you in your day-to-day job. For instance, you can use static analysis if you have a lot of tests and it takes a lot of time to run. You just want to check that there are no obvious errors in your code or stuff like that. Or if you have big legacy systems which don't have tests at all, in fact, as my first job experience, I worked in a company where we have a lot of big legacy systems, but they didn't have tests. So, using static analysis was paramount for checking that everything works before going in production. And you can use it as a form of doing reviews. Well, of course, it's not equivalent to a manual review, but it's better than no review at all. So, here's a piece of code which has a couple of problems. And I'm going to show how Pylint detects this kind of stuff. For instance, if you can see, I imported OS, but I didn't use it across my program. And I also defined a variable which is just there. It's not used. Of course, this is not a problem per se, but it leads to ugly code or unmaintainable code or stuff like that. And more important is this block here from line 8 to 10. Let's just say that I invented that code incorrectly. And that code now will not run because it's an after-run erase statement. So, basically, it's useless code. Also, at line 10, I intended to call the execute method, but I didn't. So, that's also a problem. And as you can see down here, Pylint detects this kind of stuff and properly emits that there is something fishy in your code. Okay. But we can detect more serious bugs than this, like using undefined variables or trying to access members that don't exist or calling functions that aren't functions like calling an integer. And as you can see here, Pylint detects this kind of stuff. And it's really nice for aesthetic analysis tool to tell you that you made a mistake in your program. And as I was saying before, Pylint is also a structural analyzer checking that your special methods are properly implemented or that your classes are written correctly. And in this case, the dunder exit method needs three parameters in order to work properly. And Pylint detects this kind of stuff. It emits that warning. Right? Another nice error that Pylint can detect is this one. Basically, here I have a function and then I have an if statement with func as a conditional for the if statement. And what I intended here to do initially was to call func, but I forgot to write the parents. And this statement will always be true. And that if statement, if branch will always be taken. So if you write this for this piece of code, you might, how to say, if you write this for this piece of code, you might overlook it if you didn't test it properly. So Pylint can help you in this kind of situation as well. Take a couple of moments to figure out what's the problem here. If you know the answer, you'll really, really loud. Anyone? Sorry? Var is being changed in the loop. So what's the output of this code? Who said that? Yeah, that's it. Sorry. Basically, what we did is we created a closure here and Var was looked up into parent scope when called. And Var was looked up into parent scope and called. And for this particular case, the parent scope was the least comprehension, at least on Python 3. So when the for loop was executed and callable was called, basically, Var was the last value from the for loop, which was 9. So that's a bug. And if you try to do a review, you might have missed this stuff and Pylint can help you deal with this kind of situation. Okay? A little bit of history about Pylint. It was created in 2003 by a French company called Logilab. They used to maintain it for a while for 10 years, but now they're not so involved anymore. In fact, they're not involved anymore in Pylint's development. Even though many still think that about them, which is wrong, I know that Google uses a modified version of Pylint internally for their use cases because they had a maintainer for GPylint that was also maintainer for Pylint and they used to push upstream a lot of changes. And some statistics, according to ohlog.net, we have over 30,000 lines of code, which is pretty big, but not big enough because we can detect all the problems in your code. And unfortunately, Pylint is GPL-licensed because GPL was really cool back then in 2003. It still is. Yeah. And here's my involvement with Pylint. Basically, I started working on it in 2013 when Pylint 1.0 was released and up to Pylint 1.3, I became a committer, a maintainer, the only maintainer, and the guide that pushes the thing forward. I'm planning to release Pylint 2.0 in 2016 if my time allows it and I'm going to have some advanced techniques in it. It's still right now pretty advanced, but it's going to be even more advanced than that. Basically, I'm going to use a technique called abstract interpretation where I'm going to interpret statically your code, not actually executing, but statically. And also, I'm going to add support for PAP484, what Widow stock was about, and changing the things internally, like having control flow graphs and having a symbolic execution engine for it. Yeah. Right. And at this point, you might ask yourself, okay, this is cool, but how it works and why does it work statically? First of all, Pylint is split in actually in two components. In the real checker, which is Pylint, and its inference engine and the component that understands Python, which is Astray. We're following the general pattern of building a linter. Basically, we're using abstract syntax trees, even though our abstract syntax trees are augmented with a couple of functionalities that help us in building the inference engine. And internally, Astray uses the built-in module AST, which kind of looks like this when using it. You import pars and you give it a string, and from there, you get an abstract syntax tree, which basically looks like this. So, if you can see at the top of the tree, we have a lambda, which has a couple of arguments. And the blue stuff out there, the blue thingy is a node, and the rest of them are just attributes of those nodes. So, as you can see, it's pretty structured, and you can, just by looking at it, you can reason about code. Right. Even though AST is great, it's built-in, we don't have to write our own parser, it's not perfect because it's not backwards compatible, not even for minor versions of Python releases like 3.4 and 3.5. Basically, they are going to add new nodes or remove nodes or change things out there. And Astray strives to be a backward compatible wrapper over AST that you can use for Python 2.7, Python 3.5, and as well for PyPy or Jyton or any other interpreter which supports abstract syntax trees. And the API is quite similar with the AST. As you can see, it doesn't differ too much. We have basically the same functions. Yeah. So, if you're using AST right now and you want something more capable, you could use Astray. And as I mentioned, our Astray nodes are quite similar with one from AST, but we augment them with a couple of capabilities specific for our purposes. Like in this example, for instance, you can actually retrieve the parent of a node and you can walk up in the tree starting from one node. Also, you can walk down by using the getchildren method. Right? Also, you can retrieve the lexical scope of a variable. For instance, in this particular case, the lexical scope of foo is the function where it is defined. And down there, at least comprehension, basically I'm doing foo for foo in range 10 and the scope of foo, at least on Python 3, is the least comprehension itself. So foo doesn't leak outside. Yeah. And PyLint can actually know this kind of stuff and Astray. Some nodes are augmented even more than that. Like in this example, we have a couple of classes and we're using a meta class, we're using ABC, ABC meta, and we're defining some slots for our class. And yeah. Let's see what PyLint can do, what Astray can do with this code. So we can retrieve the slots of the class using the slots method or the meta class using the meta class method or using the MRO method. We could retrieve the method resolution order of your classes, which is actually the meta resolution order when running Python over that script or over that class, which is nice. Yeah. But the most important part about Astray is not the AST itself, but the capability of doing inference. Basically, inference means the act of resolving what a node really represents. Like if you have a name and that name represents something like a variable or a function call, you want to infer that and you want to see what's at the end of the tree for that particular node. Our nodes basically implement the Python semantics. For instance, the callfunk inference rule will be the return values of the callfunk or for least comprehensions, its inference rule would be the actual list that is returned when evaluating the least comprehension. And our inference also does some partial abstract interpretation, but it's partial because it's not working for all the cases in your code. And in this particular example, we have a function which adds two values, the argument, and as you can see in this particular case, when in inferring what this function actually returns, in fact, when inferring what's the result of the function called, Astray knows that it's 48 because 24 plus 24, it's 48. This is a more complex example involving binary operators. And if you know, the rules for binary operators in Python goes like that. If you have two different objects and you try to add them, first, the Dunder add method of the left-hand side will be called. And if that doesn't return not implemented, then I think the right-hand side object, no, no, I think that the Dunder add method of the right-hand side object will be called. And in this particular example, we're having in the left-hand side a supertype, which is A. As you can see, A is a supertype for B because B has, as its base class, A. And in the right-hand side, we have B, which is a subtype of A. And the rules are a little bit different in this particular case because if the semantics would be the same as for the different classes rule, then the first rule, then the first method, A, Dunder add, will always be called, which is wrong. And in this particular example, what will happen is that first the Dunder add method of B will be called. In this particular example, it returns not implemented, so it will fall back to the other one, which is A, the Dunder add method from A. And if you do the calculations, you can see that SRAD is right. The result of that operation is actually 45. Now, SRAD is great, but they can't really understand your code. I mean, they can't really understand your full code, so we have to deal with this kind of situation. And we're providing a couple of APIs that could help you, like having node transforms. Basically, with a node transform, we can modify a part of the AST to be something else, like you have a function call, and instead of that function call, you want to do inlining, replacing the AST node with the result of the function call or anything you'd like. And we can do that with this API. Basically, that API is a function which should accept one parameter, that parameter being the original node, and should return either the node modified or a new node. And you just register that function with an internal manager, some implementation detail, but anyway, you just register your transform function, and when doing SRAD dot parse, that transform function will be called for that particular type of node. In this example, I'm registering a transform for the class node, and also you can apply a filter function because you don't actually want to change all class nodes in your code. You want to change something in particular. And as you can see there, I'm just giving it filter functions for filtering any other class that is not 6 add meta class. The same thing can be used for inference rules because at some point, really, you want to have different inference semantics than the Python offers you. So using the same API, you can provide a custom inference rule for, I don't know, you want to infer these comprehensions differently, or you want to infer function calls differently. So you're using this inference rule, and you're changing basically the semantics of Python in your AST. And we're using our own dog food in this example, not in this example, we're using our dog food with inference rules because we have inference rules for built-ins, as I'll show you immediately. Basically, with inference rules, we're understanding existence, is subclass, we're understanding the other, has other type, callable, list, set, whatever. We're understanding, not in particular with inference rules, binary arithmetic operations, really, really good logical operators or comparisons. Also, we're understanding context managers and list, set, string indexing, slicing, whatever. Yeah. As earlier, take a couple of moments to see where are the bugs in this particular code. There should be like three bugs, but if you can find more, join Pylinx thing. Also, if you know the answer, yell really loud. And as a hint, how super works, basically, the first argument of super specifies the object from which the method resolution order will be retrieved. Yeah, that's good. Because in that particular case, super C and self would be, the method resolution order in this particular case, I don't know what it is, but super C self will call B because A doesn't have Boo, yeah, and it's multiple cooperative inheritance, if I'm not mistaken, and B Boo will be called and B Boo doesn't accept two arguments. Anyone? Yeah, there's no spa and there is a full. Sorry, sorry, there are too many voices. One at a time. Full is not color. Exactly. So, Pylinx detects this stuff because asteroid knows Python really well. And as you can see, what Pylinx says, there are too many position arguments at line 14, which is actually true. Full is not callable, of course, because full is an integer. And super V has no spam number, which is actually true because it's spam, not spam. Yeah. Here's a more complex example of asteroid capabilities. And as you can see here, asteroid understands, least indexing understands has outer, callable, get outer. And at that particular line where met is retrieved, what will happen is that has outer call will return true because A has the method called met. And A met is callable. And get out from A, we'll return the met object. So it will be like true and true and A dot met. And in this particular case, the last value, which is also true, will be return. So met will be A dot met. And then the context manager is invoked. And the context manager will return real func, which has no argument whatsoever. And we're going to call that. And Pylinx will say that, hey, you use too many position arguments in your method call. You should change that. Yeah. So this kind of stuff that asteroid knows about Python code and Pylinx knows about it as well. Yeah. I don't know. I'm going really fast or how much time do I have? Okay. Thank you. Pylinx is not so complicated. Astrid is the most complicated thing. Basically, it's like a fancy walker of the AST. And it has a couple of patterns, more than a couple, because Pylinx can detect almost 180 type of errors or verifications. And basically, we are using the visitor pattern to visit each node because that pattern really decouples your data structures from your algorithms. And a small example of how visitor pattern works. Let's say we implement in our own checker, visit call func. And in this particular example, we're importing collections. And we're obtaining the default attribute from collections. There's no default attribute in collections, by the way, it's default dict. And what will happen is that visit get other will be called with the node, which will be the get other node. Afterwards, we'll infer what that node really represents. I mean, what the parent of the statement really represents, the parent will be the name, the name collections. And we need to infer it because we should know what that node represents at the end of the AST. Afterwards, we're calling the data method with the result from the inference. And if that doesn't raise any error, such as not found that yeah, collections has default to default attribute. Else, we're going to emit a no member error. But before that, we're going to have a couple of filters like the owner is a mixing class because you might not have that attribute in the mixing class. Or I don't know, the owner is a class with unknown bases, like bases from extension modules. And we currently understand extension modules, by the way. Yeah. I'll go really fast from now on because I don't have much time left. Basically, abstract interpretation will have piling for this kind of code. In this particular example, the dunder date is updated with the dictionary of, I don't know, what kind of arguments will be passed. And at line five, we will call some arguments set in dunder in it or whatever other attribute. And that attribute might not exist. And what happens right now is that piling says, hey, you don't have that attribute. But we're actually having that attribute because we just inserted it at line two. So that's what abstract interpretation is going to fix because you will just interpret every line, every logical line in your code and it will know at the end what side effects each line had. Okay. We have multiple type of checkers and errors, such as conventions for Pepe rules, refectorings, various warnings which aren't necessarily bugs in your code. And actual bugs like no member or not global, things like that. And we have two types of checkers, AST based and token based. The token based checkers are for, I don't know, line is too long or bad indentation or other similar examples. Piling has a really vibrant community. There are a lot of external packages for improving in-furnace, for improving piling checkers out there. You could write your own if you want to. The problem is that they are pure Python, so you need to write Python code in order to have a custom checker. Yeah, it comes with a couple of features, extra features like you could generate UML diagrams from your code or you could spell check your doc strings or comments. These are disabled by default. Or you could use the Python 3 porting checker, which is a checker when activated, all other checkers are disabled and it will emit stuff that it's not going to work in Python 3 anymore. Like using remove syntax or remove built-ins or special methods. Or my favorite is using map filter or reduce in non-interval contexts. As you can see in this particular example, on Python 3, map is lazy evaluated. In this particular example, download URL will never be called because at least on Python 3 because you should evaluate it first. Yeah, it will return a generator, stuff like that. Okay. Oh, there are a lot of similar tools like Pyflex, Mypy, Pychecker. Pychecker is the forefather of PyLint. I should say a couple of things. Even Pychecker is now defunct and dead for a couple of years. It was way more advanced than many static analysis tools that exist currently in Python. As you can see in this example, we are unpacking three items in two variables or we are having a constant check or we are catching an exception, which is not really an exception, true. And Pyflex, as you can see, detected almost nothing while Pychecker detected all of these errors. Yeah, there's also Jedi and Mypy. Okay. Yeah. One final note is that users, at least a part of them, actually expects this kind of code to be understood. But really, that's not actually possible. So if you want static analysis tools to understand this, just don't do it. Yeah. Yeah, and this is basically the future of PyLint. We'll have PyLint 2.0 next year. We'll have full flow control analysis, better data model, like understanding the scriptors or having the proper attribute access logic. It's not the same in PyLint as it is in Python. And I'm interested in bringing more contributors into the project. And my final slide is, hey, what your talk title is, how do I stop worrying and start loving the bugs and so on? Well, PyLint helps if you're going to use it and if you're going to write as many tests as possible. So use PyLint. Thank you. Thank you. Are we going to any questions or many questions? That's really odd. Yeah, there are a lot of questions. Hi. I actually have a comment or a word of caution and a question. The comment is, if you're going to start PyLint, it will take you some time to configure it to make it shut up about things which don't interest you. But it is well worth the time to do so. And the question was, how did you get into this kind of parsing Python and ASTs? I mean, I tried to write PyLint plugins because there isn't really much documentation for PyLint and that's really hard to get into this topic. I was a user of PyLint before and it had a bug. It didn't detect stuff like it didn't detect unbalanced topo unpacking, like unpacking three items into two items. And I had a bug in production with that stuff and I just wanted PyLint to, hey, tell me about this. So I implemented that check and I started from there. Okay. Thank you. One other question. What are your thoughts about PyLama? PyLama? Yeah. If I'm not mistaken, it's a wrapper over multiple linters like PyLint and Flake8. Yeah. So it just takes PyLint and other stuff. I never use it actually. Okay. More questions? There's one here. I'm not sure what is. Can you be more visible? Here. At the front, okay. Okay. Hello. So regarding the first question, the most common complaint about PyLint is that it has too many checks active. So maybe what if there will be like some sort of configuration wizard that will show you examples and rescue if you want this or not? Yeah. That would be nice to have. But I don't have time to write it myself actually. Sorry. Yeah. That's the question. Is there a sprint on PyLint? Yeah, there is. There is Saturday and Sunday. Thank you. And can you join if you're a beginner? Sorry. Can you join to the sprint if you're a beginner? Yeah, of course. Great. More questions? All right. I think it's fair enough. Thank you again. Closing. Thank you. Okay.
|
Claudiu Popa - 12 years of Pylint (or How I learned to stop worrying about bugs) Given the dynamic nature of Python, some bugs tend to creep in our codebases. Innocents NameErrors or hard-to-find bugs with variables used in a closure, but defined in a loop, they all stand no chance in front of Pylint. In this talk, I'll present one of the oldest static analysis tools for Python, with emphasis on what it can do to understand your Python code. Pylint is both a style checker, enforcing PEP 8 rules, as well as a code checker in the vein of pyflakes and the likes, but its true power isn't always obvious to the eye of beholder. It can detect simple bugs such as unused variables and imports, but it can also detect more complicated cases such as invalid arguments passed to functions, it understands the method resolution order of your classes and what special methods aren't implemented correctly. Starting from abstract syntax trees, we'll go through its inference engine and we'll see how Pylint understands the logical flow of your program and what sort of type hinting techniques are used to improve its inference, including PEP 484 type hints. As a bonus, I'll show how it can be used to help you port your long-forgotten library to Python 3, using its new --py3k mode.
|
10.5446/20096 (DOI)
|
Hello, good afternoon, welcome to my talk, Skill Your Data, Not Your Process. Welcome to the Blaze ecosystem. So a brief introduction. I'm a data scientist at Continuum Analytics. We have a booth out there, so if any time during the week you want to come talk to me, I'll be there most of the time. I'm from Barcelona, but I'm currently living in Austin, Texas. So also if you're from Spain, you can talk to me in Spanish, Catalan. If you're not, you can also talk to me in English or German. This is my website. I have a couple of the talks I've given at other Python conferences. You can also check them out. Just a little bit, a brief thing about Continuum Analytics. Guido mentioned the company in his keynote today. We offer a free Python distribution called Anaconda. It's very popular in the SciPy community for libraries that have C and for trend bindings. It makes it very easy to install them. We are very integrated in the open source community. We sponsor several projects, condoblase, das, bokeh, and numba. And we're a proud sponsor of a lot of Python conferences. You're a Python, PyCon, SciPy, PyData. And we're also hiring. So we're going to be tomorrow at the hiring event. If anyone is interested, also come talk to us at the booth. That's our website. A little bit about this talk. I'm going to be organizing it in three different areas. First, a little bit about what data science is and what the stack that I'm presenting today brings to the data science community. Then a little bit of what they call the data science trifers. And we'll hear a little bit about that later. And then inside the Blaze ecosystem, there are many projects. And I'm going to be mainly talking about four. Blaze, Data Shape, Odin Desk, and how each one relates to each other. You can follow the slides online. If for some reason you're not able to see it from back there. There's also GitHub repository where I have the guide and read me for reproducing the examples that I'm going to show in my slides. So you can also try that. So first area, five areas of data science. So many people have their own definition of what data science means. For me, data science is more just than just machine learning and stats. Actually, data science is just a rebranding of five fields coming together to solve data problems. A lot of people in the scientific community have already been solving large scale analytic problems. Scientists deal with large amounts of data. So they have already worked with it. Then there's this group of machine learning and stats people, the analytics community with a databases and queries. Web is where we find a lot of the data nowadays. And then there's the distributed systems community with all the Hadoop and Spark that are trying to scale those problems too. If we try to find what personas are working in each of the different fields, there's some terms. There are people calling data scientists, people that are in the machine learning stats that are actually most more concerned on modeling. We have people in the data business analysts, web developers, and in the distributed systems, a lot of architects, data engineers, and then in the scientific computing, all the research and computational scientists. If we have to find one word of what each of these personas cares about, maybe these ones are a proposal of words that identify them. So machine learning people in stats care about models, about finding the right models and solve their problems. People in the analytics community are mainly concerned with reporting, building rather than the reports and metrics. In the web development, building an application in terms of relation with data science, applications that portray accurately that your data problem. In the distributed systems, you're concerned in your pipeline, your architecture that you're building. And in the scientific computing world, in the algorithms. So if we use more than one word, what's the vocabulary of those people in those areas? We see data scientists use words like models, supervised, unsupervised, clustering, dimension, to reduction, cross-validation. In the analytics world, people concerned with joining with databases, with finding, filtering, getting summary statistics. In the web, we have scraping and crawling to gather data, gather information, things like interactive data visualizations. In the distributed systems, we have all the Hadoop Spark ecosystem, working about clusters, stream processing, et cetera. In the scientific computing, people are concerned with GPUs, with graphs, with algorithms, with computation power. What are the tools that each of these personas and fields are working with? So in the machine learning, we find R, FIANO, PsychiLearn. In the analytics community, all the databases, SQLL community, people working with Excel too. In the web community, we have all the web frameworks. We have D3. We have Bokeh for interactive visualizations. We have scrapers. We have a way to share our code with Jupyter notebooks. In the distributed systems, as I mentioned, we have Spark. We have Hadoop. We have Luigi. We have Kafka. We have all these other tools being built around it. In the scientific computing, we have the core of a lot of the libraries that are used by the machine learning and stats libraries, like NumPy, SciPy, X-ray, PyTables, Syson, Numba, et cetera. So this is to provide a general picture of what the status of the data science ecosystem is right now. So if we take a look at those tools, what are the three edges that they're bringing together? There's three things. There's data. There's the computational engine behind it. And then there's the expression. Expression is what you're asking for. So data is all about metadata of information on that data and how you store it in containers, containers meaning how the data is stored in either your memory or your disk. We then have engine that's a computation power, what gets executed. And then we have expression, meaning the API, the syntax, the language, how rich is that to allow you to express what you want to compute. So what are we looking for in each of these edges? In metadata, we're looking for semantics. In storage containers, compression and accessibility to your data. In engine, we're looking for performance, being able to do that as fast as possible. And in expressions, we want simplicity. We want to be able to express what we want to do in a language that's very close to our human language. So just to have an example, what all those things mean in other languages. For example, we have different file formats, right? hdf5, netcdf, JSON, CSV, SQLite, V-Calls. But we also have memory containers, like pandas, data frames, or non-pia arrays. In terms of semantics, we have a lot of types. We have fields. We have names. We have description of your data. We have relationships between the fields of your data. In terms of computation, we have different computation engines that perform those, like Spark, like Syson, Fortran, Python itself, or the libraries that are built on top of it. In terms of the API syntax language, I'm talking about the API, the non-pia API, the pandas MPI, the binding that we have to other libraries that allows us to express that in an easy way. We also have many of the SQL dialects, et cetera. So in the core, all those libraries, non-pipandas, databases, and Spark, have somehow each of those three edges in this triangle. Let's put a simple example. Imagine non-pi. So non-pi, we have non-pi d-types, right, that allows us to express the types of the fields in our data. We also have a way to contain the data for non-pi with non-pi and the arrays. But non-pi itself needs to compute things, needs to compute what the user is asking for. And it has bindings to C and Fortran, also Python. And in terms of the API, we have a non-pi MPI, right? That's how do you express what you want? How do you express the fact that you want to create a non-pi array? So in all of these systems, they're happy and sad faces. Non-pi and pandas mainly are limited by the memory of your laptop or your device. But people, scientists, like to express the things with arrays. It's an API that has attract a lot of attention in the scientific community. Data scientists analysts also really like the pandas API, the fact that they can deal with data frames, tabular manner. In the database world, well, we have a lot of dialects, SQL. There's a lot of overhead to set up. And kind of the Spark role, yes, it has come to expand the Hadoop ecosystem to more of the data scientists and people that are further away of the engineering side. But it hasn't quite yet reached the gap to help you in all the cases of your data. In smaller sets, you still have a lot of overhead to perform it. So let's take a look at what the Blaze ecosystem brings to this ecosystem that I've just mentioned. So Blaze started out as how do we expand non-pi and pandas to out-of-core computing, to not be limited just by the RAM that your laptop has. And from there, there's several spin-off projects that have come along the way with things that we've learned. So first, we needed to expand some of the non-pi and pandas limitations of expressing the metadata that's in them. And that's where data shape came to play. There's a data description language that's more general than what non-pi and pandas implemented. Then we had Dyn, which is a dynamic multidimensional array, which is a library written in C and that has Python bindings. We also found that there was a lot of need to move data around. Data scientists were working with different file formats. And there wasn't always an easy way to move from one format to another one or from one place to the other one. So that's where Odo was a spin-off project that came out of the Blaze repository. We then have Numba, which is a code optimization adjusting time compiler. Inside Blaze, we have what we call Blaze as a project which has been kind of the core, which is just an interface to query data in different backends. We have Desk that allows us to do easy parallel computing. Castro, that's a column store on this column store partitioned. And B-Calls, also a column store and also a query language that allows us to get data out of it. So if we place all these projects in this table that we had before, here is where it comes from now. Data shape is this metadata extraction of expressing your data in different formats. We have Dyn, which stores data in a multi-dimensional array. We have Odo that allows us to switch from one container to another container, from one from NumPyRays to Pandas to a lot of the backends that Blaze uses. We have Numba that allows us to optimize the code, Dast for parallel computing, and Blaze being this common interface that allows us to query everything in a unified manner without having to learn each of the different APIs. So if we put those packages in our triangle, we find data shape as the metadata in the metadata section. We have Castra in the storage as one of those containers. Odo that allows us to switch from one to the other one. We have on the engine computational the power of parallelizing with Dask and to optimizing the code with Numba and expressing everything with Blaze. And then Dyn and B-Calls, which also are data containers and also have computation power to resolve your, whatever you want to need to compute. If we now place those projects, I'm kind of giving you the overall picture of where they all fit. So a lot of the analytics people are interested in tabular data formats like Pandas DataFrame. So Das DataFrame is there for them. Blaze as a unified query interface. We have Odo that can be used mainly for everyone. It's just like a utility function to move data around. We have a big support in the scientific computing world. And we're solving a lot of those underlying problems that then are used by many of the libraries in the Python ecosystem for machine learning and stats. But we're kind of bringing like the solving the underlying problems. And then we are also engaging with the distributed system world with what's called Das Distributed. And then we're also going to see Blaze Server, which allows us to serve all this data in different formats through a unified API. So kind of like the ideal Blaze was this connector to all these different fields in the data size community and bringing everything in a unified manner. So if we just remove some of the rest, and just going to focus on what we're going to talk about in this talk, we're mainly going to be talking about four of our projects, Odo, Blaze, Das, and Data Shape. And here's just where each of them are. So the first one is going to be Blaze. Blaze is just an interface to query data on different storage systems. So from Blaze, you import data. And the same way, you load data the same way that you do with the CSV, SQL, Database, MongoDB, JSON, S3, Hive, whatever there is. You just call data and pass this URI. And then you can do all these squares with all those different backends. Select columns, filter, operate, reduce, do a split-by-combine operations like group by, I think, with this. Add new columns, relabel columns, do text matching. One of the features that we've just added to Blaze is the Blaze server is this way of building a uniform interface that allows you to host data in all of these backends through a JSON web API that is the same for all those databases. So you can write your YAML files, specify all the data that you want to serve, where they're located. You can also pass what we're going to see next as data shape. And you spin up the Blaze server with all of them there. And you have an endpoint that allows you to perform all the computations that we've just mentioned before through the API. So this will look like something like this. We have the data available through the API. And then we can query, we can get the fields. And we can get all the different data sets. And inside each of the data sets, we can compute the same queries that I've just mentioned. So as you see, you can just use things like curl. We have an expressive language. Compute something and return it. There's also the option to use the Blaze server to tree and just something like request. But I'm going to explain data shape first. So data shape is just a way to describe structured data. And there's the URL to the docs. And it's basically what's called unit types. And unit types is just a dimension and a d type. And that's what forms a data shape. Then we also can combine those in what's called an order structure d type, which is a record. And then that record is a very extensible language. So even if we can use it and actually Blaze uses it to express tabular data formats, you can actually combine it to have more of like unstructured or semi-structured data and express it nested fields and things like that. So for example, that bar x, y, z is going to be our fields. Bar is going to be the length of our table, which can be known or unknown. And then the types of those fields. So in our previous case, where we had several IDS data sets in different formats, we can get the data shape. It will look something like that. We have a database. Inside the database, we have a table. The table has this data shape with the different fields and types, and the same for all of them. So what's the connection between Blaze query mechanism and data shape? Well, Blaze uses data shape as its type system. So when we call data iris.json, we have access to the data shape, and we can explore it. So now that we can go back to the Blaze server and see, OK, if I know the data shape of my whatever I've put in the Blaze server, which I can get because we just saw that I can just do this shape, and I'll get the data shape of that, I can then express my query and just use request to query that. And the return is going to be a JSON with whatever I ask the computation to do. It's going to return the data, and it's going to return the data shape and the name of the field. So auto. Auto is data migration, which is like a CPU with types for data. So it has a very simple API from auto import auto, and I just have to put my source in target. So if I want to get a JSON from a CSV file, I just do auto, I will CSV, I will JSON, and that's going to create the JSON for me. And that's a pretty simple case, but it can get more complex, like putting things to MongoDB or HDFS, or moving things from one hive CSV to porquet, things like that. So how does auto does that under the hood? It's just a network of different formats and conversions. So if I want to go from x to y, the most efficient way, auto computes that and executes that to you and returns the final container that's your target. So imagine you wanted to put something that you have in S3 to Postgres. You will have to go maybe to get photo, get the file, read, turn it into CSV, put it to Postgres. Auto just simplifies that by just having this URI where you can specify your S3 where it lives, and your Postgres database. Blaze depends on auto because it uses it to handle the URIs. So the same URIs that are valid for Blaze are also the ones valid for auto. Dask, we've mentioned Dask enables parallel computing. So in data science, we have different sizes of data. We have things that are around a gigabyte that can fit in memory, can fit in your laptop. But then we move to the scale of terabytes. And that does not fit in your memory, but can fit on your disk. And you still want to be able to compute that because it does fit in your laptop. Why couldn't you just compute it with it? And then we have things that are in the petabyte scale where it fits in many disks. So with single core computing, we can compute things that are in the gigabyte scale. With parallel computing, if we use shared memory, we can compute things that fit in disks. If you use a distributed cluster, we can compute things that fit in many disks. So NumPy Pandas have solved the single core computing. And Dask is bringing the parallel computation power to the users of NumPy and Pandas. So we have the shared memory, and Dask distributed for the distributed cluster. And inside of shared memory, we have two ways of scheduling. Multi-threading and multiprocessing. So what would Dask look for an end user? So we have NumPy that looks like your image in the left. So we create a NumPy array of ones, and we return some kind of computation. We return it. So Dask does lazy evaluation. So you have to call compute on it to get your return to return it. And you also have to specify the chunk size of the arrays, how you want to partition that. You have more information on the documentation page of what are good numbers in terms of areas of the megabyte size of arrays that you should target to. So in this case, there's those two changes. You need to specify the chunks. You need to call compute to actually perform the computation. And then you have two output results. Your output can fit in memory, so you can just call a NumPy array and keep treating it like that. Or if your result doesn't fit in memory, you can actually store it to disk with an HTF file. If you're more of a pandas user, Dask's data frame looks a lot like a pandas data frame. But it allows to compute things that don't fit in memory. That's like the big change without you having to change much of the flow that you're already used to. So this pandas, we load the Nairys data set. We just do head and we query something. Imagine Dask cannot just load one CSV, but can load multiple CSV that don't fit in your memory. You can still do head. You can still do the queries. But you also have to call compute because it does lazy evaluation. Then we have also another Dask collection that's called DaskBack that allows you to work with semi-structured data like JSON blobs or log files. And imagine we have tweets that we want to load as a DaskBack. And you can just call from file names, like Sterex, JSON, Compress, Map to load the JSON files, and then query it with Take the first two compute user location frequencies and turn that into a data frame because you know the result will fit in your memory. So it feels a lot like what users are already used to and does the parallelism under the hood without you having to worry about that. Imagine that you're now in the scale of petabytes or things that don't fit in your disk. And you actually want to use a cluster of computers. We have Dask distributed for that. So you can see the only difference between using the Dask in the multi-threaded or multiprocessing single known versus the distributed manner is that you have to import this client. Tell the client where is it located. And then when you call compute, you have to pass this get and then the client, the DaskClient.get. And that's going to make the computation in your cluster of computers. The relationship between Dask and Blaze is can Dask can actually be a back end or an engine for Blaze. So you can use Blaze as your query language and have Dask drive those computations. So you make a Dask array. And then you wrap it around the same data object that we mentioned in the Blaze section. Perform the computation, get the result with compute. So right now my talk was mainly focused on users. You can know more about developers, what would be if you don't want to use those tools as a main user, but you actually want to develop with them, there's some good resources that I'm going to mention. On Blaze, there's a good talk on blazing the real world at PyData Dallas at the Cloudcape, which goes more in the internals of how Blaze works under the hood. Also, Odo, so if you actually want to know how to build your own converter, if you have one that's not already built as a back end in the system, how can you create one? That's also explained in Odo, or in both the Blaze Odo talk at SciPy last week by Philip Cloud, and then another one that Ben Zeland gave at PyData Dallas. There's a lot of good talks on Dask. It's a very six-month-old, eight-month-old project that's spin-off of Blaze. And there's also very good resources from James Christ at SciPy and Matt Rockling at PyData Berlin. And those talks go more in depth of the implementation details of those libraries. There's many talks. I have been in many libraries that are in this ecosystem that I haven't mentioned. But I have mentioned, but I have not gone into the details of explaining what they do. And there's already developers of those libraries that have given good talks, and then if you were interested. Dine, Mark, we gave a talk at SciPy also in Austin two weeks ago. Stanley Cyber gave one on Numba, on Accenture 18 Python with a NumbaJIT compiler. OK, that's the Odo. At EuroPython, we have Antoine, Oscar, and Graham, who are the Numba team. And they're going to be here all week. So if you have questions on how to use Numba, they'll be happy to help you. And because Francesco's dad is also here, and he gave a talk yesterday, and he's going to give a tutorial tomorrow. So you're interested in the trends of storing data containers, memory disk. He's going to be giving the tutorial, so you can check it out. So just to summarize, the goals of these talks were have you rethink the term data science instead of being just machine learning models, actually building the connections with those five areas and how we can bring everything together moving forward. Also, think in terms of not just one library, but inside each library, we have data, we have engines, and we have expressions. And encourage you to start using any of those place projects if it's something that you can benefit from. And this Blaze project is possible thanks to a very talented team that are working in all of these projects. Mark, Irwin, and Dined and Data Shape. Ben has done a lot of work in Odo and Blaze. Eric also. And Philip Cloud, who's also a Pandas Core Developer. He's also working in Dask. The Dask team is Matt, Jim, and Blake. And we also have some connections with the B-Colbs team and the Blossom team with Valentin and Francesc. So reach out to any of them if you have interest in any particular library. And I think I have five minutes for questions. Five or 10 minutes for questions. So thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Yes, my question is on the relationship between these projects and the other projects in the scientific Python community like X-ray, Pandas, or do you see this part of them replacing them, or merging them, or complementing them? I don't have a good view on the future. OK, so there's a lot of work on connecting those libraries inside the other ones. We have several already success stories with things like scikit image. So Dask had a pull request in scikit image to speed up some of the computations that they were doing. So there's different layers. There's the user layer that is kind of extending the use case for the limitations that some of the end user facing libraries have, like Pandas and NumPy. So right now, Pandas and NumPy cannot solve some of the problems they're faced because of how they're built. So in that case, I would say, if you are in the size of terabyte, petabyte, then Dask, DataFrame, and DaskRay are going to be an alternative to Pandas and NumPy. On the other side, there's the developer layer of improving computations that already exist in other libraries. So there, the connection is going to be in merging, in making Dask, for example, a dependency on those libraries and improving the performance of them. Does that answer your question? Then there's also, there's a good right in the documentation that, on the Dask documentation, that compares Dask to things in the distributed systems and whether it's an alternative or not. And I would say that they're targeting, I think, different users. So I would say that some of the benefits of having a low overhead to perform distributed computations is kind of the good alternative for things that exist in that world. But still, other people, it's not for, might not solve all the problems. But I would encourage you to read that comparison. So a short question. I, Dask distributed looks a bit like Spark, RDDs or data frames. What is the advantage of using one of another? Why should I choose Dask distributed over Spark? OK, so that question has been asked a lot. Actually, Matt Walken wrote, extended the Dask documentation because we were asked so many times the comparison between Spark and Dask. And of course, Spark is a more mature project. It uses the JVM. And it has a higher overhead of setting up. Dask is just a Python library. You can pip install Dask. You can condensthall Dask. And it has bring some benefits to the core Python scientific and machine learning libraries that can use it. And as an end user, I would say it brings much lower overhead, especially for people in the Python community who don't want to mess up with setting up Spark cluster and dealing with all that performance. But being said that, you can also integrate well the blade, for example, blades can use Spark. So Spark is one of the back ends used by blades. So if you want to perform a performance comparison between Dask and Spark for your specific use case, it's very easy to do that with blades because you have the same, your code is going to look pretty much identical. You're just going to change the string that you pass to your data class. But there's a very extended section in the Dask documentation that goes into all the details of that comparison. Any more questions? So thanks. Thank you.
|
Christine Doig - Scale your data, not your process: Welcome to the Blaze ecosystem NumPy and Pandas have revolutionized data processing and munging in the Python ecosystem. As data and systems grow more complex, moving and querying becomes more difficult. Python already has excellent tools for in-memory datasets, but we inevitably want to scale this processing and take advantage of additional hardware. This is where Blaze comes in handy by providing a uniform interface to a variety of technologies and abstractions for migrating and analyzing data. Supported backends include databases like Postgres or MongoDB, disk storage systems like PyTables, BColz, and HDF5, or distributed systems like Hadoop and Spark. This talk will introduce the Blaze ecosystem, which includes: - Blaze (data querying) - Odo (data migration) - Dask (task scheduler) - DyND (dynamic, multidimensional arrays) - Datashape (data description) Attendees will get the most out of this talk if they are familiar with NumPy and Pandas, have intermediate Python programming skills, and/or experience with large datasets.
|
10.5446/20095 (DOI)
|
Thomas and I'm working in the team where we process that data, where we make that available for the machine learning and where we therefore build these APIs. And that's the main topic of my talk. It's mainly three things you can see here in the title. It's building a mighty purpose platform, meaning we want to use it for different business domains. I will explain that later on. We have bulked data, so not just a few data sets, but lots of data which need to be processed in a fast way. And we do it using SQL Alchemy. These are the three points that you will see throughout the presentation. Introducing a way of building data processing applications that can be used in many business domains. That's the topic. Yeah, SQL Alchemy. We are building on SQL Alchemy. I mean, most of you, I assume, know that it's just a statement from the website which are copied here and have to say what I really like about SQL Alchemy is that it gives us so much flexibility in how to build your application. You really have both ends of the spectrum doing it much database oriented for high performance and you can, you also have the OAM where you can broke them in a much more abstract way. And within our application, we use both of these flavors depending on in what area we are and what we do exactly in the API processing part I will show here. So let's build a multi-domain platform. What do we have to do? What do our customers expect from us? Well, we have to load the bulk data via a CSV file. This is what I will show in my example or in real life. We also use XML files. We get that into the system via HTTP interface, post request. We have to verify that data. We have had several talks already regarding clean data, big data, which is not always clean, which can be quite messy. And so that our machine learning can work on that. We need to have clean data that has well-added to preferences and that's one important part of our application, what we do. Yeah, and we use it for different business domains. So what we currently do in our company is for retail, for tourism and for other areas. But yeah, what I will show in the demo I will explain on the next slide then. So there's still a lot of technical to do what we have to do. We have to create a database schema based on the business domain we are in. We have to parse the CSV. We have to save that parse CSV data to the database. We have to validate that data. Validations, there can be multiple things. We have to check that the required fields are filled. We have to check that the data is correct and that in a date field there is no time for example or no other descriptions like today or tomorrow. And we have to check that the references between the data records are correct. We want to give the feedback. We want to give feedback to the customer about the processing status of his data, whether it was accepted, whether we were able to process it, what is done with it. And it is important for us that we can separate the data that we received from the customer from the clean and validated data that we will use for machine learning. So we want to be always be able to track what was sent to us and what we made from that. Having thought about that, let's have a look at our first customer. Our first customer is a pub and what could a pub want from a machine learning algorithm? It wants to predict how many drinks are sold in the next evenings so that they can plan accordingly how much to buy, how much waiters to have. So to do that, they want to send us the drinks they have available. They want to send us how many drinks were ordered per evening for the last half year and how many visit us they had on each evening so that we can do our learning on that. How could the data model for this look like? It's quite simple. We have on the one hand the drinks, the orders reference the drinks and the visitors. That's just another table that we have for information. I told you that we need to separate the data we got from the customer to the data that we validated. To do that, we have two sets of tables. On the one hand, these are the stage tables. These are the data we got from the customer just as we get it from the customer. So maybe he sends several updates, then we have several lines in them. Maybe there are some duplicates because he sent the same file twice, then we have several lines in there and maybe there are some errors in. Then we also have them all in that stage. What you want to get out of that process is the core and in the core, we also have the drinks, the orders and the visitors but there we have one unique ID for each data record and when we have updates to the data, we will update that data record and not saving it several times. So the machine learning algorithms can use that and can be confident that they will get sensible data. How could such a CSV delivery look like? Well, let's take a simple pub. We have some beer. We have some additional information here. Alcoholic content. We have whiskey. Let's see maybe the pub since Scotland. We serve Scottish whiskey without an E and we have some coke for the people not wanting some beer or whiskey and we have these orders on that day. We sold 10 beer and eight coke on the 11th of July, 15 beer and two whiskey and on 12th of July, we sold 13 beer and one. Yeah, well, we got a new waiter from Ireland. There they're rather whiskey with an E but that would be bad for us as we only know the Scottish whiskey, but things can happen. We get that at a delivery. Now, what do we want to do with that? Ideally, what our code should be able to do. It should find references between objects. So we have that stage table here. These are the orders we get and you can see that's the external code you saw. That's the drinks you saw. That's the count you saw on the orders table and then there's this new column, the drinks reference. This is nothing the customer sent to us. This is the reference to the unique IDs of the drinks we want to find. They are available here in the core. You can see a lot core table. You see here we have a unique ID and we want to write that in there. Yes, one implementation detail I missed at the last slide. Here we don't have a foreign key relationship between this column and this column. You can be anything in so at the moment it's empty. But in the core, we defined a foreign key relationship between that table and that table. So that also the database ensures that really there's sensible data in here. So whenever we want to copy that data to that table, we really need to make sure that we find the correct references first. This we do in two steps. We have the reference finding step which writes them in here and then when they are in, it writes the validated data to the core and copies and then you can see it omits just this information but keeps the reference information with the foreign key to the drinks table. And you also see the last line is omitted. This was key. He could not process. He did not copy it in there. We have to decide in our application whether we throw an exception then whether to write some log file to give information to the customer in some way. But at least it should not come into the core. So our task as developers is how can we write the code that does these steps? How do we do that? We have several possibilities. We have plain SQL works fine and if we want to start playing around with that, that's always a good choice on just playing around with the database to be able that there's really a sensible way of doing that. We can do that in the core. So SQL Alchemy Core model which is closely resembles the SQL Alchemy and where we have here orders stage. This is a SQL Alchemy metadata object which contains the information about the stage table for the orders. We issue an update statement and we say what are the values? The values are that the strings reference column should be filled by a select and you want to select the IDs of the core table of the drinks where the external code of that core table equals the drinks name of the order stage and let me just go back to show it to you again here. We want to make sure really that this ID gets into that column and that for exactly where this name of the drinks matches here the external code of the core. So therefore, okay, this works fine. That's a nice idea and I would say maybe would be the best for implementation. We have slightly in the back of our head that we might get different customers with different models and we are thinking about, well, maybe it would be a good idea to look in the OOM so that we are more flexible there. We have here the tables as objects and we have each row as an object and it's much nicer to implement the stuff here. We can loop over the orders. We can query the table with the correct filters and update the table. It works fine also, but as we do here the single database access that might be not a good idea from a performance point of view when you really have big customers. But these are the things, the tools we have at hand at the moment. So let's assume in our team we use that statement and we are really happy. Everything works fine. We have great data. Customer is happy. Our data scientists are happy. Everything's good. Now, well, as you can see, good or bad, the customer is happy. The pub and she tells the brewery about that. So they are talking when they're getting new delivery and the brewery is quite excited because they say, well, that machine learning stuff, we read about that in the newspapers. We are thinking about our brewery. We have the machines. We have the boilers and the fermenters and we have some sensors in there. We measure some stuff like temperature and pressure. And then must be some way to find out, I mean, brewing and storing beer is quite a long process. We want to know in the beginning what will be the quality of our beer in the end. Couldn't you help us with that? And our data scientists are quite happy with that. Interesting new task. And we just need to get the data into the system. This can't be that complicated. Well, looking at that statement here, it might be. We have to rewrite all that because there are different references between the categories. We now have machines. We now have sensors. We now have measurements. All are named differently. And to make that customer happy, we would have to rewrite that complete statement. So it would work. But when we look into the future and maybe they are more interested, more interesting business domains, we might have really lots of work to do. So what could be the solution? We thought now a team. And we said, well, we could describe these things in a more abstract way. We can say we have one business domain, which is the pub at first. And the pub consists of categories. We have the drinks. We have the orders. And we have the visitors. They consist of the elements. Well, that's the external code. There's a reference to the drinks. They have some types we need for the database. And what is some of these elements are special? We looked at that reference finding task and we have seen they need special processing. And it would be good if you just could have a way to determine that these are special elements that we can inherit here from that element. And we also have, I mean, each element has a name, which we see in the CSV file, and it has a name on the database, which is, well, most cases the name in capital letters. But for this order, strings reference, you had seen there's an additional field. There's this name reference. You remember in our reference finding step, we wanted to fill that column in the stage table. So we add this here in our subclass and we say that this belongs to the category. So what does that help us if we can do that? I mean, we also thought that somehow resembled the SQL alchemy model also. I mean, a SQL alchemy model also has some tables. It has some elements. It has some types. It has foreign keys. But the SQL alchemy model is for a database description. It's not so much for an algorithmic processing of that stuff. So therefore, I will discuss that at the end. We said it really makes sense to have this in a more abstract way. How now does that look like? I mean, SQL alchemy also has parsers. Here in ORM, for example, we have that generic ORM parser. And we have here in the SQL alchemy model, we have our business domain. And we have the business domain that is also described in the application. So in both here, we need to have that business domain. What we wanted to do, we wanted to factor out some knowledge of that. So that this application does not need to know the business domain, that we really can set this here in that domain model. And that we can have specific task renderers. So for reference finding, we have here one renderer which uses that information to generate SQL statements. And we can have one for other tasks also. How does that look like? I will give you here a code sample for that pub. So we have here the domain, the category, the elements from our lingo package. We have that pub, which is a domain. And that pub consists of three categories, the drinks, the visitors and the orders. That's quite easy to write down. I mean, there's nothing more than you need. And it's easily understandable. The interesting stuff is here that reference, which you see links directly to the drinks. Having that, how does that task specific renderer look like? So this is Python code which at first checks for each category, what are the references in there? And we can just check that with an instance. And we loop over the references. We have here just one, but there might be other categories where we have multiple ones. By getting the stage and core tables from the SQL Alchemy metadata. As you can see here, we can find that by the name of the stage table. We can find the reference core table. And we issue an update statement. So this update statement here, you see update the stage with what values. These values are constructed dynamically. So because we cannot give here the keyword arguments dynamically, we construct the dict at first. This is here the update dict with the keyword values. And you see here, that's the column which should be updated. And this is a SQL Alchemy core statement on how to update that with what name. So we print that, I will show in the demo. And then we can execute that. So now let's switch to that. We can see that this really works. We are prepared in the demo, it's a simple SQL light database, and I have prepared some script. No, it's quite... Let me do it that way. Okay, great. So at the moment, we do not have to SQL light database. You see here, these tables do not exist at all. And what we want to do, we want to create them. We do that by calling our Python script, create database, and these tables are there. You can see here, by the way, we have a configuration file where we see what's the database and what is our business domain. So now we need to get data in there. So we call that Python script. We do the CSV import. We do that first for the trinks category. And you see here, that's the SQL statement created from that, and it's here in the category. So now we do that for the orders also. And you see it's in here. Now we want to do the reference finding, and for sure before we find the references, the trinks need to get into the core, otherwise it doesn't make sense. So what do we do? We are calling the core load script, and we're doing that first for the trinks. You see here, that is the generated SQL statement, and this is the core. And you see that the trinks ref column is empty. So now, for the most interesting, you can see these SQL statements are issued, and here you see that this is filled, and also that the orders core is filled. Well, it doesn't fit completely on the screen, but you can see that it's all in there. So that works fine. That's great. But that was a step. We were already five slides ago. Thank you. Now let's go back. And let's say we have now another domain. We have a brewery, and that brewery, we say it has machines, it has sensors, and it has measurements. And that brewery here, you can see the sensors referenced to machines, and the measurements reference the sensors. I would like to show that in the demo also, but unfortunately the time is not sufficient for that, but it's really as simple as we have seen it here. You just need to change the config JSON, you create the database, and the new tables will be there, and you can import the differences. We find this will be in there also. So what does that mean? When we say, what does this domain model help us? Well, it is optimized for high throughput. As you see these SQL statements that are issued, they can be processed directly on the database. So we put all things in the database into the stage, and now we have some processing of our domain knowledge, and we'll generate the SQL statements, and the things will be processed on the database. That's a really good fit for analytical models. Then when I thought about the demo here for the talk, I found it might not fit that well for transactional models where you have more complex end-to-end relations, but for analytical models this is really great and helps a lot. When I compare that to a SQL Alchemy model, which is also some kind of meta model, then we see that the SQL Alchemy is focused on a database description. The domain model in contrast can contain more information. In our team, we had also the task that we have time-dependent stuff. So some crinks are only available at several days, or maybe they were available last week, but they're not available this week, and we need to check these cross-time dependencies also. This can be done in the domain model also. We can note that there. And we can generate the SQL Alchemy model out of that domain model, so in that case, we have both. So an additional bonus is we can use that domain model for much further stuff. You can see here, we can generate the SQL Alchemy model. We can generate a SQL for our tasks. We have to CSV load our configuration, but also what we do, we generate documentation out of that, so how to fill that CSV table, and we can generate demo data and much more stuff. But that's just to show you some ideas. So that you are able also to have some questions. I will close here, and that's what I wanted to show you. Are there any questions? Hi. Is a lingo library open source and available? No. This is something we developed internally. I mean, what I did here for that talk, I prepared a small demo application, and I also thought about providing that. But I've seen it takes much stuff around to make that example somehow sensible. And yeah, before making that open source, I would have to check also internally in our company. And before going into that direction, I just want to see see whether there's interest at all in that. So if you have some questions to that, I want to get some further updates, which I can just talk after the talk. Any more questions? No? All right. Thank you, Christian. Okay. Thank you.
|
Christian Trebing - Building a multi-purpose platform for bulk data using sqlalchemy At Blue Yonder, we've built a platform that can accept and process bulk amounts of data for multiple business domains (e.g. handling retail store location and sales data) using SQLAlchemy as a database abstraction layer. We wanted to use as much of SQLAlchemy as possible, but we quickly found that the ORM (Object Relational Mapper) is not suitable for handling large amounts of data at once. At the same time, we did not want each team of developers working on individual business domains to have to handcraft their own SQL statements. To solve this problem, we built an application configuration that closely resembles an SQLAlchemy model, but also contains application-specific logic settings. In this talk I will demonstrate: - an application architecture for multiple business domains - the structure of the domain configuration utilized in the generation of the SQLAlchemy model, SQLAlchemy core statements, and other application functionality - how the domain configuration is used throughout the application (consuming and parsing incoming data, storing it in a database and ensuring data quality)
|
10.5446/20094 (DOI)
|
Welcome. Thank you. Thanks. So, hola, Euro-Python. Thanks for coming to my talk this morning. My name is Kari Ann, and I'm going to be talking about education. And this is perhaps not the most exciting keynote that you might have come to this week, but I'm hoping to give you some kind of insight into this current movement around computer science and education. So my name is Kari Ann, and currently I work for the Raspberry Pi Foundation. But if someone was to come up to me and say, if you could describe yourself in one word, what would it be? That word would be educator. Okay? I see myself as an educator. You know, I was a teacher for a long time. Whether I'm teaching in a classroom or teaching through a book or I'm educating through online resources on raspberrypi.org, you know, that is what I do, and that's what I'm about. For anyone who doesn't know about Raspberry Pi, I've had some really odd conversations this week. I've had quite a few people when I'm talking about Raspberry Pi, go, oh, you're a charity? Yes. Raspberry Pi Foundation, we're a charitable organization. When you buy a Raspberry Pi, 100% of the profits from that go into the charitable foundation. So we are able to meet our charitable mission, which is to advance the education of adults and children in the fields of, you know, computers, computer science, and related subjects. This part, related subjects, is really important to us. So things like science, the arts, there's lots of subjects that we're really passionate about. I guess you could describe them as kind of digital making. You know, we're really keen in that kind of field as well. So because I work for the Raspberry Pi Foundation, I spend a lot of time thinking about how children learn with computer science. And I'm not alone in that. Today with us is Ben Luttle from the Raspberry Pi Foundation. James Robinson is here as well today. And one of our prolific volunteers, Alex Bradbury, is here from Cambridge University too. So today, at any point, if you want to talk to us about computer science education, or in particular about Raspberry Pi, do approach one of us. They're usually wearing Raspberry Pi t-shirts. I'm not today. But they're pretty easy to spot. So this is kind of where my journey started in the classroom. I was teaching a subject called ICT. You can tell I'm teaching. Here are some children, longingly and lovingly, you know, looking up to me. Yeah, that's standard. That doesn't always happen. This is clearly a post-photograph. And one of the only photographs I can use, right, because you can't see the children's faces. But around this time, I was teaching a subject which was called ICT. And in about 2010, 2011, Google's Eric Schmidt made a speech in the UK in which he said that computer science education in the UK was really, really bad. And actually what we were doing was not teaching children how to be creative with technology, but rather consumers. And Al Press in the UK is not a huge fan of teachers. So they kind of mistook his words. And then there was a kind of a big push in the media about how all ICT teachers are terrible, how we had no skills. We certainly couldn't program. And so we were doing a terrible job educating the future, which was really depressing for me at the time. But I continued on. And I started to think about ways in which I could take hold of this new wave of thinking around education into my own class. So I heard about this thing called Raspberry Pi that was coming out. And that was around February 2012. And like millions of other people, I tried to get one on the day they were launched from the website, which crashed quite a few times. And eventually in about May, my Raspberry Pi arrived. And I plugged it into my TV and I got it started. And then I kind of thought, well, it's a computer. It's a Linux computer. I've used one before. This is nothing new. Like, what am I supposed to do this? How is this going to revolutionize my teaching? I've been told this is for children. How can I do something with it? So I found out that there was some kind of user groups getting together. And they were called Raspberry Jams. So these events where lots of people would get together and talk about what it is they do with their Raspberry Pi's. So I thought, well, I'll go along to that and see if there are any projects I can take back to my classroom. So I went to one in London. And this was in about June 2012. And it was at Mozilla Space in London. I don't know if you've ever been there, but it's quite a cool space. And there were about 50 men. And there were three women, of which I was one of them. And the three women were all teachers. And so after a while of watching quite a few talks about how people would use their Raspberry Pi's to do things like turn it into a games platform like a Super Nintendo. Someone else had put it in a big track, like an 80s big track rover. So after a few of these presentations, and anyone who's been to Mozilla will know, they give you free drinks. So I started to get quite confident. Someone said, why don't you get up and tell people why you're here? Maybe the community can help you. So I got up with the other female teachers who were there. And I said, look, I'm a teacher. I'm really excited about bringing Raspberry Pi into my classroom, but I don't really know what to do. And I've been watching your projects. And I don't mean to be kind of harsh, but they're quite geeky and retro. And I'm not really sure my students would understand that. And then I said something really stupid, which was, and also I don't think they're going to inspire any girls. At which point, someone heckled me and said, well, why don't you get the Raspberry Pi to go shopping for you, or organize sleepovers, or something like that. Which for me was a really kind of awful moment. Anyone who knows me will know I'm quite an introverted person anyway. Even standing up here and giving this presentation is a little bit scary and not typical for me. And so I kind of walked off a stage at that point with my head down. And then something really amazing happened. About 10 people who were in the audience came up to me and said, please ignore that person who's just said that. Like, we don't agree with anything he's just said to you. How can we help you? And that was really the start of my journey with Raspberry Pi. Because off the back of that terrible experience, I was invited to PyCon UK. And that was because that terrible experience was streamed online for the world to see. So not only did I have to walk off stage like this, to in a group of 50 people, but also it kind of went a little bit viral. Because everyone was like, oh, hey, look at this man heckling this woman at a Raspberry Jam. So something good and positive came out of that. I was invited to the Python conference in the UK. And they were having a teacher's track for the first day. Great. So I went along as a delegate, not really sure what to expect. And by the end of day two, I was running the education summit. Kind of just took over. Sorry, guys. And I gave a lightning talk about what it was I wanted. And really it kind of snowballed. At the time someone was there, one of the keynotes was from the University of Cambridge, which is where Raspberry Pi is kind of born out of. One of the trustees heard about me. And so they sent me someone from Cambridge. And we started a program called Sonic Pi, which I'll talk about later on. And now I'm on the board of directors for the Python Software Foundation. Like, whoa, how did that happen? Right? I went from teaching in a classroom. And now I'm on the board of the PSF. If I can do it, anyone can do it. So to give you some background about where we are in England, we have this new computer science curriculum. It's called computing. It's a computing curriculum. And this is just to kind of give you the basis of what it's about. As you can see up here, from the age of five, children need to learn how to code and program. So they need to understand algorithms, sequencing, selection, and repetition. From the age of 11, they need to use at least two programming languages. And one of those needs to be text-based. And this curriculum, it's in existence now. It started September 2014. So we've had one full year of it. And I think it's really exciting. But it only applies to England. And this is really important that I say this now, because I'll forget later on, because I'll always refer to it as the UK curriculum. It's not. This only applies to England. It doesn't apply to Scotland, Wales, Northern Ireland. It just applies to us, which is sometimes a bit awkward. And government data that's been collected said that 55% of secondary school teachers lack a qualification that enables them to be able to teach this. And a recent survey by Atez, which is the Times education supplement, found that 60% of England's teachers were not confident delivering the new computing curriculum. So that's kind of where we are. And the government in their wisdom brought in this new curriculum. And then they only invested 3.5 million pounds in upskilling of teachers. And that works out just to kind of break that down. That's about 175 pounds per school. And what that means is that a teacher cannot be released from their teaching duties, even for a day, in that per school. So it's not really helpful that they've brought in this new curriculum without any support for upskilling of teachers. This is a problem. So where else in the world are people, where else are there curriculums? We're not alone in Europe, or at EuroPython, so we should talk about Europe. Estonia have also moved towards computing curriculum. And Ben Nuttall, the Raspberry Pi Foundation, who's working with teachers in Estonia, and we've had some on our teacher training course at Raspberry Pi to try and help them with upskilling as well. Because we found out that they're in a very similar situation that they have this new curriculum and not 100% sure what to do with it. Australia very recently have now got a new curriculum, a new computer science curriculum. If you check out Pycon Australia last year, Dr. James Curran gave a really fantastic talk about that process in Australia and New Zealand as well. And Israel have had computer science as part of their curriculum for quite some time now. So this is it. We're like the pioneers of this around the world. And there is this real movement at the moment to bring digital making and computer science into schools. We've recently been to the States and had quite a lot of conversations with teachers about this. We're really super excited. Keep talking about coding. And I use this all the time when I say coding. There's this coding curriculum. We need to teach code. It's kind of becoming a bit of a fad term. I'm really keen that that doesn't happen. Last year I went to visit Denmark and I gave a talk there. And it seems like the Scandinavian countries are very close to perhaps adding this to their curriculum too. So I think we're going to see a kind of snowball effect. And you might be thinking, well, actually, why should we bother teaching computing at all? Like why do we need to teach computer science to children? You know, why can't they just learn how I learned with some books and some online tutorials? Well, the first reason is that children just are creative and imaginative. And they're not afraid of failure. And something that's really terrible that we do as educators is we train them out of this. When students used to come to me, because I was a kind of secondary high school teacher, so I would get them at around the age of 11. And by that point, they had been trained out of kind of clicking buttons and kind of just having and going and seeing what happens. It's really disappointing. Because children just are. They will press all the buttons. They will have a go at things. And that's what they will tinker with things. And that's what's really exciting, and that's how we learn. Another reason is to do with social mobility. I think we have a real problem with developers. I think that people who access computer science education and who have jobs in computer science generally have come from a fairly affluent background. I think computer science has the potential to move people socially, move them from really low paid jobs into really good paid jobs. Another reason is I think computer science is empowering. Just being able to write a program and make it do something, even if it's like print hello world is super empowering. The first time I wrote a computer program, I was like, wow, I've made something happen with this computer. And it made me feel good about myself. And it continues to do that. And I feel more confident as a person to be able to do that. This one, obviously, I'm hugely passionate about, which is about diversity in technology. We've had tons of investigations into this area. And lots of reports suggest that in the UK, 16% of IT specialists are women, and that's in the UK. It's not much better in the United States. It's something around 20%. It's usually around that percentage. And what that means is the people who are creating technology do not represent the people who are using technology. That's not right. And again, that's to do with social mobility as well. Are we representing the people from everywhere, from every type of background, if they're not the people creating the technology? And the last one, and I think is the most important, is this idea of where we're heading towards just globally. CGP Gray, he makes these really cool videos on YouTube. He made one called Humans Need Not Apply, in which he talks about how in the future, we're not that far away, most things will become automated. We think about when you go to a supermarket now, generally, you might want to use those automated machines where you self-check out. There are baristas where you just make your own cups of coffee just by pressing some buttons. And of course, self-driving cars could actually put a whole bunch of people out of a job. Just think about if self-driving cars tomorrow exist, laws are passed and actually they're used every day, bus drivers, delivery drivers, lorry drivers, all these people are going to be out of a job. And I think he quotes something like, it's just a small shift economically, that if there's like 40% unemployment, that actually it could be devastating for the world's economy. And he's quite negative about this. And I think more positively, I think if we train children today for a workforce that will be around fixing those machines or programming those machines or making those machines better, then we're winning. Because this is where the jobs are going to be in this sector. But having said all of this, I feel like this argument is kind of over now. I've been making this argument for quite a few years and I feel like actually I don't need to make this argument anymore. The focus is moving away from why we should be teaching it to how we should be teaching it and that's why I'm here today. So in the UK, before we had the curriculum, we knew it was coming. A lot of teachers were talking about what programming language should we use to be able to teach this. So scratch is an easy win. Everyone's going to use scratch in the younger years. But as you saw, 11-year-olds need to be able to use two programming languages. One needs to be text. So what programming are we going to use? So a lot of people went, well, okay, let's have a look at Python. Python seems like a good option. So why Python? Well, the first reason is that it's used all over the world. It's used in real-world scenarios. It's powerful enough for real development. You're all here, you're all developers at a Python conference. I assume you use Python in your jobs. So not only is it a good tool for training children, but you can also say, hey, this is used all the time, all over the world. It's used by NASA. It's used by CERN. So this is a good programming language. It has a really simple syntax. So we use this example all the time. If I wanted to do hello world in C or in JavaScript, I'm going to have to write quite a few lines of code, like roughly six, and it's going to have curly braces and parentheses and so on. That's going to get confusing. When I was at school in the 80s, we had a BBC micro. I have vivid memories of writing print statements in basic. That's what's really nice about Python is you can write print hello world in one line. So it's a great syntax that we're already winning. This is the most important one, I think, for educators and for me, is it has this really strong, powerful, helpful community. Community is just so important. I've never felt so welcomed by a community. The Raspberry Pi Foundation has a really great community, but I would say the Python community probably just about tops it. If I go back to my experience at PyCon UK, the reason why I took over that teacher's track was because I was made to feel welcome. When I was there, I met some developers, and after the conference, that didn't end, that kind of collaboration and friendship didn't end. I would continue to try and improve my Python skills and write code and sometimes it wouldn't work. I was able to stick it in Pacepin, send it to one of these guys I had met, and they would mark it for me. Great. Developers are marking my work as a teacher. That was really helpful for me to improve and get better. I think not just in terms of national communities or international communities, also local communities. So I've been invited several times to the London Python dojo. One time I went along and I stood up and said, why don't we just have a look at the curriculum? Would any of you guys help me try and fix some of these things with Python? Maybe we can come up with some examples and I can share them with other teachers. I was amazed that they all got really excited about it. They all formed groups. They wrote amazing programs and they shared them back with me. They were so excited about it, they went on to run a whole education special edition of that dojo. Technology is what makes Python special. I think we should really celebrate that. So I guess we're at a point now where we have lots of education summits as part of conferences. So PyCon UK has been having this teacher track for quite a few years now. When I first went, there was roughly around about eight to ten teachers there. My colleague James Robinson is going to give a talk today about his experience as being a teacher going to PyCon UK. That's at 11.45. So come to this. Have your break. Go to that. And I think it's really important that we try and welcome teachers into conferences like this. We've seen the Python, PyCon UK education track go from that beginning stages of eight to ten teachers to this year. They have tickets for 40 teachers and they're already sold out, almost already sold out. And that's coming in September. So if you build it, they will start to come along. So if you organize conferences, I really recommend that you have an education summit. PyCon this year in Montreal had a kind of education track, which was some talks here in Bilbell. We're very excited to be helping run the education summit. And straight after this, I'm off to Australia to help with the education track at PyCon Australia. So I think really what I want to say is if you try and help teachers by just starting these things. It may be today we don't actually meet any teachers here at the education summit. That would be sad. But if teachers see what we're doing, they're more likely to come next year, the year after and so on. So what I want to talk about is, okay, so Python seems to be like a great language to use to teach. But currently there are some real strong barriers. And I think we as a community can fix those barriers. So I'm going to kind of explain them to you and perhaps show you some solutions and perhaps we can come up with some ideas. So the first one is transitioning from a visual programming language to a text-based programming language. So what I was doing in my classroom was we were teaching with Scratch with the young years. And then we would move on to Python. And actually a lot of the children struggled. They really, really found it hard to move from Scratch to Python for a whole bunch of reasons. And this is something we're continually trying to overcome. Cambridge University at the moment, they have some Raspberry Pi interns and they're working on a project called Pi Land. I really recommend you go and have a look at it. And that's like a kind of game that you play on Raspberry Pi, but you have to use Python to be able to solve a lot of the problems. The next one is everyone's favorite conversation, which is Python 2 versus Python 3. I had a lot of confused teachers out there that they'll go, oh, my code's not working, my program's not working, what's wrong with it? And I look at it and I go, oh, it's Python 3 and you're trying to run it in Python 2. It's not going to work. And this is a real problem. But really, it shouldn't be a problem, right? In education, there's no reason why we shouldn't just be teaching Python 3. Really, everybody uses 3 in education. And at the Raspberry Pi Foundation, it's a decision we took early on to ensure that all our resources were written in Python 3. And I would really encourage you, if you are working with children or in schools or helping teachers, Python 3. Python 3 is the answer. Most major libraries are Python 3 compatible now. There's a few that are not, which I'll come back to later. But Python 3 is the answer. I'm just going to keep repeating that. Python 3, please do Python 3. So the next one is a bit complicated to explain. And this is probably where you're going to start booing me. So please bear with me. So I feel like there's this kind of function naming problem that we have at the moment in Python. And this is where sometimes people who write libraries probably aren't aware. Quite often I find that people write libraries for themselves, right, to fix a problem that they've found. But what might happen that will be surprising is they may be picked up by schools or by children and they start using them and they become really important. One of the examples for this is the RPI GPIO library. So a guy in the UK called Ben Crosston, he runs a brewery. And he decides to use a Raspberry Pi to be able to regulate the temperature of his brewery. And so he wrote a Python library to be able to help him do it, it's called RPI.GPIO. But the surprising thing is now it's used in all schools who have got Raspberry Pi as with children because it's the Python library to be able to control the GPIO pins on Raspberry Pi and do some physical computing, which is one of the most exciting things in schools at the moment. And it can be problematic, not specifically that library, but I find a lot of libraries that are picked up by educators that are written by people who are just using them for themselves. Sometimes there are some inconsistencies, inconsistencies in the naming of functions. And that can be a bit of a problem. One of the most common problems we find with beginner learners, especially with Raspberry Pi, is that quite often they'll create their program and they'll save that file as the same name as the library that they're using. So for example, we have a library called Pi Camera, which allows you to use the Raspberry Pi camera with Python. And that library is called Pi Camera. And so everyone, of course, when they do their first-ever Pi Camera program, save it as pi camera.py. And then they wonder why it doesn't work. And this is a really common problem we find, whether it's with an add-on board like Piperella, same thing with the Piperella library. It may seem fairly obvious to you and you may think, oh, God, what idiots, right? But you've got to understand quite often these are children or educators or beginner learners who are just completely new to this, right? I did the exact same thing at the PiCon conference in England. I was running a workshop in front of a whole bunch of teachers and a whole bunch of developers. And I accidentally named my file as I was using the Piperella library. I named it Piperella.py. And then myself and five developers stood there looking and going, why isn't it working? Why? Why is it? Five developers, not one of them figured it out until a little bit later because it's not an obvious problem, right? And then there's a lot of inconsistencies with kind of whether you've used camel case or snake case, right? If you're going to use those, like, pick one and stick with it because it can get a bit confusing whether you need to use capital letters or underscores. And be really aware of that. I mean, I guess we should be using snake case, right? But think about children who are using this. We're talking about eight, nine-year-olds who want to move away from scratch and they want to be learning Python, but they struggle with, you know, just motor skills sometimes and keyboard skills. So you've got to think about if you're going to use underscores, stick with underscores. If you're going to use capital letters, stick with capital letters because, you know, they need to kind of get used to using the keyboard in that way. And try and make the kind of the functions almost guessable names. So if I'm writing something in Python, I can almost guess if I'm trying to figure something out, well, actually, I want to set up, set this block, perhaps in Minecraft, that I might want to build a bigger kind of set of blocks. Oh, that'll be set blocks, right? That's obvious. I can guess it. And quite often we find libraries, like, it's not guessable at all and it's inconsistent. So please try and make them consistent as possible. Yeah. So think more time about naming because children might use it. So I'm just going to use this example from Minecraft Pi. So anyone who doesn't know about Minecraft Pi, it's an API that you can use to be able to program things to happen in Minecraft. Yay. So obviously educators love this because it's a real hook for children because they love Minecraft and they can see something instant happening. It kind of blows their mind that instead of spending hours and hours and hours, like, building a house, like, mining what they need to build a house, that you can just do it with a few lines of code. So here's, this is the example that we have. A really basic program you would think to get children starting. First of all, they've got to connect to the Minecraft world. Then they need to, you know, use some variables to set some coordinates. And then they need to use this line, mc.player.setposition. As you can see, there's a capital P and that's kind of their program. And I find that they fall down on the very first line, right? Just the capital in Minecraft. Like, they forget to do it. And children are very impatient. They're used to things happening straight away with, you know, tablet devices and so on. Everything's really instant. And sometimes we can lose children very quickly just with things like this. And so I'm just going to show you. So that's the same example. A friend of mine called Sam Aaron, who I'm going to talk about a little bit more later, he's been able to get this work with his system SonicPire and he spends a lot of time thinking about because he works, he quite often does a lot of outreach about how he could improve this with his own system. So please excuse that this is Ruby. But I want you to look at the names that he's used here. So you can see the difference here already, right? So mc.teleport is just a much nicer kind of sounding word than setPos. Right? SetPos, it could mean anything. It could mean set position, but we're old enough and wise enough to understand that. But children, teleport, it's like, oh, it's going to move my player. So I need to set the x, y, and z coordinates. Got it. It's a really nice example. The other one is to do with block types. So with the Python API, you have to use variables. This is really long. Here block.glass.id. Sometimes you have to use the block numbers. 108 if you're wondering is melon. Yep. TNT is 46. You can test me on those later. And then you need to say mt.setBlock, give the coordinates of where you want to set it, and then, of course, call your variable. I want glass to be there. That's, again, quite long-winded. There's a lot of capital letters in there as well that children need to deal with. And so how we thought about that with Sonic Pi is mc.setBlock. All the block names are included. So you just have to say glass and then set the coordinates. Again, apologies that this is Ruby, but you kind of get the idea of what I'm trying to say. Basically, my point is, if you are creating libraries, be aware that they may be picked up by children. Please be consistent in how you name your functions. This is an example that I'm really excited to share with you all. So that's one of the PyCon UK education tracks. I believe it was one last year. There was a lot of discussion between developers and educators. And that's one of the most exciting things about having teachers at conferences like this. So Daniel Pope, who some of you may know, is very good with PyGame. Any of you might know PyGame, the library. It's a really great library, again, to use with children because it's really visual and you can do a whole bunch of stuff. But PyGame, I found, was really difficult to use in the classroom. And I think Daniel got talking to some other teachers who said the same thing. And he wrote a very short program for him, which was very short. And a teacher said, I can't teach that in the kind of hour lesson that I have. It's just not going to work. You're going to have to explain event handlers to children. That's not going to happen. I just want them to be able to start thinking about the logic to be able to build a game. And so he went away and he came up with PyGame Zero. I believe he gave a lightning talk about this on Monday. If you catch him around and you haven't seen this, you really need to go and have a look at this. So the idea here is that it just makes PyGame much easier to use. As a teacher, I'd be able to break things into bite-sized lessons. All I need to do is change a couple of lines to be able to make meaningful progress. And that's really exciting. And it does make some of the decisions for you. And we shouldn't be afraid to do that with children, to abstract away a little bit just to get them excited. Sam Aranoise uses this example when he's talking. That with children, sometimes you say to him, OK, so let's do some coding. What do you want to do? I say, well, I really want to build a game. OK, let's make a game. Great. I can do that. Fantastic. What do you want your game to be? Well, I really want it to be set on an alien planet. And I want there to be kind of these crazy monster aliens. I want to have a big gun. I'll shoot those aliens. And I want to run around. And they'll probably be a dungeon somewhere. And because kids are imaginative and creative, they're going to go on about that for quite some time. And you're like, OK. So in Pygame, I can make you a kind of black screen, we'll say that's like space. I can probably make a circle. We'll call that the planet. That'll be gray. Right? I've already lost them. Because in their minds, when they think about computer gaming, they're thinking about Xbox. They're thinking about PlayStation. They're not thinking about 2D graphics. They're just not. So if they're going to do something with Pygame that is 2D, it needs to be something more like scratch where they're able to make quite a lot of progress very quickly using their imagination. And with Pygame Zero, I think they're going to be able to do that. So please do go have a look at that and speak to Daniel about it. So another barrier is around installing extra libraries. I see a lot of learners fall down really early on when they don't have the libraries that they need. And it's a real added hassle. If I'm going to be teaching a class for the kids, it's very easy to install a library, right? You've got 30 computers, 50 computers. It's not so easy. Raspberry Pi is even harder. And so this is something we need to think about. And quite a lot of teachers and a lot of schools, sometimes their devices aren't online because they have to protect the children within the school. And more often than not, they're behind a kind of firewall that stops them to be able to get these things in a proxy that stops them from being able to install libraries. And so I got talking to Daniel on Monday night. So this is an addition to my talk where he started to talk about this idea of having an education bundle that you would be able to download. So kind of on top of the standard library. And this could include commonly used dependencies. And this could include Pi game zero. So things like NumPy, the Python Imaging Library, those kinds of things could be in this education bundle. And I wouldn't really mind how that was delivered. If that was built on PIP, that would be fine. All I'm trying to get across here is that we should be thinking about this for educators. And so the last barrier, and this is the one I want to spend some time talking about, is Python IDE. Finding programming interfaces for children is really difficult. Just out of interest, how many of you use IDLE for your IDE of choice? Two people? Three maybe? This is IDLE with what children have first access to. It comes included when you download Python. It's very disappointed to hear Greedo's talk on Tuesday when he was asked a question, what is your favorite text editor? His answer was an IDLE. This is what children have. And so there are some really good examples of online text editors. Some really great examples out there. And I've listed some of them. A lot of you will be aware of Python anywhere, I think they're doing wonderful stuff with online IDEs. It comes bundled with just so many libraries, which is great. But there are problems around this idea of online IDEs. And I think one of the first ones is to do with the age of children using them. So one of the reasons why I couldn't use the new version of Scratch in my school was because you had to sign up for an online account. And there are issues around privacy, and in England, children under 13, you just can't do that with children under 13. So if I want 8, 9, 10-year-olds, 11-year-olds using this online text editor, I can't because they're not allowed to sign up for it. So that's a problem. In my school, I had this huge problem where the infrastructure, the online infrastructure, the internet, super slow. It wasn't built and designed for the amount of use it was getting. I couldn't use Google Docs with my children. Scratch would crash halfway through, so we couldn't reuse that. And that was a problem. So if I wanted to use any of these, I just wouldn't be able to reliably use them because my internet was so slow. And again, behind firewall and property settings probably wouldn't allow me to access some of these. They'd most likely be blocked because they'd be seen as an adult website. Or that, hey, you're going to be breaking things, so we don't want children breaking things, so we're going to filter that. Not just at school level, but higher up at kind of borough district level. It would be blocked and it would just be a nightmare to unblock them. So, and I think there's also a compatibility issue with libraries again. So these are great. And I think we're on the right track with a lot of these. Grot learning in particular is fantastic. Really designed for education for children. I recommend you have a look at that one. But they're not perfect and they're not really the solution for everybody. And if we really want to move children socially and we want to include everybody in computer science, we need to think about children who do not have access to the internet. There are children in this world who do not have access to the internet. Even in England, we have children who don't have access to the internet outside of school. So we need to think about how we can include them. So an obvious example of a kind of offline educational Python ID would be PyCharm. They have an education edition. It's free and open. Fantastic. That's exactly what we want. And they say on their website, PyCharm educational edition is not merely a learning system. It's a real development tool. Okay. We're getting back to this. It's a real development tool. Actually, I need to spend some time learning how to use PyCharm because it's not really obvious. Even the education edition is not really obvious for me to gain to. So I think that's a problem. I think this is great for, oh, my God, I've only got 10 minutes. So this is probably not great for, this is probably really good for children who are kind of 16 plus. I think this is great. If I was teaching GCSEI level, which is kind of between 16 and 20, I'd be using this because it would be great. But for my 8, 9, 10, 11-year-olds, which is the kind of key demographic I want to be hitting here, this is not going to work. And there's too many opportunities for failure with it. There's too many buttons. There's just too many things I'd have to set up first. I just want to get them on something simple and get them programming. And also, open source equals awesome source. So it's great that it's open. So I mean, you need more things like this that are open, but simpler. So going back to my idle problem, so this is what learners have. Most people who are new to Python, this is the first thing they come across, which is idle. And it has some really good, positive things about it. First of all, it's free. It has syntax highlighting. It does do some auto indentation, which is good because indentation is important in Python. It's cross-platform. We're able to include it in RaspberryM because it's so small and lightweight. And it's simple. And that's really important here. It's very simple to get started with. However, tons of problems that we find with it all the time when we're using it with learners. Anyone who's used it with teaching children will know there are so many problems. Last one, it being in two separate windows, that needs to change, right? Because that is just a nightmare game with their motor skills and so on. If they've got Minecraft running as well, it's three windows. They've got to navigate between. It's not really working too great. The error reporting is atrocious, right? It's really bad. It doesn't really tell you what the problems are in a way that children can understand. So what are the solutions? Well, something I want to show you and I keep talking about is Sonic Pi. So this guy from the University of Cambridge came into my school with a Raspberry Pi. I put it in front of me and said, I've got it to make music. Do you think we can teach computer science with this? And I said, yes, let's have a go. It sounds nuts enough that it will probably work. So on a whiteboard, I drew this really simple interface where we had a coding panel and we had a panel that was an output. It was showing you what was happening. And then there was an error panel. And since then, that was back at the end of 2012. It's now evolved into this, which you can see here that includes online tutorials, in-built tutorials. And it's now both a tool for education. So it's a tool developed for education. But it also is used by live coders to make music in a professional kind of context, which is very exciting. So I'm going to do the thing that I was told not to do, which is a kind of live demo. So just to prove that this is a tool for musicians, this is the kind of music that you can code with it. Apologies. This, again, is in Ruby. When I told people I was going to show Sonic Fire, a Python conference, they went, really? Sorry. But yes, what I want to show you is the interface. Just imagine for a minute, this is Python. Some really good stuff up here. There's a button right here in front of me. It says run. So run your code in idle. You have to click on run, run module. Right? And then you have to teach children control where five. There's a button that makes it much easier. You can stop it as well. This is really handy. Like a keyboard interrupt. This is super helpful. I'm going to blow your mind. Can you people at the back see that code? Can you see all of it? The answer is no. Oh, my gosh. I can make it smaller and I can make it bigger. That sounds really like why are you so excited about that? When you're teaching in a classroom, to be able to do that really quickly, especially when you have 30 children who are all working on their code and they're stuck, to be able to just press a button and make the text smaller and make it bigger. Great. We need that in idle. Numbers down the slide, line numbers. Really simple, really easy. Having windows together, this is obviously using QT. These things exist that we're able to put windows together. Let's put the interpreter window with the text window. It does have an in-built tutorial which I can get rid of. Other things that are quite exciting, let me see. I've written some code here. Obviously it doesn't need to be indented to work. We know that's what makes it a great starting point for children. However, again, another really cool button is this one here that says align and it automatically aligns the code. You might be thinking, well, I think children are going to be able to do that. Children should learn how to indent their code so it works with Python. That's actually really important. Yes, it is, but I think actually we can do this. We can have a function where you're able to turn this on and off, but I think it's a really good starting place. Idle does auto-indent, but really badly. Children will not write linear. Quite often you'll give children parts of code and you want them to change it up and their code just gets all over the place. The auto-indentation is not sophisticated enough to deal with it. This has these really cool pink lines as well so I can see where it's indented, when we have nested kind of loops and so on. I think that would be really helpful. That's an example of a really good interface, but obviously that's been designed for education. We spend a lot of time in schools. We see where they fall down. Sam spends an awful lot of time testing his software, all demographics. That's why he's able to make these changes. Why can't we have a version for Python that's like that? I think it's possible. You're thinking now, how can you help educators? You're really pumped up by everything I've just said and you want to help fix all my problems. Thanks. I've got a whole bunch of ways in which you can do that. The first one is to meet educators. If you run conferences and you have education tracks, eventually teachers will come to them, go and meet them, go and talk to them, listen to their problems. Add education tracks to all of your conferences. Run special education sessions at your local user groups. Teachers will come and you'll be able to talk to them and help them. Mentor a teacher. If you do meet a teacher, help them in their journey. That's what happened to me. I was helped by developers in the UK. There's no reason why this can't happen globally. And then create and contribute some really awesome libraries, which is kind of what you do already. Please make them consistent. This is what I'm really excited to talk about. Thank you very much to those of you who voted for me. I'm now on the board of directors for the PSF. Something I want to launch today is this idea of a Python education work group. I've set up a mailing list for this. This in no way replaces education special interest group mailing list. That's really a mailing list for people who want to talk about education and pedagogy and so on. The idea of the working group is that we're actually going to make some of these things happen. We're going to make a new Python text editor. We're going to make Idol better. So we want to make, you know, it's for people who want to make practical contributions to that. We're not just going to sit around talking about it. Obviously, we will talk about it, but we're actually going to make things happen. And that's really important. And I need your help to be able to do that because I can't do these things. So I need you. And so we're going to have, you know, specific goals that we hope to achieve. So before I can get this group recognized by the board, I'm really hoping some of you will join it. We can determine the governance and so on and kind of talk about what it is we want to do. And then I can take it to the board who already know I'm doing this so we can get it officially recognized. So I've got homework for you all. Once a teacher, always a teacher. Number one, join the mailing list, right? That's the first one. You're all taking pictures of that, which is really good. So I'm hoping you're going to join. Number two, so at the education track at PyCon in Montreal, a guy called Al Swagger gave a talk called Idle Reimagined. Unfortunately, I wasn't at that conference and they didn't kind of video those talks on the education track, which was a real shame because it meant I couldn't watch this talk. But I've heard about it. And I've been on to his GitHub where he has a wiki about it. I don't believe in duplicating work. I think this should be our starting point if we want to change Idle. So your second point is to go and have a look at that wiki and get involved. And then thirdly, I think everyone should read this book. It's called Python in Education by Nicholas Tolleve, who is here. It's a really small book. I believe it's free. It's no Riley book anyway, so you all got a voucher, so you should just go and get this even if it's not free and you should read it because it really does help explain about Python in Education. And so this is all due for next year, so make sure you write this in your diaries and your planners. I will be checking. And so lastly, I just want to talk about the future then. There is a real danger at the moment. Coding becomes an education fad, right? Code.org, Hour of Code, all those things are great, but I've met a lot of teachers like, oh, I've done coding. We did Hour of Code last year. It's like, well, you're not teaching programming, you're not teaching computational thinking. You've done an Hour of Code. Ensure this doesn't become a fad thing. There's a real danger that visual programming will just be the tool that Scratch will continue to be used all the way up to 16-year-olds. We need to make sure that doesn't happen and we introduced text-based programming earlier. And really, it should be Python. And it is a real danger at the moment if we don't fix some of these barriers, actually JavaScript is going to win. I had a really awkward conversation in the States with a guy who's developed a tool for teaching programming to children using JavaScript. I said, and he also has a Blockly interface. I was like, oh, why don't you have a Python interface as well? And he was just like, well, no, because nobody uses Python and children just want to make apps. No, they don't just want to make apps. Because they want to put them online and share them with their parents. I don't think that's true, actually. I think that they could do some meaningful work with Python. And let's just think about, consider if we are successful with what we're trying to do at the moment, which is to educate children and to get them thinking of computationally and be able to solve problems. We could really change society for the better. And I truly, truly believe that. Think about all the children. You know, reading, writing, arithmetic, those are the three hours. It could be the four hours if we, you know, Raspberry Pi. It's a bit of a Raspberry Pi plug. But really, computer science should be as important as reading or writing. So given all that, in 20 years' time, when the Raspberry Pi generation are older, right, we've taught in Python, we've taught them, you know, how to think in this way, you know, they're going to go into all sorts of jobs. They're not all going to be developers, right? They're going to go into medicine. They're going to go into government. They're going to go into the military. They're going to go into, you know, research. They're going to go into so many different fields. And if they are able to solve problems, I think it could be a really interesting place we live in. If we teach them about open source as well, it may, oh, my mind is just blown. Those children could make this planet an amazing place. And so I think we should be excited by the future. And just very one last thing, just out of the blue, I had an email from one of my ex-students yesterday. And he was telling me kind of where he was after his A-levels. And it's a really bad time for kids at the moment who have just left school, who haven't had the benefit of this computing education, who have like ICT kind of qualifications. He's really struggling to find a job at the moment, like really struggling. He really wants to work in the sector. He can't find a job because A, he has no experience. He's not been to university. And he wasn't taught computer science at school. So he's trying to do some kind of online tutorials, teach himself. But really he needs a foot in the door. So if anyone is in London, right, and this kid is really bright, really smart, he's just brilliant. He'd pick things up really quickly. If anyone can help him get a foot in the door, even in a free internship, I'd be really grateful. So please come and speak to me about that. But really think about how we can help young people. It's not a great time right now. We're hoping to fix that with these curriculum changes and this wave of computer science. But right now there's this kind of middle group who can't get apprenticeships, who can't get jobs. He told me he went for a job interview as a delivery driver and he didn't even get that. So this is me. Please connect with me. I'd like to thank the organizers of this fabulous conference for having me. I'm having the most amazing time in this beautiful venue, this beautiful city in this beautiful country. Thank you so much. Thank you. Have fun and silence. Hello. Great. So thanks for doing this. We've also been teaching kids with Minecraft Pi. And it was pretty cool. But the problem was we had to download the tutorial, edit the HTML source code to translate it. So if you can do anything about these tutorials, please make them translatable. You don't need to do it yourself. But please provide these options. Thanks. We're working on that one. Just so you know, we are working on them. But thank you. Any other question? Hi. I noticed you mentioned something that I've heard a lot of times from educators, which is that laws or education boards or parent groups make lots of things with computers really hard by, for example, the rule that you can't have an account under 13 or incredibly strict firewalls that are usually also broken in school systems. I once heard a story about a school where all the computers were locked in a safe because they were so afraid that they were stolen or maybe used after hours. Yeah. And what can we do about this? So the only way is to get educators change their mindset. So having conferences like this where we, in my educators along, will start opening their eyes because it takes a teacher who is willing to break the rules to change that. If I had gone to my headmaster, if I had gone to my senior management team and said, well, this guy from Cambridge is coming and we're going to make music, is that cool? They probably would have said no. So I always have this kind of argument that it's always better to ask for forgiveness than to ask for permission. And actually what we need are those superstar teachers. We need to find them. We need to help them break the rules. And at the moment, I feel like that's the only way we're going to change that. I think governments need to invest more in, you know, the equipment, not just the equipment that's in schools, but the infrastructure. I think the infrastructure is more important. Right? If we want to use these online IDEs, we need to change that. And I'm not sure what the answer is to that because governments at the moment just strapped for cash. They can't improve their systems. So that's a problem. But I think the only way is finding those educators. Sorry about my joke about the IDE, by the way. Hi. Thanks for your talk. It was wonderful. So you said there was a number of problems that you have with IDLE specifically and that you're looking for alternatives. Now, the canonical answer to make a Python interactive and accessible these days is Juniper Notebook. So what's your experience of that? And do you see any ways to make that more accessible for children specifically? Yeah. So, I mean, we could be talking about Atom right now. There's a whole bunch of things that are already out there. It's like, well, how can we just take something that exists to make it slightly better? But I think actually we need something that is completely simple, that is designed specifically for education. Right? I don't think we should be taking systems and things that exist out there already, which are used, you know, for other purposes and just trying to rehash them. I think we need something that's main purpose is delivering Python in education. So, you know, there's a whole, like I said, there's a whole bunch of tools we could use, but that's not really their sole purpose. And I think if you start from that point, but actually you want to help an eight-year-old get started with Python, that there is a simple tool that they're able to use. Any other questions? Okay. I'm sorry, we only have one mic, so I'm just going to take a little while to reach all of you. We're keeping you fit. I'm not going to run to. With respect to IDEs for Python in education, we all know that to grasp well a full-fledged IDE is a difficult task for beginners. Then what has the educational edition of PyCharm that you may, that you said that is applicable to use in education instead or in place of IDE, for example, is that it is the most simple IDE that we have for Python? Yeah. I mean, it's simple to you. When I've installed it and tried to get started with it, it wasn't simple for me. I found it. There was a whole bunch of windows I had to get through. You have to create a project and open the project. That's a whole bunch of steps. Actually, if I was doing a class full of children, okay, this is the first step. You need to click here and do this. This is the next step. That's a whole bunch of time wasted when actually when I'm trying to teach the children is computational thinking. The language is second to that. I'm not specifically teaching them Python. We're using Python because it's such a great language. The goal is not to teach them Python. It's to teach them to think computationally. I'm losing a whole bunch of time in my lesson. I've only got 50 minutes in my lesson to setting it up and getting them started with. I also have to download it and install it, which in a school system is not as simple. I have to get the network administrator on side. They have to create a package and they have to apply that to everything. Obviously, I come from Raspberry Pi. I would like something that was lightweight enough to run a Raspberry Pi. We've had a go with Python Education Edition. It's not run perfectly. I think there's just too many opportunities for failure with that for very young children who will click on everything and press everything. As a teacher, I need to find time to learn how to use that application. I'm already trying to learn new skills to teach this curriculum. That's where the problem is. Last question. Hi. Thanks for the talk. It's terrific. It's just amazing things. Thank you. I have a question about that curriculum. We mentioned it works in England, Australia, New Zealand, Israel. What about other countries? What is the problem? If there was some experience, for example, countries from so-called Iron Wall or from Ukraine or Africa or any other countries, if there are some communities or foundations who are trying to reach you and ask how to do it, what were the main problems, how does it struggle with things? I really do think at the moment this is a global movement just because it's not written in the curriculum, for example, in every country. Don't think that it's not happening. We know countries in Africa like Guyana who are rolling stuff out, just not officially with an official curriculum. What's really important to know about our curriculum is it applies to state schools, publicly funded schools. For example, we have a whole bunch of schools called academies which are funded by private sector. They don't have to use this curriculum if they don't want to. We still don't think that we've solved the problem in England. I do think in other countries this is already starting. I think if schools or code clubs are a really great way to start. If you helped run a code club in a particular area, that's where teachers start seeing the effect it's having on their children. They're more likely to teach it and then it will snowball from there. It takes governments to change curriculums, unfortunately, but I think teachers can push that a little bit. Last question, then we will thank Keryan. Thanks. That's a great idea to get school children to train and learn with Python. There are also a lot of other programs that have interfaces to Python. I learned Blender, for example, that you can model things. You have physics inside. You can combine a lot of lessons in school. Do you use Blender also? Yes. We have teachers in England who are using Blender. For exactly that, to teach a whole bunch of stuff. I really think computer science and computational thinking is cross-curricular. The stuff we showed you with Minecraft, you can teach math with it, there's coordinates and so on. You can teach physics as well. Blender is a really great tool, again, free and open. That's exactly what we want. Yes, it's great. Thank you, Keryan. A big round of applause, please, because she really deserves it. You just go talk to her after the talk. There's a coffee break. Enjoy it. Thank you again.
|
Carrie Anne Philbin - Keynote: Designed for Education: A Python Solution The problem of introducing children to programming and computer science has seen growing attention in the past few years. Initiatives like Raspberry Pi, Code Club, code.org, (and many more) have been created to help solve this problem. With the introduction of a national computing curriculum in the UK, teachers have been searching for a text based programming language to help teach computational thinking as a follow on from visual languages like Scratch. The educational community has been served well by Python, benefiting from its straight-forward syntax, large selection of libraries, and supportive community. Education-focused summits are now a major part of most major Python Conferences. Assistance in terms of documentation and training is invaluable, but perhaps there are technical means of improving the experience of those using Python in education. Clearly the needs of teachers and their students are different to those of the seasoned programmer. Children are unlikely to come to their teachers with frustrations about the Global Interpreter Lock! But issues such as usability of IDEs or comprehensibility of error messages are of utmost importance. In this keynote, Carrie Anne will discuss existing barriers to Python becoming the premier language of choice for teaching computer science, and how learning Python could be helped immensely through tooling and further support from the Python developer community.
|
10.5446/20093 (DOI)
|
So, welcome everybody. Thanks for being here and nothing with us talk. We really appreciate that and we are going to give this talk. It's called Max Real Time Messaging and Activity Stream Engine. I'm Victor. I'm a senior Python developer and IT architect at UPCNIT. I'm a Plon Foundation member and I'm Plon Core developer since 2010. And I'm also an author or a book that is called Plon 3 Interests, which was published in 2010 by Backbook. Hello, and I'm Carlos. I'm a Python and a JavaScript developer and I'm working with Python for the last years and just because this project also occasionally with Erlang. And also I'm the kind of guy everyone asks for regular expressions. And with Victor, we've been doing websites with Plon in the UPCNIT that is a private company that is owned by the Universitat Polytechnica de la Lúnea, Barcelona Tech. And we have more than 400 sites in Plon running in the university and the last four years we also make some projects with Pyramid. Okay, so again, so lots we wanted to do a demo and I don't think we are able to do that because the Wi-Fi is really crappy, but however we tried to do that. Let's see. Let's see what happens here. We wanted to do that because of course an image is worth a Y, 1,000 words and if it's not working right now maybe we're going to, yeah, to all the show. Come on. Okay, maybe we're giving up on that and we will, oh, almost there. Come on. So we basically wanted to show you the stream which is the basic UI implementation. Why is doing that? Hello. Okay. So we have this stream that is the central part of here which is the, as I said, is the UI implementation of Max in which we can post things like this. So hello, everybody. And then the posts get registered and right now Carlos has also sent me a real-time message which I can see there, like in a real-time notification and now if I go to that view then I see Carlos has already sent me that message and I can reply to him, right, like this. So this is basically what we've done. This is the two main parts, main features of Max and we just wanted to show you. So let's back to the presentation. So a little bit of history about Max. We've initially designed Max being the fact that we wanted to implement some kind of social internet which is a concept that is nowadays is very trendy for the university itself. And we initially designed that key feature that the internet should have, which is the streaming engine and after that we also added the messaging feature as well. And today Max is used by more than 30,000 students and 8,000 university staff and it's integrated mainly on the online campus of the university and also in the institutional collaboration tools used by the university staff. Basically what is Max, so you already seen it, it's an engine. You've seen the UI but basically it's a RESTful API with almost half a hundred endpoints and has these two main features. It's a multi-source user and application, activity stream because not only it's capable of receiving activity from users but also from applications and we can impersonate applications but talking in the mouth of the users too also and we can do that kind of things. And then we have the asynchronous messaging and conversations feature, we already seen it and it's our GPL license. So how we came to that concept, to the concept that they use in Max. We had a lot of forums in the university, in the online campus and in the internet also and we wanted to modernize the user experience for that forum. So we thought that maybe we can reuse some of the concepts of the forums in special from topics because if you are a user you are interested in a special topic but you always have to go to the website and check that topic and we wanted that the information came from a modern point of view to the user. So we basically introduced the concept of a context. So in Max we basically is the way that you, we, hello. So in Max we have context which are mapped directly to topics, to the old topics which are the main features that they have, they are related to a unique URI so we can map it to virtually any page even if it's not related to Max directly but in this context we can aggregate posts that have social features like commenting or likes or the favorite feature. And we added also the concept of subscription so I can subscribe myself or someone can subscribe me to a context and this person can give me, can grant me some permissions that it allows me to interact me with the context in different ways. So we have the subscription, the context that is mapped to the topic so I can, as a user, I can subscribe myself to all the context that I wanted. So what are the other features of the context? As I said, they are identified by unique URIs and we can assign permissions per context which are basically now we have these eight permissions which allows us to model different kinds of context based on that permissions and that gives us almost 6,000 kinds of context which is a pretty useful, useless fact however. And we can also override the permissions of the user in a granular way because we are able to grant or revoke specific permissions per user for allowing users to do more than the context does. So what are the real life examples for that? For example we have now these two scenarios where we have the community sites and the online campus and the community sites is kind of, as I said, the social internet at the university where we have a lot of communities which are directly mapped to one context which can have any of these like institutional events or institutional news or in the online campus because we have all the subjects that are mapped to each one to a context. In the community site we decided to have these three community types just to show you what we can do and we have the open communities where everyone can join and can live at will. We have the closed communities where the owners should invite me in order to join but I can live also whatever I want and the institutional community type where the site admin decides who subscribes but no one can leave. This is the example of the institutional news or the institutional events where I want that everyone be subscribed but nobody leaves. As a summary of features we have the activity stream, you already seen it, it stores activity from user applications and we have that social feature too like comment slides, favorites and we also have support for attaching images and files. The conversations feature is almost the same because in kind of way are sharing some of the basic features of the activity stream. Now later Carlos will show you and we basically allows us to have one to one to best conversations and group conversations as well and it supports also images and files. Then we have the JavaScript UI widget which you already seen and it's the reference implementation of the UI attacking the API and you can use it as example for implementing your own UI if you wish. So we also have the notification engine for the real time features that it's platform specific push notifications for Apple and Google and also these allow us to have kind of internal notifications too which we are exploring right now that we want to be able to send notifications for example to the desktop, to the users not related with the real time messaging but also for other information that the user could have. We also have an external source aggregation we are currently monitoring Twitter, hashtags in Twitter which we can push it inside the stream also. Also a key feature that we are able to deploy max on Premiere or whatever we want and this address some security concerns that we could have if we are using a more popular tool like let's say WhatsApp or iMessage or things like that and if you are worried about privacy and ownership we provide a way to have like a corporate WhatsApp or corporate iMessage so it's also a good feature. And this is the summary of them. Okay I will try to explain you a little bit some of the components that we will to make this possible. We have kind of three different kind of components. On the left you can see the components that are the ones that have a user interface, the ones that the user really uses and sees which are the mobile apps and the integrations that we have done with other systems like clone and model and the JavaScript widget that it can literally be put anywhere. In the center in red there are the main software that we developed to do this. We will go through some details about it and at the right some of the well-known software products that we used as the backend title of Max. To start what we call OCDIS is an implementation of the server implementation. This is built on top of Pyramid. This is a very simple implementation. We only support what is called the resource owner creations flow because as we only are using UALF from trusted systems we don't have the need of the more complex flows of OALF for identifying and giving permission over resources. So as the systems where we are using this are trusted by the final user, it's that system that provides the credentials to obtain the tokens and do the validations. We're using MongoDB both in OCDIS and in the API for storage. One of the reasons that we developed our own system is because we know that we will have some clients that have specific user directory systems or single sign-on systems and we wanted some flexibility to adapt to those clients so we made OCDIS pluggable so which client can get its own plugin to adapt to its features. But we have a base LDAP implementation and we offer also an LDAP service to our clients if they not already have a directory service. Max is the API part. It's also built on top of Pyramid and why we use Pyramid? We were using Pyramid for the last four years. We know pretty well what features it has and it fitted well with what we wanted to achieve with this API. I will not go in deep about all the endpoints that we developed on Max, only a few of the implementation details that we used. For example, on the routing of all the endpoints that exist in Max, we have a centralized place in the code when we define all the URLs of all the endpoints and this is used by Pyramid to actually define the routing and we use it to generate automatic documentation based on the string of each implementation of each endpoint. We also use these traverse. It's a combined way of routing that using URL dispatch for reaching the URL and traversal to retrieve the object that is the object referenced by this endpoint. So once we enter in the definition in the actual code of that endpoint, we already have all the initialization, authorization and authentication done and so the real code of each endpoint is very simple event and very short. We also use the feature of Pyramid that is called twins. And this is actually a decorator but is managed by Pyramid so you can define a lot of twins and you can define in which order the twins are executed and basically you can perform tasks before and after a request enters the system. So you can modify the request that has arrived and modify the response that will be shown to the user. As an example, one of the twins that we have implemented is a check that is done on a special value on a header that we are using because we thought that maybe in some point in time we would reach a point that we will have no option that makes some breaking change in the app. So all the mobile applications that we have distributed across all the client devices will be broken. So the mobile devices and the IPI server agreed in this value that it's just an entire number that we are incrementing over time and each request is checked on this number. So if we have to make a breaking change, we just increment that number and then mobile apps know that a change has been made and that probably a new version of the application is ready and it tells the user to upgrade the app. So we have a control about this scenario but we try not to make that kind of changes. It's just in case we have no other option. And the last thing about Macs is how we implemented the exception handling. We wanted a unified way to show the errors to both the final user that is using the API and a way to catch possible handled exceptions and bugs that occur in the system. So what we've done is that every known exception that we have in our code is rised as a custom exception. And then this is handled by Pyramid catching the exception and rendering a customized JSON message for each kind of error. And for non-handled exceptions, we record the trace back of the error. We record the information that was in the request and we built a little user interface to be able to inspect that debug information and quickly act on possible bugs without having to deal on contacting the user, trying to reproduce the error, et cetera. On the real-time messaging part, we have used RabbitMQ. This is really the third attempt that we made while we were exploring different possibilities. First, we tried with Pyramid and Socket IO. Then we tried with Node.js and Socket IO, too. And finally, we discovered one of the plugins that come with RabbitMQ that is a storm with Socket's plugin. I don't know if anyone had heard about it. What we can achieve with this and also with unique routing features that has RabbitMQ is that we can directly map a client that connects it to the system to the internal Qianic change system of RabbitMQ. And also, we have developed an Erlang plugin for RabbitMQ to connect with our authentication. This is a diagram of what we've implemented to wrote the messages. One feature that we are very proud is that the structure of this design holds the security of what users can read and write from. So each of the exchanges that you see here are connected by some bindings that are maintained by the API calls. So any time a subscription is made, the correct bindings for each user are populated. And so a message can only go through the system if that bindings exist. So the security is implicit in the design of the bindings. We also developed a kind of protocol message that is what messages in Socket RabbitMQ contain. And that message is designed to be packed and unpacked to use the minimum size possible. So we have a kind of specification to unpack these messages based on static values that we know. So we can go from this human-reavable format to this nerd-reavable format. So it's not unreadable, but it's easily debuggable if we have to plug in and see some of these messages. And finally, this is the MaxVani component that is the Kiwi consumer. It's a multiprocess Kiwi consumer, so we can start many, many consumers for each of the types we have developed. And it's in charge of taking the messages, doing the appropriate task for its message to fill in the APIs and so on. And that MaxVani component uses what we call MaxClient, is a Python wrapper for the REST API that wraps all the functionality from the API. And this is what is used from MaxVani to call the API. I will skip this. We also developed a special version of the MaxClient that is the Whiskey iMatchClient, which is a way to, from the client, execute all the code base from Max using a special fake Whiskey server so we can make a request in the same way that is done by the real API, but without using the real servers. And at last, the Twitter external aggregation is already explained by Victor. So I want to show you now the current integration, which you already seen in the demo, which is the site of communities, which is this social internet. We have the Moodle integration, which is what we call the Ulearn campus, and also the iOS and Android apps. So this is, you already saw it, the social internet with the key, the JavaScript, which is in the middle. This is the Moodle site, or one of them, because we have a lot of themes of also, and with the integration of the stream here, which allows the students and the teachers to have a very tight communication, even with the stream and with the conversation part. These are the iOS app, and this is the Android one. So they basically have the same functionalities. And potential integrations, almost virtually whatever you are thinking of, we are also thinking in put this stream also near PS, or in, I don't know, any HTTP web app that you are thinking of because of the JavaScript. You can put the JavaScript with very little effort and make it work with Max anywhere. And we are already exploring that, as I said, with maybe putting in our ERP or something like that. We have a huge list of to-dos, because there is so much to do. We wanted to add social integrations, like follow, like Twitter-ish feature or the share, like the Facebook one or the G plus one. We definitely have to finish and polish the documentation, maybe explore the utilization of each little service and make it a real microservice, exploring also some other kinds of cache, like Redis. We also wanted to port it to Python 3, because we are using 2.7 right now, and maybe use also I think IO and improve our API by adding some JSON LD or hyperlink technologies there, and of course, adding some kind of end-to-end encryption, which is a very tough one, but nevertheless, it's on the list. So we are here also, because we also want to explore if we can build a community around Max. We are very, let's say, happy and proud of what we've done, and we know that there is a lot of things of improvement, but we also definitely want to explore what if we can do community around it, and please contact us if you are interested or you find it interesting and pull requests are welcome. This is the resources where you can go and take a look at it. The first one is the documentation one, and the other ones are the repos, the GitHub repos of Max. So thank you. Thank you very much. Do we have any questions from the audience? So you said you tried several technologies besides RabbitMQ when you were choosing it. I'm interested in knowing which ones, and especially if you took a look at MQP 1.0 as a protocol to use. Can you repeat the final question? So I'm interested in knowing what technologies you guys looked at instead of RabbitMQ, and I'm especially interested in knowing if you guys took a look at MQP 1.0 as a protocol to use rather than RabbitMQ with 0.9. Okay. I think I already said it. We used Pyramid with Soquetayo. The first two approaches were without thinking in a key we broker inside the system, but we realized that we couldn't achieve a real time with a little load with a system like RabbitMQ in behind. So RabbitMQ is using EAM MQ, but we are using some of the extensions that RabbitMQ has done over the standard protocol. So we are a little tight to these features. Any questions? No more. Let's thank the speakers again then.
|
Carles Bruguera - MAX: Realtime messaging and activity stream engine What if I told you that we’ve built an open source “WhatsApp”-like RESTful API on the top of Pyramid? We've developed MAX: a real-time messaging service and activity stream that has become the key feature for a social intranet at the BarcelonaTech University We will show how we designed and built MAX with performance in mind using state of the art Python technologies like Gevent, WSGI, and multi-threaded queue processing. We will also show you how we've managed to design a simple architecture guaranteeing both high scalability and performance, achieving connecting ratios over 30.000 students, teachers, and university staff. The API is secured using oAuth 2.0 resource owner password credentials flow powered by our own oAuth server implementation (Osiris) also Pyramid-based. We are using MongoDB as general storage backend and RabbitMQ over websockets to support realtime and messaging.
|
10.5446/20092 (DOI)
|
Hi everyone, welcome. Thank you for coming. Now before I get started, who here was actually involved with AdoptPytest Month? Alright, so most of my little AdoptPytest Month crews over here. And who here is a contributor to an open source project? One or more? Alright, so nearly everyone. Great. So it probably won't surprise you to know that I love Pytest. I've been a Pytest user for a few years. I've contributed a few patches and I generally find it a total joy to use. And it was founded by Holger Krekel, who gave this morning's keynote. It's been around since 2009. It's a very mature, stable project, full of features, and has around 170 plugins to match all kinds of frameworks. And so in January at the FOSDEM conference in Brussels, we had a little Pytest meetup and it was lots of fun. We got to talk about all kinds of things we were excited about doing or fixing in Pytest. And I thought to myself, why doesn't everyone use Pytest? Like there's some libraries, for example, requests. Everyone uses requests and everyone says, just use requests. But that's not really the case with Pytest. It has many, many fans and many people strongly recommend it. But it's not really in the situation that people say, just use Pytest, end of story. I mean, we even heard, even Guido, you know, sticks with the unit test. So why doesn't everyone use Pytest? Well, there's a few reasons that I can think of. And some of them are out of our control. We will never be able to satisfy them. But there are some that the Pytest community, the Pytest contributors, could do something about to encourage more people to use Pytest. So at FOSDEM, I had this idea that we could pair up open source maintainers who were interested in getting started with Pytest with experienced Pytest users or developers. And they could work together for a limited period, so there's not too much of a commitment, to help them really get started and get off on the right foot with Pytest. And that would help overcome knowing where to start, knowing if you're doing it right, and really taking advantage of all the best features of Pytest. And with any luck, at the end of a month, we'd have a host of new Pytest users, which would be great. I decided to organize it for April, and so I wrote up a page on the Pytest website describing what we were planning, and I made some surveys to sign up the Pytest helpers and also the open source maintainers. And first, I tried to sign up the Pytest helpers because I thought, well, if we are inundated with open source projects, we want to make sure that we have helpers for everybody. And so I asked the Pytest helpers what was their experience. They didn't have to be a Pytest developer, so they did not have to have contributed code to Pytest, but they had to feel that they were experienced a new, most important features of Pytest. And I asked them about their experience with other areas of Python programming, like frameworks like Django or Pyramid or scientific programming, so that I could try to pair people with appropriate interests and experience. And we had 26 Pytest helpers sign up, which was a lot more than I was actually expecting. You saw at the Pytest meetup, there was six of us at Fosdham, and so I was really surprised. We had 26 people volunteer to help other people learn Pytest, and only nine of those people were people who had actually contributed code to Pytest before. So I'd found 15 new people who were keen to advocate for Pytest. And then next, I tried to sign up the open source projects. I wanted to reach open source projects that were fairly well established, and so initially we had a requirement that they should have more than one core contributor. And after a couple of weeks, we did not have many signups, and we thought, okay, maybe we can relax this requirement. And so projects with just one contributor, that's fine as well. And the main way that projects actually found out about the Pytest months and took part in it is by being approached by Pytest users. So we asked the Pytest community to look at other projects that they were part of or aware of, and maybe ask them on their GitHub issues or on their mailing list if they would be interested in taking part in Adopt Pytest Month. And that was by far the most effective way to find people who were interested to take part. And in the end, we had eight projects that kind of officially took part in Adopt Pytest Month, and there was another four that kind of informally took part in that they said, yeah, sure, if you send me a pull request, I'll accept it, but they were not really committed to kind of working closely with the Pytest helper. And so the projects we had were Bidict, which is a library which provides a bidirectional dictionary. Yesit is a command line utility to extract information from media filenames. Calathea is a source code management system similar to GitLab, which supports both Git and Macorial. Qt Browser is a keyboard driven browser based on PyQt5 that you can control in a similar way to VIM. Trump is a framework built on pandas that aims to centralise the management of data feeds. Arquestra is an extension to Django that provides semantic web publishing and is used by a bunch of organisations including the recent DjangoCon in the UK. Nefrotari is a REST API framework sitting on top of Pyramid and Elasticsearch. And Coursera DL is a command line utility for downloading Coursera.org videos. And Trump and Nefrotari were both open sourced by different companies specifically in order to take advantage of AdoptPytest Months. So there were companies that were thinking about open sourcing their project or planning to eventually do that, but AdoptPytest Months was really the trigger for them to actually do it. And informally we had Jason Pickle, Cookiecutter, Django, Ginger2 and Pelican. And the last three actually accepted pull requests before April even began. So they kind of, as I said, informally took part. They did benefit and Pytest did benefit from contacting them, but they did not really work through the month. And so given that every project would be in a different place with regards to their testing, I left it up to the Pytest helper and the maintainers themselves to decide what activities would be appropriate, how much would they try to get done during the month. But I just suggested to work in incremental steps that built on each other rather than trying to change everything in one big bang. And to do lots and lots of code reviews, like very intensive code reviews. And to spend a lot of time just talking about the Pytest features and kind of wondering out loud, is this possible? Can you do this? And having an opportunity to learn through discussion. And so after the month was over, I sent a survey out to all the participants to ask them what they did and how did it go. So the Pytest helpers did things, of course, like updating tests, changing unit tests, sort of equal statements to plain assert statements, changing the layout of code and tests and docs to, well, I'm not sure if that was for Pytest or just kind of incident or cleanup, removing boilerplate code that they didn't need anymore, fixing minor bugs that they actually discovered through this process, converting tests, tests to use parameterize, making custom fixtures, custom markers, and introducing the use of various Pytest plugins. And so self-reported, I asked them how much time did you spend on adopt Pytest month? Initially, I advised maybe two to four hours a week was the time commitment, but I'm not really knowing if that was a lot of time or not very much time. And really varied. I'm not really sure how accurate these numbers are because estimating, remembering back how much time you spent can be quite difficult. But five of the Pytesters said they would definitely take part in it again. Five said they might take part, and one person said they probably would not. And I asked them how do you feel about the Pytest community now after the month? And five said they were more positive and keen to be involved in the community, and five said they will have a similar level of commitment, which might already be quite high in fairness. And they also reported learning about different features of Pytest that they hadn't been exposed to before, from various plugins to how to run doc tests, how to support PyTest in various continuous integration servers, and more about PyTest's support for unit tests and knows test suites. And then I asked the maintainers to tell me what they did. And of course they reported discussing the plans over email or chat. One project used a Gitter chat room, which is built on top of GitHub. Others used email or just the issue tracker in GitHub itself. They wrote new tests and put pull requests for the PyTest helpers to give them feedback. And one enterprising maintainer actually updated a PyTest plugin to work with Qt5. And one project, CourseraDL, made a PyPy release for the first time. So CourseraDL is actually a very popular project, but the number of users is not proportional to the number of contributors, which is not uncommon. And the maintainers, in one case 80 hours, I thought it was really going above and beyond. And that time varied a lot. So five maintainers responded. Others did not respond to my survey. And four of them said they would recommend taking part in adopt PyTest month to other open source projects. One said maybe. And they reported how they felt about PyTest now that they had a good grasp of the basics, but they were also aware that there was really a lot more that they could learn. And they felt very positive about PyTest, actually. And they all indicated that they're going to keep using PyTest in their project. And two of them told me that they have actually introduced PyTest in other projects that they contribute to, which I think is really cool. So even though we directly only worked with eight projects, because many open source contributors work in multiple projects, we have a multiplier effect. And in general, the people who responded to me told me that they really enjoyed it and they were really grateful for the opportunity. And so maybe some who didn't reply didn't have such a good time, which is not a surprise that they didn't fill out the survey. But just based on my observations of looking at the activity in the various projects, of course, they were not all runaway successes. So this is my observations about some of the areas where people struggled. And they're not really surprising. They're the kind of problems that we face in open source in general all the time. So you get excited about something, you think you're going to devote a lot of time to it, but real life intervenes and you end up not being able to do anything. So some PyTest helpers didn't really show up. And I mean, I knew nothing about them except they have completed a survey for me. And some of them showed up, but we're not really able to make much progress in their project. And some of the maintainers were extremely busy and were not really able to work with their helpers either. So I wanted to think about what are some things that we've learned from this that if we do it again, we could take into account to try to have a better experience for everyone. Because even though I've told everyone, you know, it's all volunteers and we hope that it works out, that people were very understanding, but at the same time it's kind of disappointing if you think you're going to work with somebody but then they don't actually show up or you don't get done what you wanted to get done. So how can we have realistic expectations about what we're doing? So the first thing I learned is that you don't really know how many people are lurking in your community. So for those of us who are experienced in open source, we kind of know that the way to get into it is to just march right in, sit right down, start working on something, submit a pull request, and that's fine, that's actually the way it works. But when you are outside the circle, it's not very clear how do you get in, how do you start, and it's kind of intimidating. So giving a foothold to people to say, this is the thing that you could do, and this is like a very clear expectation about how you take part in this activity was a really helpful thing to do. And one of the Pytest helpers said this, they said that it was the first time they contributed code to someone else's open source project. So I think that's super cool. So they were able to bring someone inside the circle. And to me, this was a really important revelation of doing Adopt Pytest Month, that there's a lot of value in making a space for users to be advocates of your software. So by advocates, I mean people who are excited about bringing other people on board to your project. They don't have to be contributing code. In companies, they often have positions called evangelists, developer evangelists. They'll come and give talks at conferences and tell you how great this library is that you should totally use. And in open source projects that are not corporate backed or company backed, that's usually left to the developers. But there's real value in recognizing the work of advocates in our open source projects. It's not only good for the community, but it's good for the users of your software. And so the third thing is that code review is maybe not just important, but actually amazing. So I called my talk the realities of open source testing, and this is actually what I'm talking about. Because the majority of open source projects, probably the median number of contributors is one. So there's a real long tail distribution of open source projects. And at the head, we've got really popular projects like Python, Django, even PyTest is fairly far up the head of the graph. But there's just this long tail of hundreds of projects that have one contributor. And even a popular project like Coursera DL, I mentioned, only has one person who's really developing code. So for them, it's incredibly valuable to have someone reading their code, engaging with it, and able to give them specialized advice about the way they're using a particular library. And a testing library is something that you really rely heavily on. Your test code might be 40% of your whole code base. So you really want to make sure that you're getting the most out of it. I mean, the parameterized marker in PyTest is a good example of that. You can easily write very repetitive tests without realizing that it even exists. But if someone tells you, hey, this exists, you can condense 500 lines into 50 lines, that will just blow you away. And so the last thing is that something I really underestimated is that this kind of project, I was thinking a lot about how we are teaching or mentoring the open source maintainers about PyTest. But it was really also a case that for each helper, they would need to be mentored or onboarded onto their specific partner project. And I didn't really emphasize this to the project maintainers. Of course, in hindsight, I can say, of course, it makes sense that you need to understand the domain of a piece of software, the concepts, and the functionality, most of all, before you can effectively test it. And you need to have documentation in place to support all that, which is more than just having doc strings. So this is something that I would be aware of if I was to do this again. And so my question to the open source maintainers in the room is, can you make a space for advocates in your project, or do you recognize the work that advocates are doing in your project? And could you offer a kind of targeted code review program to users of your library? And if you were interested in doing something like that, this is a few suggestions. And something important is that you don't necessarily need to do this with new users, as we did with AdoptPyTest Month. It would actually work probably even better with existing users, just helping them really make the most of using the library. And I put here, encourage people to say, sorry, I haven't done that yet, or I'm not sure where to start. And I think it's very common that when we are blocked, or if we procrastinate and we don't finish something that we told someone we would do, and they write an email and they're like, hey, how's that thing coming along? It's kind of, it's very hard to reply and say, I haven't done that yet. And it's much easier to just kind of ignore the email, and then another week goes by. And it's a real loss of time in a short period of a project like this, which only runs for one month. So making it, trying to tell people it's really okay to say either I haven't done that yet, or I don't know where to start would help to overcome some of those communication barriers or uneven contributions. And I think it's really important to recognize your advocates and try to recognize people who are doing work in your project that are not, that's not contributing code. So something important for me was I made the point of writing LinkedIn recommendations for the advocates as a way of acknowledging that they've done really good, really valuable, and really important work for PyTest. And definitely don't schedule it in the same month as PyCon US, which is what I did. Take note. So to conclude, it was a very interesting experiment, and I think it's been really exciting to meet some of the people who took part in it here at EuroPython. And one of our project maintainers actually delivered the PyTest training that we had on Monday. So in this space of less than two months, he's gone from not knowing, or not really knowing PyTest at all to delivering training on it, and updating plug-ins and all kinds of things. So I think it was really valuable and interesting and positive experience for PyTest, and I'd encourage other projects to think about if it is something that they could do in some way. And so thank you very much to the PyTest contributors who kind of supported trying this in the first place, as well as the helpers and the project maintainers who had a lot of patience as we were kind of trying this thing and we're not really sure how it was going to work. And I just wanted to mention that tomorrow there are two talks from the two PyTest helpers talking about slightly related things, and there's more information including a PDF report about PyTest Month in case you want to get all the details about it at those addresses. So thank you. APPLAUSE We have a manager two for one or two questions, if there are any. Well, first of all, thanks a lot. I'm one of those guys who do use PyTest, but I don't contribute, sorry. But again, thanks a lot, that's just inspiring. My question is, how do you feel about the effort that could be spent on advertising a product, especially at DevLibrary through social media? For example, if I search Twitter for PyTest, I got Twitter.org with not as many followers and so on. So do you feel that is important to maintain such a channel as Twitter, Facebook and so on for developers or GitHub is just enough? Thanks. I haven't thought about if it's important in general, but I started the PyTest.org Twitter account about the time that we started thinking about Adopt PyTest Month, and I tried to use it to advertise the concept on Twitter, which worked okay. But actually something really great about the Twitter account is that many people just post tweets saying about how much they love PyTest, or they're giving a talk about PyTest at their user group that we had no idea about. It's just been given in Portuguese, Japanese, Thai, all around the world. And just having a search set up for any mention of PyTest has revealed all these things to me, which is really, really so nice to see that people love something that you are part of. And we had a meeting yesterday, we're actually talking about setting up a blog to gather that kind of information as well. So I don't know if I would say it's essential, but I think it's a positive thing. And I think having a blog where we can post links to tutorials and things will also help users to see that there is a vibrant community around PyTest. And I think that's something that people take into account when they're evaluating whether to use a library or not. It's good to be able to show that you have a lot of users, you have people who are keen to help you if you need help. Okay, I'm afraid we're out of time now, but thank you very much, Brianna, again. Thank you.
|
Brianna Laugher - The realities of open source testing: lessons learned from “Adopt pytest month” Ever feel like your open source project could be better tested? Lack of tests holding you back from contributors but you don’t know where to start? You’re not alone. [“Adopt pytest month”] was held in April 2015. [Pytest] volunteers were paired with open source software projects, to find a path to better testing with pytest. Projects varied from libraries/command line utilities, to a browser, to a complex Django app. In some cases converting existing tests was necessary, in others writing the first tests in existence for non-trivial amounts of code. Two projects were open sourced specifically to take part in “adopt pytest month”. What began as an experiment in increasing software audience proved to be an interesting exercise in strengthening community and most valuable of all, provided a newcomer’s perspective to veteran contributors. This talk will discuss what worked well with “adopt pytest month”, what didn’t, what we learned about pytest and what you could take away for your open source project, be it an improved testing environment or an improved contributor community. A basic knowledge of testing and pytest will be useful.
|
10.5446/20088 (DOI)
|
Welcome. Great to be back at Yorah Python. This is my talk, physical computing with Python and Raspberry Pi. My name is Ben Nuttall. I'm from the Raspberry Pi Foundation. We've got a series of talks today, so we've just had Karyan's brilliant keynote. And we've got a few talks lined up in this track, related to education. And it's the Education Summit. This expands to this weekend as well. We've got some sprints in education stuff, trying to crack on with Karyan's list that she's given us all to do. So a bit about me to start. I'm an education developer advocate at the Raspberry Pi Foundation. I do software and projects, kind of develop stuff internally, and do some, I write learning resources that go on our website, the free and creative commons learning platform we have there. We run teacher training courses, but it's something called Pi Academy. And I do a lot of outreach, go to a lot of conferences and stuff. We're based in Cambridge in the UK. That's me on Twitter, Ben Nuttall. If you were at Yorah Python, Picon UK, PIS, Picon Island, or Euroside Pi last year, you may have seen me speak. So just a quick update. So since I was here last year, the news is related to hardware. We've now got the Raspberry Pi 2, second generation Raspberry Pi. It's now a 900 megahertz quad core on V7 with a gig of RAM. Its back was compatible and exactly the same shape and size as the previous model, the B plus. We kept it at the same price, which is $35, and reduced the old ones down to 25, and the A plus is 20. So we are an educational charity, as Cary Ann's just been talking about. It's not something that came afterwards and bolted on to a computer company or hardware company. The idea was originally to help education. That's why we wanted to create something to empower young people and give them a possibility to learn programming skills and learn about computing from an early age in a sandbox environment. They've been on general sale since 2012, and they've been made available worldwide to everybody who wants one for the same price. We're not limiting this to just children or just the developing world. It's the same price for everyone because that's really helpful to have that whole community around the Raspberry Pi. I'll give some examples. We write free learning resources for makers and educators, and as I say, we have this PyCademy, our teacher training, CPD course, professional development course, currently in the UK. We run a few around the UK, and we're taking it to the USA next year, and we'd like to see more things like that around the world. So a couple of stories to start that are kind of about our community and about our origins and how things have developed over the last three years. Glenn Crosston is a guy in the UK who brews beer, and he wanted to use the Raspberry Pi to control the beer brewing process and monitor it from his phone and have a web app running and that kind of thing. He chose the Raspberry Pi and chose Python to write it in, so he started to develop a platform which allowed him to communicate over the GPIO pins. I'll explain more about those in the next few slides. That project became kind of output a library, it became a general purpose library for that's how you talk to the GPIO pins from Python. Another example, Dave Jones, also in the UK. His wife's a paleontologist, and she wanted to use a microscope to do some fossil photography and that kind of thing. The problem was that the university's microscope was a real pullover to use and it sort of stored it to an external drive somewhere else on the network that you had to then get permission to access and it was a real pain. So he created, he said, just stick a Pi camera on the top and then you'll get the picture straight out and be able to use them. So he built a little web app that allowed her to kind of monitor and deal with all the pictures that came out of it and archive them. Again, this became the general purpose Pi camera library and it's a fantastic resource for people to anyone who wants to use the camera for anything and he's been working on it ever since. In fact, I told him recently, because I kind of talk about this in a lot of my talks, and I said, oh, by the way, I always use you as an example of when I'm talking about community at the conferences about how you created the web app for your wife and he said, oh, did I? Is that why I started? So Raspbian is our Linux distribution that we use. It's a foundation-issued, Debian-based distribution. Currently based on Debian Weezy. Jesse is now kind of stable, so we're going to be moving to that shortly. The image that we provide supports Pi 2 and Pi 1, even though there are different architectures on 6 and 7, but we make sure that we still support our old users on the old platform. We pre-install certain software on that image, so you get the Python interpreter, you get idle, unfortunately, and a bunch of other stuff. There's Ruby on there, Sonic Pi, Java, Mathematica, and stuff like that, that we kind of put in the image and create it so that you can use it out of the box. We have other things as well that are kind of community-contributed, so the GPIO library and the PyCamera library, for instance, are examples of what's already installed, and once we move on to Jesse, it will be a lot easier for us to ship with pip pre-installed. There's alternative distributions available, so Ubuntu, now we're on V7, also works, and there are other ones like Arch Linux, but we support Raspbian ourselves and maintain that ourselves, and all for educational and general purpose use, we recommend Raspbian. You can get that from raspberrypi.org. The GPIO pins, so this bank along the top there, we refer to as the GPIO bank. Strictly speaking, not all general purpose, they're called GPIO, general purpose input output, strictly speaking, not all general purpose. You can see this little accessory somebody made. Our community is full of cool stuff like this that people create and put on sale, and there's whole businesses around this. This is a little pin diagram that you just stick over the top of the pins there to tell you which pin is which. You can see here there's some three-volt pins and some five-volt pins, some ground pins, and then these ones are GP, and they have numbers. The GP-19 is general purpose pin 19, and those ones you can control. You can have inputs or outputs, listening for inputs or sending outputs. You can wire things up to those to do physical computing. Analog, so there's no native analog pins on there. They're all digital pins, so ones and zeros, but there are options if you want to use analog inputs with raspberrypi. You can use an ADC to convert analog into digital. There's various Arduino compatible add-on boards available, and there's also PySeries or the Python module for reading Arduino pins over USB. Let's have a look at the RPY GPIO library. This is included in Raspbian, and the implementation is in C. The features are you configure pins, so you give a pin number, and you define that pin to be an input or an output, and then you can change them in your script. You can read the inputs, so you can say, is this one high or low? You can just get the value of it, and you can send outputs, so you can say, turn this pin on if it's an output pin. There's also a feature called wait for edge, so an edge is when something changes from high to low or low to high, and you can have a script that pauses and waits for something to change, like wait for the button to be pressed and then continue, and there's event detection with callbacks, so you can run a function on the event of something happening, like a sensor going off or a button being pressed. This is kind of fundamental, so everything you might need in such a program. Just a quick example, we've got some stuff on a breadboard here. We have a circuit connecting, putting three volts through an LED and through a resistor to limit the current, then we have it going back to ground, so that just completes the circuit. There's lots of little bits of basic electronics that help you with these sorts of projects. I don't know an awful lot about this, but just enough to get by. This is just a complete circuit with an LED in it, and putting three volts through there that gets limited through the resistor just means that this is always on. I can't change this in my program. It's just connected to always be on, because it's sending three volts through. If I move that from the 3V3 pin to GPIO2, then I can write a program which turns pin 2 on and turns pin 2 off. This is what the code would look like. We import the GPIO library. We set the mode, because annoyingly there are two numbering systems for the pins, because one isn't enough. We set the mode to BCM, which is the actual numbers that they are according to the chip. The GP2, GP3, GP4. We set the mode, tell it which mode we're using. I'm using a variable here just to say which pin represents the LED. I set up that pin as an output. Then here's just a little bit that just goes turn it on, sleep, turn it off, sleep, in a loop. That's just flashing the LED on and off. This is really empowering, especially for young people or for anyone. Instead of just having something saying print hello world or fry in range 10, print I, this is something flashing, something in the real world, something physical. This is another example. A simple circuit where we've got a button going from GPIO17 through to ground. Just complete a circuit there. We can listen on pin 17 and wait for that button press. This is what the code would look like. In this case, I'm actually using the pie camera library. The button is going to trigger the camera to take a picture. This is just the setup. We have our set mode. Button is pin 17. Setup pin 17 as an input. This is the pull. This is, that's wrong actually. That should be pull up or down. This is, so we set it up. We tell it which pin, what it's going to be. Then here should be GPIO pull up or pull down. This is an electronics, bit of an electronic which determines whether a floating value, like a digital input, is you kind of send it to high or low. Then you wait for it to fall. If you've pulled it up, that means it's being pressed. If you've pulled it down, you wait for it to rise. This is a mistake. Here we, and then the following on, so that's our setup. Then following on, we have with the camera, so with an instance of a camera, pie camera object, we start the preview, which means you see what the camera sees in real time. Then we have a loop which just says wait for that button to be pressed. Wait for that to fall. I would have pulled it up on the previous slide. Then it just sits and waits for that button to be pressed on 17. Remember, this is 17. Wait for that button to be pressed and then continue. It would go onto this line and capture a frame. Then it's right around incrementing the frame number, which would go into this bit here. Then this is pointless because it's a wild tree, but you could have something to catch that to exit like a second button or just a keyboard interrupt. Another example here, we've got, this is for a GPIO music box. Instead of wiring up just one button, we wire up two. We have two buttons. We've got one going to GPIO 2, one going to GPIO 3. Then I create an add, use add event detect on that button, looking for a falling edge because I've pulled it up. The callback is the function I've defined as play and the bounce time, which means how long it takes before it lets another event come in, a thousand milliseconds, so one second. I would have a function here called play, which I've defined as my callback. All I'm doing here is saying when this is called and it passes in the pin number that actually triggered the event into the function, I look up which pin in here and I get this is a pygame sound object and then I can just go sound up play. I know which pin was pressed, two or three, according to which one. I would do this on both buttons, then look up which sound to play and play it. Really simple, straightforward example of using events for multiple button presses and doing different things according to different inputs. Something great for these types of projects which are really small and simple but really empowering and really can be really interesting and engaging. They're called the cam jam kit. They're just a collection of little bits, really useful bits for lots of these types of projects. This is kit one which has a breadboard, some buttons, lights, and resistors and wires. This is kit two which has a load of sensors. It's got a passive infrared sensor so you can use that as a motion detector. It's got some LEDs, a buzzer, and some other bits in the breadboard. These are really cheap. These cost five pounds and this one costs five pounds and this one costs seven pounds. It's a really pocket money electronics. They come with a bunch of resources that you can follow along. These are just put together by one of these Raspberry Jam, the community events in partnership with one of the Raspberry Pi community resellers and accessory sellers. They just got together and made this really cheap and tried to put it out there for more people to get involved in this sort of thing. Once the Raspberry Pi came out, there was a lot of add-on boards and accessories have been brought out. One of the first ones, one of the early ones, a guy called Gert, he's actually one of the chip architects at the company that made the chip that we use. He's very knowledgeable about Raspberry Pi and he's teaching me to solder while I'm wearing an elf hat. He made this board called the Gert board which is, he's an engineer and he thinks this is exactly what people need. This is an amazing board that gives you loads more potential to get out of your Raspberry Pi. It's got all this stuff on it. He was right that it's a really cool board and it can potentially do all these things but he didn't provide any software for it. He probably just put, here's a C program which you can use for this mini project. Another barrier to this was that initially it came unsoldered. You would buy a kit like that and you'd have to solder up all of that entire board and then there was no software and no examples really to use it. A lot of people get turned off by that sort of thing because they want to get going. I'm going to go progressively through good examples. This one is a young lad actually, he was about 17 when he created this called Ryan. He created a little lad on board that just does one thing really well. It's just a little motor controller board. You can drive motors so you can use it in robotics. He didn't provide any software for this but he showed you how you can use the GPIO library to drive the pins forward. So you can, as long as you know which pin is which, so which pin it's connected to, you'll just say pin 17 is the left motor forward or pin 18 is the right motor back. Then you've got enough to go with. You just write a little bit of code, similar to the examples I showed you earlier, just turning pins off for a certain amount of time and it will drive forward or spin around or whatever it is that you've asked it to do. It's just kind of cool and a nice learning, not too steep learning curve. This is another example which is kind of further along the spectrum. It's a little add on board that has three LEDs, colored like traffic lights, has a button and a buzzer and a bunch of extra inputs on the front as well. They provide a software library for this. So import Piper Ella, you've got a full abstraction away from which GPIO pins things are connected to. You just refer to them by the light and the color. So you just run this function, Piper Ella dot light dot green dot on, the green light comes on. Instead of GPIO output 17 true, you have turned the green light on. There's a really nice abstraction, especially for young children to get involved, even less of a steep learning curve. There's plenty you can do with this as well. They've even got event detection. So you can create a function that flashes it. So turn it on for one second. So this is actually all the lights because I didn't specify a color. Sleep for one second, then turn them off. And then I can say when the button is pressed, run that function. So when I press the button, flash the lights, all the lights on for one second. And there's loads of places you can take this. This is an add-on board called Energenie. It's a company that makes remote control switches. So you can buy a remote. You can turn off the Christmas tree lights without having to get right behind it. Then they made a Raspberry Pi add-on board. So this just sits on top. And they provided you a massive file with about 50 lines of Python saying this is how you turn switches on and off. We had a young girl with us on work experience who's about 15. We said, can you take a look at this code, kind of tidy it up, work out the logic, and then put it into a function that you can turn them on and turn them off? Because 14, 15-year-old girls are better at programming than the people who make this thing. And then I packaged it and took all the credit and put it on PIP. So you import Energenie, switch on and switch off. You can pass in the number of it. If you've got a bank of four, one, two, three, four, you can turn on the sockets arbitrarily with numbers. I also made a little web app to show that you can do this sort of thing. So this is just a little basic Flask app, which is on the documentation, that you just press that on your phone to turn them on or turn them off. That was running on a local network. But if you configured your router correctly, you could do that over the Internet. And there's plenty of services to help you do that as well. So a bit over a year ago, we moved from the original Raspberry Pi, this one, the original Raspberry Pi model B to the B+, so this was still Raspberry Pi 1, but it was a plus version. And we extended the pins from a 26-pin to 40-pin, so you get a lot more to play with. So the examples I've shown you so far of hardware would sit on the 26-pin bank, and they still work because the first 26 pins are exactly the same. But we also defined a specification for add-on boards. People don't have to follow this. They can still make 26-pin ones. They can still make 40-pin ones, not following the rules, but we call these hats. So it's hardware attached on top. So Raspberry Pi hats are the name of the official specification for add-on boards. It has features such as you can use ribbon cables like this if you cut it out, and if you put it in exactly the right dimensions, it will fit nicely on top of a B+, or an A+, or a Raspberry Pi 2, and you can use mounting holes because we've got user mounting holes with screws to just hold it in place. One of the projects we worked on last year was looking at sous vide cooking, which is where you get a vacuum-packed food, such as steak like this, and you cook it in water at a very precise temperature. So we decided to design some hardware to do this. We did some prototyping first with breadboards and stuff, and then we had somebody create an actual hat for this. So if it's a hat for cooking, it's called the Chef Hat. So it was like this. We used the Energenie to remotely control the cooker. It's kind of dumb, really, and we're just using a temperature sensor in the water, and if it's too high, if the temperature's too high, we turn it on. Too low, we turn it on. Otherwise, we turn it off. There's a little library that I'm working on. It's a work in progress. We have another board called... These are just some kind of internal projects that we've been working on. There's one called the Dots board, which uses conductive paint, and there's a load of little holes on this board, little dots on this board, that you use the paint with a paintbrush to draw a pattern from to connect the dots. We tried this out at a couple of events. We went to Make Affair and South by South West and got really young children engaged in doing something more interesting than just typing on screen. They sit there for five or 10 minutes just drawing out their plane really, really intricately, and then they take it over to the Raspberry Pi, stick it on top, and then they run a Python program that draws a plane in the colors that they chose along the top, which is kind of a really, really novel idea. This is some of the part of the software, and all we do here is... We have a pin is active function. We just check whether, given the list of pins we know to be the aeroplane pins, then we just count how many of them are active and not. Just by putting paint over them, connects the outside of the dot to the inside, which shorts at the ground, which you can detect in software. Then we just have how many of them are connected. If enough of them are connected, we show the plane. It was a little Pi game app. Another one, the Plant Pop greenhouse. This is a beautiful looking board designed by Rachel Reigns from Raspberry Pi Foundation. This bit, you've kind of clipped all of these bits out. This bit sits on top of the Pi that's kind of hat-like. Then you have a wire running to the sensor board, which is this stick here. You stick that in soil. It's got a soil moisture sensor, a light sensor, and temperature and humidity. You just use it to monitor the conditions of the plant. You put it inside a little acrylic greenhouse thing. The Pi is inside there. Then it LEDs on top of the board glow, according to whatever you've programmed it to do, whether it needs watering or something like that. This is one made by an amazing American company called Adafruit, the capacitive touch hat. You stick this on top of the Pi and you wire up crocodile clips to each of the pins, to each of the holes on there. Then you can, a bit of this bill like Makey Makey, where you can build a piano with fruit or something like that. You wire that up to anything and then in Python, just read which ones have been pressed and then have different actions according to that. There's a bunch of really cool hats made by the community and made by these companies that have accessories. Pi Moroni is a brilliant example. They were one half of the team that built the Pi Brella that I mentioned earlier. This is the Unicorn hat. It's a nice 8x8 NeoPixel grid that you can program really simply. The Skyrider, which is you can hover your hand over it and you can assign different events to gestures or how close you are and things like that. The Propeller hat, which is a little prototyping board and the Explorer hat, which has got a lot of capacitive touch things along the side and sublights here. You can control motors from these extra pins you have here really easily. This sort of thing is really helpful to put out there because it means a teacher can pick that up and really simply put together a simple robot or anything that uses motors or capacitive touch. The great thing about Pi Moroni is that they provide the software libraries in Python. Always in Python 3, unfortunately, we always nag them about this, but it's great that we can have this sort of thing readily available. We have something, another internal project. We've made this weather station. It's a board that sits on the pie that has some stuff going off to some other sensors. We've produced this in-house and having it manufactured and we're sending 1,000 of them to schools around the world to collect weather data and feed into a big database. That's going to be really cool. James Robinson, my colleague, is going to be giving a talk about that later on today. Again, this was prototyped on a breadboard, just components and looking at different sensors and then some software running that just reads the sensors, logs it, uses a database, that kind of thing. We have a, this is kind of our headline project of the year. AstroPi, if you haven't heard of it, is that we're sending two Raspberry Pi's to the International Space Station at the end of this year. British astronaut Tim Peake is going to be going up in December and he's taking these two Raspberry Pi's with him. He's going to be doing some educational outreach as part of his six months in space. There's some kind of general education stuff, they're sending some rocket seeds into space and then bringing them back and then distributing them to schools for them to experiment with growing the seeds to see if, to compare seeds that have been in space with regular seeds that haven't been in space to see if there's any difference in their growth and for them to start thinking about space more. Another program is this AstroPi competition which is, we've built this sensor board which sits a hat that goes on the pie and there's a competition for kids to write code for this sensor board that they will pick the winners and run that code in space. Primary school and secondary school kids in the UK have had a chance this year to have their code run in space. The AstroPi board is what we call the sense hat, this is what it looks like. We've got a little 8x8 grid, that's an RGB LED matrix. It's got a bunch of sensors so you can get the temperature, humidity, pressure, accelerometer, gyroscope, magnetometer and it also has a mini joystick for input. We've been through a lot of testing of getting this hardware suitably acceptable for flight in space and it's gone through all sorts of vibration testing and all sorts of things to make sure that it's safe to be in orbit. It's going to be mounted on the wall in one of the modules on the ISS and it's going to be there after Tim Peake comes back so we're hoping to do more things with them once they're up there. This hardware is going to be really handy for a lot of people as well, not just for the competition so this should be going on sale within the next couple of months. It'll be really handy for a lot of projects. Just to finish off, again to remind you that we've got this education summit so have a look at the rest of the talks on this track for the rest of the day. My colleagues and I and some other people from the community are going to be talking about education, Raspberry Pi and all sorts of other things. Also at the weekend we have some sprints so if anyone wants to come and hack with us on different educational projects and hopefully we'll have some teachers there and see what they have to say as well. So thanks very much and I will close for questions. So do we have any questions? Hi Ben, thanks, that was a really good talk. So I'm quite obsessed with space just as much as I am about programming. Could you tell us a little bit more about the tests you had to actually put the hat through in order to be able to take it up into orbit? Yeah, so I don't know an awful lot about this because I wasn't part of the testing team but I kind of hear a lot from it that we've got a developer called Dave who's been doing all these tests and he's really keen telling us all about it. So for the actual launch, so for it to the rocket, the shuttle to go up, it vibrates a lot so they have to make sure that it doesn't break and break apart during the launch. We have to make sure that there's no kind of radioactive materials coming out or anything bad leaking out into the air that could be harmful to the astronauts. They have to kind of shatter one with a hammer or something and see what bits come off and what could be harmful and test it again after that. And also things like heat that it doesn't gain excessive heat. I've actually built a special case for it which is probably the most over-engineered case besides the one that we made recently for general use. This ever been made, it's a great big giant aluminium case that has the matrix, the LED matrix shown, a little hole for the camera and it's designed in a very specific way for the heat not to radiate so make sure that it doesn't get hot to the touch or even radiate out. I like your talk and I was just looking at the Raspberry Pi homepage and so I have a question, are there any Bluetooth possibilities with Raspberry Pi? So there's no built-in Bluetooth on the Raspberry Pi but you can use a USB dongle. So if you have some Bluetooth device that you can compare with, you can do that. I think we use, I think the remote for that is Bluetooth and I've seen people do projects with that, waving the weave around, driving the robot forward and that kind of thing. So yes, it's definitely possible you just need a USB dongle. Yes, the code that you showed is actually quite simple and it's meant for kids to be able to use of course. But I don't know if the Raspberry Pi has got any visual programming environments like the, for instance like the logo on the OLPCX1, I don't know if you've ever seen that but this is very simple, just drag and drop things and you then make the turtle do stuff. But I was actually thinking that with something like physical computing, it would be even nice to have a physical block dragging or block aggregation programming environment. Do you know of any projects on that direction? So the Raspberry Pi comes bundled with Scratch. So Scratch is a popular visual block based programming language. Very good for beginners and really good for young children as their first programming experience, not having to think about syntax or even bothering to type anything but just being able to click and drag and drop. So that's the Scratch is pre-installed and there are possibilities for extending that to be able to use GPIO so you can actually drive robots forward with your drag and drop programming Scratch. We tend to, in our mind, we tend to look at Scratch as the first step. The second step is really good move above and beyond Scratch called Sonic Pi which Carrie Anne discussed this morning. It's based on Ruby so there's not as much precision required in the syntax, not quite as much as Python even though we all like to think Python is great for that but we see tabs and we see spaces and we see colons, kids don't really see that stuff. So Sonic Pi is a nice step forward before they get to Python and they can do more. You can, and there's also Turtle in Python so that's available as well. That's another good one. Okay, there's no other questions so thank you very much, Ben. Okay? Thank you. Thank you. Here we go. Okay? There we go. There we go. Okay.
|
Ben Nuttall - Physical computing with Python and Raspberry Pi With the Raspberry Pi, it's easy to do physical computing directly from Python code - rather than usual embedded hardware engineering in C or Assembler. In this talk I'll show examples of physical computing projects that use Python on Raspberry Pi and demonstrate the sort of code used in such projects. Physical computing with Python is very popular in education - as it's so engaging, and more interesting than printing to the screen. This will be an informative session with learning possibilities to give those new to physical computing a change to get started.
|
10.5446/20087 (DOI)
|
Great. Thanks for joining me, everybody. So for today, we're going to be talking about my team, and the product that we build, and basically how we bought into the promise of microservices, and admittedly, maybe a little bit of the hype. We're going to be kind of going over the lessons we've learned over the first year or two of working on that, and tools for what you can do, how you can approach problems after the shiny coat of new paint wears off, and you're left with the remainder. So first up, let me give you a little background. Make sure you know what Yelp is. So Yelp, if you're not aware, is an app, or if you're feeling a little old-fashioned, a website, where you can go and basically figure out what local businesses near you are really good at what they do. And it's not a small number of local businesses, so we have about 70 million reviews, and 140 million monthly unique users. So really a lot of content, a lot of depth there. And as well as depth, we've also got a fair bit of breadth. So most of you might associate Yelp with restaurant reviews, and that is a very popular use case, but we actually have a lot of other businesses on Yelp and a lot of other content to look at. So shopping is our largest category, but we have a fairly widespread, and in general our goal is to make sure that if you can find a business on the street and it has a storefront, we want to make sure it's on Yelp and you can find out information and find out good things about it. So for myself, I'm Scott. If you're inclined to use Twitter, you can find me at Scott Trillia on there. I've been at Yelp for four years now, working on backend systems, spent a lot of that time on the search team, but have worked on a variety of different applications, machine learning, a lot of work with locations and geocoding, and I've also spent a lot of time both developing infrastructure for our service stack at Yelp and being a consumer of that service stack. But the team I'm going to be talking to you about today is called the Transaction Platform Team, and it corresponds directly to a product that Yelp provides called the Yelp Transaction Platform. So let's talk through a little bit of what that means. And in general, if you hear me referring to just platform, this is what I'm talking about. All right, so what is platform? Our goal, in short, is to make sure that for all of those great businesses that we have on Yelp, as much as possible, we can actually connect users directly to their goods or their services. So if you're at home hungry after a long day and you're looking for some delivery, you might search on Yelp to figure out which places are good to go. And if you decide that what you're really in the mood for is a next crispy tacos and you go to that business, we want to give you the opportunity with the platform product and the platform team to go ahead and just click and get whatever that good or service that next crispy tacos provides. In this case, hopefully, next crispy tacos is going to be giving you something like a carne asada taco or maybe some chips and guacamole, whatever is appropriate. However, we don't only have places that serve tacos on Yelp. So of course, we also partner with a number of other third parties to provide other types of goods and services. So you can find some clothing. You can also find some of the shops or you can find a hotel booking we partner with Hitmonk, as well as a number of other partners and different types of product verticals. And then the goal is, regardless, you can go ahead and check out on Yelp, pay with some stored information so you don't have to be re-entering it on 19 different websites and establishing all those different logins and get that thing that you were looking for in the first place. So that's a little bit about the team. The only other point of order is making sure we're all in the same term with microservices. Just a quick show of hands. Has anybody in this room not been to at least one talk with the word microservices involved this week? Okay. That's kind of what I thought. So, yeah, to say it's a hot trend might be a bit understating it. It's pretty recent, at least in this exact term, if not the general concept. But honestly, it's hard to avoid at this point. I'm probably familiar with it. That said, let's go over basic terms so nobody's wildly off base. So I like Martin Fowler's definition. You can read words yourself, but I want to just point out a couple of them that I find particularly insightful. So he focuses on microservices as being a suite of small services. So you're going to have a few of them, and you're hopefully aiming for them not to grow out of control. He also wants them to be in their own processes and communicating with lightweight mechanisms. So in our case at Yelp, this is things like HTTP, JSON, separate web servers, things like that. So this is mostly by contrast to monoliths. And this is sort of the traditional term for the single app. All of your code in one place, all of it being deployed together atomically and working from there. I would say if you ever hear the word monolith and you hear people talking about their monolith or your monolith or anything like this, I would really encourage you to clarify what kind of monolith are they talking about. It's a very generic term for a lot of different kinds of systems. If I do a little hobby, you know, Rails app, that's technically a monolith. And in Yelp's case, if we have a million lines of code in a single code base, that's also a monolith. There's some slight qualitative differences between those two. So in Yelp's case, our monolith was big. Like I said, million lines of code, million lines of tests. We had hundreds of developers working on it. And you can imagine the sort of problems you have when it comes to a monolith of that size and amount of code of that size, that much cooperation in a single code base, things get a little messy. And that's basically what we saw at Yelp. And that's what led us to go to a service-based architecture. So what is bad about these monoliths? I don't want to just assume this is a given. I'm going to say a few brief points, and then we'll move on on the assumption you've already heard this a few times. First off, one big problem we had was monolithic Python code was really resisting decoupling. We all know that we would like different parts of our application to stay different parts of our application, to not be tied together accidentally or on purpose. Unfortunately, despite the best efforts, trust me, we found this was extremely hard and basically completely unrealistic to attain in a single code base. You might be objecting, you might be saying, well, I know, Scott, I know about all these things, Zope interfaces and all sorts of great tools that you could use to decouple your code base. Trust me, we thought of it. We in fact even thought we didn't need to go to microservices or to service-based architecture at first, and we basically found through a very expensive series of tests that we actually did. So it is surprising how much it resists decoupling and best efforts in that direction. In addition, it has the bad habit of catering to the lowest common denominator. So we had the case in our monolith, we called it YelpMain, where there were a number of old libraries. We've been writing code for a decade in this application. It's got some age to it, and so you might imagine the occasional upgrade needs to happen to something like SQL Alchemy or what have you. And it turns out that the oldest and scariest and darkest corners of your code base that nobody likes and nobody wants to talk about and certainly nobody wants to change are exactly the things that are going to hold back the parts that actually need to move quickly. And so we found that we were being limited on libraries, we were being limited on choice of languages. Wouldn't it be nice to run in Python 3? Well, it turns out when you have to port a million lines of code to do that, it's a significant undertaking, and you're being held back by all these parts of your code base that you don't really want to be held back by. And finally, monolithic Python code is really the opposite of being agile. We wanted to be pushing multiple times a day to production. We wanted to be running tests quickly and nimbly. We found ourselves doing neither of those things and moving in the opposite direction that you would want to be as far as whether those are getting better or worse. And basically this was what really drove us to say, it's fine, it's time. So we started doing services. Circa 2011, actually about when I joined the company. And this is just the cumulative graph of number of services over time. You can basically, my whole point here is it started pretty rapidly, it continued rapidly, and it hasn't stopped anytime soon. This includes green field development, this includes slicing pieces off of the monolith and pulling them out into services. And this feels awesome. Like this is especially the first part of this graph is an amazing part of sort of the best feelings of microservices, the best benefits, all the things you see in a nice slide deck when somebody is presenting why you should move to this architecture. That's what you feel in the first six months to a year of this. And it was amazing. Fast code pushes, actually having isolated systems and being free of these ancient dependencies that we didn't care about. So my team, when we built our system, said, man, that sounds like a pretty sweet deal. We can have this horrible old creaky monolith idea that has been proven to be really painful and terrible. Or we can have this really nice new setup as microservices, everybody's talking about it. Admittedly, it sounds kind of cool. I like to write cool things. So we really bought into this. We had this idea of what it was going to be when we finished, and it sounded pretty excellent. So it was a little bit surprising, a little bit concerning and upsetting when we actually built that, scaled it out for a year, and we looked back and realized, you know, that thing that we built, what we actually came up with, it didn't really look a lot like what we had been promised. Right? So you can kind of see, you can see where we were in the right direction. We had the rough concept. You can even recognize individual features. Maybe our pushes got slightly faster, or maybe we had some good luck with getting some isolation. But the whole picture, having it all work together exactly as we were told, that was not happening. And that was pretty frustrating. So let's talk a little bit about why that is. Very briefly, we saw a lot of problems as we grew larger and really got embedded in this service-based architecture. Our API complexity increased. So those nice isolated services that made sense and did one thing well, started to do a few things kind of okay, which is a lot less compelling, it turns out. We perversely saw coupling rise, right? The whole reason we did this in the first place was because I was tired of being coupled to something I didn't care about. It turns out it's kind of really hard to build a product, continue to evolve your product organically, and not accidentally couple yourself back to everything that you tried to get away from. The interactions between the systems get really murky. When I call a function, I can read that function, I can dig into its contents, it's running in my process, I have some guarantees about what it's about to do. When I call an external API, third party, even worse, who knows what it's going to do, right? Being able to say confidently what it does or does not do turns out to be extremely hard. And finally, all those processes that we talked about, pushing, deploying, testing, being confident that things are working in production, none of what we were used to doing was scaling up to this system of services, and that hurt. In fact, that was actually exactly the opposite of what we were trying to do with this whole idea. So that kind of brings us to the topic of this talk, which is great. I bought into the hype, right? I bought into what I was told was going to be awesome, and I was convinced that it was going to be awesome. I implemented it more or less as told. And a year later, I'm finding myself left with the exact same problems. Why is my code coupling to things I don't care about again? Why aren't my tests fast? That was the entire point of this thing, right? Why can't I deploy regularly every day? And so I want to kind of give you some hard-won tools that we've built, some that we've learned to sort of attack these problems and to be able to leave you with the good parts of services and hopefully be able to work through and improve the bad parts. So to do that, we're going to talk about four main areas. Curiously, you might note that all of these areas are kind of in the bullet list of benefits of services in the first place. It turns out there's a really nice initial spike when you move to services where these all get a lot better than they used to be in the monolith. And then there's a really nasty plateau, or even dip, depending on the details of your system right after that. So we're going to kind of talk about how to avoid that plateau, how to get over that dip with each one of them. So first up, let's talk about decoupling. There's this old boring problem, right? I have a monolithic ball of spaghetti code. What are you going to do about it? Well, if you've been paying any attention, there's an easy answer here, right? The solution is microservices. What are you even doing here? Like, this is so well-worn at this point, it's not worth going into more detail on. Unfortunately, from the perspective of coupling, we have other problems now. My services are absolutely going to share concepts. You can't get away from this. That's arguably the entire point. So when they share concepts, how can I do that in a maintainable way? And when I now have all of these distributed systems, some of which I control, some of which maybe my colleagues control, what in the world do I do when I need to refactor something? Now you're in a real spot. So we're going to kind of walk through basically a case study of one particular problem we had, and talk a little bit about where we ended up on it. And that concept is called a service type. I will give you the convenient and mostly true definition, which is a service type tells us what product a particular business provides and how they provide it. Okay? And for a team that does fulfillment for a variety of different verticals, this is a pretty core concept. So when we launched, we launched with food. These were the early days. We're just going to start, we're going to keep it simple, get a nice MVP out and see if the idea has merit. We're not going to make it too complicated. We're not going to over generalize. What are our service types? Well, pick up and delivery, right? That seems reasonable. You're either going to pick up food from the business or you're going to deliver it, and that's going to be the only definitions of service types that we have. Okay? Our product did well. We're growing. We want to expand. Let's do something new. So we're going to do appointment booking. Okay? So we're going to pick up and deliver. These have a couple options. So we're going to name them booking it business and booking it customer. We've learned, you'll note, we don't call it generic pick up and delivery because it's not generic. It's booking it business and booking it customer. So here's the learning curve. We're really accelerating. And you go through another year of growth and success, honestly, and what you find is that this is a complete mess. Like what's going on here? Let's just pick a couple of these. If you look at hotel reservation and booking it business, I'd be curious if we took a poll of the audience, if anybody, like if the majority could successfully tell me the difference between those two. And why do we have goods at customer and delivery? Isn't delivery goods at customer? Like this stuff is nonsense, right? And this is where we found ourself. And we honestly, we kind of got there because it was convenient, right? And it made sense to us. So I want to digress briefly into an architectural diagram just so that what I'm about to say makes a lot more sense and you have context for it. This is roughly what the high-level architecture looks like of our system. We'll quickly walk through it and then go ahead and explain how it relates back to service types. So we have front-end, right? These are things that face users' web pages, many views, checkout pages. It communicates with fulfillment in the center. Fulfillment is essentially a keeping track of order state. Are you ready to be charged? Are you ready to be delivered, et cetera? The important bit on the right is payments, kind of essential to the whole process. And then we have a big loop in the system, essentially. Fulfillment decides that something needs to be done to an order, and it does it. It notifies the partners, third parties, which I will say this a few times, we do not control, so these are external companies with opaque implementations. They're going to call back through the partner API and say, cool, your order is ready to be charged, we do some work, and we kind of keep this loop going, feedback between us and the partner. So that's the rough architecture. And a couple things of note. In yellow on the right, these are systems we don't control. These are internal systems we don't control. Partners, external systems we don't control. The front end here is actually all back in the monolith, because a dirty little secret of migrating your code out of a monolith, you usually don't completely migrate your code out of a monolith. So front end, for reasons, is still locked in that code base. So we have complications here, right? We have a lot of sort of history that would take me an hour to explain to you properly. And this whole system is messy. It's messy from the start. So highlighted in red here are all the pieces of the system that knew about service types. And when I say no, I mean I saw the code and they all cared a lot, right? We're not talking about like I'm passing it along to my friend downstream, we're talking about I'm making important functionality breaking decisions based on this data. Partners aren't red, but that's just because I don't know what partners do. Realistically, we have no control over what they do with service type, and that's equally terrifying if a little different. So this concept in general was really confusing. It was pervasive, alarmingly so, far more than we ever intended it to be. And it was really convenient, right? This was something that we used all over the place to make decisions, to use if statements, you know, this was very useful to us. But it wasn't designed, and that was really the core problem. So I want to kind of talk about how we approached this, and I want to preface right off the bat with saying we needed to do a refactor, we needed to do it across all of these systems, and we knew we couldn't completely eliminate the concept. Those were the constraints we were given. So it's not a great spot to start from in a hard refactoring project. And our approach was basically we're going to draw boundaries, we're going to introduce some domain-specific concepts, things that make more sense to the system they're in than this big service type thing that I couldn't even explain very well to you. And we're going to make sure those concepts are tied to functionality, right? And this is sort of the domain-driven design concept that if in a system you're using words vocabulary concepts that are consistent with the system, your code's going to look a lot nicer. So I don't want to go into detail of the refactor, because it was long, painful, and boring, right? This is five different services with interfaces across to all the other services. Service type was in an alarming number of them, and it was just hard mechanical work. What I do want to note is our eventual solution was to basically corral it into the smallest space we could in the system, and then write little adapter layers to concepts that made more sense for those systems. So we don't control partners, we had to keep service type for them, although maybe someday we would deprecate it. But in the rest of the system we were able to basically transform this into something that made more sense locally and was more maintainable in the long term. So what lessons can we draw? First off, and this might be like the most enlightening thing we discovered, interfaces are not just your APIs, right? Interfaces are everything that's shared across system, and it's alarming how many things are secretly shared across your systems. We got bit, not because service type was a resource on five services, but because it was that little add-on parameter on the end of 15 different service calls. This was something that was convenient to us, and we didn't really examine it closely. So be very aware of what you're sharing and just be intentional about it. And if one day you have to refactor, know that it will be very expensive. That's unfortunately what comes with a service-based architecture. In addition, sacrificing this dryness was a really good choice in this case. It is not always the best choice, and I think by default most of us are pretty uncomfortable with repetition in code. But it's a tool you have in your tool belt, be aware of where it could help you out. And one of the places it helped us out in this case was at service interface boundaries. It's a natural place for decoupling, so if you're going to do something, that's a pretty convenient place to do it. All right, let's talk about defining, and by defining I mean our interfaces in particular. Have you ever needed to understand something and been told to go read the source? I assume yes. It's a pretty common retort, right? I'm busy, go read the source. Okay, maybe it has its place. How does that work in a network of services? How does that work if I don't own the service I'm supposed to be reading the source of, or if I don't speak the language it's written in? It kind of gets a little gross. What happens if I want to know your services interface, but I don't actually know if you validate it properly or at all? Does that mean your interface is the one you wrote down, or the one you told me, or the one you actually validate in practice, or all of them or none of them? That's a pretty common retort. The fact is coming from our Python monolith, honestly our interfaces were bad. This is a cultural thing. This is a habit thing. Python is not the strictest language when it comes to defining interfaces clearly. If you want to write bad interfaces, Python is thrilled to let you do so. Say what you will about that, I think that's just a fact of life that we live in, and we need to be aware of it when we're moving into situations where the importance of interface increases. I like this example. This looks like the kind of Python code you might show to somebody. This is pretty Python. Look how elegant it is. We have some Quark usage. I'm getting points in my expert Python user column. We've got some validation. Everybody loves validation. We have very nicely named methods. This is an important interface for us. We want to be very clear about what is or is not happening here. Unfortunately, I'm willing to bet nobody in this room can say with any certainty what's actually happening here. What is order? It has an attribute. That's what we know. What are in those Quarks and what in the world are they doing to the users? And really overall, what are we staying with this interface? We don't know. We have gotten into from the monolith from working in a single Python code base. We attacked this problem with something called Swagger. Swagger is an API description language. It is essentially a spec for what your API is. If you heard Lin Roots' lightning talk a couple days ago, I think, on Raml, Swagger and Raml are sort of slight variations on each other. Different implementations, different design choices. We chose this with three main goals in mind. We wanted to make sure that we documented our system for ourselves, both for people on other teams and for the people who are writing the interfaces. We wanted to make sure that we made our clients smarter and as much as was smart and why is to do so. And we wanted to make our servers smarter. So what is required for Swagger? In short, you need to write a gigantic spec. You need to describe your API. What endpoints do you have? What arguments do they take? What do you have to do? What is the best way to do it? What do you have to do? What is the best way to do it? You might be laughing to yourself and thinking, why are you making a big deal out of this? I was astonished at how hard this was. This is really hard to do. If you give your best developers on a service the job of writing a spec to this level for their service, they will not know how to do it. They will have to look it up in code. They will have to back it out. They will have to reverse engineer it. They will have to reverse engineer it. They will have to reverse engineer it. They will have to reverse engineer it. You might think you do, and that is more terrifying. This is the work you have to do. It is not a lot of fun. I wasn't a big fan of it when I did it for our API. There is a hint for you. What do you get out of it? That is the real question. First up, Swagger has a great set of tools. Swagger UI is essentially a pretty view on that same data we just showed you. Swagger UI is basically a tool that you can use to create a web app. On the client side, we have a library called bravado. The goal of bravado is to be, again, consuming that API from some remote service, learning about what that service is actually offering in terms of its end points, in terms of its types, and doing all the annoying mechanical work for you that you don't want to be doing by hand. We maintained client libraries, Python packages for a long time, that did this themselves, and they did it wrong, and they did it incompletely, and they validated it, and they had many other issues. It's a hard job to do, right? This is aimed at basically making all the mechanical difficult parts go away. On the server, we also have something equivalent. I have a library called pyramid swagger that essentially hooks into the pyramid web framework. If you're using that, if you're not, there are equivalents for other frameworks. Its goal is to do the equivalent on the server side. What does that look like? You're serving the swagger schema at something like slash API doc so that other things can access it. That looks like applying validation, because you've defined your entire spec, remember, and you don't have to rewrite that validation or do it incorrectly. There's also a variety of other smaller tricks and goodies available. What lessons can we pull from this? First off, your interfaces should be intentional. Don't patch them on piece by piece. Don't build it halfway, and then slap together the rest of it. If you don't, it will become a complete mess. I guarantee it. We've seen this like clockwork for every API that doesn't get a regular, basically redesign at least, whether or not you actually implement it, thinking about it from top to bottom, figuring out what the new concepts are and what concepts are outdated. It's an important process. Your interfaces need to be explicit. This is a thing that sounds very attractive when I say it out loud, and people don't really like doing when they have to actually implement it in code. There's no shortcut here. If your implementation is not explicit in what your interface is, you're going to still have an interface. You just don't know what it is. You can't make smart decisions based on it, and you're going to eventually get bit and bitten hard when you accidentally break backwards compatibility because you had no idea what you were actually changing. Finally, find the mechanical things about this process and automate them mercilessly. I say mechanical pointedly here, because one thing that we didn't do was automate away the network. We didn't hide the network in that swagger client, and that was due to some hard lessons where we kind of automated away the network, and we discovered that that was a dead end and a very dangerous area to go down. But all the mechanical stuff, all the stuff you can completely know and unabashedly automate away you should do so. Okay, so let's talk about production. We've kind of talked about design, decoupling. What's it like getting this thing actually running in practice? This is a real customer bug report. Your customers might give you bug reports that look similarly. It says, I was using your API. I'm a good bug reporter, so I tell you which API endpoint I was using, and I saw a 504. It happened yesterday. There you go. Go fix it for me. This is your job. My job is to report the bugs. Your job is to fix the bugs. What can you do with a system like this? What can you honestly go for? There are a few approaches in ancient times. You might pull in your most experienced developer, stick them in a room, tell them they saw some 504s coming from that away, and hope that they can just kind of grab whatever logs are lying around and fill in the details, right? And I don't know. My poor analogy here is that this is kind of like picking up a random bookshelf and hoping it's going to help you on a research project. Like, okay, it might. Yeah, you might get lucky and find some things that are relevant, but you're also going to pick up just whatever this person had lying around. And if you're using whatever logs your company or your team happened to create previously, you're going to be getting stuff that was built for a different purpose. It doesn't really work very well. Okay, fine. Let's build our own logs, right? Let's log all of our APIs. We have a bunch of services. We're going to know every request that comes in, every response that goes out. We're going to nail this. Great. Well, so there are some obvious upsides to this problem. You no longer don't have the data, and there are some obvious downsides. This is your job, right? Go find me the pattern in all of this data and do it by tomorrow because I'm not paying you to just sit around and look through data. This is technically possible if you have some command line gurus. You might be able to whip out a crazy one-liner that involves five different uses of grep and manage, but I don't really recommend it. What we've kind of settled on is both logging everything we know about and getting tooling to basically examine it efficiently. And for us, that's ended up being Elasticsearch and Kibana. This may be an old story to you. It's certainly a very popular choice. You'll often hear it in a connection with this guy here, Logstash, and collectively, this is called the ELK stack. If you're completely unfamiliar with it, I'm not going to really do a good job of explaining it, but I would encourage you to go investigate it because it's, frankly, pretty awesome. So what do we use it for? Well, Kibana lets you build dashboards. Elasticsearch stores the log data. That's the long and the short of it. So what we've mostly found is it's tremendous what you learn about your system when you log data in production and you actually look at it. It's not a lot more complicated than that. We did nothing more complicated than showing ourselves a graph over time of all the API hits broken down by partner. That's all this is. What we've done is we've looked at, and we've looked at, and I think we've looked at this in a way that's a little bit more complicated than it really is. But what we've looked at in the previous video is a partner that has decided this is their pattern of hitting our API. This kind of blew our mind. If you had asked us all beforehand, we would have said they're all doing exactly or more or less what we told them in the API documentation. I'm sure there's not anything too weird in there, right? Because we were really specific when we wrote that documentation, and it's totally clear what they're supposed to do. You can also learn great things about your errors. So when you push out that bad code and you see a spike of I can't talk to your service errors, you at least have a way of viewing it now. And this is definitely a step up from tailing a log somewhere and piping it through grep and making sure you don't write the wrong regular expression. So this kind of stuff is permanent and you can share around links and it's very nice. But how does it help assault that mystery 504? So you're not going to be able to read this, but this is a dashboard you can create quite easily. And the real trick here is even though by default, these dashboards are going to be a view over all of our log data, we can write queries to whittle that down. We can add filters to clarify what we're looking for. So in quite small text up at the top, we've basically said I want only results that have the method path, you know, user info, that they're hitting that part of the API. And I want everything that doesn't have a 200 status code in response. Okay, not complicated. So we're going to go ahead and do a quick look at the data and we'll see what we can do. So we've got a lot of data that we've got out in this particular day. We were doing pretty well and we got exactly one of those and that was the customer complaint. So this is the kind of thing that would have been, I'm not going to say impossible, but it would have been particularly unpleasant and inefficient under raw data formats. And it's quite nice when you have tooling like this. In fact, for this problem, we dug a little deeper and using basically the same setup, we were able to say, hey, if we split out timings by data center, really it's hard to overstate how much value you're going to get out of just logging what's happening in production and being able to view it efficiently. And realistically, we don't want our customers to have to be telling us these things in the first place. That's a little bit inefficient and it leads to its own problems. So we have a monitoring tool that we've open sourced at Yelp called Elast Alert. And what Elast Alert essentially does is it sits on top of Elastic Search and it does three things. You tell it first, what is the Elastic Search instance I'm looking for and what's the index named, where am I looking basically. You tell it what sort of constraints you want to apply. So in our case, we're saying I want to see at least 20 errors in two minutes, although this language is rich and you can say many things. And then finally, we take some sort of action. So what's going to basically happen with this rule is if we see 20 errors in two minutes, we're going to go ahead and page on call. We're going to say, hey, there are more errors than we'd like. And maybe we can insert a nice graph link that says, here's something that will help you understand more about the problem. Okay, lessons learned. First up, logging, it's honestly a superpower. This stuff is awesome and it's going to be amazing for you. I can't really overstate how important this has been for us. And I would say it's night and day since before we started using it extensively. That raw data is not enough either. You do need to visualize it and ideally you need to be proactively monitoring on it just so that you can be spending your time doing what you care about rather than digging through mountains and mountains of raw text logs. And overall, these approaches have made a world of difference. These took our incident responses from days. I mean, we had to wait to get an email from a customer for that one. Down to minutes, right? We're getting paged the moment this stuff happens. And it took the investigations figuring out why something went wrong from basically arbitrarily long, because maybe we just never could figure it out, down to something much, much quicker. Okay. I'm going to go ahead and skip the last one, because I'm straight out of time. But I want to talk overall about some sort of lessons we've learned and how to wrap this up. There we go. So overall, how can you approach problems like this? Frankly, the first step is you really need to understand where you're coming from. You need to understand the system you came from before. In our case, this big monolith, a million lines of Python code, and all that implies. And you have to factor that into the decisions you're making. If you're ignoring that, you're going to make mistakes based on those weaknesses. This isn't anything new, but it's an important thing to keep in mind. Services aren't exempt from this. They're not a magic cure-all for all your problems. In fact, they just exacerbate some of them. Second, be explicit. You want to be straightforward and be really clear about what your system is, how it interfaces with the other parts of the world, and with what you expect. Being explicit is going to help you in a lot of ways that are hard to quantify while you're just thinking of it abstractly. But it's sort of the first step toward any automation. If you don't write something down, if you don't document it, ideally in a format that's programmatically readable, you're never going to learn from it. You need to measure everything. If you're not measuring something, you just don't know it, and you certainly are never going to be able to automate it, right? This is the kind of stuff that is basic to say, and it has a really profound effect in practice. And again, it's taken us from feeling like we sort of knew what we were doing to feeling like we're on top of any new issues that come up and that we can even proactively respond to things that are going to be a problem in the future. And finally, scaling. Microservices are more complicated than monoliths. This is just a fact. You're introducing overhead, you get some nice benefits from it, but you can't just pretend that the way that you treated a monolith is going to work in the new world. And automation is the obvious way to get around it. It feels really rewarding to pull off. And it's going to turn your team into a bunch of people that are able to really focus on what they care about, building an application, delivering whatever the value is that you deliver, rather than a team that sits around and patches up after broken services. A couple of resources to just leave you with. Those GitHub repas I mentioned for actual projects, Bravado and Pyramid Swagger for Swagger Integration, and Elast Alert for working with Elasticsearch and doing monitoring there are all on GitHub. We have a long-form article on our transition from monoliths to services, written by the tech lead of our services team. Really well-written, has a lot of background information. I haven't dove into even most of the issues that came up in that, so it's very interesting reading. And if you're more in the mood for something bite-sized, we have what are called our service principles. All of the senior members of the teams that had been working on services for a long time at Yelp basically got together and said, here's what we're going to write down and here's the summary of what we know. And with that, we may have some time for questions, and if you want to get in touch with me after those work, and also the hallway, I'll be at the Yelp booth. Thank you. APPLAUSE OK, so we have time for a couple of questions. Hi. I know you said APIs need to be, you know, you need to get them right as much as possible from the start, but you don't always do that. And so how do you deal with API versioning? Do you have a tool for that set of best practices, something like that? Sure, sure. Yeah, you're absolutely right. Obviously, we did not get it right at the start, and we continue to version our APIs. We're still solving this. I would say that the best suggestions I have are, we really appreciate at least documenting the interface you have. That's a great first start, because it will make you realize when you need to version your API. And as far as interacting with them, we've been mostly just treating them like any other endpoints. You know, you need to make sure that you make a V2. You can have clients then switch over to use it. All these logs that you're collecting will let you monitor when V1 is actually completely dead, rather than you just don't know of any consumers right now. So those tools can help you out. I don't think I have any sort of plug-in solution for the overall problem, though. Yeah. So, thank you for your talk. In the beginning, you said that not only did you want to address complexity, but also wanted to interact with different languages and decouple that as well. Now, all the tooling and everything you've shown was in Python, which is great, because we're at EuroPython. I was still wondering, did you end up using other languages in your system? We do use other languages. Yelp is, I would say, 95% Python, ballpark. We have Java services for the more high-performance stuff and for the search stack. And we don't, to my knowledge, currently have any cross-language talk, but the beauty of Swagger in particular is its data, right? It's not code. Your schema is written down in data. Your schema can be read in JSON, and there do exist plenty of Java-based clients for Swagger. So we definitely focused on making sure that would work cross-language. I don't want to misspeak and say that we have yet. I think we're in the process of doing so. Yeah. Did you look into the rest constraints and more specifically hypermedia formats in order to get a more decoupled system? Did that help? Yeah. I can say that we talked a lot about hypermedia. I can't say that we went much beyond that. I think for our money, we've been busy with much more practical. That's sort of like all the way on the top of the pyramid of needs. When you've got everything else nailed, maybe you'll do hypermedia. We're still working our way through some more fundamental issues. And these kinds of low-level issues are way before you get to the point of nice hypermedia APIs. So that's kind of where we are in the evolution, yeah. With Alastel, what's the lead time from bad requests starting to hit your servers to them being fed to the log processor, you know, searched and then getting an alert? Sure. Do you mean theoretically or in our system? I mean in your system. In our system. So Alastel doesn't care about how the data gets to Elasticsearch. So depending on what your data generation scheme is, that matters a lot. For us, we use scribe logging. We have some bridges into Kafka for persistence. And I think the overall lead time to get into Elasticsearch is on the order of 30 seconds, 10 seconds, like low seconds on good times and it can get delayed. In practice, we find we can rarely react faster than that. So that's been pretty excellent for us, yeah. And there's no reason you can't make it faster. We just haven't put in the effort to do so yet. How do you handle registration and governance of the services and the granularization? How do you know which back ends to talk to? More about how do you manage, you have many services inside, I guess, like different teams and how you prevent, for example, reinventing the wheels, reusing the services and so on. Yeah, so how do we make sure that we aren't constantly redoing each other's work and all that? We have a few ways of doing that. We have a service infraternity that focuses on building tooling. Inevitably, there are also a source of a lot of expertise, so you're very much encouraged to go to them and discuss. We have a sort of, by similarity, something that's like PEPs, we have basically SCIF service reviews, basically. You're designing something new, you put it out for review, you write a formal spec and you say, see if anybody likes it or has problems with it. So we have a lot of human code review kind of processes like that. And we've generally tried to make sure that services always have one coherent owner, a team. We have one service sort of attached onto the side of the monolith that lets us access that data. And we've had a lot of trouble in the years that it didn't have an explicit team with it just sort of being a tragedy of the common. So making sure someone owns it is very important. Okay, I think that's probably all the time we have for questions now. So just a round of applause.
|
Scott Triglia - Arrested Development - surviving the awkward adolescence of a microservices-based application The potential upside of microservices is significant and exciting. So much so that Yelp's Transaction Platform committed from the start to an architecture of small, cooperative microservices. This talk explores the inevitable complications that arise for Python developers in as the services grow larger and stretch both their own architecture and the developers responsible for them. Come hear tales of terror (tight coupling! low test coverage!), stories which will warm your heart (agility! strong interfaces!), and everything in between as we follow the adventures of our plucky team. The talk will be focused on the functional, cultural, and reliability challenges which occur as a microservices-based project evolves and expands over time. Particular attention will be paid to where these diverge from the utopian way microservices are often described, and to the particular difficulties faced by Python developers trying to implement such systems. My goal is to share with attendees some mistakes we've made, some successful methods for growing gracefully, and Python-specific tools/libraries which can help with these problems. To enjoy this talk, you should be aware of the basic vocabulary and concepts of HTTP-based services. Any additional awareness of distributed systems (and their failure modes) will be helpful.
|
10.5446/20086 (DOI)
|
Hello everyone, I'm Abraham Martin. I work for the University of Cambridge, as you may have guessed from the huge logo on the screen. I was thinking about making a little introduction and I thought everyone knows what the University of Cambridge is, so I think I have little to say about that. And I thought I can show you some pretty pictures so you can invite me where I live and this preplace and where I work. This nice architecture, this nice rivers you can pond with your friends during summer, well the two days of summers we have a year. The classic math bridge, as you may know, we have a lot of clever people there, some noble prices, people that usually work around the city dressed like this, which are gongs. It seems weird but you really see, when you go there, you really see these people working around academics. This seems pretty classic, pretty old, but we also have pretty nice new buildings like this one, which is the new University data center, which is one of the top data centers in the UK. It's pretty big, it's green, we have a lot of innovation inside it. We even have HPC service, which has, in 2003 was the second most green computer in the top of 500, so it's not just classic building and architecture, we also have some cool things. This is the computer lab, what I used to work. I did my relationship with the university, started with the computer lab, I was doing my PhD and then I worked there as a postdoc. This building is called William Gates Building because Bill Gates paid for the half of the building and I work on this building now, which is the university computing service, where we provide IT services for the rest of the university. Both buildings share a common history. They used to be the mathematical laboratory where this machine, brownie points if you know what it is, it's the exact computer, one of the first computers in the world based on the von Neumann architecture. Those built there, we still have some pieces in the computer lab as a museum. But also other things were built there and in the part that I'm working now, which is the university computing service, like Exim, which you probably know about because still 50% of mail servers use it. We have pretty cool people working there, I'm not of any of them, but we have some pretty cool people that also work in a lot of open source projects. What I want to explain to you today is one service that was proposed a lot of years ago, which is the managed web service, and was born to solve a problem. We have a lot of researchers in the university, as you may know, and a lot of them usually hold a conference, they do research, they want to do some simple website with the statistics, show statistics, show results from the research, or either do questionaries, et cetera, et cetera. They end up using their own web servers under the desk. It was a cheap computer, it was running under the desk. It was not maintained usually because the academics usually use that computer for the conference and then they left this computer under the desk, the software was not updated, and then we get security problems. We get servers hacked, et cetera, et cetera, you know. The proposal that the AT services and the university did for solving that problem was centralizing these web services. The solution was to provide a service where you don't have to worry about maintaining the OS or the software, you only have to worry about maintaining the web application. We maintain the OS, we give basic web hosting capabilities like external services does, you don't have to worry about backups, and you have some dedicated resources to your web app. That's very, very old, and when I say gold, it's like 15 years ago. The first version of the managed web service was using a Solaris 7 running in the Solaris machine, so you can see that it was using a very old version of Apache, PHP, and it was using a true system to maintain the separation between these different web pages. The second version came sooner than the other one, provided the new software, XLRS 10, Apache 2, more new software, and started to use Solaris zones, which is a kind of virtualization inside the Solaris machine. It's kind of a container, so we were using containers before it was cool, but it's still pretty old. It also had more enhanced features like database, driven script, so you could do a script based on some information in database, so it's centralized, it's very to manage, some needs, and NFS server, very classic, set-afes file system, so it provides also snapshots, which is good, and the users were able to create Vhost, Ali-Assist, et cetera, but the problem was everything is manual, so they had the users send us an email saying, I want this, and we make the changes, we execute the scripts, the scripts make the changes, but everything is manual, so we need a lot of human intervention. So when it started to grow, we currently now have more than 200 users and more than 400 websites, it started to become a little bit difficult to manage, because it requires a lot of time. So before we ended up making a new version of the managed web service, another solution was in parallel, which is the Falcon service, which is a clone-based, you only get a clone instance, you don't get access to a server or anything, it's just a CMS as a service, and we also have like 200 websites there. So if you go to any university website, you probably will end up in either Falcon service or a managed web service service, so for example, one of the most visited websites inside the service is the Stephen Hawkins website. So we decided to make from scratch a new service, so restart what we have done, because we don't have more Solaris machines, Solaris machines are routing, they're pretty old, we don't have a replacement for that, so we also thought, let's do more automation, we can do more automation, so it requires less time from us. So we decided to go to the classic dedicated VMs, but still maintain the same things that was proposed by the previous ones, like no root access for users, and everything is maintained by us. I want to say by us, I mean by Ansible, because we don't touch anything, but we will see that later. And so to end up these emails that come to our inbox saying, can you please install these packages, can you please install these, we created a web panel using Django, where we delegate some power to the users, so the users can do things without having root access or anything. So architecture is basically a devian aid machine, we installed the basic packages that we have been installing up to now, like the Apache, which is the most common feature of the month, but we also support other Apache available, like if you want to install Python, Django, et cetera. We have a list of system packages that you can install, and they are pre-approved, so you don't end up with a machine with a lot of packages that you don't need or that are strange to need. And we give them all the power to do authorization to the sites, create the host, apply for domain names, install TLS certificates in the machines, do the back-ups from them, password reset, power management, et cetera, et cetera. So we give them the power to do a lot of things that we were doing before. So they have a panel, don't blame me for the design, it's an in-house design, if you visit any Cambridge website, you will see that all of them look exactly the same. So it's just a panel with some options to manage your site, so when you create a site you get this web panel based on Django, you have some options to create Vhost, ask for domain names, et cetera, and you get an extra VM, which is a test server, so you can clone your production server to a test server, you can test things there without having to compromise your production server, so that's good, especially for people that have Drupal installing the manage web servers when they Drupal, really bad things happen. So it's very, you can test it before, and if it goes right, you clone it back to the production server. So architecture looks like that, we will go one by one, see how we build it. Be aware that this is not a talk about OpenStack, neither about talker, so don't expect any of that. But we use all the, most of them are Python technologies, and we did that project in a few months, it's still not finished, we are still working on it, but most of it we have done it by using one and two, 1.2, 1.3 FT, so it's not much people, so the amount of resources that you require for doing it, although it seems like a huge service, it's not that big. So we have here the VM architecture, here, the VM is separated, the VM service is separated from the rest of the stack, so we start describing the VM architecture. The VM architecture is just a VMware solution, you may know these VMware solutions, it's just ESXCI servers, and you can manage these XCI servers using VSphere control panel and some APIs, and we have a standard backup server where we do the backups, but it's not replicated, so if something happens, then we rebuild the VM and recover the things from the backup server. So the flow is easy, a user enters to the Django web panel, authenticates, so we know who he is, and then he asks for a new managed web server. A host name and an IP64 are allocated to this site, the VM API creates a new VM, the VM API installs the OS, and when the OS is ready, Ansible is executed, and Ansible is the one that configures the whole machine, so we are using Ansible as a configuration management and it does everything we need. So there's no Ansible, it's just a bunch of things together, they are easy to script, so it's very easy to understand what they are doing, they separate into folders, which is very good, and you can find the file that you are looking for, and there is separation of things in the different files that you can see. So it's pretty good to use, it also has inventory, so you can define all your servers based on dynamically or static, so you can have a file with all your servers, or you can inject the output from another API as a list of servers you have, or even the database, et cetera. So it's pretty nice, it works really well, and it based on playbooks, playbooks is just a bunch of roles linked to a bunch of targets, so you have a role and definition of role is things you want to install in this role, in these machines that have this role, and then you have targets, and then you say this target, these machines, I want to install this role, for example, a web server, web server can be a role, and the web server role has a lot of tasks, then install Apache, configure Apache, et cetera, et cetera. This is an Ansible playbook, as I said before, you define the host where you want to install things, and then you define the roles, does these machines in this list will have. For each role you have tasks, templates, which are changed to templates, scripts, handles, and variables, you can have also global variables or variables entering into the script, and this is how a role looks like, just a bunch of tasks inside the role, you can see that here we're installing packages, it's a Jamel file, as you can see, and it's pretty easy to understand what you are doing, so if you're working with more people, it's easy to modify the file, change other configuration, et cetera. You can see here that the templates can be used with variables, are Ginge to templates, we use variables there, and therefore we use the same templates for all the configurations, all machines. You also have Handlers, which are basically callbacks, when some function in Ansible is executed, then you have a callback later, and you can, for example, if you have updated the Apache configuration or your Django app, you can restart Apache, and the callback is cool. So this is from the VM part, we use this VM for infrastructure, we use the APIs, we launch it, we create the VM, everything is good, after that, Ansible configures the machine, and then we can offer the service to the user. So if we start from the top of the stack, we see authentication part, we have our own authentication, we use Raven, Raven is our authentication service, so you can see that we have a lot of services interconnected using a lot of APIs. It is based on a web of API, and we have to build a custom Django web again, but this could be substituted by any authentication that you could use. You can use the Django one if you want, you can link with your own enterprise if you want, et cetera. So the second layer is authorization, we have a kind of a Nellabish service, it is called lookup, and then what we have there is just a list of users and a list of groups, we can see these users if they belong to each, which institutions they belong to, which groups they belong to, so the end user can configure their MW server based on this list, they can search for other users, authorize them as administrators, authorize by groups, et cetera. So it is just a basic list. We use that instead of using the Django groups, because it is more useful for us, because people use in the university to this L-Lab service that we have, so they create the groups there and they are automatically updated if someone leaves, et cetera. So, when the user has authorize the user, another user to enter the machine or use the service, we need to still install the user into the machine. So we have another service over there called Django, which provides more information about users, it is like user identity management. So we get from Django a unique UID from the user, we need that, because if we install the same user in different machines, we still need to identify the files that belong to him or to her in different machines, so they have the same unique UID and we use this unique UID in all the user installs. Users install using Ansible as well, it is installed on all the VMs where it is authorized, and we have periodic refreshes to refresh the lookup groups we have authorized, so if the groups change, people different change, and we allow people to upload their own SSH keys and the SSH keys is also installed in the user configuration, so they can enter either using the password, which is checked with this L-Lab server or either the SSH that they have installed in the panel. So once they can access to the Django panel and they have the user installed in the VM, they can already access the machine, everything is configured for them, so they can start using it. For previously, we had also another communication with the IP register API, which is on the bottom there. This allows us, this is another external service, so as you can see, we have the main service and then we have a lot of other services that we communicate with, which provides the university registration for KAMACAC domains, so if you want to register a new KAMACAC domain, for example, important studies.KAMACAC, you can launch an API, we launch an API request using the same Django panel, Django panel sends the request, and then we get the main name alias for that site, so everything is configured automatically, the user doesn't have to worry about the internal processes, this API tells us if the user is authorized for that domain name or if this domain name is already in use, et cetera, and the same API also provides us from IP addresses. We have to pre-allocate some IP addresses together with the host name, so when the user requests a new site, he can access directly using the domain name without having to wait for a DNS refresh, so what we do is just pre-allocate some addresses, and then when the user gets their site, they can access really using the host name without waiting for a DNS update. We use two IP addresses, we have one address, one host name as well, as a host address, so the communications for the host and another for the service, so if we want to move the service to another machine, we can do it without having to modify the host name or the host of the machine, so we can separate what is the service and what is the host, and we can move the service IP, and you will see this is useful later. Additionally, we have SSHSP records and DNSSEC, anyone knows what SSHSP records stands for? No one, good. This is to forget about that, I'm sure pretty much you all have seen this screen like that, and what it does is you can upload SSHSP record into the DNS with your public host key, and then if you have DNSSEC activated, you don't have to check the host fingerprint manually because the DNS server does it for you, it gives you the fingerprint, and when you check that the fingerprint is the one that is in the DNS server, which is secure communicating with you, and then you don't have to check manually if the fingerprint or the machine you are connecting is the one that it's claiming to be. So that's pretty useful, as I said, we have a lot of services in the same architecture, we have an inventory there, which is using another API, it's based on JSON API, on pool, consume, so we give to this service called best the data of all our services, so we can use it as an external database as well, when we know where all VMs are, we know where they are located, the IP address they have, etc., etc., so it can be used even as an inventory for Ansible, or it can be used for other purposes. So as you saw, we have a lot of APIs, different ways of accessing APIs, we use SSHAPIs, REST, non-REST, HTTPS, using JSON, non-JSON, but we have to deal with a lot of them in an asking way, because we don't want Django, the main thread of Django, to be stopped by that, so what we do is executing as a background process, using a Chrome jobs, which is the easy way if you don't need the API executed after just after the user has launched the petition, or if you want the execution scheduled, you can use salary and credit, which is what we use. Sellery is pretty good for us, because it works pretty well with Django, it's very easy to configure, you just have to add on the top of the function, you just have to declare that it's a shared task, you can use different templates, like this task with failure, you can define the number of retries, you define the template, so you can define if it fails to log something or send you an email, et cetera, et cetera, so it's pretty easy to configure, and it works pretty well, and you can also execute Chrome jobs from the same salary, it's called salary beat, so it's different, but the jobs are just configured as it was a Chrome job, so it's pretty useful for us, it works well, and these architect, these salary, these APIs and these services are supporting all of these, all Ansible driven, so the changes are then in Django, Django stores these changes in the database, and then Ansible is executed, takes these changes from the database, and then it executes these changes on the VMs, so we have this service, we went to the community in the university, we made a workshop, and we said, we have this for you, we thought that you will like it, and they said, hmm, we will like it, but what about if the service fails, and then we thought, well, we have a backup, you can recover for the backup, it won't take too much, we create the VM, et cetera, and then they said, hmm, but I need an SLA if I want to switch to you, but we didn't have an SLA because we had a backup, we have a plan, but we didn't have, we didn't thought about what happens if 300 VMs fails at the same time, which requires a lot of time to recreate and a lot of time to take from backups, so some of the people were saying, hmm, we are thinking to change into MWS3, but only if you provide high availability, so lucky for us, we designed the application so it can cope with different VM architectures, which is good because you don't have to worry about the VM architecture that you are using because you are creating the VM using an API, which maybe this one that we are providing or could be an Amazon EC2 server, and then we execute everything through Ansible, Ansible only needs an SSH connection, so it's pretty easy, so we just need to replace this component, which is the VM architecture. So we thought, okay, let's update VMware with high availability, and then we saw that we need replicated vSphere, replicated the storage, which we didn't have, and replicated the storage for a lot of servers, it's very expensive to maintain because you need to do huge file storage that is shared between all the VMs, so we had a lot of things pending from this architecture, so we thought that's pretty risky for the low time we have better take another architecture because it's still, it's also expensive to acquire all the hardware and software that we need, so we decided, okay, we don't want to maintain a huge share of the file system, so what we do is replicate each one of the VMs file system to another one, so we thought we can use the VMware, still use the VMware infrastructure, we had a pacemaker cursing, which is basically a cluster that checks that all the VMs are in contact with each other and then can change the service network configuration, this is why it's useful to have service network configuration to any of these two production VMs, so we have a replicated VM, the second one, the second column is just a VM that is waiting in something fails to be changed and start acting as a reactive VM of the cluster, and then we replicated the storage individually for each one of the VMs using the RVD, which is basically a driver that sends all the rides of a machine into the other VM, so the storage is replicated to the second one and the pacemaker takes care that if some of the components fail, the switchover is made automatically, so we don't have to worry about it, but then we thought, oh, this is maintaining a lot of clusters, we may end up with one cluster for each one of the VMs that we're going to have, because for each one of the VMs we will have to have a pacemaker cluster, and this is very expensive to contain, and it may fail if we need to execute an C-ball, we need to execute an C-ball in both sides in the two VMs, so they are synchronized, so that's a lot of work, and it's going to break quite easily, so we thought, okay, let's start from scratch. We move away from VMWare, and we decided to use Shen, Zen can be configured very similarly as the VMWare, so you can see there are two Zen servers there, they are also executed, pacemaker and coursing, but the difference is we don't do clustering for each one of the VMs, we do clustering for each one of the Zen servers, the Zen servers have a lot of VMs inside, so if something happens with one of the servers, the whole cluster, the whole server and all the VMs that are inside one of the Zen servers that you can see on the top, all the Zen VMs that have a Zen server are automatically migrated, live migration, to the second one, and you don't notice anything, even the sockets are kept open, you don't notice that the switchover has happened, with the VMWare solutions you will have to wait until we restart the VM, for example, and that kind of stuff, with Zen, you don't notice anything, you just don't see that nothing has happened, so inside you have changed your VM from one Zen server to the other, but it's completely transparent to you, so for doing that it's a bit more complicated, because this is the file system that we had to use, this is a very complex file system where you have all of these, these disks are tied together in a physical volume, and then you have the left one is the DOM zero, which is basically the operating system that is running in the Zen server, and then each one of these are the V, each one of the individual Zen hosts, so you have all the storage replicated to the other Zen server, which provides this live migration, so we had a happy transition, it's working well now, but the architecture is the same, because we designed the architecture, so we could change the VM provided, but just changing the API, the API is a middle API, so we only had to write an API for the Zen server, and then execute everything was exactly the same, so we are happy with that, seems people may be happy with that as well, and we changed it from the VMware solution to the Zen solution, which is three node clusters, and all clusters are in different locations, we can do live migration so the users don't know anything, it's still using Ansible, and we also use Ansible to deploy more clusters, so this is an example of the Zen server clusters, we can deploy many of them, it is easy to deploy because it's Ansible, so if we want to create more Zen server clusters, it's just as easy as create, well, get the machine, the physical machine, start it, and launch it, so it's pretty easy, let's talk one minute about security, because we like security, well, I am not an expert in security, we like to enforce security to our users, so we then end up with problems, we decided to not use root passwords when we create the Zen host, so we don't have to manage root passwords, which is difficult to secure a lot of root passwords in a database, we don't want to manage that way, so we only use keys, we connect to the machines using keys, Ansible connects to the machines using keys, etc. We have a separation of privilege, for example, we need to regenerate the host keys of the Zen host, the host needs to be generated previously, because we need to upload the SSHFP records before even the machine is created, so we need to have a pool of host keys that we can use in the future, and we can install in the machines. We use user service, which provides a useful interface, so users can execute commands from root to users, or other more privileged users based on some filtering and some templating. We also provide a TLS certificate service, this one, an additional one, because we want to follow one of the new initiatives from the EFF and the Mozilla, which is HTTPS everywhere, or let's encrypt, it's an open source CA, which provides you with free certificates for your web page, HTTPS everywhere, EFF trying to force everyone to have HTTPS web servers, and even the HTTP2 specification doesn't include to enforce HTTPS, but a lot of people are saying, when we move to HTTP2, everyone is going to be HTTPS. This sounds really true, but because the specification doesn't say it, but all the implementations from Microsoft, Internet Explorer, Mozilla, and Google, and Chrome only implemented HTTP2 if it uses HTTPS. So we like to test our servers. I encourage you to do the same if you enter to the SSL labs, you can get a qualification of how secure your web server is, which is pretty good, because it gives you some hints of if you have any open back, all the specification, all the version of open SSL, et cetera, et cetera. Then apart from security, we use also some metrics and logging systems, so we can give users some information about how their host works. For example, we use a metrics service, which is basically a statsD and collectsD in each one of the machines installed, also using Ansible. We have a cluster of metrics brokers that get this information from all the hosts, and then we have a cluster of carbon graphite, which is stored, this information gathered from all the machines. So the user can see these graphs in their panel, the web panel, and they can see how the machine is behaving, et cetera, et cetera. And we are now trying to implement LogStash, Elasticsearch, and Kivana, which also provides information about the hosting, the web server, and how it behaves, whether you have the VCs, how it behaves during different periods of time, et cetera. You can have a lot of logs gathered by LogStash stored in Elasticsearch and then showed in a Kivana. So that's pretty much all I want to talk about. I hope you like it. Thank you. Thank you, Abraham, and we have a few minutes for questions. Any questions? Yes? Why did you choose Shen instead of KVM, for instance? What made you choose one thing and not the other? This was a long discussion we had between the developers, which is basically three. We didn't have any strong reason to choose one or the other. We saw doing a little bit of research that Shen worked a little bit better with the IPD, which is one of the main companies we wanted to use because we wanted to do the replication of storage from one server to the other. So we thought that it was integrated inside the same server. We decided to go that way. But we could have chosen KVM. It was in the list of bugs that we had to research and decide. Hi. Thank you for the talk. It was really fun. But to balance, I didn't really understand the subtleties in the last architecture because you had several hard drives and then you have several Zen servers that actually overlap on several drives. So were there virtual servers? Yes. So this is the picture of the VM architecture on the top. This is more the view of the file storage. This is a single machine with a rate of disks. And this is a file storage for a single machine. So you have the physical volume and then you have the first column is where it used to be called DOM zero which is the operating system that manage all the VMs when you access to the Zen server. You access to this. It's also a VM but you access to this and it has direct access to the hardware instead of the VM which have access to the hypervisor. And all the other columns, each one of them are a Zen host which are one of these VMs or DOM view as it's called in Zen. And each one of them has the DRBD device which is basically a virtual block device which is replicated from here to another DRBD server. So you have each one of these DR devices are inside this list of virtual block devices and they are replicated through the network in real time when they are written or read or only written because it's a sync to the secondary Zen server. So it's done automatically through the network when it's executed but the difference is each one of them has a DR device. Okay. Thank you. Okay. We have maybe one minute if anyone has a really quick question. Okay. Please join me in thanking Abraham once again. Thank you.
|
Abraham Martin - Architecture of a cloud hosting service using python technologies: django, ansible and celery The talk will show the architecture and inners of a cloud hosting service we are developing in the University of Cambridge based on python technologies, mainly django, ansible, and celery. The users manage their hosts using a web panel, developed in django, with common options: ability to create a vhost, associate domain names to vhosts, install packages, recover from backups, make snapshots, etc. Interaction between the panel and the hosts are made using ansible playbooks launched asynchronously by celery tasks. The VM architecture has been designed to be VM platform agnostic and to provide disk replication and high availability. The University of Cambridge central IT services also provides other services to the rest of the university like domain name registration, authentication, authorisation, TLS certificates, etc. We link all these other services with the hosting service by using APIs while keeping a microservices architecture approach. Thus, enabling the use/link of other services within the same hosting service web application.
|
10.5446/20081 (DOI)
|
to add the number of the JIT decorator on it. And it only optimizes that function, or those functions, if you have many of them. The nice thing with that is that it allows us to react to semantics. That means we are not bound to Python's user semantics. We can cheat a bit, and we cheat actually quite a bit in order to optimize your code. But also, your high-level code around that, all your classes and meta classes and so on, they can still use all kinds of complicated things. That number doesn't support that. It's not a problem, because since they are executed in the regular Python environments, then number doesn't care. So as I said, specialized. So really, well, right now, it's specialized for number crunching. It's really, how to say, it's tailored for NumPy arrays. That is, NumPy array is the dominant data type in scientific computing. It has a lot of features, and we try to support them. A bunch of other things. So we are slowly trying to extend the range of things that we support. But still, right now, it's more specialized in number crunching. So the main target is the CPU. We officially support x8664. Ideally, LLVM provides us with support for many other architectures. And we also have a target for NVIDIA GPUs using CUDA. So this means you write Python code and you can execute it on the GPU. But we've limited features that because there are some limitations on what you can do on a GPU, of course. The reason to your runtime, you will be able to do memory allocation, but it will be quite slow. And the allocated memory will run in the GPU's global memory, which is not very fast. We also have potential support for other architectures thanks to LLVM. So one of my colleagues tried a number on the Raspberry Pi, and it actually works. But we don't support it officially. I think LLVM takes several hours to compile. We have some support going on for HSA, which is something I aimed to call the 8-minutes heterogeneous system architecture. It's an architecture for what they call APUs. So the goal is to blend the programming model between GPUs and CPUs. You write one implementation and it can run either simultaneously on the GPU or CPU, or you can run it on either of them. And supposedly, there's some memory sharing and so on. Let's talk a bit about the architecture. So number, if you compare it to other jets, is quite straightforward. It's not very exciting. It works one function at a time, which is a constraint which we're going to relax because we need to relax it in order to support recursion. But right now, it's one function at a time. It starts from the Python bytecode. So we don't have a parser. We just use the bytecode emitted by C Python. And we have a compilation analysis chain which transforms it slowly across various steps to LLVM IR. So LLVM IR, it's LLVM's internal representation. It's a kind of, let's say, supportable assembly. And it allows you to specify a lot of things. The difference we see, for example, is that you can specify some behaviors in a very granular way. For example, you can specify if signed overflow on integers is well-defined or undefined. If you have, for example, in this example, if you have undefined behavior on signed integers overflow, then it allows LLVM to do further optimizations. So after the LLVM IR is shipped to LLVM, everything is delegated to LLVM itself, including low-level optimizations and also executed in the function. And on top of that, we also generate some Python-facing wrappers because each function gets low-level implementation, which takes some native types. And you have to marshal those from and to Python objects. So this is the compilation pipeline. You see there are two entry points there. You can see the wavy rows. So the first entry point is the Python byte code itself, as I said. We have an analysis chain from the byte code. First, the byte code is analyzed. We build a control flow graph, a data flow graph. And we produce something which is called number IR. So number has its own intermediate representation, which is quite as high-level as byte code. But it's a bit different. It's not a stack machine. It's based on values. The second entry point is when a function is actually called. When a function is actually called, we record the types of the values. And we do type inference of those values. We try to propagate all the types across the function. I'm going to talk about the number types just after. It's a bit more complicated than just mapping some classes to some types because we have more granular typing in number than in Python. After the type inference pass, there's a pass which deals with rewriting the IR. So it's an optional pass which has some optimizations. The next pass is lowering. Lowering is from the LVM jargon. It means that you take a high-level language, which is numbers IR, and you lower it to something very low-level, which in this case is LVM IR. And then we ship everything to LVM, to the LVMJIT, which produces machine code, and we execute it. So there is a small rectangle named cache, which is grayed out because it's not implemented yet. But ideally, we will be able to cache either the machine code or the LVM IR in order to have faster compilation times. OK. So number types. As I said, the number type system is more granular and more precise than the Python type system. We have several integer types based on the dependence on the bitness, on the signness. We have a single precision and double precision 13-point types. We have a tuples are typed, which means that you don't have a single tuple type. You have a different tuple type for every kind of parameter that's in the tuple. So tuples are typed based on each element's number type. So you have a different type or type, for example, for a pair of int and for fluid64, for a pair of fluid64 and fluid32, and so on. NumPy array themselves. So they are a very important part of number and of scientific computing. They are typed according to the dimensionality and to their contiguousness. So the lowering path is what really takes the type inferred number IR, and it transforms it into LLVM code, LLVM IR. So this is very straightforward and not very exciting part, but it has a lot of code because we implement a lot of functions. We implement all the operators. We implement math functions and so on. And if we are careful enough with what we generate, we can allow LLVM to inline and do other optimizations here. So what's supported? NumBa supports a rather small subset of Python, at least. Unless in tax field, unless in tax front, it supports quite a bit. Not all. It supports all control flow routines or constructs. It supports raising exceptions, but not catching them. It supports calling other compile functions. We have recent support for generators, but only the simple kind of generators, that is not those to which you can send values, not coroutines, but just syntactic iterators with yield keyword. So what don't we support? We don't support well, over rest. We don't support exception catching code. We don't support context managers. We don't support comprehensions. And actually, we don't support list, and sets, and dicts yet, although it will certainly come. And we don't support yield from. As for the built-in types and functions, we have support for most types which are useful for scientific computing. So all the numeric types, integers, floats, and so on. Topples unknown, which are quite basic. And we have support for the buffer protocol, which means you can address, you can index over byte, byte raise memory views. And the other thing which supports the buffer protocol, which also includes, for example, memory map files using the M-map module. We have support for a bunch of built-in functions. And we have support for most operators, but of course, only the types that we support. So all the numeric types. We are able to optimize several of the standard library modules, mostly those which are specialized for numeric computing, so C-map and math, of course. We have support for random number generation. We actually use the same algorithm as C-piphone, so the mouse and twister, except that we have a separate state. We have support for C-types, which means you can call row C functions from number code, which is a cheap way of actually calling C libraries. And it generates very fast code because it calls it from a native context. Similarly, we support CFFR, which is just a replacement for C-types most of the time. And we support mostly NumPy, at least a large subset of NumPy. So what we support in NumPy is really the objective of the whole page in the documentation. So I'm not really documenting. I'm talking a bit about it here. We support most kinds of arrays in NumPy, so most dimensionalities from 0D to ND. We support arrays of various D-types, scalar arrays, numbers, and so on, structured arrays. We support arrays with sub arrays in them. The only thing we don't support, and we won't support in a long time, I think, is arrays containing Python objects. Because the whole point of NumPy is to generate native code, which doesn't go through the C-pyphone API. We have recently added support for constructors, so we can do memory allocation, allocate memory from number functions. Various operations on arrays, such as iterating, indexing, slicing, so there are various kinds of iterators we support, such as v.flat operator and more or less fancy ones. We have support for reductions of the products, cumulative sums, and so on. On the scalar types front, we support daytime 64 and time delta 64, which are weird. And I think little known types which allow you to do low level computations on some daytime and time deltas. And we support NumPy.random in the same way that we support the random module. So the limitations, apart from what we don't support in terms of syntax, in terms of syntax, and in terms of types, we don't support recursion. So that's because we're compiling one function at a time and we'll have to elevate to change that. We can't compile classes. Again, that's because we compile one function at a time, so we don't have a way of specifying a structure and several methods operating on a user-defined type. And the other limitation is that type inference is really has to succeed. So if the type inference pass fails to infer a type for a given variable, then the whole compilation fails. Ideally, we would have a way to say, well, this is a Python object, but the rest is still inferred, so we will be able to bridge it. But right now, this is not possible. And actually, when type inference fails, it goes into a mode called object mode, which is not very interesting as far as performance is concerned. So as I said, the fact that it's opt-in, it allows us to relax the semantics. So as you have understood, perhaps, it has fixed size integers up to 64 bits. So for example, if you have an addition of two integers and the result overflows, then you would just see no truncated results. You don't have an overflow or anything. We take the liberty of freezing the global and outer variables, so we consider them constants, which makes it much easier to compile. And it allows us to generate more optimized code. For example, if you have math.py, then usually math.py won't change, so it's only fair to consider it a constant. But of course, in your module, you have a global variable whose value changes. Then you won't see it in your compile function. It will keep the order value. So we don't have any frame introspection. Basically, we don't have any debugging features right now, neither from the C level nor from the Python level. So this is something which, at least at the C level, we're going to work on it, because we want to expose the names of the JIT functions to LLVM so that you can fire some JDB and have a nice traceback. So how to use it? So basically, the main way to use it is to use the JIT decorator. It's very simple, so you have a function, and you just tag the decorator on it, and hopefully it will be able to compile it. So the default way is not to pass any argument to the JIT decorator, and it will lazily compile a function. This means that it will wait for a function to be called, and it will do the type influence thing at this point, and it will generate the native code. And since you're calling the function, it will call the native code on the fly. And there's another way to call it, which is to manually specialize the arguments. Let's say you really know you want some 32-bit ints, and you want some double precision floats, or some single precision floats. And so you are able to pass an explicit signature to number.jit. But this is not really recommended. It's mostly for us to test. So there's an option to remove the guill, which is quite easy for us, since we are not calling any CPython API from the generated native code. So you just pass no guill equals 2, and the guill will be removed. So the guill is a global interpreter lock. For those who don't know, it's a lock which constrains CPython execution to a single thread. If you remove the guill, you can call your function or your functions from several threads and have a parallel execution on several cores. But of course, you have no protection from these conditions. So you are in the same position as a C++ programmer who has to be careful about not having several threads accessing the same data and mutating it, for example. As a tip, instead of having your own thread pool, you just use concurrent.futures on Python 3. Another feature is the vectorized decorator. So NumPy has something called a universal function. To explain what a universal function is, it's better to take an example. So if you take the plus operator between arrays, for example, which is a shortcut to the np.add function, the np.add function is basically doing an element-wise operation on all elements of its inputs. And the way it's implemented is really to have a loop on the element-wise operation internally. The nice thing with a universal function is that you have several additional features. There's something called broadcasting in NumPy. So if you are adding, for example, a scalar, an array, actually the scalar will be added to each element in the array. So really, the lower dimensional argument is broadcasted onto the higher dimensional argument. So this is handled automatically by the Ufang framework. And the inner loop doesn't have to care about that. And it also gives you, for free, some reduction methods. So you have some reduce and accumulation functions. So NumPy comes with a fixed set of universal functions, so add, multiplication, square root, and so on. Traditionally, if you want to add a universal function, write your own, you have to go in C. So you write your inner loop in C with a specific CAPI provided by NumPy. You compile it against the right NumPy version, and you get your universal function. So it's not very convenient for users. And the users don't do that. So using NumPy, you can write the element-wise function in Pupy function. And you can put the vectorized decorator on it, and it will generate the Ufang. Another more sophisticated feature of NumPy is a generalized universal function. So this is an extension of the idea of a universal function. Universal function works on one element at a time. It doesn't see the neighbors of the rest of your arrays. A generalized universal function can see the whole arrays, and you have to specify exactly what the layout of the inputs are. So it's almost for some more sophisticated functions, such as a moving average. So NumPy also allows you to generate a generalized universal function using the geo-vectorized decorator. So here is an example. It's called the Ising model. So it's something which is used, apparently, mainly for benchmarking, but it seems inspired from some physics model. The basic idea is that you have a two-dimensional grid of Boolean states, either Boolean or binary states. And you can think of each element having either a value plus one or minus one. And it starts from a random state, basically. And at each iteration, you make each element vary based upon its neighbors. So at the end, it's supposed to converge towards something which is quite stable. So this was generated with number, this animation. So if you look at how it looks like, well, you have an inner function which processes each element in the array and which updates it based on its neighbor's value. So there are a couple of operations. It takes its neighbor's value and combines it with the actual value of the element. And it takes a decision based on that and a random number. And the outer loop is just looping over the whole array. And it updates all elements. So the outer loop which we see in Update 1 frame, it does one iteration. And then if you want to make the model converge, you have to call it a number of times. So if you measure that, well, you get 100 times speed up for number of a C Python, which is less than you get with Fortran. But still, it's within range. In this case, it's twice lower. And we know why, actually, because array indexing in Python is more sophisticated. For example, well, the main reason is that if you Python allows negative array indexing, you know that if you have a negative index, you are indexing from the end. So you have to have a runtime check of the negativeness of each number. And in some cases, LLVM isn't able to optimize it out. So besides that, we have QLED support, as I said. So the main API for that is the QLED.Jet decorator. So we don't try to hide the QLED programming model. The QLED programming model is based on the notion of a grid of threads. So you have blocks of threads, and you have a grid of blocks. And the GPU executes all those threads in parallel, more or less. But you have to tell the GPU which is the topology of threads. And besides that, there are two types of functions. There are kernel functions, which are called from a CPU, actually. So the kernel function is not able to return a value. You pass it some arrays, some input arrays, some output arrays, which are marshalled automatically by number to the GPU. And you write the results in the array from the GPU. And there's something called device functions, which are really sub-functions. And they are called from the GPU to the GPU. So these ones can return values. When you're using the QLED support in number, you have a limited array of features, because, as I said, you don't have a large long time available on the GPU. So it also requires the program to have not only some knowledge of QDA and how a GPU works, but also to have some intuition of how to optimize the code for execution on the GPU. Because it's not usually arranging your algorithm in the same way on the GPU or on the CPU, especially except in trivial cases. So here is an example. It's a very simple example of this one, just to show you how it works. We are trying to compute the cosine of an array. So we're using the QLED agenda decorator. We have a function which takes two arguments. The first argument is the input array. The second argument is the output array. So there is no convention. It's just a choice here, for example. We must, the idea is that each GPU thread which will compute one value over the array. So it will take one element in the array and compute the cosine and put it in the output array. So the first thing is that you are computing the index. So to compute the index of a current thread, you call the QLED grid function. And then you have to just call math.cos on the input and write it in the output. So this is the definition. Then you want to call it. You have to, so GPU cos defines the GPU function. Then you want to instantiate the kernel, actually. And instantiated the kernel means that you define the grid topology. So this is the thread config here. It's a two-element tuple. The first element is the number of blocks in the grid, I think. And the second number is the number of threads in each block. So you define the topology based on the length of the output. And you call the GPU cos function with the topology and the input and output. So in this example, the numbers are better on the GPU. But it's not very important because you won't use a GPU just to compute a cosine. You will do something more complex. So if you want to install number, since it's open source, you can compile it from scratch if you want. But you have to compile LLVM on a specific version of it because LLVM has backwards and compatible changes in each feature release. So the current version of number requires LLVM 3.6. And you will have to fetch LLVM 3.6, compile it for your platform, or get, if you can get them, some binary development packages. And then you have to compile LLVM light with a sufficiently recent C++ compiler, which is not trivial at all. So we really recommend you use Kanda, which is a continuous-owned package manager. So it's an open-source package manager. And it comes to with a default distribution of binary packages called AnnaKanda. And if you have Kanda, you just type Kanda installed number and you have it on your platform. So let's wrap up. So you can find documentation on the web. We have, of course, GitHub account with code and issue tracker. You are very welcome to come to a number of users made in list, either as a user or as a potential contributor. I must also mention that number is commercially supported by Continuum Analytics. So if you want to buy consulting enhancements, support for some architectures, you can write to sales at Continuum.io. And there's a last thing called number pro, which is an extension, a proprietary extension to number, which provides bindings to some specialized libraries for the GPU, various scientific specialized libraries. And it also has, I think, it has extensions to allow it to parallelize the code easier on the CPU. So that's it. Thank you. Thank you. So two questions about your use of LLVM. First, it sounded like you supported only a subset of all the platforms that LLVM supports. Why is it that you don't just have the same support requirements and platform list as LLVM? What did you say? We support a subset of what? Do you support everything that LLVM supports, or do you only support a couple? You mean as architectures? Yes. It's a matter of validation, because ideally, each works. But who knows what it will give actually, you know? OK. And I was also wondering a couple of years ago, I had an attempt to marry C Python and LLVM together called Unladen Swallow. I was wondering if nothing ultimately came of it and Unladen Swallow died. But I was wondering if the work that they had done was helpful at all in the development of Numba. I don't think so. Well, not directly. Then at the time, they said that they had helped LLVM improve the support for JIT compilers, perhaps indirectly benefited. But we didn't take anything from them, because we use our own wrapper, one LLVM called LLVMLight. And then Numba is pure Python. The big difference with Unladen Swallow is that Unladen Swallow did everything in C++, which I think was a very, I mean, it's necessary if you want to compile very fast. But it's also much less flexible. So pure Python allows us to experiment and develop very quickly. I have three questions. First question is, does a number compile, do the JIT compiling in a separate thread? No, it's in the same thread. So you actually have to wait for the compilation to finish before it gets fast? Yeah. Well, what will you do anyway? Because it's lazily compiling. So if it's compiling when you're calling the function, anyway, you must wait for it. Yeah, of course. But I mean, well, anyway, sometimes in some JIT compilers, they do it in a separate thread, and it just continues with the slow version until it's done. Oh, right. No, we don't do it. OK. The second one is, do you have any support for storing the compiled code? For? For storing the compiled code on JIT? Oh, not yet. No, as I said, we want to support caching, but not yet. So that's what you meant with caching. So you actually not like PyPy, which has the problem that it cannot store the compiled version? I'm not sure if PyPy does that. PyPy can't cache anything. OK. So they have to redo it every time you run the code. So it would be more efficient if you just do it once and then store it. Yes, of course. Yeah, that's an obvious way, not the thing to add. But right now, we don't have it. OK. And the third one is, how do you do error handling? Because you said you don't have any way to catch exceptions? Yeah, so we have a way to raise them. So if you raise an exception for number code, then you just catch in when it goes outside of a number code. So you can communicate errors to the user, but you can't handle it in the number code. OK. Maybe an extra question. So you support, you're working for the support of NumPy. Do you plan also to support PyPy? Are there some plans for that? Not yet. We mostly support NumPy right now. So every kind of pure Python code which relies on NumPy arrays may be perhaps accelerated if it intersects with a subset of things we support. But we don't have direct support for anything other than NumPy right now. I suppose we are, someday, we want to support Pandas. We have no more time for questions. Thank you. Thank you. Thank you.
|
Antoine Pitrou - Numba, a JIT compiler for fast numerical code This talk will be a general introduction to Numba. Numba is an open source just-in-time Python compiler that allows you to speed up numerical algorithms for which fast linear algebra (i.e. Numpy array operations) is not enough. It has backends for the CPU and for NVidia GPUs. After the talk, the audience should be able to understand for which use cases Numba is adequate, what level of performance to expect, and have a general notion of its inner working. A bit of familiarity with scientific computing and/or Numpy is recommended for optimal understanding, but the talk should otherwise be accessible to the average Python programmer. It should also be of interest to people who are curious about attempts at high-performance Python.
|
10.5446/20080 (DOI)
|
Hello and welcome everybody. Thanks for making it through this day to the last round of talks. My name is Andriy. I work at the company called Celera One. I would estimate my Python experience around five years so far, but this is the first time I'm giving this talk. So I also hope for some feedback after you, from you afterwards in the end. And yeah, let's start. So first of all, I will give some words about the company, who we are, what we are doing, then introduce the architecture of our platform in a course-growing level, give some information on how we use Python and Pyramid in general and our software, describe in detail our analytics subsystem, and we finish with some overview on a general development process in our company. So first of all, who we are, company is called Celera One. For short, we are calling ourselves C1. The company is relatively young. It was established in 2011. It's based in Berlin. The company is quite small, I would say, for now. We are around 25 people, but it's already quite international because we're coming from nine different countries. And the main product of the company is the platform for doing the paid content, content recommendation, real-time decisions on content access for users, and of course analytics. We are also developing our own programming language. It's called COPL, as you might correctly guess, that stands for Celera One Programming Language. It's a functional language and strongly typed. The main customers of our company are big media and publishing companies in Europe. So try to represent the infrastructure of our software in a layered level. It's somewhat hard because in reality they are quite often interconnected, but this is how it looks. And we go from bottom to top. So first of all, the first and maybe the hard layer of our system, it's in memory what we call Engine. And Engine is a custom solution implemented in C++. This is a no SQL in memory database. It's a bit special. It's not only a robust storage. It also provides some business logic. So engines are usually coming in pair, where one is the master and the second is the replica, and they're connecting to each other using zero MQ. And this is the point, actually, where all the real-time real-timeness of our system happens. And it stores data in the form of events and streams and indices. And the typical use case would be the real-time user segmentation. For example, when request comes in, we can define already to which user group user belongs. And this is not trivial task because it usually, each user action brings him actually to the different groups. So basically each action can take him to different groups. And Engine is quite fast in this case. We can compute the user group membership just within a couple of milliseconds and provide the result. The next layer would be analytics system. This is a scheduling application written in Django and a set of workers. Django was chosen, of course, for its admin panel. And workers, what they actually do, they connect to the engines, collect systematically some analytics, metrics, and statistics data, and store it for later usage by the upper level. And the upper level is where we actually use Pyramid finally. This is the level of RESTful API. And this is somewhat a gluing layer because it's used for integration of all third-party customer systems into our platform. So basically it exposes API, which are then used by the customer systems like SAP, and so on and so forth, to interact with our system. Its Pyramid application, served by USGI, could be served as one big melodic application or running in several USGI processes. And the topmost layer, it's what we call communication and proxy. It's implemented in open REST framework. Basically it's a bundle of Engine X and Lua code. We also wrote our own extensions in Lua. And because it's super robust, super fast. Well, let's be honest, the Python sometimes could be slow. And open REST was super fast. So part of API is also implemented in this layer. For example, the endpoints for event collections. This is the most frequently triggered API where we get like 10,000 requests per second, for example. Yeah, it's implemented in this part. Also it's together with the Engine is responsible for making these real-time decisions, for example, on content access. And it's also responsible for requests forwarding to different sub-applications if they are running in separate USGI processes. So before installing our software on customer side, we are usually doing some assessment and somewhat we sometimes face challenges. And the biggest challenge is, for example, at our biggest customer, we face that we need to serve at least like 10,000 requests per second. For this, we kind of tweak our system. And yeah, depending on the customer workload, which is expected in SS, the setup can come in different ways. The most typical ways is when we have two front ends and two back ends, back ends, I mean engines. The biggest cluster so far, it's up to five front ends, which are running our Python applications, also serving JavaScript and open REST applications. And the back end could contain up to nine engine pairs, so totally 18 machines, 64 gigabytes from each. And the data would be, some part of the data is sharded all over the cluster, some part of the data is copied for availability reasons. And this makes us store billions of events in memory, providing super-fast access to this data, and it's giving us possibility to serve around 10,000 requests per second. We also use three, sorry, two Mongo replica sets. The first one will be used as a storage for the application data for Python applications. And the second one is the persistence layer used by the engine internally. So the logic is that engine keeps data for sliding window 30 days and then starts to back up this data in the persistence layer for availability reasons. So how does the Python software stack look like? So first of all, this is USGI as a web server, running in usually emperor mode, then the Pyramid as a web application server. Then we are using some plugins together with Pyramid, notably this is a colander and corny. Here is the library used for data serialization, we are using JSON, but it is also suitable for parsing, for example, XML. Some basic validation of the incoming data could be also used implemented in colander. And then the corny is a plugin from Mozilla. It actually simplifies our developer's life to implement the restful services. It's also quite useful because it's integrated with things and helps to generate documentation. Then we wrote a couple of wrappers on top of request library because we are interacting with the engine over HTTP and we have just some classes which wrap up requests to interact with our engine. Then Pyramid itself is built on top of the ZOB component architecture and we are also reusing these components in our code to implement so-called team plate points. I will talk about this in a moment. Then the build system is build out. As I also mentioned already we are using Django for the web application of our workers management and the robot framework is used for testing. So hopefully this is readable. This is an example of some Hull world application which is using Pyramid, corny, and colander. I'm going to explain step by step what it's on this slide. So first of all, we are defining the data schemas. They describe the parameters that handlers would later expect. Would it be the query string parameters or request path parameters or incoming data payload parameters, they could be parsed and treated as a specified type. So the first schema would be used in the get handler. It specifies one parameter called username. It should search for this parameter in the query string and treat it as a string parameter. We are saying that this parameter, if it's missing, it could be discarded from payload. That means that after trying to access this parameter in our handler, well, it would be missing. We should keep in mind this. And yeah, going to second schema which is used in a post handler, we are describing some basic JSON structure consisting of three fields where each of those fields should be also treated like a string. We are saying that they should be found in the request body. And at this point, we can already use some basic validation. For example, we say that the field message should be from five to 20 characters length. And the full field should be one of the valid values bar or bus. Providing this, Cornish, quite good interacts with colander. Providing this information during this data deserialization, these basic validators will be already checked and the error message is generated and propagated to the client, to the requester. So you don't actually need to treat these special cases in your handlers. The Cornish plugin would do this automatically for you. Then for some more custom validation, for example, you need some dependencies between the fields in the incoming payloads. You can do custom callable validators and pass them later to the Cornish. I'll talk about this in a moment. Then we define, finally, we define our rest service. It would be called hello service and available at the path hello. Then we decorate our handlers, get in post handler respectfully with the created service and we pass schemas and if we have custom callable validators, we also pass them. So at this point, we defined handlers for get in post. If a requester calls, for example, put, again, the Cornish handles this by its own and the error message would be generated like 405 method not allowed. So I would say that this quite simplifies your life and especially if you keep in mind that if you're doing the parameter application for each of the handlers, you need to, during application config, include its path. You need to add this line for each of the handlers that you plan to write. Instead of doing this, you only need to include Cornish during application boot time and just define services like shown. I think it's much simpler. Okay. And this would be an example of robot framework test. So we defined the two end points for get in post. And now, yeah, robot framework. It's a keyword-based test suite for mostly integration testing because as I mentioned earlier, our business logic is somewhat split between Python application and the engine itself. That's why we are mostly doing integration tests. So user can combine their own keywords to implement more complex keywords and stuff. So our tests would boot up the MongoDBs, the USJ applications, the engines on the background and this exact test case would then test if our response to the put method is what it is as expected. So just showing you that running this test passes. At this point, the engines are started locally on my machine. Then the test is executed and it has passed. Then it's generating a nicely looking report where you can see the logs which, what happened during your test if there were any failures, in our case, everything is great. So we are happy to go. Okay. Let's continue. Yeah. Also we're implementing our application in a way that we distribute the logic of application into different submodules so that we are making sure that different features that could be put on output to the customer. So we have, for example, the SSO module, the SAP integration module, the analytics subsystem depending on the customer demands, where they are developing and shipping these modules to the customer. So they could be all served as one monolith application or each of them is running in the emperor mode and served separately in a separate USJ application. Also one of the challenges is how to keep, because customer base is quite big. We have around eight customers, some of upcoming customers. So we need to keep our code similar, but also we need to provide custom solutions for our customers because the demands could be different, their systems could be different, and the best example maybe is the SAP. It's quite inflexible. It's quite slow sometimes, and for those we need to, for example, sometimes develop some custom code. And this custom code would be then placed in a separate package, and we are trying to keep our generic code base as generic as possible. And for this, we are implementing so-called template methods in our code, template points, and then the custom hooks which implement this custom logic would be overwriting the generic behavior in the runtime. And thus we are able to deliver the custom solution to our customers. The example of such case, as I mentioned earlier, is a SAP integration, and this is an example of real existing API for importing SAP catalog. So we are reusing the Zope interface in this case. We are defining the interface called Custo catalog transformer. So the whole idea that it has to transform method which would take the catalog in the format of which customer defines and do some transformation and transform it into the internally acceptable format. And yeah, store it. Then we have a generic implementation which actually is doing nothing. Well, it's called default transformer. It lives in a generic code base. It's doing nothing. It just assumes that the incoming payload is already in the internally acceptable format. Then it's during the application boot time, it's registered by calling register utility. And in the meantime, in the customer-specific code base, in the custom package, we are defining the catalog transformer called sophisticated transform, which actually does some magical transformation, whatever, and brings the catalog payload into internally acceptable format. And then in the customer code, this would override by register and this utility, also during runtime, and then include in this custom component into the generic code base, the behavior would be already tailored to the customer. This brings a benefit that the API's handlers, they all stay the same. So they don't change and you don't have to split your API between different packages. They all still live in the generic code base, but it still gives you a possibility to implement custom solutions for your customer needs. And yeah, time to speak about our analytics subsystem. So essentially, schematically, it would look like this. We've got engine pairs, the data which we are going to collect, the analytics data, it's started between the engines, so we need to query each single of those, then merge this data and store it for later usage. So workers, they connect to the engines periodically, query the data, do this preaggregation and cache it for later usage in the MongoDB, then the metrics API, we also call it analytics API, this is the Pyramid application which would then later read this data and according to the incoming request from our single page JavaScript application, which then should later read the analytics data, it would filter this data additionally using Mongo aggregation framework, produce the result and based on this data, then nice graphs and charts would be drawn. And as I talked already, we are using Django, this is a scheduling application, so it manages the errors, it's possible to see if there are any failing tasks, if the task should be restarted, how does the whole execution process looks like? Yeah. So at this point, I was trying to do some showcase of our analytics, so basically this is our demo system and the graph stands for page impressions. And it's possible to see the time span, for example, of one week using the data, different time resolutions. Time resolution usually means how real time the metric is. So this view shows currently a time span of one week with the resolution of five minutes, then we can see the resolution of one hour and even more cause-grain resolution of one day. So the time span stays the same, but the total numbers represent different time resolutions. Okay. No, I'm sorry. Good. This is how our Django admin panel looks like. Here is an overview on the completed tasks, failed tasks. You can disable metrics collection for time if there is any deployment happening or stuff like this. And on the right side, this is the configuration of the metric job itself. So on the left bar, you can see this would be the time resolution for which we collect the data. This column means how real time the metrics collection should be. Okay. And for the end, this is the last slide. Given overview on the typical development process in our company, so there is a developer. He makes his changes, commits them to the code review tool that we are using, Gerrit. Then the code gets reviewed after some time. The changes are merged to the Git and Jenkins, keeps overview on the Git repositories. And after the code is merged, it starts all the different tests. When your tests are green, we are always trying to keep that our master branch is ready for being bump versioned and released. So if the tests are green and all okay, you can then bump the version of your package. Then it gets packaged into an egg, put into internally hosted Python X server. Then the documentation would be built. And yeah, it would be ready for release. And when the release time comes, we have a wrapper package. So all the versions would be included by the build out, both from internally hosted X server and also from PyP that would be combined into the DEP or RPM package, depending on the customer operating system. And then OBS guys are doing their magic, putting our installation on the servers. For availability reasons, we are usually doing it in the halfway. So first we upgrade one half of the cluster. And then the second one. So this brings virtually no downtime. And it's not visible to the end users who are using these systems. Okay. So thank you for your attention. Thank you for coming to this talk today. And questions. So I understand you use both Django and Pyramids. Can you clarify what exactly does Django and what exactly Pyramids? So maybe you can share some experiences. What is better for which use case? What are the strong sides, weak sides? I have quite some experience with Django. I think it's really nice framework. But mostly I think that everybody loves it because it has its magical built-in admin panel. And Django is only internally used for us. It's not visible to anyone. It's just for us to oversee how our workers are doing. If there are any failed tasks that we need to restart, if there are any problems, so it's only like an internal tool. And the Pyramid, it's more flexible. It's used to implement the restful API as I described and showed the examples. And this is what actually is visible to our customer systems. So that if they have some legacy SSO systems and they want to connect to us, they would be using our Pyramid API. Thank you for the talk, Atrey. That was one of the questions I wanted to ask. So thank you. But I have another one. What is your development effort now at the moment? Is it on the analytics part or is it on the data scaling deployment on the largest scale? If you have any more customers, would it be easy to do? Sorry, can you repeat the question once again? So the development effort you have at the moment, is it on scaling the existing system or is it coming with new analytics, new algorithms? So I would say that, well, we have two teams. One is C++ team which develops the engine. And I would say that most of the computation efforts are there in that part. And in the Python team, we are mostly working on bringing the different metrics data. So we need to do different aggregations to optimize it. Usually, this is actually the layer where we consume probably the most of the memory. So it's quite memory intense always. We are trying to use different techniques. For now, the Mongo aggregation is doing fine. But the workload is distributed somehow between bringing new features which are demanded by the customer and implementing more different kinds of analytics and views, like the charts which would be shown to the customer. Because those are then which are used by the business analysts. And based on this data, they are doing some decisions which can impact the income and stuff like this. More questions? Right, thank you again.
|
Andrii Chaichenko - Building a RESTful real-time analytics system with Pyramid CeleraOne tries to bring its vision to Big Data by developing a unique platform for real-time Big Data processing. The platform is capable of personalizing multi-channel user flows, right-in time targeting and analytics while seamlessly scaling to billions of page impression. It is currently tailored to the needs of content providers, but of course not limited to. - The platform’s architecture is based on four main layers: - Proxy/Distribution -- OpenResty/LUA for dynamic request forwarding - RESTful API -- several Python applications written using Pyramid web framework running under uWSGI server, which serve as an integration point for third party systems; - Analytics -- Python API for Big Data querying and distributed workers performing heavy data collection. - In-memory Engine -- CeleraOne’s NoSql database which provides both data storage and fast business logic. In the talk I would like to give insights on how we use Python in the architecture, which tools and technologies were chosen, and share experiences deploying and running the system in production.
|
10.5446/20078 (DOI)
|
So, I would assume that we hear all of Python or maybe your company offered to pay for the trip and you couldn't say no. We love Python because of its expression power and nice learning curve and its coding style. The fact that many language features look like native languages, the general straightforwardness of Python and the lack of hidden tricks makes it appealing both as the first language for people who just started programming and as a new language for experienced developers. Practice shows that you can quickly start doing useful things in Python if you are already an experienced programmer. I used to program for Microsoft.net and C-Sharp for five years before I got into my first Python project. I didn't really know that much about Python. I didn't have any formal training in Python or read any book in Python. My approach was purely practical. I was only interested in getting my task done and not really getting too deep into theories. That was three years ago. Since that, I learned a lot about Python. But what I noticed is that many tutorials, books, blog posts that feature advanced concepts, they usually use really impractical examples. Like for example, how to generate Fibonacci numbers. Who really generates Fibonacci numbers in production? That's why I decided to make this talk and provide some practical examples on usage of advanced language constructs in Python. So I'm not really here today to teach you but more to provide interesting ideas, examples, to inspire you to practically use those constructs. So quick check. Please raise your hands. Those who use yield keyword in production code within last year. Wow. I'm surprised. A lot of people. Good. I'm going to introduce himself as a decorator in last year in production. Okay. Great. Who wrote context manager within last year? Okay. So it's good. You're more familiar with the theory so I can not really focus too much on it and jump to examples. So for my examples in this talk, I used code from my project that I worked on and I collected the code from different open source libraries. So what is iterable in Python? All the definitions I took from official glossary. So well, it used a pretty vague abstract definitions. So what it practically is a terrible. It's a terrible object that has an iterable method. Where can it be used in Python? It can be used in for statement in the loop. It can be used in list comprehensions or in generator expressions. Or you can pass it to functions that expect iterable like all any of some filter or whatever. And I would like to draw your attention to turtles for those who don't know. It provides a lot of very useful functionality to work with iterables and iterators. So I said that iterable returns iterator. So what's iterator? It's a writer. Basically the whole functionality of iterator is to keep the current state of iteration like for example if you were iterating over list is to remember the current element and that's being returned. And basically only thing it has to do it has to provide a next method which basically produces a new value on every iteration or raises stop iteration exception if the iterator exhausts it. Well, how it can be created? You can create a class implement method next and instantiate it. I would like to point out this completely impractical way and I won't be really talking about it. Then generator expression. Generator expression looks very much like list comprehensions except that it uses around brackets. It's an important difference. Basically in list comprehensions you immediately force the creation of the whole list and production of the whole values. So it takes memory and it's not very actually useful and well. So my strong opinion is that for most of the time you should prefer generator expressions over list comprehensions because it saves memory and sometimes you don't even need the whole list of the values. Then generator functions. So generators. Generators is a function that has a yield keyword in it. Well, many of you raise hands so I won't go too deep. So this is the most practical way to create a complex iterators in Python. It was introduced in the same version of Python as iterators because yeah, author of Python immediately understood that creating classes for iteration is not very convenient. You can read the discussion about it in PEP 255. It grew into something way larger into coroutines. This is outside of the scope. So quick glance how it works. So if function has a yield keyword in it, it's marked as a generator. When you call it, it doesn't actually execute a function. It just creates iterator object. When you iterate over it, the control flow will actually step into function and go to the first yield, then return it to the loop outside, and then stop until next iteration. So let's quickly run through it. So this how it will be executed. It will print yield. It will return. But on next iteration, it will start from the next line and go until next yield or end the function. So this will print the zero to. So what's the difference between the regular function is that it remembers its states between the calls. So what can it be practically used for? Because many people, when they study the subject, they don't really know, well, okay, this seems cool, but what can I do? One of the very common usage of generators is basically to create a view over a collection, do some filtering and mapping. So this is the code from a project I worked on. We had some list of fields, some of them were considered dynamics, started with some prefix, and basically we needed to get a list of the dynamic field names in different places. So we wrote this simple generator. What's good about it? First, you don't need to copy paste this for and if everywhere you need the list. Second, it looks clean. It's easy to read. Well, it's multi-line list comprehension, so generator expressions are, well, ugly. So this is the, well, this one's some filtering and then we want some mapping. So another very common way to write generators is to flatten lists. So this code comes from Django framework. So flattening lists, if you have a list of lists and you want to iterate over all sub-elements as if it was a single stream of values. So it's basically usually done with nested force, and so they created this iterator so they don't need to copy paste this to force everywhere. It's actually a good and bad example because iterator module provides a much easier way to flatten lists with iterables. That's how it can be done in one line. So it's a very popular question on Stack or Flow, so I would assume that, well, yeah, people have difficulties with it. But there can be more complicated flattening list logic. This code comes from Django, or from JINJA. It's a templating engine often used with Plask framework. So they have a list of items, and item can't be a node or a list. So if it's not, we return it, we yield it. If it's a list, we go into sub-loop. So this is, here you can't use really iter tools. So this is an example of more complex flattening. So what it can be also good for generators is to save memory. So this code comes from request library, is so-called HTTP library for humans, whatever it means. But it's very convenient library to do HTTP requests. So somewhere in the depths of it, they have a socket object or I know some sort of wrapper, self.roll, that can, that, well, it's a network stream. It's a sequence of bytes. But responses from servers can be fairly large. We don't really want to load them in memory completely. So they wrote the iterator that breaks it into chunks, and you can iterate over those chunks without actually loading the whole content of the response into memory. That's very convenient because it saves memory. And it's convenient basically because users of this function may not really need a whole response in memory. It can be used to directly write those chunks into file. So you can save it locally, and it's good for memory and performance. So it's a little bit more complicated. So this is their internal usage of this chunks iterator. So what's they doing here? So it's iterator that takes iterator of chunks and produces iterator of strings. So chunks have fixed lengths, but strings can be arbitrary lengths, and they're divided by delimiter. So what they do here, they break a chunk into strings. But it can happen that the remainder of chunk is incomplete line, and we need to get another chunk to complete it. So here they introduce a state, a local variable, pending, which carries over, a leftover of chunk that is uncompleted string. They keep it between yields, and they add it to chunk so that they can, they remember it between calls so that they can connect it with the rest of the string and return it. So if there are no more chunks, then they assume that pending is complete string and return it. It's a convenient example how also you can save memory and create an iterator over iterator with the generators. So generators are also convenient to traverse complex data structures. So this is example of first code of standard OS module, the function walk. I simplified it significantly. Well, it's like 200 lines of code, but this is the core of it. So it uses recursive generators. So it's very convenient in this case because file system is tree data structure. So walk returns a tuple of current path, the list of deers in this path and the list of files. Then it goes over the list of directories in current path and calls itself basically and yields the results. So in the end, if we iterate over walk generator, we will get a flat list of hierarchical data. That's very convenient. Also iterators can be infinite. So some people may wondering why would you need an infinite iterator? You can't even iterate over it. You can't do four over it because it will eat all your memory and burn CPU. This is example. This is Django templating language for those who are unfamiliar. It just basically copies the four tag, copies its contents. It's here, it generates a rows in a table. So we would like to often, you know, in web, we have tables. The rows come in different colors, one after another. So Django has this cycle tag which actually will produce on every call row one, row two, row one, row two. Internally it's implemented with the iter tools cycle iterator, which is infinite iterator in the lower part. In the lower part you can see, well, it basically takes it terrible and then repeats it one by one, items one by one forever. So every time this tag is called, it calls render method and it just calls next item on the iterator. So it's not a problem that the iterator here is infinite. It will never try to iterate over it completely. It will call it exactly as many times as four iterates. So this is very nice and clean code. You don't need to maintain any state like which value was returned previously. You just call next on cycle. So again, I advertise you iter tools module. So those are the most practical examples for using. But I recommend you to research deeper because iterators, generators is very complex topic in Python. It has way more advanced usage. So I recommend you to look more into iter tools, iter tools module and master it because it can improve your productivity. Then actually yield is not the statement. It is expression. So it can be used as a mechanism to pass a values from color to generator. This can later led Python to even more advanced stuff. So then I recommend you to read about yield from its so-called generation delegations. There those two features are used to implement coroutines in Python, which is also very interesting. And based on coroutines, there is a new, relatively new module, the whole concept in Python called Async or X project. It's for asynchronous input output. They are heavily based on coroutines, generators and iterators. Actually don't need to photo the slides. We'll put them online. So yeah, you can, don't worry if you miss something, you can check it out later. So another advanced feature that's often not used in practice by, well, maybe starter developers, it's decorators. So again, many of you raise hands. So I would assume that many of you know how decorator works. Basically decorator wraps your function with some inner function. What it basically all it does, it assigns to your function the decorated, the wrapped version of your function. What is it good for practically? With a decorator, you can modify input arguments. You can modify return value. Well you can do things before functions called, after functions called. You can actually not call the function at all. So your decorator may decide not to call function for some reason. You can modify some global states, some outer variable, some thread locals and stuff like this for this function and then set them back to the values that were before that. It can use, decorators can be useful to assign metadata to functions of all sorts. So this is example from Flask web framework. They heavily use decorators. So they demonstrate two things. Here it is, hello is a web view. It's basically a function that is called when you call your web server slash hello or slash hello name. So they do here, they use decorators to parse URL into function arguments. And another usage of decorator here is that it basically makes it possible to discover the function. So the framework knows which functions are web views. So in this sense it provides a metadata for your application. This is the small snippet from my current project. So we have this. We want to make extendable list of, well, some filters, whatever that means. And we would like to show them on UI so users can select from a list. So we need some sort of label for those fields. So I created this decorator that provides a human readable description that is used on UI. Also, I can query all the functions in the modules and check if they have this decorator. So it's also good for discovery and for providing metadata that can be used, for example, in UI later. So color or not. So with a decorator you can make a decorator that decides maybe actually we shouldn't call the underlying function at all. This is again simplified example from Django. They have a decorator called permission required. So you can apply this decorator on a web view function and check if the currently logged user has certain permission to call this web view. So what we can see here, they check if the current user has this permission. If user has this permission, they call the actual function. If user doesn't have permission, they don't even call the function. So that's how you can check whether you should actually call function or not or raise error or whatever. This is decorator rate limit. We created it in one of my projects. Basically it's also for web views. It counts in a cache how often the page is called by a certain IP address. And for example, if it's called more than 10 times in a minute, we will not generate the response. We'll show an error. It's very simple mechanism from prevent of abusing of service. So if the user calls it too often, we can prevent call completely. Caching is very big feature that can be done with decorators. So this is in our project, we created really simple and I have to confess a little bit stupid caching decorator that we apply on properties. Exclusively, that's important point. So we have this numpersons and it counts, it's a Django ORM. So basically just counts related objects. So this function is called, this property is called a lot in reports. So we have reports that call this function 20 times. And we can just drop this decorator cache result here and it can very conveniently prevent too many of the calls. So how we implemented it? It's also very straightforward. So you can see here we get the func name. That will be numpersons in this case. So if the object has this attribute, we do get attribute. If it's not in the object, we actually call the property, calculate it and then set corresponding attribute and the subsequent calls to this property will return the cached value. So what's bad about it? It's basically that if save underlying object or data changes, well, there is no way to enforce recalculation. But this is good for simple cases. Like for example, if you're sure that your object won't change. So the proper way to cache with decorators. There is this awesome, I really like this, library called doc.pl. It actually creates proper and very sophisticated way to cache functions and method everything. So the key decorator is cache arguments. You can provide different backends, memcache, local memory, anything. And they will actually cache the functions for corresponding arguments as well. So the cache key will include the function arguments. It's very smart. You can refresh values and everything. I suggest everyone to look into this because that's probably the most sophisticated caching library with the decorators I've seen so far. My personal award for most creative user of decorators goes to Ansible. Ansible is a very powerful tool for IT automation. You can use it for deploying your code and servers, manage your infrastructure. I also recommend you to look into it. It's very powerful. But inside of it, they have a very interesting decorator called timeout. So you can just apply to function. So it does, if the function takes longer than, well, by default, 10 seconds to execute, they stop the function execution and raise an error. How they implemented it? So they implemented it using unique signals. So before the function is called, they set alarm with a callback. Then they set the alarm to seconds. So what happens? How does it work? This kernel feature of Linux, well, any other POSIX compatible OS, it will set this alarm internally and when this alarm goes off, it will call the handler method. And handler method will rise timeout error and this timeout error will appear within a function and it will exit the function immediately. So if it takes longer, it just stops the function and exits. So this is a pretty nice way to do a function by timeout without using threads. Obviously, this will not work on Windows. Sorry. This is mostly for unique operating systems. Digging deeper. So decorators can be way more complex than examples that I showed. There was a previous talk today, unfortunately I missed it, but they covered, I hope maybe someone of you were there, in the program, they wrote that they will cover advanced cases of decorators. So in decorators, you can decorate classes and decorators itself can be class that maintains the state. So you can create way more complex things. But I would say in for more practical cases, the simple decorators are enough. But I recommend you to research more into this. So context managers. Context manager, this is how we call with statement. It is just a very simple thing. It's basically it calls enter on a context manager object and returns a value and signs to variable. Then it calls the actual code that is within the statement and finally it calls exit. So what it can be good for, it's good for if you want deterministic release of unmanaged resources, for example, files, everyone must use with files in Python. Because it will close the file immediately after you stop using it. It's also good to modify some global states. For example, you can select a piece of code that will be executed within transactions or you can set some setting just for this piece of code and then restore the settings. And it can be used a lot of interesting tricks for logging, for more advanced logging and debugging. So a couple of examples, Django. So they have session transaction. So on exit, on exit, basically they check if the current scope was within a transaction. If yes, then we need to commit the transaction or, well, on error, roll back. So that will put it in a transaction. The important here is that finally it's called always. So whatever there is error in transactions or not, it will call the exit method always. So for example, in request library, they have a session, context manager, when you exit it, it closes the session. It's also good because you should close it as fast as you stop using it and not when garbage collector picks it and calls it. This is a small context manager that I wrote for debugging. Basically when it enters, it stores internally the time when you entered into a function, that function of the piece of code and when you exit, it just prints the amount of milliseconds it took to execute your piece of code. This is very simple and very easy to do micro-profiling small chunks of codes without doing too much calling profiler and all this stuff. So digging deeper. Again, I of course didn't cover all the uses of context managers. So Django does a lot of stuff with context managers, especially when it comes to transactions. They also have, it's both decorator and context manager because it wraps some piece of code. So also you can implement with decorators and context managers, you can implement database locality. So you can select, for example, let's say that this function or this piece of code, depending if it's decorator or context manager, all the queries will be executed against slave database. It's, for example, useful for reporting. You have a function generating report. You can create a context manager saying, okay, use slave database for all queries in this. Also more sophisticated log and debug, not just stupid milliseconds, which is actually very, has very low precision. And yeah, you can do some more sophisticated profiling. So that's it. Thank you very much. This is my email. You can send me suggestions, questions, whatever. I registered that stupid domain for my personal blog. I'm going to put the slides and there, and I started a series of posts, which basically I cover the same stuff, but in more details with better explanation. So if you miss something, you can come there and yeah, all the content from the slides will be there within the blog. So thank you very much. Do we have time for questions? Okay. One question, did you already see this LRU cache module in the standard library? That's kind of doing the same thing as one of your examples. It just caches the result code, the result value for specific parameters sets. So maybe you can also use that. One question, did you already discover a good pattern to handle open files? Sorry, it's covered what? To handle open files, if the context is not as local as in sometimes with, so you basically need to give around open file handle and still you need to make sure that it is closed at some time. Is there a good pattern for that? Well, yeah, well, with the context, well, answering the first question, yes, there is a lot of this caching, there are a lot of caching done with the decays in very different libraries, but I recommend to check out DocPile because as I said, it's the most advanced one. It has the richest functionality. Coming to the next one, well, if you want to use file in different place, well, then it's a little bit more complicated than, yeah, you can use it. Well, context managers are good for small chunks of code. So if you have few lines, but yeah, you can write your own class that will store a file handler and once this, and you can call close on your own class when you finish doing with this. So as far as I know, there is no straightforward thing to do with the context managers. Hi, hi, is the order of the decorator important? If I have, for instance, three decorators, sorry, is it going to grab the following decorator? Yeah, this is, you're right, this is a tricky part. Yes, the order of decorators affects it. So some decorators, for some decorators, it's important. The order is important in which you do it, but for some it's not. So yeah, order matters in general case and more specific. So you can write the decorators that, for example, I will give you a very small example specific where was my wonderful cache. So here it's important a property should come before a cache result. Otherwise if you do it the other way, it won't work, it just won't work properly. But because the property is a special, it's a building decorator. But for simple decorator, you can write the logic in a way that it wouldn't matter. So it really depends on, yeah, particularly the decorators. But evaluate the full amount, right? I think it actually, no, it actually, first it will call cache result decorator and only then it will call property. So that's why it's important that property will be like outermost. Yeah, yeah. This is a tricky part here. Any other questions? Comments? We still have a few minutes. Hello, thank you for the talk. Just one observation about the wrap decorator that the standard library offers. So I saw that in one example that's present and in the others when you are writing a wrapper function you are not using that. So I think it's worth mentioning that it is better to use because it copies the doc string and functions name and those things. Yeah, you're absolutely right. You're absolutely right. And practically you should use wraps from Funcutils module. That's absolutely right because it will keep the function name and doc string and all this stuff. But it was just like too much for the talk. But you're right. Thanks for the note. Can we push for one more? Questions, comments? Or if not, let's thank the speakers again. Thank you. Thank you.
|
Andrey Syschikov - Practical usage of advanced Python constructs Python is a language of choice for developers with wide range of experience, for some it is a first programming language, others switch to Python after years of experience. Python provides friendly syntax and smooth learning curve. This sometimes leads to developers lacking comprehension of some more advanced constructs. It happens that experienced developers jump into using Python and sometimes miss less known Python language constructs. On the other hands people who purposefully learned Python sometimes lack practical ideas for how to apply those constructs. This talk will be specifically focused on the practical usages of advanced Python constructs like iterators, generators, decorators and context managers. Goal of the talk is to share ideas about how those constructs can be used for practical purposes in real projects. Prior knowledge is not required, there will be a brief introduction to every construct being presented.
|
10.5446/20077 (DOI)
|
Hello, my name is Andreas Klostermann and in this talk I will talk about brain waves and how we as Python hackers can use our tools to explore our own brains a bit. Last year at EuroPython I held the 1.0 version of this talk. This year I have only 30 minutes so I'm going to make extremely short thrift of the theory behind brain waves but we have to cover a bit of it so that everyone of you understands what we are doing here but the talks don't have a lot of overlap so this is really 2.0. Now first of all what are brain waves actually? Brain waves are an electro physiological potential that you measure on your scalp that basically is a voltage measurement which is measured over time with electrodes that are on the scalp and what you are measuring there is the summed potential of most or even all your brain cells which give off electrical signals to communicate with each other but also a bit of waste energy gets transmitted outwards and so you can measure it on the skin without having to go inside the scalp. Most of what the most useful analysis of brain waves usually is the Fourier transform which basically only means that we assume that the signal is composed of several frequencies and we want to know which frequency is dominant and by doing that we can infer something about what the brain is doing because we know when a certain frequency range is overrepresented and the subject is for example concentrated or relaxed or that kind of information. Now to measure this current we need an electro and a phallogram and these devices can be quite cumbersome and quite expensive but what I have here is basically the Neurosky Mindwave Mobile which is a Bluetooth connected headset and is really low powered and very optimized for developers and hackers. It does all the amplification stuff inside this headset and digitalizes the data. It also does some preliminary analysis of the data so that you could even connect this to an Arduino and the Arduino would be able to tell if you are concentrated or not. Aruinos are not really sophisticated processors or software in any case so it's quite nice. I wrote a Pyzeology framework and that is a framework for physiological experiments and analysis. It currently mostly only supports Mindwave EEG but eventually I also have code for ECG from Vitalino but that's not activated currently. It's powered by Psypydita and it's iPyzen Jupiter enabled so what I want to achieve is an experimentation platform that works for do-it-yourself people who don't have a lab fancy equipment or something but I want to use mainly these quantified self-type devices with Bluetooth connections or other near field stuff. It's currently Python 3.4 only and I'm making heavy use of Async I.O. Now you've probably heard of the Internet of Things. The basic idea is that you have local computers connected to some kind of sensor and there is a pattern that I see emerging in these data acquisition applications so you have a local component where it very much matters whether the computer is like a Raspberry Pi or an Arduino or something because it has to be right next to the device because of the low range of the sensor. Then there's a cloud component which actually analyzes the data and that does not matter where it is just that it is connected over a network and the user interface is connected also to cloud and local and displays the data and user can interact with the analysis. In my case, the local component is just an Async I.O. data server that I've written. It communicates over Bluetooth with this thing here and it also babysits the connection so if I switch it off and switch it on again then it will reconnect and try to fix errors. That's quite nice because I have that separated from the iPython kernel which is the cloud component because it's difficult for the user to deal with all the exceptions that occur in Bluetooth and we have most of the data processing inside the kernel then. The kernel pushes data to the user interface which is a browser. In an iPython notebook like this one you have the server, the kernel and the user interface and they are somewhat separated but you need a server for Python anyway. Now I'm going to quickly show you a real-time demonstration. This is junk data because nothing was connected and now if we are lucky I will be silent for a moment so that you can appreciate normal brainwaves. The problem with EEG also is that the brainwaves themselves are so, so weak that pretty much everything going on inside or in your head, all the muscles, the facial muscles and eye muscles all contribute artifacts to the data and are stronger than the actual brainwave data. I can show you that I can clench my teeth which are like the strongest muscle in the head. I can also use my eyes to... Yeah, and when I blink. And that makes EEG signals very difficult to analyze. Also this is a bokeh graph. I can use iPython widgets to push data from the kernel to the notebook and also the kernel is running an I think IO loop internally so that it can communicate over the web socket with my data server which handles the Bluetooth stuff. Now I have stopped and then I can go to the next slide. Now let's do some data science. Data scientist is who data science does so we need some data for that and to get good data we need experiments because discovery requires experimentation. And this is a central idea in the physiology framework that is this experiment stuff. The experiment is here initialized and then immediately used as a decorator much like the iPython notebook interact one. And this decorator when I start it, it will wait for data from the server and whenever there is new data this handle message is called and it can then do its thing to do whatever. In this case it just display clear output means the output is cleared and then some HTML is written with the last attention data. So attention is a value that is computed by the newest guy mind-face itself inside the headset. And I can show you how that works and that is really all it takes. So this is a value between zero and 100 currently it seems I am totally inattentive but I can push that higher. So in terms of time it doesn't work. Now experiment classes have these attributes like attention as a time series. So every second or so the such a value is computed by the device in send over Bluetooth and I try to figure out when that was but it is a bit difficult to do that. The time series stuff is really good and so it is just accumulating this experiment. Also the raw data, the meditation data and several other things are also communicated and stored. Now that was one thing running experiments in real time but we also like to record data. That is done by giving it a file name and that is a HDFI file. In this case I think I have a bug because I have like one or two gigabytes of data and that is completely crazy. I have to fix that later. In any case this is the same as before but it displays the amount of raw data which is here, then experiment raw over 512 because the raw data is at 512 samples per second. Now let's record some raw data. I think it is not working currently. In any case, now to resurrect the data we need a so called batch class and the batch class can resurrect the data then. Here we have recorded data from another session which just is much the same as the experiment. I forgot to tell you what actually a batch is. Every time I run the experiment I record a different batch. First time batch 0, second time batch 1 and that way experiments don't overwrite their own data and you can also try to vary the experimental conditions if you want to do that. Now I would like to show you a simple experiment. I don't have time for a lot of them so I would only do one. The new Sky Mindwave computes these ESense meditation attention values. They are patented and sort of back box algorithms where they don't really tell us what they are doing there. Here I am transforming the time series data to a table of features. I don't have time to really explain what is going on there but I take windows of the time series data then I compute how strong different brain wave frequencies are and in the same row I also have the tension and the meditation values. I wrote separate library which is table cleaner and we need that to clean some of the data up because some of the data is bad. As I told you before we have lots of artifacts in the data and we don't want to analyze artifacts. We want to analyze real EEG data. So this table cleaner library is inspired by jungle forms and cleans tabular data. The nice thing is it outputs both a validated data set and a data set of the errors. So on the bottom we see a table of the grouped validation errors and the most rows, 206 rows were deleted because there was too much variation in the EEG data. So when there is an artifact then we try to remove that row of the table. Also poor signal is computed by the brain wave itself but that's not so important in this case. So if you need data validation thing maybe you should look into it. Now I'm doing the linear regression analysis. So I said we have features of brain waves correlated with attentional meditation data and the linear regression all it does is it tries to explain the tension value or the meditation value in terms of the frequency strength. So it multiplies each frequency with the coefficient and then they are all summed up. And that way when you figure out the coefficient you know how important the different frequencies are for this value and you probably can't read the labels on the graph but I didn't have time to make that any better. What you may actually see is that we have here the best or the highest and most significant coefficients are at the... What was that? Are in the high beta or in the beta frequencies from 17 hertz upwards. I can't see that but just believe me when I tell you that there is a certain range of frequencies which is called beta frequencies between 17 and I think 30 hertz depending on the definition. And they are strongly associated with all sorts of attention and concentration stuff. So if you ever heard about the prefrontal cortex and the tension circuits of the brain also ADHD that has a lot to do with with beta brain waves. Now I'd like to meditate a bit about meditation. There's a similar graph but we have more going on. Here we have the alpha values which are between in this case are between like 8 and 10, around 8 and 10 hertz which are significant. But also some beta values which are a bit undecided. I think that's because most of the time I or others too need to do some concentration to enhance alpha values. During baking stage you need to have a somewhat calm mind to actually exhibit alpha waves. Or you close your eyes which is also easy but that's not what we want to do. In any case these values are used in newer feedback. So you can write an application that looks for attention or meditation and feed it back to you. And then you can learn how to consciously manipulate these brain states. That's what I was talking about last year at Europe, Pison and Berlin in 2014. And there's also a YouTube recording of the talk where I go into more detail about the psychology and biology behind all this newer feedback stuff. So this all sort of implies, so the linear regression implies that the strengths of the frequencies are not dependent on each other. And I don't know why that. I'm not going to get it to work. In any case these are correlated and we see that the alpha and the beta values are correlated among each other. But not between the alpha and the beta waves. It has some biological meaning but I'm not going to go into that. I'd like to make some technical remarks. Bluetooth is really bad for timing. So if you have Bluetooth data coming in, they often don't come with timestamps. And these timestamps, you have to give them timestamps as the local component in the server, data server, and figure out which are the right timestamps. And if you do it wrong, then you have like overlapping data and crossing your own timeline is strictly forbidden except for cheap tricks. Yeah, I think I was very good for waiting and problem resolution. I do like yield from weight and I yield from sleep and so I do so I can write these troubleshooting algorithms with reconnecting and I can wait for a few seconds and then reconnect again without disturbing anything else. The combination of Bokeh and notebook widget really rocks because this notebook has its own web socket push communication channel which I can use to stream data to the JavaScript site and update the book charts. And in general, Jupiter notebook is a really great tool to tell computational narratives. I rather skip that or rather went hastily through my little computational narrative about the brainwaves and about attention and meditation, but I didn't really have time for more. But you can imagine that you can do real time experiment or show how the computation works and also explain or narrate how the data really works. Also, it's a lot is the the biology framework is an example of a library for the notebook which really is bigger on the inside. So I just had this experiment decorator and entire heights all this web socket and async IO and other magic. Mostly from the user. I'd like to thank you for your attention. And my Twitter handle is page in horse which was a few years ago supposed to be a pun or Trojan horse or something but now it just sounds silly. The library is by the ology. I just published it. It's very raw and you probably won't get it to work. Especially unless you have this device by the way resistance is futile. This table cleaner library I try to make it a bit bigger because I have had a lot of trouble over the times validating CSV data. And that is sort of my attempt to fix that problem. In this case, I wrote also this notebook assets library which I use for for turning coffee script into JavaScript and serving it's from the actual library itself and not from the extension stuff which I find a bit troublesome because the notebook profile directory has no clear one on one direct connection to the Python library installation stuff. And it's basically all. Thank you for your attention and we have time for some questions maybe. But you when you want to go to the lightning talks. I tried to do some stuff with with the mine wave years ago. But I'm guessing that your framework didn't exist then because then I had to connect via some some proprietary Bluetooth executable that I had to wrap stuff up around so so does your Python code do the Bluetooth stuff. Yes. Also, the current documentation for mine wave points to my older libraries for mine wave so it's a bit difficult to do and I think what the newest guy stuff is doing isn't really that developer friendly on the library side. But the protocol is very friendly. So if I want to get in one of these nearest guy things is that I just had a brief look at the web page is the nearest guy mind wave mobile which is the one to get. Yeah. Yes I think that's the best device as one that is a bit more expensive and more variable but I think the new sky mine wave mobile is the best one for like 100 or 120 euros. You can also get it on Amazon and other shops and you only need the device itself. You don't need any of their software or anything. You just use that all. Okay. Great. Some people even wire up Arduino directly to the sink your ship. So as we as we have some time. Would you mind to show the code for the demo part. Maybe the demo itself. Yeah. By the way, this I want to publish this the notes but I wasn't able to to push it yet from this network and it contains a lot of notes of what I was talking about which are not visible in the presentation itself. So here's a hidden code with the real time raw demo and but to to know what's going on here is this normal book a let's make a plot stuff and then it's the raw source book a model source column data source which is important to have a reference to it. And then my touch bed. So in the handle message one I am doing some re sampling and then I push the data as a very convert the data to this raw source data on this Python side. But the book a knows it's this book a objects know the ID on the browser also on the Python side so I push this ID here. And then replace book a data source and it was a widget which has a communication between the Python side and the notebook side. And then I have just this function replace book a data source which pushes this this data source with its ID and the new data to the JavaScript side and then it deconstructs it and puts that into the data source into the book a data source on the browser and then magically through different callbacks. And then I have this new graph is just redrawn. Thank you. I have to reconnect again. I'm asking to show the demo. I'm sorry. I'm so excited. I work on the key to working now. What is it? What are the lines? Can you most code me? It's actually an easy abuse of this technology to do more scouting like I can also bring. Yeah.
|
Andreas Klostermann - Brainwaves for Hackers 2.0 This talk is a sequel to "Brainwaves for Hackers" and illustrates some experiments you can do with a Neurosky Mindwave headset, a bluetooth enabled EEG device. I'll also talk some more about how to integrate the device with the IPython Notebook for real time viewing and how to use the Mindwave with the Raspberry Pi.
|
10.5446/20074 (DOI)
|
Good afternoon everyone. Thanks for having me. So today on the summit we have talked about Python education and it seems to be that I'm in the other end of the spectrum here because this is university education. So my name is Anna Sleeman and this is a compact view of my 20 years as an electronics engineer. So I've had the luck and honor to have a very diverse career I think. But now I work for the University of Orhus where I am teaching bachelor students in electronic engineering. So the talk that I'm going to present to you today I'll go through these items here. Just short of what we are doing now. What kind of education we have now. And what a little bit about how the Danish university educations look. A brief introduction to the online education landscape if you could call that. And then I'll go into more about what are the challenges for education, for online education. Mainly what can we foresee as challenges and how should we address those challenges. Then I'll present what we are trying to do in our transformation of the existing education. And lastly how Python can fit into this online education of university education. So the existing education that we have today is we call it electronic design engineer. It's an accredited bachelor education. It takes three and a half years. Where there is a six months internship included. We have a very good cooperation with the local industry. So we are able to place almost all our students in companies for these six months. And it tends to be that after they have been six months at a company. Most of them will write their bachelor thesis at that company. From the tasks that these companies give them. And a large percentage of those who have solved problems for the companies. Will continue in employment in these companies. The education is located in Herling which is in the west part of Denmark. It's not a large city. But we are lucky to have a quite large number of industries and some of them are quite big. The most, the largest one is probably Siemens wind power. Which is located less than 20 miles from the education. And of course there's a lot of wind power related companies that support Vestas and Siemens. Which are the largest wind power factories or industries we have. So it's a kind of exclusive education. We only accept 40 students each year. So it's kind of a small education. So a little bit about University of Ous. So University of Ous is the largest university in Denmark. Even though Ous is only the second largest city in Denmark. But over the last five years there has been a lot of mergers in the Danish university sector. So actually Ous has swallowed up a large number of smaller campuses. Among them the campus in Herling. So in Denmark the tuition is free on the university educations. And the students will get an allowance while doing the studies. And they can apply for low interest student loans. The allowance is almost enough for the students to live off from it. It's around 800 euros a month. But they are also allowed to work. There's a limit for how much they can earn in these student jobs. Sometimes when our students go abroad and talk about their conditions as a student. People won't believe them but it's actually true. But I actually think it's a good idea. I'm happy with this system because it allows people to focus on their education. And that's a good thing. So of course we have all heard about these new guys in the education sector. And of course they are a can university, open edX and Stanford. Everyone seems to have offerings in online education. And I think that the traditional university education are slowly coming around to seeing that there's a need for it, there's a demand for online education efforts or offerings. But what we have seen in herning, what we were trying to do in herning, we're not going to emulate these offerings. We are trying to take the best out of them and then provide it in our setting. We want our education to still be accredited. We want our students to be able to obtain these allowance. So that should be the same level of quality and control in order for the education to still be accredited. That's very important for us because that's how the university gets paid. If the education loses its accreditation, we cannot be reimbursed for the work that we are doing. So it's extremely important that we don't lose that accreditation. So as we see it, the online offerings that we see is mostly focused on single subjects, single topics, a single course. Eventually, maybe a kind of certification that can be used to document your skills. The way that they are taught is through short videos explaining the topic. And then there will be some problems that the student can solve. And after each larger section, there will be some tests, either that the student will score themselves to see how they are doing, or there will be, if there are also offerings where there will be a teacher or a system that grades the problem solving skills of the students. So this is what we see that the way that the online education is going. And of course, there's a lot of good things in this. It's nice for the students to be able to study when they have the time and the drive for it, and they can do it at home and in their own pace. But we also see that there are some challenges to this approach, especially if you want to have a full-time education based on online offerings. We see a possible challenge in how can we keep the online students focused, how can we avoid that they spend too much time on everything else that you can do when you sit in front of a computer. And if we have full-time students that only doing their study through online offerings, there is a real danger that the people get isolated, and they won't feel as if they are part of a larger group or that they are all alone and stuff like that. We are actually fighting these things about isolation, even though we have a campus and we have people coming to classes. In those situations, we actually see that there are some students that for one or another reason get isolated or feel alone. So I think that's a real issue here. Also, when you are doing online learning, you are the driver for the motivation. So online learning will, to some extent, favor people who are good at motivating themselves. Of course, people who are good at motivating themselves also have an advantage in traditional education, but I don't think that the motivation part is going to be smaller in online offerings. And finally, there are also some challenges for the teaching faculty. We need to plan differently. We need to be better to anticipate questions, because if there is no direct interaction with the students, we have to prepare in a different way. So I will briefly go, this is not a talk about learning theory, but I would want to emphasize a few things in the way that we look at teaching and learning. Of course, if we talk about teaching, then the important actor is the teacher. So if we look at education as something that we have to teach, then we will focus on the teacher. And maybe that's not the most beneficial way to look at education or learning, because maybe the outcome of the teaching should be that the student learns something. Also, when we as teachers feel that our students didn't learn what we had hoped or planned, then we could very easily fall into this trap that we could just say, okay, I did everything that I should, I planned, I had all these nice problems, but I made a very good lecture, so if they didn't understand it, they must be stupid. So that's a very dangerous thing to do, I think, to begin to look at the student body and divide it into groups, the good students and the bad students, because the good students, they are going to thrive anyways. If we kind of give up on the bad students, we shouldn't call them bad students, because they don't exist. I think that we have to accept that there are no good students and bad students, there are only different ways of learning. And if we, instead of focusing on the teacher, focusing on the learning, so the outcome of our teaching, then we will focus on the student and we will begin asking questions like, how would you prefer to learn? Would you prefer to read or listen or look at videos? Do problem solving or what is the best way for you to learn this topic? Are there, sometimes, that are better than others? And are there some places where you would prefer to be learning? These questions we can't really ask today, because how is really determined by the teacher? When is determined by the classroom schedule? And where is, that's the on campus. We can't really change that in the traditional settings. We have the possibility in the online situations, we have to make, we can make more allowances for people to choose their learning environment. So as a teacher in this setting, our job will be more to provide the tools and the curriculum to allow for the students to have a learning process. So what we're going to do in the new education is that we will offer this as an online, this bachelor study as an online study, but we will kind of, we will try to mix it. So there will be both on campus students and online students and they can mix and match as they want. Some on campus students can choose to stay at home some days or a week or whatever. And some online students can choose to come to the campus to engage with the other students. So, but we want the start of the study to start with a week long boot camp, gather all the students, either, or both the online and on campus students should come to the campus so that we can set up the computers and form the teams that we all need during the study. And well, introduce them to the teachers and get a relationship going between the students internally, and also between the students and the faculty. The semester we will use this, we call, it's not our invention, but the idea of the flipped classroom, that will be the main way that we are going to teach. So we will have, the students will prepare for the sessions, which we all also will have. Again, by looking at videos at home, doing small problems, and so there will be video lectures and they have to read stuff and things like that. So they are prepared for the session, the online sessions or on campus sessions, where there will be discussion about what was the important part of the lessons for today. And there will be problem solving and stuff like that. And there's something that we haven't tried yet, we don't know if it's going to work, but we want to gather people, both online and on campus, in the same time, so that people on campus can interact with the online students via the internet. So a little more about the flipped classroom. In the top we have the traditional class, the teacher is going through the topic for the today and the students are listening and then they go home one by one and study the, primarily the text for the class. In the flipped classroom there will be a preparation part where the students are prepared for the lessons at home. And then when they come to the lesson, the teacher will be more a facilitator than a lecturer and the students will engage with the topics and with each other, so that they use the preparation in the classroom. And they can use the discussions between the students to further improve their learning. That's the idea. As I said, it's not our invention and it seems that there have been quite successful stories about it. So we are going to use Adobe Connect for the collaboration part and LiveSize for streaming. It's not very important. This is an electronics study, so we have made this box with all the stuff that the engineer will need. So there will be a breadboard, a PC oscilloscope, an embed computer, and all the tools that they need so that they can do exercises at home. And they will get this the first week and they will take that home so they can use it by themselves. The on-campus students, of course, will be able to use the laboratories, facilities that we have on the campus. So final topic, and I think I'm almost out of time. So how can we use Python in online education? Of course we can use Python for teaching programming. That's not really any news, but I think it's actually a very good way of teaching computer science also. But since this is an occasion which focuses on embedded systems, we need to supplement it with some C so that students will have that on the CVE as well. I want to use the iPython notebook as a MATLAB replacement. I think that's perfectly reasonable. But we'll see. I'm not the only teacher on my education. Okay, so this is just an observation that when we have these internships, we go and talk with the students while they are at the internship to see if they are all right. And then we actually more times than not discover that Python actually is used in the industries for different things. So I don't think it's necessarily a downside to be able to write on your CVE that you have used Python during your education. I think that's perfectly valid to have that. So I'm at the end. You can read this. I have spent all my time. Thank you. Thank you, Anders. Any questions? Hi, thanks for the talk. You mentioned students potentially becoming disengaged and isolated if they're online only students. How do you know if that's happening? Do you do any analysis of their usage? And if you do see someone who's gone like that, what do you do about it? What I didn't discuss here was how we are going to follow up throughout the semester. But we will have one-on-one Skype sessions with the online students. We do that already with the on-campus students. We talk to them more times throughout the semester to see how they work. If the curriculum is suitable for them or if the teamwork is okay and stuff like that. So we try to interview our students in order to make sure that they are okay. That's something that we value, that there should be a very close between the students and the faculty. We want to have that. And we want to be approachable as well. And that's also why we want this boot camp to actually show ourselves as teachers. But we are also human beings and we are not dangerous. Any more questions? No? Thank you, Anders. Thank you. Thank you.
|
Anders Lehmann - Online Education: challenges and opportunities for Staff and Students From september 2015 Aarhus School of Engineering will offer the education Bachelor of Electronic Engineering, as a combined online and on campus education. In the talk I will describe the technical and pedagogical setup, we are working at to meet the challenges of having both on-site and remote students. I will also touch on how IPython Notebook, will be part of the technical setup, and how it can be incorporated into the teaching.
|
10.5446/20073 (DOI)
|
Good morning. Thank you. So, I call this how to GIS in Python with a subtitle of A Tale of Two Cities. It was supposed to be about the city of my university, Ahus and Istanbul, but due to time range I have chosen only to talk about Istanbul. It's the more interesting case anyways. So, the subject is a little bit wrong, but it was kind of funny. Okay, let's go on. So, about me, my name is Anas Lieman. This is a compact edition of my CV. Well, it's not very important, but now I'm working at Ahus University in the Engineering School of Ahus, and I'm teaching electronic engineers physics and programming. So, and part of my work is to finish my PhD. So, in one year's time I have to deliver my dissertation, and I'm doing the PhD in the context of a funded project called Ecosense, which is about collective mobile sensing and modeling of emissions from mainly traffic. So, and the emissions is, there's two kinds, two kinds of emissions you are, we are interested in. The one is the climate gases, the CO2 and the methane and stuff like that, and the other kind of emissions we are interested in are the more pollutant, the more, more acutely toxic pollution coming from car traffic. There's also a visualization part of that project, which I'm not a part of. Okay, but the idea is that you ask people to use a special app on the mobile phones, and then we get data about how they are moving around in the cities, in the country, and we can use these data to make models. So, the contents of this talk is like this, so a little short introduction to what GIS is. I'm not going to go into all these, there's a lot of detail in GIS, and I'm going to just hand wave on that. I'll look at a couple of applications for GIS. I think I'm more interested in applications, so that's my forte, you could say. A little bit about where you can find data to use on your GIS systems or applications, something about the patent tools available, some examples, and then I will go and if I have time to talk about my research on building a transportation model for Istanbul. Okay, so, yeah, let's get on with it. So, what is GIS? Well, GIS stands for geographical information systems, and it's all about maps. There you go, that is simple. Well, not just a moment, because we live on a sphere, our Earth is spherical, and we like our maps to be flat, two-dimensional. So, how do you fit a sphere on a piece of paper? That's where all the nitty gritty things are coming from in GIS, how to project the spherical things on our Earth into a two-dimensional map. And it turns out that there are a lot of different ways to do these protections, and each of these ways have certain properties which are good properties, but there are no projections which can have all the good properties that you can do. So, some projections retain area, some retain length and stuff like that. So depending on your application, you will need different kind of projections. But, luckily, this has mainly been solved in my view. There's a database containing 4,000 different projections, and there is a standard way to convert from one map type to another map type. So, if you have legacy maps, you just need to know which kind of projection that map is using, and then you can turn the data into the map projection that you need. The most, well, most modern maps are using this WST84 projection. And so, if you start a new project, then it might be a good idea to try to use this projection, which is kind of standard. So, the nitty gritty part of GIS is mainly this projection part. There's a lot of mathematics and stuff involved in that, and it's kind of hard and confusing. So, I'm just going to punt that to someone who knows about it and go on with the applications. So, there are quite a lot of different applications for GIS. This is just one application. This is from Denmark. All the black dots here are the street lamps in my area. So, someone, the municipality has to keep track of their inventory of street lamps. They have made a map where they have put in all the locations of the street lamps. So, that's a way of keeping track of your assets, and it's a way that they can use for planning. They know how many street lamps they have and then know where they are. And if they have to be serviced or something like that, you can plan in how many cars you need to service these lamps and in which order would it be a nice thing to do. So, that's one application. The assets track of things that you put out in your environment. Another way, another asset tracking application is to keep track of where you put down your sewage lines, your underground cable, your electric cables and stuff like that, so that you hopefully can educate contractors so that they're not ruining anything when they have to dig into the ground. Fleet management, this is more a dynamic tracking thing, but you still, you would like to know where all your taxes or cars or trucks, where they are and where they are going. So, you need to put them on the map, maybe dynamically, but still on a map. And if you're a city planner, you would like to have accurate maps so that you can plan for different zones in the city. Where should the industrial facilities be and where should the residential area be and stuff like that. And in these kinds of maps, you also need features of the geography. Is this hilly? Are there swamps? And what kind of, what is the ground made of on these specific sites? So, that's another way of putting information into maps that you can use for GIS things. Okay. And of course, routing, which I think is very important and very used nowadays, very used application of GIS. We do it on our smartphones, we do it on specialized GPS route finders in our cars and stuff like that. So, it has been a very successful application of GIS. And of course, we call it GPS receivers because, but the GPS only give us a position. We need the maps to actually find a way from one point to another and preferably the fastest one. I already mentioned city planning, but there is a specialized version of city planning, which is traffic planning, traffic modeling. This is where I have done some research in, and the problem is of course that we want to plan our city so that we get as little congestion as possible in order to plan for congestion. We need to have a model for where is the traffic going and when is it going there. So, all right. Where can we find data for our maps? Well, there is a very important resource called OpenStreetMap. I don't know if you know it, but I think that it might be undervalued and underused because Google Earth is of course very, Google Maps is very good. But the OpenStreetMap has this further, it's an open format and you can actually go in and change stuff if you want. If you live in an area and you build a new road or the closer road, you can actually go in and change that in your map, in the OpenStreetMap and it's very easy to do that. So, for researchers I think that's a very important thing that the data is available and it's open and free to use. Of course, I can use Google Maps for just for static maps, but I can't use it for getting the features of the map. So, I can see a map where there are roads, but I can't really get the coordinates of these roads. It's a little bit harder to do that actually. But there are other sources, at least in Denmark we have national data centers for a lot of different maps. There's a center called Environment GIS which has all kinds of strange information about where the rivers are and where the where they are deposed for toxic waste and stuff like that. So, you can go there if you want to research that. The municipalities also have these sources and that can actually be sometimes a problem because they tend to be quite old, some of these databases and they there you can actually risk to have a non-standard projection mode and you can, it can be quite hard to convert these legacy databases to a more modern projection so that you can use OpenStreetMap for instance with the old data from the legacy database. So, I just wanted to mention, I found this database the other day about where all the cell towers for mobile phones are in Denmark. There's a database so you could find the nearest cell tower if you knew where you were. I think there must be applications for that also. Okay, so this is not easy to see but the patent tools that I have used the most is this, it's called Q, quantum GIS, this is a visualization program with an embedded Python interpreter. So you can make all your scripts in Python. I'm sorry about the colors here, I thought that, yeah, but maybe we could just make a little demo here. So, now it's the same image just build a green color instead of the light blue. This is actually all the roads in Istanbul. So there are 300,000 roads in Istanbul and that's the only thing that's covered here. So you can actually see the Pospers Street where there are no roads and there are some islands here with some roads on. And then there are some mountains where there are no roads. But this visualization program is actually fast enough to accommodate this quite large number of roads so you can see it redraws quite fast. So the other tool I would mention is the Arc GIS, it's a commercial product which also has an embedded Python interpreter. This is, it seems very popular in research. I've read several papers where they're using Arc GIS. But I think that I found QGIS a bit more approachable from the Python side so I chose to introduce that. So, yeah. So I'm going to talk a little bit more about tools. These are not Python tools per se, but they are very good bindings to them so you can accept them through Python so that they are not really Python. The Postgres database with the extensions called PostGIS and PG routing are quite good. So the Postgres is of course a very good SQL database and the PostGIS extension gives a lot of relevant functions to manipulate GIS data. So I'm not going into the details but because that's a whole new talk I think. And on top of PostGIS this PG routing extension has been built so that you can, if you have, if you put your data into the database you can start to find the shortest paths from one point to another using this extension. In order to do all the GIS conversion from, there are a lot of different formats. Not only projections but also data formats. But there are these libraries available, DIL and especially OGR to OGR which can reach almost any relevant format and convert to most of the other relevant formats. So they are very generic tools but they are not that hard to use actually. And there are Python bindings for both of them. Okay, so I talked about OpenStreetMap and I'm just going to show you hopefully. So I should, there you go. So when you OpenStreetMap you get maps of course and I've zoomed into BitVal. Maybe you can recall or recognize the river and venue. What is quite hard to see with these colors is that there's actually some problems with this map because there is a, you can walk like along here and actually you can go under the bridge and you can continue on this footpath here. But there is no connection from this bridge to this footpath. But that's not hard to change. You can just press edit and if you have logged in and stuff like that, you are presented with this view and you can then choose to connect this footpath here. Oh, sorry. I have to press with this footpath. And then you just have to do that once more. So now I have put in a new line here and I should put in the metadata as well that this is a, this is, it's a path I guess. So when I have put in, I don't know if it has a name or if I should prepare the surface and stuff like that. But there are a lot of metadata that you can put in to OpenStreetMap and when you're satisfied you can just press save and then it will be updated to the database. So and I want you to recognize that it is very easy and it's actually could be beneficial for the, for, for at least for me as a researcher if the OpenStreetMap is as correct as possible. So I want to urge you to, when you've noticed something changed in your local environment, please try to, to go into the OpenStreetMap and we can take it later, right? And change it. It's very easy and it's very nice when it's working, when it's correct. So let's go back to the presentation. So now we come to my, to my stuff, the traffic model that I've been working with and yeah, of course this might be very boring for you because this is of course my pet. So maybe I go into too much detail about this. But the, the reason why we want to have a traffic model is in, we want to be able to predict how the traffic flows through a network. We want to be able to predict where the congestions are and, and we, especially, we want to predict how, if we change the network, how would that, how would that change the flow through the network. So what I've been working with is a model that, that considers how people actually choose where to drive. So it's kind of a selfish margin model for, for drivers. And in, in the research, it's usually called the assignment problem. It's how to assign drivers to different routes in a network. The, the basic principles are derived from econometrics and it's based on, on finding illiter, equilibrium values, and steady state solutions. So it's, it's kind of, it's kind of a static model. So you, you, you create, you, you find out what is the demand for, for going from one place to another. And then you try to find the steady state solution for, for that network. So it's not dynamic per se. But it's the, the, the big problem with, with these kind of models is that there are actually a very large number of different possible routes. And if you want to make a computer, find the, the, how people are distributed through all these routes is actually a, a large problem. So this is a, this is a picture of Copenhagen, or the central part of Copenhagen anyway. So at the bottom right, we have the Copenhagen airport and there is a motorway coming down from the north. So this is an assignment to, to see how will the, how will traffic flow through the Copenhagen area in, in the morning when people come from the north and I want, they want to go to the, to the airport. So the large path here is the motorway. There's a motorway all the way around Copenhagen and it goes to the airport. So of course most people choose to, to go about it like that. But as, as many as more and more people come onto the motorway, it, it turns out that the, the traffic flows and then it's actually, so there is other ways that are quite as fast. And they are going to through the center of, of the Copenhagen, which might not be that nice for, for the, for the morning, morning traffic in the central Copenhagen. So, but the idea here is that we try to find out how many are going at each route. And when we have an equilibrium, everyone would have the same travel time. That's the basic idea because then we have what we call a user equilibrium. Okay. So how, how do we model congestion? Well, we, we know that, that roads have a capacity. They are designed to, to have a certain amount of cars per hour. This is the C A in the bottom here. And when we reach the capacity, we know that the traffic flows and we model that with this simple power formula. So the time it takes to go through a link is given by the free flow time. That's the flow we would normally, we would experience if there were no other cars on the road. And then we have this power function where when, when the volume of the cars are approaching the capacity, then we, it takes long, long, long time. So this is in a very old formula, but it has worked for, it still works. And it's 50 years old or something like that. It gives actually a very good approximation on how traffic flows. And we can use this idea if we, now we have a two routes between our origin and destination. And we want to assign traffic to these two routes in such a way that the travel time is the same on these two routes. And we can find the equilibrium point where these two congestion curves, across. That's the basic idea. The user equilibrium is the assignment where everyone has the same travel time. Because if we all have some travel time, we can't, we can't have a better travel time by choosing another route. That's the basic idea. It's, I think it's attributable to Nash equilibrium, if you heard about them. So it ties into all this game theory and economic physics and selfish users and stuff like that. User equilibrium, the different term, inistic part here is only applicable in, in very small networks because you need to consider all, all possible routes. And as in even small networks, there's a lot of different possible routes. So it's, it's not really feasible to do deterministic user models. So people have invented stochastic methods where you can look at, well, the basic formulation is that instead of being certain that you have the best possible route, the shortest time, we can, we can put in some stochasticity and then we can, then people think they have the best route. So, but there's a bit of uncertainty and this uncertainty we can model and we can, we can use that to, to drive the assignment process. So it turns into, since we are looking for the smallest travel time, it's extreme value problem in stochastic. And we can, there are different formulation for it, but the one I've worked with is this path size loaded where you, where you look at how, how path are looking, if they are looking like if different routes use the same links. So, so the stochastic user equilibrium part is nice because you can use stochastic methods and, but there are still, the problem with the stochastic user equilibrium is that there will be a non-zero probability to use every route. So it might be a very small probability that you use a very stupid route, but it's, it will be there and it will be calculated. So you, you, you kind of underestimate the good parts, the good routes. So that's not so nice. So in the Istanbul case, I have used data from, from, from Open Street Map. So you can just, there's, there's a way that you can actually, you can, you don't need to have the complete, the complete data set. You can just pick what area that you want the data from and then it's, it's converted into a poly, so that you have a network where everything is connected and the, and it's turned into segments instead of roads. So yeah, this is, it's a bit of detail. But there are 300,000 road segments. I have done a simulation with 2000 origin destination pair. I used the Postgres thing and have implemented a combination of, of the deterministic user equilibrium and the stochastic user equilibrium so that we, so that I can assign traffic to different routes. And I used QGIS for the visualization. So, and I think it's kind of nice. You can see on the left-hand side here that you can see the elephant. That means that there's a direct connection from the QGIS to the Postgres database. And I can just point QGIS to my database and then ask it to show all the routes in my solution set. And this is from, this is how I have, how my algorithm has assigned the traffic to, to the different routes. So, of course it's, yeah. There's a lot of detail that I can't really, I haven't time to, to cover here. So, that was my demo. Okay. So, in conclusion, so the way that I have used Python is to drive all the other tools. There are, there are libraries for doing routing directly in Python. But it was actually very easy and fast to use Postgres. So, I just, I just did that. So, there are many tools. And now I'm prompted to end my talk. Yeah. It's confusing, but it's doable. I think that's my main conclusion. So, thank you. Okay. Do you have any, do you have any questions? Comments? Thank you for the talk. You, you do a lot of theoretical research and traffic congestion, but do you actually get signals from drivers? And do, and my question is mainly because companies get that from their devices like Nokia or Google, they get that from, from users. But you as an independent researcher, where do you get the signals from? Well, I'm part of this Ecosense project. And so, there were, in this project, there were three PhD students. One should do the, the mobile applications. And one did the modeling, that's me. And one did the visualization. So, we have, we have made this library for, so it's easy to make applications that send us data. And we have made some applications for, to get users to use it. So, there's an application called are you e ready, where you can download this application, use it for a month. And then there will be an analysis if your traffic pattern is fit for a e-car. And if it is, you can borrow an e-car for a month back after that. So, that have given us some data. There's another project called the, Herning drives to the moon, or bicycles to the moon, where we, but the municipality of Herning, it's a small town where I teach. They want people to bicycle to, to work. And then they have, they ask people to use this application, which will measure the, the length of the bike distance, and then accumulate it to hopefully get it to, to, so that they could bike to the moon and back. That also gives us the data, also when they are driving in their cars. So, so there, there have been several small scale applications like that, that gives us some data, which we, which I can use to, to extract routes and see where people are actually moving. And I, I get the speed information as well. So, I can also guess where there is congestion. So, so I can, I can use this data to also see if my, my models are correct. Okay. More questions? There's one here during the talk. I have one question, one comment, a question. Okay. I use Postgrease for large lists of, of points, geographical points and areas and lines. Is there something directly in Python that, what would you recommend if I have a few thousands of points and I want to find the closest ones or to group them? What library would you recommend? To store longitude, latitude and so on? Yeah. Well, I look briefly at the PyRoute, which is a, has a routing facility, I guess. I'm not, I don't remember the, the data format that they use for storage. But there are, you could look for Dijkstra. There would be several implementations of, of, of Dijkstra, Dijkstra path finding, route finding algorithm. So you could use that, of course. Okay. Thank you. And a comment to OpenStreetMap. Yes. It's actually not a map, it's a database of geographical data. Yes. Fully available. You have just now, you did an update to this path here in Bilbao. In between the increasing number of contributors to OpenStreetMap and the increased complexity of all the routes you can do, you can make a lot of further attributes is some, sometimes it's too much for new users and not inviting them because they, oh, I have a lot of things to learn. In OpenStreetMap, there is a nice feature where you even don't have to, to sign up. Mark a problem or a bug, submit a bug. It is directly on the page and you can just put a marker anywhere on the map and write in your language. Yeah. There is some change to it. This is false and there is always someone just like you or me or anyone else who goes to the area and then looks at this and fixes the problem on the best way possible. Yeah. I had to do it for the Istanbul case because I noticed that the routes that my algorithm found was, well, it was not how the taxi drivers and the bus drivers drove. So I had to find out why and then it turned out that there was a missing roundabout. So I just put that in. So, yeah. There's also another feature which I think is a good feature for OpenStreetMap. They are trying to automatically find problems. So for instance, if a footpath crosses a motorway, that should not happen. So if they mark it with an arrow, so there are a lot of automatic features to find problems. So if you are bored, you can also look at all these in your own area. You can look for the errors in OpenStreet. There is a special map where all the errors is pointed out. So you could just correct them if you are native to that area. Okay. Okay. That would be the time to move to the next talk and let's thank this speaker again. Thank you. Next talk will be in four minutes.
|
Anders Lehmann - How to GIS with Python In this talk I will present some tools for working with Geographic Information Systems in Python. Geographic information Systems are widely used for managing geographic (map) data. As an example I will present how to use Open Street Map data, in routing, traffic planning and estimation of pollution emission. For the purpose of the project EcoSense, GPS data from users smartphones are mapped to OSM roads. The map matching algorithm is written in Python and uses data from the database PostgreSQL, with the PostGIS extension. One of the goals of the EcoSense project is to devise methods to improve the estimation of air quality in urban environments.
|
10.5446/20068 (DOI)
|
Thank you Alexander. I'm from Mannheim in Germany. I'm a developer of my own company. I'm an organizer as well. I'm a speaker, sometimes a MongoDB trainer for our local community and for the Python community. I've served as a program workgroup co-chair building this conference. So if you have any comments, suggestions, what we could do better or so, I'm around. Just grab talk to me. I'm very interested in your intro. My talk today is MongoDB and what's the MongoDB aggregation framework. We're going to cover the pipeline bottle, the pipeline stages and map reduce in MongoDB. And at first, who knows MongoDB? Just, okay, awesome. And who's actually working with MongoDB? Okay. And who has worked with the MongoDB aggregation framework? Okay. Cool. Okay. So let's bring everybody up to speed with orientated document, okay, orientated databases and set 50 seconds. Basically, we work with the document as a JSON-like object. We can store it in our database. We have no shamer enforcement. A collection is basically just a collection of documents actually and multiple collections make up our database. So it's a pretty simple concept. The MongoDB aggregation framework was introduced about like three years ago with MongoDB 2.2. It's a framework for data aggregation. Basically, the documents are processed through a multiple stage pipeline and it's giving aggregated results. It's basically designed to be a design to work straightforward. So no unions like in SQL. And technically, this looks like this. So we have our documents. We do a match, which is a find. We get less documents because we found some, the cell phone set, and then we do some grouping and we get even less. But actually, I thought it's a little bit too technical. So basically, I think it's more like a relay race. You know, like relay racers, they go and they pass the button to each other and basically, this is how the MongoDB pipeline works. So we have our match, which is find. So we say to please, little doggy, get the button. Please pass it on to the smart fox doing something smart, which could be like a grouping. And then we want to present our data a little bit more nicely. And so we pass it on to the projection space. And let me tell you a little bit about the data set we're going to work. I'm going to do present the things and what we're doing. And I've also prepared some live demos. And this is built with MongoDB. Yes. We're using MongoDB 3.0. The new wide-tired Tiger storage engine with compression, PyMongo is obviously the driver we are going to use. It's maintained by MongoDB themselves. And it's pretty well maintained driver. It's always up to date. It's really good. And we're working on a data set of 37 gigabytes, which gives us compressible wide-tiger about 9 gigabytes of RAM. Basically, as you might remember, IT is my second career. I used to be in the record industry with a technical house startup business in the 90s. So everything I do still in IT is very close to working with music. And so we have, from a project we're doing, it's called ChartGuys. We have some collection of a playlist from the Aikons music store. A playlist is basically all the information about the release you can find in the Aikons music store. It's a set of playlists that appeared in some charts somewhere around the world within the last three years. And basically, this is what it looks like. So pretty cool. So don't worry. I've narrowed it down to what we're going to work with today just to give you an impression about our document structure for the demos. So basically, this is a document. An info is all the release information. So like the album artist, the album name when it was released, how much is it in store. And the children, it's what we call a sub document. It's a list with objects. And that's basically the songs, each and every song we have in our playlist. And I was wondering, actually, which artists to use for my demos, because it's really hard to choose music artists making everybody happy. And I thought I found something neutral because I chose Taylor Swift. And it's not because I like her music. Actually, I don't know any songs of her, but she did this great blog post making Apple pay for the trial period for the new Apple music servers. So artists get paid more money for people using a new service by Apple. So she did a good thing. And I think that's really worth mentioning her even at EuroPython. So, okay. So let's build our first pipeline. I've commented in some notes for the SQL guys to make it easier. So basically, this is a pipeline. A pipeline is basically passed in as a list into PyMonga. And match is just basically a find, as you might remember from our document. It's an artist name. So we were looking for the artist, which is a variable I've already stored as Taylor Swift, of course. Then we're going to do a project that basically it's a select and basically all we want to do is print out the, yeah, all the releases by Taylor Swift sorted. So then we switch to this. And go here. So basically it's just an import. We import PyMonga. This is just like a simple database connection. And so let's see our database, which is live on this MacBook. I must say it's only assigned two gigabytes of RAM for this database. So it's not usually we work with a lot more RAM in MongoDB. So we have 1.3 million playlists found. And it's about like 17 million songs covered in our data set. And usually you could like with your match, it could also like just like a query and our query says, okay, we found 40, 93 releases of Taylor Swift. So with the aggregation framework, that's the same code I've shown you on the slide before. We do a match find, project. Basically, just like we're just projecting here. It's just like a renaming of the attribute actually and then resort by release ascending order. And basically that's looking like this. Okay, you see we have many releases. She's quite busy artist, famous karaoke. Okay, so we have a lot of duplicates. And what else can we do? We can extend our pipeline. We can do a grouping. So now I want to group everything by name, which is basically the album tab. As you probably see, we have a lot of duplicates. You have done some duplicates in our dataset, which is because albums are released by different companies worldwide. So at the iTunes store, they get a new ID. They're basically different products, although it's the same contents from the music. So the passing in, the name as underscore ID. Underscore ID is in the grouping operator, basically, what we want to group by. It's mandatory and it's always called underscore ID. And we want to count how many albums are there. Account, we don't have the count operator in the aggregation pipelines. So basically we're just summing one for each and every document in our group. And then we project and sort just to make it a little bit more nice. And this is what we get. We still have some different versions. Okay. So, okay, now we have this nice pipeline. We've got results. And I think it's so nice. Let's print out what we found again. And we get this. And it's so nice. I just want to print it again. And oops, what happens? I just actually just rest is where we stored our query to the aggregation framework. And I just wanted to print it again. And what's happening here? Why doesn't it give us any result back? That's like the first track, I want to show you is MongoDB aggregation framework. It returns a cursor. So basically the cursor, it points to the data in the database. You get back to from the MongoDB aggregation. So once we call list, the cursor is exhausted. So all the data in and then they're printed and then they're gone. So you can't just use them again unless of course you stored them in a new variable. All right. So these are all our aggregation stages. I've put the like their SQL brothers on the right-hand side. So basically a match is where our having operator sort. Pretty obvious is order by limit. It's also, I think, no explanation necessary project is a select. We can also use it for renaming as an SQL form. Our result group is group by unwind. And we're going to go into that very soon. It's somehow a little bit of a join, not really. And redact, we're not going to cover. And out is basically just an operator. Please send the result of the aggregation back to a new collection of MongoDB to store it. So to make things a little bit easier for you to fall, in the next we're going to do something with the artist name and name is the album title. And we're making our pipeline a little bit even bigger, like with the group operator, what I've already shown you already. This is how it looks like. Print out a little bit more nicely. And next step is how can we work with lists of soft op human. So as you see, we have a list here, just with all the songs on that album. And we want to do something with it. And of course, the natural thing would be I query the database and just iterate over it with MongoDB Python. But that's quite an extensive task. We can do it in the database and there's this unwind operator. And it's basically from my experience, it's at first a really confusing step because it's quite unusual for what I've seen. So it confuses people. And so I think just like, let's just chill you. I think it's probably the best explanation because what unwind does is basically take all the sub documents in the list and for each object in our sub document list, it creates a new document. And this sounds really like an expensive operation, but I can assure you that I did a really good job and it is not expensive at all. It's really handy and basically, let me show you, this is like what we're doing now. So, and of course, I'm really sorry. Let's go to that later. Here we are. I'm sorry. So, let's do this. So, now we have all three hundred and thirty-two songs by Taylor Swift. Found them. We can immediately work with them here in our grouping stage. And as you see, the path has not really changed. Although this used to be a list before and we don't do, there's no need to do anything about like iterating over like a list index or anything like that. And I've prepared a little bit more. This is basically what's happening. So, it's basically we just get one release, limited, we get one playlist and then we do unwind and then I'm just renaming it with the project parameters. This is basically what we're getting. We're getting all these are single documents, new documents just like rated on the fly. We can immediately work with. So, basically, it's basically like just yeah, it's an unwinding of the data. It's a little bit unusual concept, but it's basically really simple. So, but, okay. So, let's go back up. Okay, another one, which is quite obvious one. Okay, that's, we have also like a sort, which is also like an obvious pipeline stage. And I want all the releases just sorted by count, descending and release ascending. And basically, it looks really simple. And it returns us something like this. And what's going wrong? Something's wrong. Because we said I want by count descending over all our data. And then I want to have it sorted by release and ascending order. But our result is basically by release and then by count. So, something's going wrong here. And I can assure you it's not broken. It's actually like a trap because in the pipeline, we pass in a Python dictionary. And the Python dictionary, of course, is in unsorted. And of course, so we just pass in something which is not ordered. And of course, our results get a little bit unpredictable. But of course, that's like, this is a solution. And I can encourage you always to use some from the BZON collection. Or you can also use collection order and pass in all sort of parameters in an ordered fashion. Because otherwise, your sorting order won't really work. And so this, oh, wow, it works. Okay. So this was just like a really quick introduction to our stages. There's a lot of pages mentioned before. It's a skip. It's just like skipping documents out, write your results to a new collection. There's a Neogear which just gives you all the documents around geospatial point. Redactors, I don't know. Some people use it to restrict document access on a document level. I've never really seen it in production. And these are like the stages. This is like our race. And now we have some data. And basically, this is very limited from what we can do. Basically, it's just like mangling around a little bit with the data. So we have more. Of course, there's like a minimum, a maximum, first and last operator. And this is what we're going to work on. Again, we're searching for an artist. We're using release date and a release date epoch. And the subtle distance is that a release date epoch is actually a date. And the release date is string. It's no date. It's a string. And we're building a new pipeline. We're doing no grouping by what we want to find out. I want to find out what's the earliest release of Taylor Swift and what's the latest release of Taylor Swift. So we do a grouping by underscore ID. As you see, underscore ID is empty. It's our new primary key. How can that be empty? Yes, it can be empty because we want a group of our complete result set. So we can just put none in there or leave it empty. So there's no need to look for an attribute which is the same on each and every document. Just leave it empty. And we introduce two new attributes, mandate, max date. And basically it's a really simple operation. We just walk the path info to the information, min, max, and project it. Okay. So let's run that. And yay. Now Taylor Swift is around things 2006. I think she started really early. She's releasing stuff. And she's been around for a while. And so what's first and last good for? I mean, we have min, max. It also would work actually on min, max, would work actually on array. But it's just like a little bit different. And it can save you some extra calculations. What's the difference? The difference to our previous pipeline is we have a match and then we do a sort by release date. And then we do our grouping and our grouping instruction is first and last. And what does first and last do? It's really simple. Get the first document of the group and last is get the last document of the group. So there's no need to iterate over your complete set within the group to find min or max values. Basically you just can say, okay, I want this document. I want to look at this document and what's in the middle. I don't really care. So this can be really effective. And as expected, same results. So and with dates, we can even do more. We have some nice state operators. Pipeline. We do basically the same result by release date. We do a grouping and I want to have releases grouped by year. I'm a fanboy now. I've talked so much about Taylor Swift. I want to really know everything. So I want to have see which year which release. So how many releases per year. Sorry. So we actually extended our ID a little bit. And now it's an object with our dollar year operator. And we pass in the date epoch, which is the date. And we just pass it in. And the dollar year, we'll basically just grab the year from our date. And this is then our ID. We want to group by. And this works like this. It's really easy. Makes it really easy if you have some data with timestamps. And so we see okay. Count. So you see she's like a bee. She's releasing every year. A lot of releases. She's hard worker. And so. But what if I want to dig even deeper. I'm not interested in getting the releases by year. I'm also interested in getting each release count for each and every month. She has released something. And of course, I wouldn't mention it if we couldn't accomplish it. In the year, we also have a month operator. And the next thing what happened is now the ID, which is our primary key, can also be a multi key. And so we have a multi key year month. And basically, we do the same. As before, we get the year and month new attribute. Our ID key has two, it's a builder of two attributes year and month. And let's just run that. And wow. We see. I haven't checked. That's probably not a month. Hardly any month, she didn't do anything. So, well. There's a lot of more data operators. As you can guess, there's also like a second many more data operators. We're not able to cover them all in this small talk. Of course, with that. But I'm getting a little bit bored now with Taylor Swift. Because I have, it's early in the morning and we want some more tension. So actually, I thought about who could and else, who could join. So I thought, hey, I just Google Taylor Swift Nemesis and Google says it's an alien space robot called Katy Perry. And so let's bring in Katy Perry. And it's really easy. We can extend our match operator. So Katy Perry is now stored in our Nemesis variable. And basically, we can also do searches with a dollar in operator. And it's basically just the same as in Python. So I think it's not really necessary to explain to you guys. So, and of course, now we have big competition. I'm wondering who delivers more song value for my 99 cents. Is it Katy Perry? Is it Taylor Swift? So I want to see the average playtime of their songs. I'm interested. Who gives me more songs, longer songs I can enjoy for my money. It's not a good thing, but it's just like a nice example. So what are we doing? As you see, we now have three unwind stages. So basically, the first thing is we unwind the songs. And then we unwind the song offers. The song offers, and within those song offers assets is basically the playtime stored. So we want to access this information. So that's why we have a pipeline of one, two, three unwind. It's unwind, unwind, unwind. And then we can group by just going down the path by the song name, which is a child name. And then we just do an average of the path of the duration we have stored within the assets. And show you like this and something's wrong. Something's really wrong. Okay. Sorry. And just fixed that. Okay. Something's broken. I'm very sorry. Won't waste any time to fix this now live. So basically, what I can explain you, basically, it's just like the same we did before with the releases here. And counting the releases. And the next step, of course, will be getting the playtime. So I hope my notebook didn't break. Yes, it didn't break. Okay. Sorry again. Of course, we have our group, our playtime, and we just projected and as a result, we can see, okay, Taylor Swift gives us more music, like about like 10% more music than Katy Perry for 99 cents. And it's a really easy operation. Okay. Now something, something a little bit more challenging. I'm interested in, thank you, I'm interested in getting the prices of the releases of the artists. And my, it's basically scraped data. So it's not probably as clean as I would wish. Basically, we see a formative price with the currency in front and the price, but it's just like in one attribute. And I'm interested in getting the prices in US dollars. And that's easily to solve with a string operation and a compare operation. So I have to speed up a little bit. Basically, we do a project phase. So just focus on the things in bold. They're important ones here. And we have US dollar. Basically, it's a comparison of the lower strings of the first three characters in our price formative, which gives us back US dollar or some currencies or numbers or whatever. And the comparison is basically, is this US dollars? It's pretty obvious. And then we just do a new match for US dollars zero. Okay. So also feels a little bit wrong, but compare. Parameter gives us zero back when it's a match. And it gives us minus one back if the value is higher and one back if the value is lower. So it's pretty handy one. We could also do, is equal. There's also an equal operator which could give us like one or two back as we expected as a Boolean true false. Then we sort, group, and we can even do something else. We can also go and push every release we find in our group into a new list with a price and a product that's basically very similar to JavaScript. It would be like actually like an append and Python. So let's go here. Back. And here you go. And you see Katy Perry's products. And here's Taylor's next object, Taylor Swift, and the list with all the products. So there's really a lot more operators. And just, I can suggest if you fend the application framework, it's probably useful. Just go to MongoDB documentation. It has a lot of examples. It's quite easy to get into. And one more, it's a variable operator. It's a map operator. And as you can imagine, it's basically the same as a Python map. And what do we do here? We're getting the ratings count, which is actually how many users have given some stalls to the product, one we scraped the data. And we want to adjust it a little bit because our management is our back and we need to make it a little bit, look at it a little bit nicer. What we don't really want to do, but it's just like a good example. So basically, we can pass in to a dollar map and input the ratings count as value. And then we can just reuse the value on our list. And we just add 10 to each and every object in our value we find in our list. And then it's applied. Then another thing, which is not probably obvious, we cannot use some operator on a list like we can do really handy in Python. We have to unwind first. So basically, for each and every value in our list, we have mangled with, we unwinded to a new document and then we can do a simple grouping as we've done before. And yeah. And there we go. Which brings us back, of course, to the next thing. You can also do map reduce in MongoDB. And how many of you guys work with map reduce? Who knows map reduce? Yes. And who actually works a lot with map reduce? Okay. So bring everybody up to speed. Map reduce, basically, it's a really simple concept. We have all these documents. We map them. Map them is basically, we just go through and we find key value pairs, which is actually, in our example, we find to find the most popular words in our release titles. And we just emit them as tuples, as you can see, to the reduce phase, which is run by a producer. In our example, it will just like sum up the counts. It's really, really, really pretty easy operation. Basically, we just will use our name operator. And you might wonder, why would we use map reduce in MongoDB? Because we have this great aggregation framework, you see in substrings, we can do so many things. So what's the point, actually, in using map reduce? And basically, for most of the time, you can work with the aggregation framework. In most of the cases, it's faster. It's more accessible. Thank you. But, however, map reduce gives us more power, because you can actually pass in JavaScript there. You can build much more complex queries. For example, our example was splitting up the release titles in words to count them. This could be quite challenging in the aggregation framework. So let's do it. So, okay, this is map. Okay, let me show you a little bit more. Okay. Okay, this fits. Okay. Sorry, this is our map. And here's our map reduce. So for map reduce, we're just using from B some import code, which we can just pass on text. This is JavaScript function. And it basically stores the name of the info which is stored in the name attribute. It just stores it and splits it. It's a really simple operation. And then we just check for some punctuation and stuff. We move it. It's not the best way to do this. It's just like for the simple example. And if we actually find the word, we emit it. So basically, if there's something like teenage, we emit teenage one. And if the album is called teenage dreams, we can also emit dreams one. We send it to the reducer. And the reducer has really simple code here. We just take all the keys and basically just count, just do some how often the key would actually appear from our emitter. And here we get a result. And this is, and now we're going to do a little bit more because I want to remove stop words, which is not really part of the aggregation framework, just to make it a little bit nicer. That's why I've added the natural language kit, remove stop words. And this is like the most popular words in Katy Perry and Taylor Swift's albums. So you see, they probably have a younger audience with dream and teenage and one and boys and fearless and speak, kissed and stuff. So, yeah, it's really easy. Of course, unfortunately, we don't have enough time left. We could also run this operation across the complete data set and to see what's basically the most popular words in album releases being sold at the iTunes music store. So, to finish, I want to give you some more best practices and tips you can use with the aggregation framework. First of all, database, think about your indexes, especially if you do queries on them. Of course, if you have a huge data set and you don't have an index, MongoDB has a collection scan and if it's a slow computer, it's, of course, taking time and probably frustrating for you. Think about probably getting your data set, your database to your RAM. You can just touch commands in MongoDB, which actually do something like, yeah, similar to Unix touch. You touch it and then it fills up your RAM as much as possible, as much as you will ever get from the system to store data. You can work with life and memory. You have to mind that the result can be only like 60 megabytes because that's the maximum we can store in a BJ's and document, but I mean like 60 megabytes is still huge. Pipeline operation has also a limit of 100B, but you will hardly ever, sounds not much, but you hardly will ever really hit it. On your queries, you can improve your queries up front. There's this nice, sorry for the break here, that's a nice explain operation, which will basically give you information what would MongoDB do when query, doing your query. You get some results, you see how many documents scanned, if indexes were hit, if it was all index, and then you can really go and say, okay, I can really optimize all my work with just like introducing a new index. Hardware is, of course, really important, especially RAM. More is better, and it's really simple equation here. Mind the disk performance, of course SSDs and cloud computing makes it really easy. And yeah, and you can also think about working about a dedicated server in case you have something like a replica set and a right heavy database. So you can also say, okay, just do another copy and work locally and do your aggregations without having to worry about if you have a lot of traffic in your database. And the last slide is some useful resources, of course, as I mentioned, MongoDB has a very good documentation, it's pretty updated by Mongo as well. And I also want to mention Asia Kamski, she works for MongoDB, also as a trainer, and she has always like awesome tricks and tips. And here we go. Thank you. We don't quite have enough time for Q&A, so if you want to ask Alexander questions, then try to find him instead. Yeah, I'm around. Just ask anytime. No problem. Thank you.
|
Alexander Hendorf - Data Analysis and Map-Reduce with mongoDB and pymongo The MongoDB aggregation framework provides a means to calculate aggregated values without having to use map-reduce. While map-reduce is powerful, it is often more difficult than necessary for many simple aggregation tasks, such as totaling or averaging field values. See how to use the build-in data-aggregation-pipelines for averages, summation, grouping, reshaping. See how to work with documents, sub- documents, grouping by year, month, day, etc. This talk will give many (live) examples how to make the most of your data with pymongo with a few lines of code.
|
10.5446/20067 (DOI)
|
Yn ymddiwch, Alex. Mae'n adegwyd i'r paswyd. Yn ymddiwch, Alex. Welcom i chi. Yn ymddiwch, mae'n 4 munud ar y ffordd yn ymddiwch. Mae'r paswyd yn 1996, mae'n cyfnodd Cessum. Mae'r paswyd yn 10, 10 o'r paswyd yn 10, 10 o'r paswyd yn 10, 10 o'r paswyd yn 10, 10 o'r paswyd yn 10, 10 o'r paswyd yn 10, 10 o'r paswyd yn 100, 10 o'r paswyd yn 100, I've since lost count. Some days I go, go through thousands of bits of entropy in a single session. With your help, I hope I can get through this problem. So, I won't dwell on this, we all know that passwyd suck. Everybody, myself included, and I'm certain a lot of you, Ielodd yn ymddangos, a rydyn ni'n ddod i'r wyf yn ymdyn nhw arall, ychydig i'r pethau o'r 1, 2, 3, 4, 5, 6, i'r pethau o'r rai cyfnodd yn ymdyn nhw, rydym yn ymddangos, rydym yn ymdyn nhw, rydym yn ymdyn nhw'n ymdyn nhw, rydym yn ymdyn nhw'n ymdyn nhw, ac rydym yn ymdyn nhw'n ymdyn nhw'n ddwy'n ddwy fwy o ddysgu'r ymddangos. Rwy'n gwneud y pethau o'r ymddangos, y byddai'r pethau o'r ddwy fwy o ddwy. Rwy'r cyfleidio'r rhywbeth yn ymddangos, mae'n cael ei ddwy fwy o'r cyffredin. Rydyn ni'n ddwy'r pethau o'r rai cyfnodd yn ymddangos. Rydyn ni'n ddwy'n ddwy fwy o ddwy fwy o ddwy fwy o ddwy, a gael ein Paid Ypwyr yn teulu gwych chi'n ei ddwy fwideg. Felly unrhyw rai dwyn ymdyn nhw, rylyniad y byddai錯. A yddo ni i f accustomed effaidd o gyfynd drwy'r Paid y пятьadau'r wil seed소�u gyda. Felly roeddech chi fel dech newspaper.ати bobl fel würdwyn sydd gael bod chi'neringangor r 어�ain addysgu. Gorong gwaith'n byw pedag ymate procession darne ar gyfer 500 o 1000 Eu Passwords cyd- astr Mae Passwords' enw'r tro cymdech, beth bywch Cymru yn rhan f auxiliary Deadman. Maen i chi'r Password syniad i newid shell i fi'r 쓸 אבלau a mae'r shifts yddy wonderful o sophisticated, To top it all off, we've now got games consoles, phones, thermostats, all manner of crazy internet of thing devices asking for your password and yes, you still do need to have three symbols, a mix case and a digit. Mae'r traddwl i'r bydd yn ymddangos y bydd. Bill Gates yn ym 2004, IBM yn 2011, Y Wyrdd Magosin yn ymddangos y bydd yn 2012, ac yw'r Gwgol yn 2013, Y Wall Street Journal yn 2014. Mae'n fwy o'r cyfnod o'r bydd yn 1970, ymddangos y bydd yn ymddangos y bydd. Fy fyddw'n o gyff task. Fyddi'n af learnol yma, ond.. Nid cartf theory oareddurau y dyfanc Youser Iwn, a mun—— yma'r ystod yn ymlaen. Mae'n ystod yn ymlaen. Mae'n ddysgu o'r ddysgu o'r ddysgu sydd yn ymlaen. Rwy'n ddim yn ymddangos ymlaen o'r ddiogel. Mae'r log-ins ffederau. Rwy'n ddysgu'n ddysgu. Ond mae'n gyfragment. Mae'n gwyllwch. Mae'n gwyllwch. Mae'n gwyllwch. Mae'n gwyllwch. Rydyn ni'n arwine smells gwwlethau. Mae'n eich cyf Augusto, mae'r an abnormal, rhythmol o'r gysylltio, ydych. website, web frameworks. Federated login is, it requires a slightly more effort, but it leaks tracking data. Every time you log in with your Facebook account to a wired magazine, wired knows, but also Facebook knows. And so does everybody, and so do the advertisers that partner with Facebook. And you can't use your Facebook login to log into your thermostat. I checked. The distinct advantage that federated logins have over passwords is that the reassuring of credentials is somebody else's problem. If we compare passwords to hardware tokens, hardware tokens are still weird. It's just banks and big businesses and enterprises that use them. So I wouldn't be happy giving a secure ID token to my uncle Horace, let's say. In addition to that, their proprietary, if you want to use an RSA secure ID, you have to use it with RSA secure ID software. And you can't substitute another token if the secure ID becomes expensive or unavailable. They're hard to deploy. You have to physically post them out and they're hard to reissue because you've got to post out another one. Software tokens have become a bit more popular in recent years. Think of Google Authenticator. They're often also paired with the SMS technique of sending a six digit pin via a phone call or as a text mail. They require a bit of training. Again, they're fragmented. There is a standard in theory. You don't have to use the Google Authenticator application. You can use any application that works with that OAuth standard, but it's a bit hit and miss. You usually don't need a third party to use such an application, but some of them do and they do leak metadata when you log in. Biometrics, they're familiar. Everybody's seen iris scanners and fingerprint readers in the movies. So you don't need to train people how to use them. The problem with biometrics is that's about the only good thing about them. They're proprietary, they're expensive, they're hard to deploy because nobody has a fingerprint reader unless it's on an iPhone. They're impossible to reissue until cloning comes along. With the exception of the iOS platform with Touch ID, they're almost nowhere. So we're stuck with passwords. What can we do to mitigate this travesty? The first one is that you might want to switch out your password generator. I'm sure you're all familiar with a certain XKCD cartoon. I won't mention it by name, but if you'd like to use it and you can't be bothered to code your own, I highly recommend horse phrase. It's pip installable. It comes with a command line tool and you don't have to give it any special incantations to give a nice familiar XKCD style password. Studies have shown that this style of passphrase is easier to use than the mixed case symbols, digits type. If you need extra entropy, just add another word. Similarly, horse phrase, just as it's usable from the command line, does, of course, also a module that you can import. If you have to use the mixed case digits, at least one uppercase, then you can make that a bit easier to type on phones and games consoles and thermostats. I'd like to try a quick experiment with you all. Could you all get out your phones or your tablets and try to type that password? You don't have to type it into a password field. Just any text editor or note application will do. I'll give you all a minute or two. Raise your hand when you're done. Fantastic. How many times did you have to switch keyboard? Did anybody make mistakes? On average, depending on whether it's Android, iOS, different versions, it takes 24 taps or 9 keyboard changes to type that password. It's a typical randomly generated, considered strong password. Now I'd like you to try this variant. Sorry? Yes. OK, so we've got a few raised hands. Was that easier? OK. So the improvement is you save about seven taps and you only have to change keyboards twice with this variant. This comes from some research done by the US National Institution for Science and Technology. They find that on average you save a certain number of taps and a certain number of keyboard changes by permuting passwords like this. If you'd like to permute your own passwords, I've created a Python package for it. You can use it from the command line or in your Python software. The bottom will take you to a summary of the research. By permuting the password, you do lose some entropy, but you can typically regain that lost entropy by just adding one or two characters. The research summarises the bits of entropy lost and gained by the permutation and adding extra characters. I hope to be able to get this integrated into keypast or to provide it as a plug-in in the coming month. If you're running a server and you're fed up with your users choosing 1, 2, 3, 4, 5 yet again, the classical way of combating this is to ask for mixed case symbols, digits. No repeating, not the same as your username. They all add Hulk rules. What you really care about is entropy, but it takes a bit of maths to measure entropy. Very few people coded up. Luckily, you don't need to. It's been done for you. If you're writing a Django app and you'd like to get rid of those annoying rules, but you'd still like to encourage your users to have strong passwords, then you can use Django ZXCVBN password. Django app is a mixture of Python and JavaScript. It's based on an underlying package called ZXCVBN. It measures the entropy of the password based on the username and the current day and a few other signals. As the user types a new password, it gives them an interactive strength meter. Again, there is research that has been done to show that if you provide this interactive feedback, it encourages users to pick a strong password without antagonising them, without driving them away as much. When someone is creating an account, you don't want to drive them away with repeated validation errors on their new account. If you'd like to use ZXCVBN in other frameworks, then the Python package index page for this Django app also links to the underlying package that you can use yourself. Another thing you can do to make your users' lives easier is to let them see the password. We've taken it for granted that you should hide the password when people are typing it for years. That made sense when the password field would always appear on an office computer in the middle of a load of cubicles with people passing behind without you being able to hide the screen, or on a shared terminal on a main frame. Those days are passed. More times than not, you'll be typing a password on your phone, or a tablet, or your laptop, and there won't be anybody else around. So why hide the damn thing? The safest thing to do is to not scare your users away and make them think that the form is broken and insecure is to, by default, still hide the password, but provide a little tick box or clickable icon so that they can show it. The link below goes to the page from which this screenshot is taken and shows more examples. I'm afraid I don't have a pre-packaged Python solution for this. Shouldn't this be a standard problem feature? It is a standard feature of login dialogues on Windows Internet Explorer, passed a certain version. That has a slightly unusual implementation in that rather than that i symbol being a toggle, you actually hold your finger down on it on a touch screen, or the mouse button, and it shows as long as it's being clicked on, like a dead man's switch. So this is ongoing usability evolution. There's a very good chance that browsers will implement it like this, by default, but it's not happened yet. Ideally, if this ever becomes a reusable HTML component or some form of HTML5 thing fall back, it should take into account that it could become a native feature. The final mitigation that I'd like to suggest is please, please, please don't disable auto completion. Don't disable password managers. The second example you see there was called out, British Gas, was called out quite recently for doing this on Twitter. And thanks to a, what's the word, Twitter, not hate mob, a Twitter mob brandishing pitchforks and torches, British Gas will be reconsidering their practice of using this. Unfortunately, umpteen banks and city councils, county councils still think that they shouldn't allow you to save a password. Unfortunately, the only way to get them is one side at a time. So that's more or less all I have to say about passwords. The next thing I'd like to introduce to you is a new standard or a new pair of standards for authentication. They come from a body called the Fido Alliance, which was set up about a year ago to fix strong authentication. As I said earlier, the problem with all the alternatives to passwords is that they tend to be fragmented, proprietary, unfamiliar, and generally unstandardised. The Fido Alliance's mission is to fix all those things. They've released open specifications. They will do, if you pay them money, they will do certification testing for you and give you a logo and a trademark that you can use and stamp on your products, but you don't have to do that. It's based on internet and it's based on public key cryptography. Think of the Wi-Fi Alliance a decade or 15 years ago when wireless internet was still a weird thing. They are the Wi-Fi Alliance, but for authentication. The aim of the standards is to allow multiple authentication methods such as dongles, fingerprint readers, pins, smart cards, mobile phones, whatever, to all be usable against the same API and all be usable no matter what the transport is, be that USB, Bluetooth, NFC, something we haven't thought of yet. It's a single browser API no matter what combination of those gets used. Because it's based on public key cryptography, no sensitive data leaves the user's authenticator. So there's nothing sensitive on your server to be leaked out. There's nothing sensitive in the browser to get phished. So if, God forbid, your user database gets leaked, the only sensitive information that will get into the wild is things that you've specifically chosen to gather. The Wi-Fi Alliance standards don't require that you gather an email address. You can still gather an email address and if, God forbid, your server gets breached, an email address is the worst thing that will get leaked. There are no password hashes within the Wi-Fi Alliance standards. Like I said, it's an industry consortium. The big backers are Google, Microsoft, RSA, Samsung, PayPal and a host of others. They're the internet companies that are as fed up with authentication as you are. It's ever-growering. Recently they accepted their first government members. So the US government national institute of science and technology has joined and so has the UK home office. There are two standards that have been announced by the Fido Alliance. The first is universal authentication framework. This is the one that is designed to replace passwords. The second is the universal second factor framework. That's the one that's designed to standardise two-step authentication. So you still have a password under U2F, but it doesn't have to be as strong because you've got a second factor. It could just be a pin if you wanted to. No matter which of those two you choose, the way it works is, the very first time you open an app or visit a website, you'll be asked to register. It can, but it doesn't have to ask for a username and email address. But when it comes time to set up a secret, you would activate your authenticator, be that your mobile phone or a USB key or something else. Your authenticator would generate a new private key that private key is only used for this pairing. It's not used for any other. So if two different websites are used against the same key, there is no way that those two websites can know it's the same key. It's just a different private key they see. Private key is stored in a secure element on the authenticator, and the authenticator sends back to your website or to the app a public key and a key handle for use in future authentication sessions. Once the registration is complete, the server stores the private key against that user's identifier, username, email address, whatever, and can be sent back to the authenticator as a challenge next time. Authentication looks almost identical to registration. The only difference is that when the user is asked to activate their authenticator, it uses the key that it previously generated to sign the challenge that was sent. So to give you a concrete example with some code, this is a pseudo code Python web server that's handling the registration for a new U2F device. The library we're using is an open source BSD licensed package provided by Ubico. It works with Ubico keys. It should also work with U2F authenticators from any other provider. The app ID you see there is just the domain of our website. It must be HTTPS. You cannot do any FIDO Alliance authentication over an unsecured channel. We simply generate a challenge, store it in some session or against the user's username, and send that challenge to the client. The challenge that gets sent is just a short snippet of JSON. It's all base64 encoded, so it's safe to stick anywhere that you don't have to worry too much about binary encoding of it. The challenge that you see there at the bottom is just a randomly generated string. There's no structure to it. You can generate as many as you want and throw them away. You don't have to worry about losing them. That challenge gets sent to the browser. The simplest way of doing that would be to put it in a hidden field on your registration form. The browser provides a JavaScript API, which we call U2F.register, with the challenge that we generated server-side. That takes a callback that accepts the response generated by the authenticator. What would happen when this JavaScript gets run if this little authenticator was plugged into the laptop, little green light starts flashing there as part of your page, you put up some text saying, please activate your device now. The user touches that, which allows the authenticator to respond. It's had a proof of human presence. It generates the new private key, returns it to this JavaScript, the JavaScript sticks it in a text field and submits the form. The response is just some JSON. The challenge is exactly what was sent by the server. The client data and the registration data are generated by the key and they are cryptographically attached to the challenge. If somebody changes the challenge on the wire, the client data and the registration data will not validate. The app ID also is what was sent by the server, has to match if that gets changed, if somebody tries to do a man in the middle attack, all the signatures will fail. It doesn't matter that it's a UB key. It just has to be a U2F authenticator. Final step of registration on the server. We received the response that was sent by the U2F authenticator. We retrieved the challenge that we generated earlier. We call U2F.complete register with both of those. That takes care of all the cryptography for us, checks the signatures and returns a registration object. That's just a Python dictionary that contains a key handle, the public key of the authenticator that was freshly generated and the app ID that we generated all the way back. We save those three things for future authentication. That's just an example of the registration that is returned when we call the complete register function. The thing that I'm going to skip over there is the attestation certificate. That is so that you can say, I will only accept signatures from brand X. If you want to restrict yourself to just UB keys or just Samsung phones or just some particular brand, they all contain a manufacturer certificate. The manufacturer signs that certificate and there is an online database that you can check that attestation certificate against. For simplicity, we are just going to skip over that right now. Authentication is almost identical. On the server, we generate a challenge. The difference is that instead of start register, this time we use start authenticate. Again, we save the challenge that we generate. The challenge looks like this. The app ID is the same as it was during the registration. Challenge is a freshly generated random key. The key handle is what was returned by the authenticator during registration and it's how we identify this key. Again, on the client, receive the challenge, submit it to the authenticator. The authenticator starts flashing or doing something bingley bingley beep so that it can get your attention. The user activates the authenticator and the authenticator then returns a response that you can transmit back to the server. On this, you just press the gold circle. That doesn't prove it's me, it proves I'm present. This is a U2F device. The idea of it is that it is a second factor. Obviously, this is cryptographically unique. It's got a private key inside that was burned in at the factory. It's tamper evident, that sort of thing. It's not only useful as a second factor, not as a universal UAF device which would have something biometric in it so that not only was it proof that it was the person present but by the fingerprint or their retina, it would prove that it was them. Did it prove that you have the device and generally you can assume that you have the device on the person who registered it otherwise can't see the person? Yes, but the point is you couldn't use this as the sole authentication factor because effectively it's a bearer token. If someone steals it and it was the sole factor, they would have everything they needed to pretend to be me. Final step of authentication. This is a U2F device. This is the second factor of two. For a login with this, first as usual we verify the user's username and password. Assuming that's correct, we get the associated device that was stored during registration. We then call verify authenticate which gives us a counter and a flag as to whether touch was asserted. In the current version of the standard, touch is always asserted so that will always be true. The point of the counter is that although this device is tamper evident, it could have a vulnerability of some sort, the counter is an indication of whether this device has been cloned. If somebody could get in there with a microscope or a very powerful magnet or an x-ray machine, they might be able to clone this. If they did, and we were both using it, there is an internal counter that would be... It should always increase, but if there are two of them, sometimes you would see a counter be returned that's less than the counter that you saw last time it was used. If that happens, it means someone has cloned the device and you need to take some sort of action. It could be send the user an email, lock the account. It depends on the particular application. Assuming all is correct, you then just store the new counter value, save the device and the user is successfully authenticated. This being a second factor, you don't have to use it just at login. You can choose to do it when the user starts a sensitive action, like entering the admin section or transferring $100,000 to Nairobi. You can ask for additional authentication at any point. That brings us to the demo, but before I do that, does anybody have questions about the code? Thank you. Not entirely about the code, but if you have an authenticator on your mobile and your register, how do you authenticate on your laptop, by instance? Jumping ahead a bit, the vast majority of deployed hardware at the moment just works with USB, but the standards include transports for Bluetooth Low Energy and NFC. I would have to connect my laptop to my mobile phone using Bluetooth. It would be Bluetooth Low Energy, so you wouldn't have to pair the two devices, but you could. Basically, the browser, if it supported that form of UAF or U2F, would have a Bluetooth stack in it. It would ping for local devices. Your phone would be listening. It would send a message that says, yes, I'm here. Yes, I am a UAF authenticator. The browser would then send back, I have these key handles. Do you know about any of them? The phone would say yes, and the phone would go bingly bingly beep and ask you to authenticate. OK, thank you. Any others? OK. Time for a demo. How do you... Oh, it's on that side of it. Yeah. Strange. Goodness sake. Some sort of sticky edge on the screen. I mean in software, not physical sticky. So this is the demo application of a Django application called Django 2 Factor Auth. It currently supports the Google Authenticator app, SMS, phone call, or plain old UB keys. I've extended it to also accept U2F devices. So I'm currently not logged in, and there is a secret page that I cannot view. So I'll just log in. Saved password. I thoroughly recommend password managers. I haven't yet enabled 2 Factor Authentication on this account, so I cannot view the extra secret page, but it requires additional verification. So I need to set up 2 Factor Authentication. Here I can choose between the different versions. In your own Django application you probably choose... You probably provide... You probably offer fewer choices just to keep things simple. I'm going to go with U2F. Insert the dongle. I'm sorry you can't see it, but I do promise that there is a little green blinking LED here. I touch the dongle to prove that I'm present. Sorry? So that JavaScript that I showed you earlier, the callback when it is called just inserts it into the input with that ID. In a real world application you wouldn't have that box there. What you'd show the user is just an animated thing saying, now please activate your device. And the second that they did, you submit the form and return it. I'll come on to that at the end. So I'll complete registration. I'm done. I can add a phone number at this point. This is a standard feature of Django 2 Factor. This is just an example app so it's not actually wired into Twilio, but in a real application it would be. So I'll just register that phone. Cheat, yeah. So there we go. This accounts now registered to Factor. I can view that secret page now that requires two-step authentication. If I log out, log in again. I have to choose a password and I can choose whether to use phone authentication or not. I'm going to touch the device. And I'm in. That is two-factor authentication without having to dig a phone out of your pocket and type in those six bloody digits. APPLAUSE So to answer the question about what support it requires. Damn it plays my mouse. If you want to see any of the code for that demo, those of the URLs, I will be uploading these slides, tweeting the URL, and also putting the URL on the description page for this talk on Europe Python website. I'm afraid the code that you've just seen running isn't available on PIPI yet. I still need to convince the maintainers of the upstream projects that my pull requests are worthy. So to browse the support, I'm afraid this is the bad news. The only thing that supports U2F at the moment is Chrome. It's Chrome on any platform, which is good, and Chrome EM. Firefox have a bug open to add support. It just needs somebody to do the work. They're not opposed to it or anything. And in a few days' time, Windows 10 will be released, and the new edge browser will support U2F. And that won't just be USB tokens. That will be biometric devices like fingerprint readers or iris scanners built into laptops that support Windows 10. For the hardware support, the only thing that the browser singular supports at the moment is USB. Standard for Bluetooth and NFC was released on the 1st of July this month. So expect support for more devices in the coming months. In terms of phones that support it, if any of you have a Galaxy S5 or a S6 or a Galaxy Note 4, congratulations, you have a UAF authenticator, and you might already be using UAF to log into PayPal. A PayPal application support uses UAF on those devices. Android M phones. Google are making a push for fingerprint authentication in Android M. About time. Touch ideas been stomping on Android's lunch for far too long. And you will see more phones with fingerprint readers in the next few months being released. Qualcomm are one of the Fido members and their new hardware supports it out the box. Possibly even it will be ultrasonic fingerprint reading so you can touch anywhere on the glass. If you'd like more information, the specifications are available. There is a tutorial that goes into the steps that I have in a bit more detail. There is a nice video that gives you a history of Fido Alliance and the sales pitch. And if you'd like to use any of Ubicose open source libraries, they were until recently GPL, but they are LGPL, but they're now relicensing them as BSD. You can use U2F with Ubico tokens or in theory any other token for SSH login for PAM on Linux. There are Python bindings, there are Go bindings, there are JavaScript bindings. You don't have to be in a browser to use it. The client application just has to implement the wire protocol that is standardised. That's it. Thank you very much. APPLAUSE I think we do have time for questions and we have questions. How do I look at for you? That's fine. Right. My current impression is somehow that the only proof that you are willing to authenticate is your super secret password. You enter manually. Is this correct? So there are two standards from the Fido Alliance. I speak of U2F. So U2F, it is two factor authentication. You are proving that you know a secret password and you are proving you own a particular device and have control. That's understood, but a proof that you are willing to authenticate is the password, right? Sorry, could you repeat that part? Your willingness to authenticate is that you type in your secret password that you have remembered or in a password manager. I would say that's part of the proof that you're willing to authenticate. The proof that you're willing to authenticate is that you complete authentication. The act of doing it is the indication that you're willing to do it. There's a second proof that you're willing to authenticate and that's pressing the button. The button on the U2F device. When you show it again, the Yubikee and Neo, I guess, that little button in the middle that is blinking. As long as this button is blinking, nothing happens. This device will only respond to the challenge if you click this button. That doesn't need to be me. That doesn't need to be me, exactly. That's why you should use this Yubikee as a second factor. I won't recommend you to shorten your password. I would recommend you to... Pass it anyway. Yes, exactly. In my opinion, you should use a strong password and then use this as a second factor only to increase security. But if your password is secure, why do you use this? Because your password could somehow get cracked, probably. And you keep J.K. in his soul and then you can see... But not the device. So, yes. U2F is not a replacement for a password. Depending on the application, you could choose to allow weak passwords, but you don't have to. The true promise, at least from my point of view, of the Fido Alliance is when the UAF standard comes about. And that mandates that the authenticator not only proves that I'm present with just pressing a button, but also proves that I'm me by some form of biometric or typing in a pin or something else. So, I agree that biometrics are not the greatest thing. But they are mitigated in this case because the biometric data doesn't travel over the wire. It never leaves the authentication device, just like Touch ID on iPhone. I'm happy. No one's ever happy authenticating. I think we'll take all the questions anyway. So, if I understood that standard correctly, it's required that you have a physical device, right? So, you can't have, let's say, a third-party application on your PC. So, the specification strongly encourages you to make it a physical device, specifically if something that is tamper-evident and has a secure element. But you have to anyway because you can only connect using USB, right? Well, if you're at the kernel level, you can pretend to be a USB device. There is a Chrome extension that is a soft U2F authenticator. It has a particular key. When you, as the server, receive the registration data, that would be evident in the attestation certificate. So, there's nothing in the wire protocol that says it must be a hardware device. Wire protocols can't enforce that. But you, the server, has access to a metadata service that includes the manufacturer's signatures of various different devices. So, Ubico has signed their certificate. Samsung has signed their certificate. Other providers have signed theirs. Any software implementation of an authenticator would have to have a manufacturer ID. It would have to be available in the metadata service, and you could choose to blacklist that, or whitelist only Ubico devices when registration occurs. Thank you. Hi. I have a habit of losing things. And if I'm using one token to access everything on the web and I manage to lose that, it sounds quite... Anyway, how would you deal with that? So, two parts to that. Any website or application worth its salt that implements two-factor authentication will either force you, or strongly encourage you to set up some sort of backup method. That'll typically be his eight, six-digit codes writing down on a bit of paper and keep it somewhere safe. Or it'll be, please give us your mobile number, we'll text your messages back up. Or both of those. The Django two-factor authentication app lets you do both. The other part of it is that if I lose that, it's... I've lost it. It's small, it's black, it's not going to turn up. If I lose that, it's got onboard GPS, it's got an onboard radio, I can ping it remotely, and it can't ping it to ring itself. If you're going to keep all your eggs in one basket, that's a pretty good basket. I had two questions, one of them was already answered. Still, I'm a bit confused because you start by saying, well, passwords have issues and whatnot, but then in the demo, the first thing you do is you type in a password and then you use the password manager to store it. So basically, you put your memory inside your computer. And then you use the second authentication using your token, which, well, I'm not convinced, never the... like others, that it's a suitable solution. Now I have a question. I mean, how is this any different from Google Authenticator? In the sense that the Google Auth actually is convenient, it's a six-digit number that is changing all the time and is synchronised with the clock somewhere, and there is actually a Python memory that does exactly the same job. It's called windtime password, I think. And there's whole math behind that to justify how good this is. So I don't see how a physical device is better than something which is purely software. So Google Authenticator, on the one-time password standard that it's based on, do require that the server stores the seed that generates those six-digit numbers every 30 seconds. So if there is a server breach, there is secret data in addition to the password to be lost. With these, the server is not storing anything secret. I find this more convenient than typing in a six... going to the Google Authenticator app and typing in a six-digit number. But you have the choice of what you want to use. You can use this with your Google account, by the way, either as a UB key or as a U2F device. And if you want a dedicated U2F device from Uico, it's cheap, about 15 euros. Sorry, I forgot the other part of your question. We still have time for one very short question. And this is a yes or no question. In the FIDO standards, is there any capability for duress passwords? This is the case where someone is putting a gun to my head and I need to log in to my bank, but I want to indicate to my bank. I want to be able to successfully log in, but also indicate that I have a gun to my head. So it's a second password that also works, but indicates a bad condition. So in the case of U2F, your duress password would be... you would type that in as the first factor, I guess. That wouldn't be part of the FIDO exchange. In terms of UAF, I haven't seen any reference to that. This is speculation now. The only way I could think of doing it would be if you tap the device, it does a normal authentication, if you press and hold it does duress. But I don't know if this could be reprogrammed to do that. I don't know if there's anything in the FIDO standards for it. Thank you again.
|
Alex Willmer - Taking the pain out of passwords and authentication Passwords are a pain for us all - programmers, users and admins alike. How can we reduce that pain, or eliminate it entirely? This talk will - Review research into techniques that improve the usability of password systems, and mitigate shortcomings - Introduce the new standards Universal Authentication Framework (UAF) & Universal Second Factor (U2F) - Describe how they streamline authentication, even eliminate passwords entirely - Show how to integrate UAF/U2F in Django and other Python frameworks - Summarize the state of support for UAF & U2F in browsers, devices, and the wider world
|
10.5446/20066 (DOI)
|
Thank you. Welcome everybody to my talk. My talk is about the tracking alliance with pattern, citizen science application. Shortly I introduce myself and the company I work in. I'm Alessio Sinis Calchi. I'm an Italian engineer and I work in Rome and my company is Be Open Solutions. I'm from Rome, Italy. We use the mission and philosophy of our company is to use open source software as much as possible as for Italian companies or for large industries and international institutes. One of our main subjects is oriented to develop geographical data such as environmental data and image recognition and image plotting and so on. My contact information is displayed here and for you all at the end of my talk. My talk is about alliance. I don't know if you read the abstract of my talk. This is a little question for I think the big question, definitive question for we all is are we alone? Can we collect and enhance instruments to track aliens and non-flying objects and so on? Okay. No. Of course not. I'm joined with you and we are not talking about Martians. Of course we are talking about a system to register, to collect data about marine alien species, marine alien species that is to say a species invasive, non-autotonous. So of course we are not alone and stakeholders of this kind of application is not telescope operators or radiometers but is scientist, marine biologist and then of course citizens, divers, professional and non-professional fishermen. Okay. Now we speak about this application, the importance of to register, to collect data about alien that is to say non-indigenous species, marine species. Why? Because the presence of these invasive species is one of the major cases of the decrease in biodiversity. So not only the pollution of the sea we consider. I underline that we don't speak strictly about fish but only jellyfish, crustaceans and living sea plants and so on. You can see in this picture one fish with three eyes pictured near the coast of Springfield, USA. Okay. But how they travel? They travel in the ballast waters of commercial ships or the island. We have an intentional human introduction of this fish, an annotation for example from aquariums, also intentional introduction to balance the habitat of these species. Of course marine biologists do this introduction but they should be strongly monitored. Okay. When a marine species is interacted in a new habitat they begin to multiply their presence, increasing predation and competition. So consuming and co-systemal resources and so the deep change of the habitat change the species itself, the DNA we know. So we have, this is a very important question. So what this kind of problem deal with Python? Okay. This is some characteristic of this kind of problems are very typical. We consider in this case geographical data, geolocated data. Of course you know that scientists, the dream of each scientist and to have a wide, very wide area collected with collection of data in a homogeneous way. But of course it is not always possible. You know that resources are very scarce. Also scientists are few and their work is expensive. I don't know here but in Italy we have a few resources, a few refunds for researchers. So usually reports are very few. They are focused above all in the most populated areas on the most interesting ones from the point of view of the environmental habitat and biodiversity. And so this is in this picture we see a case of this shell. The signalation of this shell is around the Venice lagoon above all. So the reports are very few. And what the solution of this kind of problem is the application of the citizen science paradigm. What is the citizen science paradigm? Citizen science means public participation in scientific research. This is a Wikipedia definition. It is one century years old. So what it needs in my opinion, of course to be applicable, it needs the supervision of professional scientists. Why? Because we cannot consider the collected data from non-professional user citizens and so non-professional scientists as good. So we have a first phase to validate data collected. Then of course we have to do a survey over these collected data. And at the end of this survey we have the need of a person, a scientist that has the responsibility to publish the survey. So in words we need a workflow, the definition of a workflow. Of course non-professional users should be emotionally involved in the project. So we need to give them a feedback, some information that should be important to the public participation is that it should involve a public important issue. So also users should not be scared by long forms, long subscribing procedures, complicated work. So they should compile user friendly forms. We have to balance two needs. They need to have a lot of data, but poor information over each report with a lot of information on each report, but very few reports. And we prefer to have a lot of reports, but each report should have very, very few information. So we need to have homogeneous data in a wide area. Of course, citizen science paradigm nowadays is used because there is a big diffusion of media devices. Of course users should be equipped with tools for the report of the observations. So we can use this kind of paradigm and these are some examples of citizen science projects and some of success stories. And not only environmental projects, they involve, for example, this one from the National Geographic, they send you a little kit to get your DNA for tracking history of human migrations, for example. Okay, now I speak in general, now I'm going to speak the more specific case of my project, SMAA, System for Tracking Marine Alliances. The client, the client was ISPRA, ISPRA is the Italian Institute for Environmental Protection and Research. And this is the main infrastructure, system infrastructure. We see on the left the input data, input data is coming from anonymous user or authenticated user via a Facebook authentication. Facebook authentication API is defined in Facebook, it is not such a complicated work. Of course feedback can be applicable only for Facebook log users. And then this HTML5 forms to be visible in a media device with a web view, web view plan instance. We see that we have two plan instances. One plan instance is designed to be visualized by media devices. And the second one for desktop workplaces. So we see that these kind of projects have three databases. Note that I have marked two different colors, pink for developed software and light green for only configured software that is available. So what about database? Zodib is the hierarchical database, I don't want to speak about this. I designed a Postgres with Postgrease extensions to collect data and also a file system to collect the pictures of the species for a local database of marine species and for signalations of marine species. So I need to consider a project scalable because you can't predict in advance for citizen science application, you don't know in advance the amount of the participation of citizens. So with file system storing, you can add space, storing space, simply mounting new parts of the file system. Of course this project have open layers. Open layers used for users in the plan instances both. But one on the left is used to mark the position of the signalations and the second one on the right used to manage all the signalations by biomeric experts. Of course at the end we have on the right the interfaces to use such data for bi-external services. So we have an XML or CSV exporter and a geoserver instance to make data available with very famous protocols such as WFS. Okay, this is the simply crowd friendly form to make a signalation. You see the top on the left and the bottom on the right. You see that it's only one page, one blown page, one language, one layer on the map. It's very simple, should be very simple and with very few information. The picture and position are compulsory and so we have other data to report, this is the date of the observation depth and some additional notes. So there is a validation, a type of validation client side and after of course there is a validation of the expert server side of course. We see the workflow definition, this is an example of the work for the definition but it is very typical in this case studies. We have two kind of experts, expert level one, we call the level one that makes the validation and the creation of the survey and so the attribution of the right species to the signalations and expert level two that have the responsibility to publish the survey. Any way signalation reacted or published generates a feedback to the citizens so the citizen should be involved and should be considered. And so now I quickly show you some Python piece, three little Python pieces I used, note that I use Python non such in eccentric or strange way but as a general glue to make all tools working together. This one is a piece of code from the database model, I use SQL Alchemy declarative base so this is the class signalation, you know that this with one to one relationship with the survey, use list is forced. Then we know that latitude and longitude are two floats but position is a geometry so I use a geo Alchemy two and a simple function to make the translation of the position to be used with open layers. Another piece is the piece of code for the XML exporter, it uses chameleon and we have to translate a ABCD DNA template, this is a template used for exchange biological data so I think it's enough simple but I translate a page template file, I fill it with data. And third is a little bit interesting, there is a connection with worms, worms is the word register of marine species, the fact of standard for the taxonomy definition of marine species all over the world. We know that worms offers a service to query its database via SOAP protocol so I use SOAP pie to query this database. With this schema I reach the possibility to have the classification of each species automatically and so the biologist scientist should not insert and insert the rank, the philum, the genre of the species. And now Sma now is available at this link and it is not sponsored yet but should be published soon, I hope. And okay and this is all, thank you for listening. We have time for two questions, if anyone has a question please raise their hand and we'll get you the microphone. Hi, so why did you choose blown? To use blown. Yes, I choose blown because the final client, one of the reasons, the final client is already skillful with blown also the East Prane Institute has a blown site so we didn't have in this project an after deployment assistance. So this is one of the best solutions for the client. Of course also be open, the company working have a strong know how about this content management system so I choose that. Time for one more if anyone has one. Okay, thank you very much, Alessio. Thank you. We'll be back in five minutes with a talk on TDD.
|
Alessio Siniscalchi - Citizen Science: Tracking Aliens with Python! The talk discusses the challenges of implementing a Citizen Science Paradigm in a Python-centric platform, and the solutions devised for the System for observation and monitoring of Marine Alien Species, currently used by the italian Institute for Environmental Protection and Research (ISPRA). "Alien" Species means species introduced into a natural environment where they are not normally found. Topics includes strategies for crowd-friendly forms, work-flow definition for collected data, choice of the best technologies for its components: app for android devices, web application for citizens and experts, webGIS for data browsing and web services for data exporting.
|
10.5446/20065 (DOI)
|
server-side related. We use of course Objective-C and Java for mobile applications, but for everything which relies on a server, it's in Python. I've been a member of the TurboGears 2 web framework development team for the last four years. If you don't know it, it's one of the oldest web frameworks together with Django. And I contributed to various Python web libraries like the MongoDB Object OpenMapper Ming, which is used at Swissforge.net for everything related to MongoDB. I have been the big maintainer since this year, and I worked also on Tosca widgets and former code, which are libraries related to validation and forms for the web. So most of my work has been related to the web world for the past years. What I'm going to talk to you is about a project that really happened, and we had at our company, which started just as a plain proof of technology. The customer came and said, hey, I want to try my idea, see if it can work, if it works properly, if people can use it, and it's not a huge mess and something like that. So we started with a really simple code base that then became the final product. It became what the customer launched. As usual, it happens always like this. The customer came with something, it's just an idea, a test, and then it became the real Frankenstein. And the core part of this product was that it saved a lot of files, mostly images in this case. So we decided that as it was just a proof of concept and we were really short on the budget, it should be done like in two days. We decided to not rely on a cloud storage because it would involve more time to bring in any library to store the files and more money to actually pay for the storage itself. So we just decided to go for storing files on the disk and letting NGIN serve them. So the most simple solution because it was really simple and for a proof of concept was good enough. The issue is that the customer had a technical guy on his side and this guy was in charge of deciding how to deploy the solution, which servers, which infrastructures, and so on. And here started the real problem because the customer provided us the final decision with where the software is going to run just three days before they go live. So we didn't know where the software was going to run until three days before the public launch. And the issue is that as they were obviously short on budget because at the beginning it was just a proof of technology, they decided not to rent a real server. And this was actually my face when they told me because they decided to go for the most possible solution in this case. They went for a free solution on Heroku and Heroku doesn't support storing files on disk. Well, you can store files on disk. They will just disappear whenever the application launches. So actually we couldn't deploy the software on that platform because we stored a lot of files, we stored them on disk and we knew that whenever the application was started, the files would just disappear. So that was a huge hope. Right before the launch, remember that we had like three days before they go live of the world software. And so we decided to rewrite everything we had from scratch. Everything related to storing files, generating time, making them available, serving them, everything we used, just plain. We just relied on engines to serve them. We just saved the files on disk to serve them. We had to switch everything to another solution which could work with Heroku. In this case, we decided to go with GridFS, which is the file system storage of MongoDB. I don't know if any of you know what it is. Actually, because the application relied on MongoDB for the database, and MongoDB has support for storing files in MongoDB itself. And it's actually a really good support because it scales through MongoDB and it's pretty fast to serve them because it's just a key value storage. So you just put the real file and MongoDB will serve it. And usually it's really fast because it's going to serve it from memory if the files is able to stay in memory. The issue is that it was just a huge hack. We didn't have time, maybe we could have time to write it properly, but as we were in total panic, we just started to look for the fastest solution to make everything working. And so we monkey patched all the classes that were going to save data and replace them with something that saved on GridFS. And then we monkey patched our whiskey server to actually whenever a specific path was asked, it went to GridFS with the data and serve them back. So it was actually a huge mess and it went online with practical no testing because we finished it like the day before. We tried it on our testing environment, but we didn't try on the real world deployment. So we didn't have time to try it on another Eroco application, for example. And so we went online with just that solution. After we went online and thanks God everything worked. So we didn't have any major failure because actually what we did was pretty easy. We came together and thought that we actually needed a better solution. It was obvious for everyone in the team that this kind of thing should not happen anymore. We knew that the customer changed idea. We knew that we did the best possible things with the budget time and knowledge we had at the time, but still we had an issue. Still we did the wrong choice. So we wanted to find a solution that could work independently from the budget constraint, from the customer change of requirements and ideas. And we decided that this solution should be a tool that our developers could use and just rely on the tool and don't care about how and where their files are going. Everything related to storing files should be moved to the production, to the deployment phase, to the configuration phase and not to the coding phase. So that's how actually D-Pod born. We created D-Pod for that purpose to make our life easier to store files and be able to just say, hey, D-Pod store this file. I don't care about where you're going to store it. I just want you to be able to give it back to me when you need to serve it to the client. Actually, we wanted it not only to be easy, but of course to be fast enough for most web application use cases. And here starts the interesting part because I started to think how it was the best to design a framework that should be used in a web application environment and was related to storing files. There are a few things I learned by working on TurboGear Steel for a few years. TurboGear Steel has been used like since 2007, if I am not wrong. And so it evolved a lot. We saw a lot of changes. We started with a template engine which was named KID. Then we moved forward to Ganshee. And now Ganshee is not supported anymore, so we are going to move forward to Kajiki. And of course, every one of our users needs to be able to continue to run these applications. And for example, some of our users didn't like Kajiki and Ganshee and KID and used the Jinja too. Some used Mechro and so on. And we needed to be able to support all of them and let the users work with all of them. So what I learned is actually that web applications, at least for the part of developing them, are much like a little kid. They have a lot of issues. They want things like they want them to be. And they might change their mind like every five seconds. Okay? Whenever you are working with developers on the web world, the web world is really fast. So your infrastructure might change anytime. You might start small. Then you have like 10,000 users the next day and you need to scale and change everything in your infrastructure. And I started with a specific technology. You decide to go with storing files on disk. And then the next day, you need to change to MongoDB for storing files because you need to scale or your developers just don't like the previous idea anymore. Or maybe the library you are using has died, like in the case of KID when we switched to Ganshee. And so everything you do for the web world requires to be far more able to change on real time while on production because the web world environment changes pretty often. Okay? For various reasons. Not all of them are good. Sometimes it changes just because it's cool to switch to a synchronous technology or things like that. But whatever. You user want to be able to change what they are working on. And the third point is that automatic testing is actually something which is done for real on most web applications because it's easy to simulate the environment. It's easy to perform a request and check the response. So most web applications want to be able to provide automatic tests and test with. So wherever you write a framework for the web world, it should make really easy to monkey patch the framework. Well, monkey patching is the wrong term. But to drive the framework in a way that it's good for making easy to write tests. So to simulate the production application without needing the world production infrastructure. I'll make you an example. SQL Alchemy is really good. And one of the reasons why it's really good is that it's able to work on SQL light. Because when you write tests, you don't need to set up a whole SQL environment or Postgres environment just to run the test suite on your computer. You can go with SQL light or you can even go with SQL light in memory, which doesn't even need to store your database at all. When we decided to choose a MongoDB support library for Turbogears 2, because whenever you start a new project in Turbogears, you can choose to go for SQL databases or MongoDB. We decided to go for Ming because Ming had a feature which is called Mongo in memory implementation, which made it possible to write test units without needing MongoDB at all. It simulated the world MongoDB server in memory. So you can create a record, check them and so on without needing to even start MongoDB. And the approach should be able to do the same thing. I want to be able to save the files without needing to actually start the file storage itself or without needing to actually upload them on S3 if I'm going to use Amazon Web Services. So, and the last point is that what I learned is actually that making things really simple and easy to use, wins over providing them a huge amount of features. Providing a huge amount of features requires a real big investment in trying to keep them together, moving them forward, keeping them in shape and so on. And usually you are not able to cover all the use cases on all the features because maybe you are going to use just 20% of the features, but there will be one of your users which will rely on the other 80%. So just focus on the really important features and let your users write extensions over them. If the good foundation is solid, then people will start relying it for writing their own extensions. This is one of the reasons why, for example, DIPOD doesn't have a file system structure. It doesn't have directories. It doesn't have the concept of collections of files. You just store a file. You want a directory. You want a hierarchy. Write it yourself. It's not hard to set the pointer to the file somewhere where you can have the hierarchy and so on. And in fact, there is a guide which brought DIPOD FS, which is an extension for DIPOD that provides support for file system-like because it works also on things like read FS which do not provide the file system at all. You just can save that file and you cannot say, I want to have a group of files in any way. So the first thing we focused on is to allow for infrastructure changes because that was our first problem. We faced that problem, so we knew pretty well what we needed to check and what we needed to do. So the first thing, three things we decided to do was to allow to configure multiple storage engines. So whenever you use DIPOD, you can say, hey, I want to save something here, something there, something else there, too. I want to have three different storage engines because I want to use locales and also read FS and also Amazon Web Services as free. And we wanted to be able to switch storage engines at runtime with a graceful restart, of course, not that you can actually switch it in your configuration without starting the web server unless you properly write some checks. And it didn't have to, it should continue to keep working on the previous uploaded files. So you can say, hey, from now on, upload files on read FS, but everything I uploaded on the disk should continue to work. And DIPOD will do that. And we wanted, of course, to be able to rely on multiple storages concurrently. So not only you could have read FS as free and whatever, but you could also use them in your application at the same time. And this is because actually it happened for real. One of our users came and said, hey, DIPOD is really cool, but I want to store my avatars here, my items uploaded on my social network there, and whatever is a temporary file for my own use should be on disk, too. So how can I use three different storage engines at the same time? And this has been like the second question we had on DIPOD. So it has been a real need from one of our users to be able to use multiple storage engines concurrently. So whenever you upload a file, if you do not specify anything, the file goes to the default storage engine. If you specify something, you can drive the file to be uploaded on a specific storage. And storage are actually identified by a name. So that storage right now can be on read FS. But if you configure a new storage, which is named the same, but is on S3, your old files continue to be served from read FS, and whatever you upload new will be served from S3. Because DIPOD knows that they are on files on read FS, and the new files are on S3. And you are still using the storage which is named avatars in case of user images. And then you can, of course, use multiple of them during runtime. And that's made possible because DIPOD, as I told you, has no concept of a file hierarchy. So it's able to identify files by an ID. And the ID is paid to the storage name. So every file is uniquely identified by an ID and the storage name. So as far as the storage has the same name and the file has the same ID, it will be able to look up for that file, even if the underlying storage changed. Okay. And the other part we wanted to do is provide a really easy way to use everything. So we provided something which is called the DIPOD manager, which is in charge of actually doing all the configuration so that it could work on practically any web framework. And we were not bound, for example, to using the INI files, which is what we use in Turbogears for configuration. You could use Yamle or whatever you want for storing configuration, or you can even write the configuration in Python itself. Because the DIPOD manager is the one in charge of keeping the real configuration and is able to load it from various sources or from dictionaries or from whatever. And it keeps track of what you have currently active and configured. So whenever you need something, you go to the DIPOD manager and say, hey, DIPOD manager, give me this storage. I don't care where it is, how it's configured, and how it works. Just give it to me and I will save a file there. And if you don't want to get any specific storage, you just ask for a storage and it will provide you the default one. So this is an example from the documentation of DIPOD, which is the most simple case. We are just getting, configuring a storage, getting the storage itself and storing the file on the storage. So you can see that the configuration in this case is made through a dictionary and we are configuring a default storage, in this case in name default. And the storage uses the gridFS backend and provides some additional options which are related to the backend itself. So in this case, it provides the MongoDB URL. Then we get the storage itself. In this case, we don't specify any specific storage, so we are actually getting the default one. And then we just created the file. Whenever we create a file on the storage, we get back the file ID and we can look back for the files through the.get method of the storage. So you see that the interface is pretty similar to dictionaries. Just create something, you get it back by key. Nothing more, nothing less. This is the core foundation of DIPOD. And over the core foundation, there are more advanced things, more complex things. We focused on providing a solid foundation on which we could actually implement more advanced features. And one of these features is the support for database systems. Like in this case, we have support for SQL Alchemy. So you want to store a file which is somehow related to your model, like in the case of a user, you have the avatar. And you want to store the avatar inside the user. You just declare a column which is of type, upload a file field. And you can specify the uploaded type. In this case, it's an image with a thumbnail. So whenever you upload the image, it will also get a thumbnail too. And then whenever you save your document or user, you just assign the photo to the file and DIPOD will upload it on whatever system or whatever storage you wanted to or if you don't specify any on the default one and we'll link it to the actual model itself. So I told you that one of the things we learned is actually the web application changes so often. Maybe the developer changes, maybe the technology improves, whatever. So it should be easy to support different technologies. So in DIPOD, we focused on making everything a layer over a layer. For example, we have support for SQL Alchemy attachments. We have support for MongoDB attachments. We have support for storing files on S3, local files, and grid.fs. And we have implemented everything as plugins. So if you want to support storing files of your own system or whatever you invented yourself, you just write the plugin and everything else in DIPOD continues to work. The SQL Alchemy support will continue to work even if it's written on your own plugin because you just need to implement the storage engine and nothing else. And there are even files made by a whiskey middleware. So you can use it with any web framework. We use it with turboguils, but if you have a flask user, you can just attach the framework to flask and go on. Actually, most of our users are actually flask users because currently it's what most commonly used for web APIs, I suppose. And then it works together with your database. If you don't know this, it's actually a real query. It's called the query of despair. It's a really, really long SQL query. And what it means that it works with your database? It means that it copes with your transaction, for example. You uploaded the avatar of the user by saving the user phase, updating the user phase. Your transaction gets rolled back as far as you have a transaction manager properly working. And then the text that your transaction rolled back and will recover the previous states of the files. So if you try to save a new state of the user, and the state includes a new avatar and a new name and surname, and storing the name and surname fails for whatever reason, maybe a dialogue on something in your query or whatever, the port will detect it and will recover the previous states of the avatar too. So you didn't save things, alph, only the avatar but not the name. Your models will change in a proper way. And whenever you delete an item, it actually deletes the attachments only if the deletion of the item properly worked on the database. If you fail to delete the items, you don't end up with an entry which is in your database but you don't have the avatar anymore. So the text, the transaction failed and will recover the files that wanted to delete. And the last thing is that it should be really easy to extend. So we focused on two types of extensions to provide additional behaviors over the depot. One is attachments themselves. So whenever you provide an upload file field, you can provide an upload type. The attachments are actually in charge of changing the file itself. So whenever you want to replace the file with a new file, you want to go for an attachment type. And then you can, over attachment type, you can also provide filters. Filters do not replace the file itself. I'm not able to change the content itself but they can add additional information to the content which might be additional metadata or additional files in this case. And you can, of course, apply multiple filters. For example, you might have a filter which generates thumbnails and you might apply four of them because you want the small, medium and big thumbnails. And you just declare the same filter three times with different construction options. And you will end up with three different thumbnails. Let me show you a real case of an attachment which is took from the documentation of depot. And the interesting part is actually that they not only can change the content itself but they can also add additional behaviors to the files. What does it mean? It means that whenever you recover the file from your file system, it will be converted to that upload type. So if you upload a type provided additional methods, like, for example, I don't know, give me the histogram of the image, you can call them on your already stored files. So depot will know the original type of the upload and we'll be able to recover its state and provide all the additional features and behaviors to your file set, not only to just change the file itself. Or, for example, if you want to add additional information, like you want to store not only the file but also, for example, the primary color, for example, if you want to look for the images which are red, you can store that inside the file as a metadata because depot keeps tracks of the files and all the metadata over the file. So you can add additional details over your files. And this is the example of a custom attachment. In this case, it's uploaded an image unless it's bigger than a specific resolution. In case the image is bigger than that resolution, it gets shrinks to that site. So the first thing we do is getting the content itself and its data. And this is done to help our functions because we don't know what the content is. We know that depot is going to save files but we don't know what the user is going to provide us. For example, it might provide a file, it might provide bytes in memory, it might provide a byte IEAO, it might provide a GGI file field if it was something uploaded from the web. And we have these pretty convenient functions, file from content, that whatever is the content will convert it to a proper file. And it's pretty efficient because users in memory storage for files which are smaller than the sites and then it stores them on this only if the site is bigger than the maximum site. Then we open the image, check for its sites. If the site is bigger than a specified limit, we create a new thumbnail for the image of the maximum site and we replace the content. We see that in this case we replace the content variable with a spool temporary file which is that kind of temporary file which stores everything in memory until you make the data bigger than the maximum sites you specify. And then you save the image itself inside your spool temporary file and then gone and provide to the process content the replaced content. So you just call your parent method with the new content and in the middle you can do whatever you want because the real logic of saving the files is inside your parent implementation. Moving to filters, we already know that we already know that attachments can have more than a filter and we already know that they run after upload. So while the attachment itself runs during upload, this call runs before the file gets uploaded. So in this case this is by design because if we fail in generating the thumbnail, we do not want to go on and store the data in the database for example and we end up with a user without avatar again. So if the avatar for user fails, the depot crashes and you won't have the user created at all. So not only if writing on the database fails, the depot recovers the files but also if creating the files fails, you have a proper exception before saving the data to the database. So we try to do the best we can to keep in sync the few things. If any of the two fails, you don't have done anything. You haven't done anything at all. Then so in the case of filters, actually you do not work before uploading the files but after. Why? Because filters usually provide additional behavior-saving information. So in case it fails a filter, it will just go on and provide the details that the filter failed but you already have the files. So you can recover the additional information from the existing file. So even if the secondary thumbnail fails, the medium-sized thumbnail fails, it's not a huge issue because you can recreate that medium-sized thumbnail from the original data. And as I told you, you can add additional data to your files but in the case of filters, not behavior. So you cannot add additional methods to your object through filters. And here is a simple example of a filter which actually saves the thumbnails for a specific resolution in a specific format. And you see that we just receive the on-save event and inside the on-save event we have the uploaded file and at the end of the code, which mostly just creates the thumbnail, we just add to a uploaded file any information we want. In this case, we add the thumbnail ID, thumbnail part, and thumbnail URL to the uploaded file. So uploaded files work like dictionaries. You cannot add anything you want to them and you have the file itself, so the content and all the metadata you added to the file. When you look back at the file, so you query it back from your database, you just have the thumbnail URL property because we added it here at the end of our code. So you just get it back and look for that property. If the thumbnail URL is none, probably if your thumbnail failed and you can recreate it from the original file. And one of the core parts of the point is that it's meant for the web, specific to the web. So we wanted to make easy to use content delivery networks and we wanted to make easy for people to rely on the pod for saving data to the web. So everything which is needed for saving files themselves is provided by the pod itself. So when you specify the pod already gets the content type, the last modified time, the content length of the file itself, and the file name. So when you save it back, you can properly add the either of the HTTP either for that file without having to work on them yourself. And we already know that whenever you want to save them, you just rely on a whiskey middleware. So you just create, make the middleware and wrap it around your application and the pod will do the proper thing to save the files. And if they can, you are storing the files on support, HTTP itself, for example, in the case of S3, you can be sure that the middleware will not serve the file itself, but will redirect the user to the middleware itself. So in case of the content delivery network, you will end up serving the files from your content delivery network. So please try it. If you have questions or anything, let me know. If you find bugs or anything, I'll be more than happy to fix them. Everything is supported from Python 2.6 to Python 3.4. We haven't tested it on 3.5, but it should work. Everything is fully documented. So if you find something missing in the documentation, let me know. We will cover it. And everything is tested with 100% coverage. So you can be pretty sure that it works. And we are already using it in production on various environments. So try it and let me know. Thanks. APPLAUSE. Questions? Okay. Do you have a microphone for them? No, for them asking. But, yeah. Yes? Okay. Okay. He asked how much cost, how much effort would be required to make it work on an asynchronous framework. Well, we used it on production on G-Event, but G-Event is not a really asynchronous framework. It's far different from a Twolip or Twisted, because it's implicit and not explicitly. So I'm not sure how much it would take to adapt the middle of itself to something like Twolip, which would require to move from function to coroutines and so on. But it should be fairly easy, actually, because it just gets the files and sends it back to the content to the browser. So it's a pretty good use case for an asynchronous framework. And the middle itself is just one underline of code. So even if you have to write this from scratch, it would take like two hours, not more. Okay. So the middle is already divided in utility functions. So the code itself, the file is like ten lines of code, which you can probably move to Twolip or something like that. But I haven't tested it. Only use with G-Event. And I know that only G-Event works well for... She works before. Okay. We have it. So you mentioned that in case of a rollback, you restored the files. So do you need some sort of storage for the depot itself or some metadata? No. Actually, what happens is that the depot generates a unique ID for each file. So if you create a new version of the file, you actually end up with a different ID. And the old ID gets deleted only when the new one, when the transaction gets committed. So for a time, when you have the transaction rolling, you have both the files and they have two different identifiers. If the transaction goes on and successfully commits, it will say, hey, this new one is the proper one, delete the old one. If the transaction rolls back to say, hey, the old one was the proper one, delete the new one. So it just keeps both the files available at the same time and then decides which one to keep at the end of the transaction. And you mentioned it is transparent to switch from one type of storage to the other. So when you get the request for a file, how do you know if you need to store it, to serve it from the old storage system or the new one? Okay. That's actually stored in the file metadata itself. So every storage engine needs to provide support from some kind of metadata. In the case of GridFS, it stores the metadata together with the file on the DB. In case of S3, it stores the metadata as HTTP of the file itself. In case of the local file system, it saves a GZOM file with the metadata and so on. Every storage engine is in charge of providing a way to add metadata to the file. And then the report will rely on the metadata to know from where it should serve the file itself. But when you get the request from the user, you only know the file name. So how do you know what's the storage? Not really because when you store the file at the low level, we only know the file name. But if you bound the file to a column of SQL Alchemy or MongoDB or whatever, inside the column it gets actually stored the JSON with various information, including from where to look up for additional details of the file. So if you use DIPO at a low level, yes, you have to provide the fallback to your self. But if you rely on the high-level APIs, they already provide it for you. Okay. Does it out of the box support uploading to a temporary URL on something like S3? Sorry. I didn't understand that. On Swift and I think S3 as well, you can be provided with a temporary URL to upload. Directly from the client. Does it support that? Okay. I understood. Now, currently, as most of the logic happens in DIPO itself, the client needs to upload the file on your server which processes the data and then uploads it on S3. You cannot directly provide the data on S3. As otherwise, you will lose all the metadata that DIPO calculates for you. We will need to provide the data on the server. We will need to provide some kind of DIPO support in JavaScript itself so it can get the metadata before uploading that. I don't know if we have more time. We can ask outside of the room. Thank you. Thank you.
|
Alessandro Molina - Why storing files for the web is not as straightforward as you might think. DEPOT is a file storage framework born from the experience on a project that saved a lot of files on disk, until the day it went online and the customer system engineering team decided to switch to Heroku, which doesn't support storing files on disk. The talk will cover the facets of a feature "saving files" which has always been considered straightforward but that can become complex in the era of cloud deployment and when infrastructure migration happens. After exposing the major drawbacks and issues that big projects might face on short and long terms with file storage the talk will introduce DEPOT and how it tried to solve most of the issues while providing a super-easy-to-use interface for developers. We will see how to use DEPOT to provide attachments on SQLAlchemy or MongoDB and how to handle problems like migration to a different storage backend and long term evolution. Like SQLAlchemy makes possible to switch your storage on the fly without touching code, DEPOT aims at making so possible for files and even use multiple different storages together.
|
10.5446/20063 (DOI)
|
Well, good morning and thanks everybody for attending this. As Pina said, my name is Alec Han-Roydow and I'm not a double... Let me introduce you to Fabian Croetz. He's going to give his speech on... So I was saying I'm not a top notch developer, neither a top notch system administrator. So that's the reason that my job title is the option engineer. And today I'm going to talk about extending and embedding Ansible with Python. Please raise your hands. Any one of you who has used Ansible before? I'll have a brief introduction, like five minutes, just talking about what Ansible is for people that are not familiar. I'll continue explaining how to do some hacks and how to do some leverages, improvements and integrations, which is the part of extending and embedding. First, the introduction, the 101 to Ansible. What Ansible is? It's a configuration management tool. You may know many of those tools. There's also Puppet, Chef, Saltstack, CF Engine. The basic idea is to be able to automate the configuration of hosts. By hosts, I mean virtual machines, bare metal, containers, everything can be managed. There are many alternatives. Basically, the strong point of Ansible is that it works over SSH out of the box, and it is agentless. That means that you do not have to install anything in the host system, apart from Python 2.x. Another very strong point is that it is quite readable, and it has a smooth learning curve. It's not steep because it's DSL, its language is Jamel-based, so it's quite readable, and it's easy, very easy to get on and start it. This could be the example of an Ansible architecture in which there's a central node called the management node of a controller, which could be a laptop, could be a server, in which you run the Ansible software. Then you will have some inventory files that describe the IPs and DNS names of the hosts and the groups of the hosts that you are managing. And then there are the hosts, which are remote machines that can be accessed over SSH, and there are a series of playbooks that are like scripts that the Ansible machine executes in each one of the hosts that is managing. So this could be a very basic host file in which we define two groups of servers, one of web, another of dv-servers. We can define variables that are common to a group of servers, for example the username of the dv-group, and we can define meta-groups, that group, another group, like this one, which is infrastructure group, that joins the web and dv-group. There are two ways to use Ansible, basically there are other commands and using playbooks. So other commands is the easiest way, it's the hello world that everyone does, and they are just throw away one-time executions in which you use the Ansible command in order to execute a module in a group of hosts. And the syntax is very simple, just use the Ansible, name of the group, as described in your inventory file, minus m, name of the module, and minus i, the options. This could be the hello world that everyone has done. We are using the module called ping, the only thing that does is respond with a ping in the server. So if we issue ansible all minus m ping, and all is a magical group that groups everything in the inventory file, then we would have this response in each one of the servers, we would have the answer that is ponged, and also there's a value that says that stays if anything has changed in the remote host. And the other way to use Ansible is to use playbooks. Playbooks are a much more structured way to do configuration management. They are basically jumble files that specify a list of plays, and each plays a series of tasks that are applied to a group of hosts. And each one of those tasks is an execution of a module with some parameters and a description. This could be one task that is ensuring that the NGINX server is started. There's the description, this is the module name, and those are the parameters saying that the service is NGINX, and the state we decide is started. This is an example of configuration management concept that is that we define the desired states and not the actions to be performed, so that this module is clever enough to know that sometimes it has to start the service and sometimes just make a no-op and go on. This could be a playbook to deploy the NGINX server. In this playbook we are targeting the group of hosts named web. We define some variables, and this is the list of tasks. As you can see there are modules to add repositories, APT repositories, modules to install APT package, modules to manage services, modules to create or remove files. In this case we are removing a sim link, and modules to template out things like a Jinja 2 template that we are going to template out in the remote host. Also there are handlers that are special kind of tasks that are run once independently of the number of times that they have been notified during the playbook. This could be the execution of the playbook. We will have an output that sequentially shows the results of the task in each one of the hosts, and finally there's a nice recap. Just to finish this introduction, the way to manage things is using a kind of encapsulation that is roles. Roles is a little bit more advanced concept in Ansible that is a group of variables, of tasks, of handlers, files and templates, and maybe even dependencies to other roles. So we have finished our introduction. Now we will be talking about how to hack Ansible. There are two ways, embedding and extending Ansible. So what do I mean by embedding? I mean calling Ansible modules and playbooks from your Python code. This is basically possible because Ansible is based on Python. It's written in Python, so it has some kind of a Python API. It's not very well documented, but it's quite easy to use. If you read the source codes, you will see how to use it, although you can just continue and see how we can use it in this presentation. Before we start, a few disclaimers. This code is valid for Ansible version 1.9, which is the current stable one. Scroll in version 2, which is the current development one. Things are going to change. This API is going to change. And probably even things like the plugin mechanism is going to change. So this is only valid. I can only guarantee that it's valid for the current stable version. And everything in Ansible is Python 2.x only. It hasn't been ported to 3. You can find the examples in this GitHub repo. So first thing that we may want to do, we may want to run an Ansible task. This is a simpler kind of automation. It's just calling a module from your Ansible code. How do we do it? Very simple. Just import some classes, basically run around the inventory from the Ansible library. You will build your inventory, for example, directly in code. We may be just targeting our local host. And then making an instance of runner and calling the run method. Just passing the module name, the arguments, the inventory that we have just created, and the pattern of host that we want to target. And just a little bit, a little discretion. I'm talking about facts. What are the Ansible facts? Ansible facts are information that is retrieved at the beginning of the execution of the playbook in each one of the hosts. And there's a bunch of information available for you to make things in your playbook. So you have the host name, IP addresses, the hardware information, even the installed version of the software. So I'm going to make an example, which is the Flask factor, in which there's a proof of concept of embedding a module. I'm running a program that will create a REST API using the Flask RESTful module that parses URLs in this format, getting a fact that we want to know about the system. And it will run the module setup and then pass the JSON that we have showing the fact in our system. It could be just not very practical, but it could be an idea of how we can retrieve information from a remote server. So if we run it in localhost, I can retrieve the version of the software, the base system that I'm running, the version of the kernel, or things a bit more complex like the network card information. Sometimes we need something more complex than running a module. We need some kind of orchestration. So we may need to run a playbook, which is a more structured way. So you have to import a little bit more classes, basically classes for the callbacks and UTIMS. You have to build your inventory. It's similar to the other case. In this case, we're also using host variables, because in this example, I'm going to attach some information to the inventory, the list of users and APD packets that we want to install and create. You have to put some boilerplate code just to set the velocity of the execution of the playbook and also to register the typical callbacks for the runner and the playbook. And finally, creating a playbook instance and using the module run, specifying the playbook. In this case, the playbook is the installer.gamel. In this example, what I'm going to run... I'm going to run a playbook that install some APD packages and create some users, because it's a typical task when you arrive at a system to begin creating users and installing software packages. So I've just created a proof of concept script in which I get... First, I get via the interactive console the information of the name of the users and the list of packages, and then I call the interval playbook, which is called installer.gamel, and I pass the list of packages and users via inventory of arrivals. So that could be an example of this kind of integration if I run it. I can create users... and I can install packages. And when I finish, it runs in localhost, and it has changed that... It has shown that although the current user Alex is created, it has done anything, but for the new user that I've created, which is called George, it has changed the system, also installing another package. So I have called an Ansible playbook from my Pethen code. Well, one way to hack Ansible is by embedding, another way is by extending Ansible. And by extending, I mean adding more functionality or customizing its behavior. Basically, it's done by three ways, creating modules, creating dynamic inventory scripts, and creating plugins. So let's start with creating an Ansible module. Ansible ships with tons of modules ranging from creating user names, managing databases, spinning cloud servers. But sometimes we need something more specific to our business. So we have to create a module. And what is an Ansible module? It's just an executable file that you will find in the... some folders in the dot library relative to your playbook or in the Ansible library path. And basically, it has a JSON interface, so we expect a JSON for the input, and also it emits a JSON for the output. So it's language agnostic. You can do it in BASH, you can do it in every language that you want. But if you are using Python, then it's easier because there are some helper functions that will make it really easier. This could be the structure of a typical module. I borrowed some parts from the existing modules. And basically, it's a file in which you have two strings. We will come later to them that are the documentation and examples. Then this part of the code is what I call the Python part, in which we do the hard work for the libraries, and we run our business logic. And then from that point, this is like a template, let's say, in which we find a main module. We do some instantiation of the Ansible module class. We call our Python part from this function, and then we emit a different kind of JSON depending on the result. So let's get more detail. Commutation and examples are two strings variables that are very important because they are used by Mac WebDocs to generate the HTML of the documentation. And some important parts are the version option part, which will specify the values of the arguments. And it's important to keep it in sync with the real code, which is the argument spec dictionary. You have to specify the requirements, even some more Python libraries to be installed in the controller. And you have to use the node section if you need that some environment variable, for example, is present at the controller. And in the examples, just put code, please, that is tested and worked. Especially if you want to submit it to the Ansible repository. In the part that I call the Python part, it's usually a good idea to put pure Python code in all the libraries that you may need for dealing with the problem. But try not to use the Ansible part because it will make your code more robust to change this in the Ansible API. However, there are some helper functions, for example this one coming from the Ansible module, from the Ansible library that allows you to easily run commands and find executables in the remote machine. And just some tips. Try to return a value and a message, a meaningful message, and code every information on those values. And don't print to standard output or error because this mechanism won't work. And in the main function, you have to create an Ansible module instance. And when you instantiate it, you pass an argument spec dictionary which defines the module arguments, defining the series of modules that are required, the ones that are optional, the default values, possible choices, aliases, and then there's a section option which you specify the module exclusion of the parameters. And also if you support check mode, this is the mode in which you do a drive run and you do not mess with the real system. It's just a test. So when you have created that, you magically have a dictionary in which you have the parameters of the invocation of the module. You have just to take those parameters, call the Python part of the module with those parameters, and then pass the results depending on the status and the message. You have to create some kind of JSON. And if you use some helper functions like this one, exe-jation and fail-jation, then your life will be easier because they manage everything and create a JSON for you. And finally, probably you have seen those two lines at the end of the file. If you are a Pythonist, that you are. Probably you are now in tears and in pain because this is the Pythonic anti-pattern, the import asterisk. But please resist the temptation to make an explicit import because really it's like a define, a preprocessional in C because Ansible will substitute those lines with the code of the helper functions. So if you do not put it, it won't find it and you won't have an import. And if you do not put it at the end of the file, it will change the line numbering and debugging will be a hell. So some creation tips. To make a module you will love to use, a module that is edam.pdent that supports the check-mode. Test your module. There's a very handy tool included that is the test module script. And if you want to submit it, please follow the module creation checklist that is available online and it's much more comprehensive. I've created a module, an example, which is called the Tiger issue. Tiger is a agile project management system that has a REST API and there's a library called Python Tiger that allows to manage it and to play with it. So I've created a very simple module. That is called Tiger issue. This module is just for creating issues in this platform. For example, if we are deploying a system by an anteval and we find some problem during the automation of the task, we can create an issue so that the team will see it, will have the information, will have the logs, will have everything. So I have submitted it to Ansible extra module repository. And basically this is the documentation part. You have to specify the different options that the module supports. And you have to put some examples. This is what I call the Python part. I import the Python Tiger and I'll do the issue management here. It should be the pure Python code. And finally, this is the main part in which I specify the restrictions and the parameters. I part them and I call the Python part with arguments and depending on the return status, I issue one kind of JSON or the other. And if we see the playbook, we can just use this module to create Tiger issue in our project. We can manage the description. We can put also Ansible variables like the host name and distribution. We can attach to the issue a file. For example, we can attach the playbook. Then I will pause and when I continue, it will delete the Tiger issue because the module supports the state present and absent. So this is the test project that I've created with no issues. And then if I run the demo, it calls the Ansible playbook of the playbook that we have seen. First, it has run the creation of the Tiger issue and it has stopped. If I take here that I can see the issue. This issue has some tags. This is my Ansible distribution and this is the playbook also. So if I resume, it will go and delete it when possible. It's not there. So creating a dynamic inventory script. There's another way to hack it. If you are managing cloud servers and cloud infrastructure, probably you know about this. Dynamic inventory scripts are a way not to have to deal with a long list of servers that probably are changing. Their IPs are changing. You can scale things. So dynamic inventory scripts are a way to deal with that complexity. And basically, they are just an executable file that supports those command line features. Minus, minus list that will return a JSON dictionary with the name of the groups. And each group is a list of hosts and minus, minus host in which you specify a host name. And then you get a dictionary of the host variables. Just pure JSON. So just for the fun of it, I've created an example which is a shelf inventory. Shelf files are basically a key value store that Python supports natively in the serialization. So in the example, we are just using a shelf file that I've created. And we can open it and set the groups on the host version of it. I can get past through the demo because we are running out of time. I want to run another demo. Plugins. Plugins are a way to hack Ansible in very different ways. Basically, the common thing is that they run in the controller node and there are different kind of plugins. We're going to see some of them. And basically, the way to add them is just drop them in a folder. And if you want to use another folder, just treat the Ansible config file. So the callback plugin is just a kind of plugin that reacts in the controller to the Playbook and Runner events. In order to create them, you just have to define a callback module class and then override the method you want to. And this is an example of the list of methods available for this class in the repository. So just a brief example, I'm going to use to create a callback plugin that reacts to the failed event of a module. And then it will call the notify send binary, which will create a popup in my system with the name of the module and the result of the module. So, I'm going to use the callback plugin. I'm running this Playbook with just output message. The first task will go up. The second one will fail. So, I'll get a popup saying that the second one has failed. It's just an example of things that can be done. You can make Slack notifications. You can make HipTag, everything you see it useful. You can create connection plugins. Basically, the connection plugins allows the controller to connect to the remote host and Ansible ships with lots of them. But you can create much more. You'll just have to define a connection class and override the basic methods that are just connecting, disconnecting, executing commands, putting and affecting files, which are basically things that Ansible is doing in SSH mode. You can create lookup plugins. Lookup plugins allow to expand the functionality by accessing information in external sources from the controller, for example, in databases, in file systems. And the same task is just calling the lookup function. And also, if you define a lookup, you'll get a width expression that allows you to loop over some results. So, you just have to create a lookup module class and basically define an init method and a run method, which will pass the terms, which are the functions, the arguments of the lookup, and then return a list of results. Following the example of the shell file, I can define a shell file lookup that will search for a shell file and retrieve a key from that shell file. Basically, I just implemented the run method. And just add demo. In this paybook, I'm opening the book.dv, retrieving the current book name, page and book author of the book I'm reading now and print them. It's a good book, if you haven't read it. And you can create filter plugins. And you can define some other filters. The syntax for using them is very simple. Variable followed by a pipe and the name of the filter. Those are example filters that are simple comes with them. But you can define many other. So, it's very simple. Just create a filter module class that has a filter's method that returns a mapping from the name of the filter to the Python function that implements it. For example, we can create a filter to rotate 13 positions. This is the C-star rotation. And just if we see an example of playbook, we can apply once and then apply twice. And we can see that C-star becomes PNRFNA and then becomes C-star. So, it's working. Finally, just other plugins. Very fast action plugins that are plugins that allow separation of actions between the controller and the remote host. For example, when you're templating out a file, there are some actions that need to be performing the controller and some other. Different in the remote so that you can just implement an action module class with a method run. And when you're dealing, doing things in the controller, then you execute this method, run your execute module to call the real module on the remote host. You can define bar's plugins. Quite an undocumented part, possibly going to change with sensible version two. And it's the way that we can retrieve more information about the host by some external source like the host vars and group vars directory. Basically, this is the template that you will find in the repo. And you have to implement the get host vars and get group vars for a new plugin. You can create cache plugins, very modern functionality from mcvill version 1.8 in which we can retrieve the information of hosts that have not been contacted in the current playbook execution. Just using a backend like a redis or memcast, you can run once, get the facts and then runs without having to retrieve the facts of the rest of the host. You have to tweak your answalt config file. And basically, if you create a new one, you want to create a new one, you just have to implement this template. And finally, those are the references that I have used. These are very good books, not just for developing answalt but just for using them, especially the first two ones. And some articles talking about that. So that's it. Thank you very much. And now if you have any questions. Thank you very much. Well, thanks for the talk. And my question is, you mentioned that while using Ansible as embedded module, it's very hard to find the documentation of API or is it very poorly documented? So actually, how do I get this information to use it as embedded module? Okay. If you want to embed a module like in the first example that I have said, I have done. I had to do some research. I had to go into the source code and see how the Ansible command is using the API and try to mimic it because we are using the same API that the command line version that is just an icon file is executing. Also, I have seen some other examples which are here in the reference that are much more pragmatic examples. So I have been inspired from them. I think that in version 2.0, they are going to document it much better. They know in the Ansible project that there's a lack of documentation in that part. I don't know if intentionally, but there's a lack of it. And it's changing. So when Ansible 2.0 comes as a stable one probably, it will have changed. It's not going to be, I think, not so simple, but more flexible. So what about the way of testing modules? Do you need to write tests in some specific language, like or library for testing the code or standard tests in Python? Well, testing things in Ansible is quite a tricky aspect because even the Ansible code base is not very well covered with tests. If you see the project, there are just like 100 and something assertions on the base code. So they are not very concerned about the testing. And testing a module is even much more difficult. I would suggest you to test the Python part separately. Just those kind of part, which is pure Python. You can use PyTest. You can put the tests in another files. But there's no standard way to run programmatically tests to a module. And you do not have even to make tests to put tests in a full request to get the module accepted. They are not very strict with this part. So I would suggest using testing only this Python part, the business logic of your module, using external library like PyTest, something like that, putting in a different file and importing the functions and the classes that you are defining. So maybe it's easier using something like Test Kitchen. Like? Sorry? Like Test Kitchen. Test Kitchen. Well, with Test Kitchen you are not using, like, I think it's unit testing. It's more like, I don't know if it's functional testing or something like that. Test Kitchen, just to focus, is it a tool developed in Ruby for testing a puppet or things like that? Actually, I don't know. I was wondering if you can do a test of it. I see that it has a lot of parts. I didn't know that it was possible to reuse. Thank you. I'll have a keep an eye on that. Any more questions? Okay. If there are no more questions, then please thank Alejandro.
|
Alejandro Guirao Rodríguez - Extending and embedding Ansible with Python [Ansible] is the _new cool kid in town_ in the configuration management world. It is easy to learn, fast to setup and works great! In the first part of the talk, I will do a super-fast introduction to Ansible for the newcomers. If you are a Pythonista, you can hack and leverage Ansible in many ways. In the second part of the talk, I will describe some options to extend and embed Ansible with Python: - Embedding Ansible with the Python API - Extending Ansible: creating modules, dynamic inventory scripts and plugins Previous experience with Ansible is advised in order to get the most of this talk, but beginners to the tool will also get an overview of the capabilities of this kind of integration.
|
10.5446/20062 (DOI)
|
Hello everyone, my name is Alejandro Garcia and I'm going to talk about Python game development. Here you have my Twitter and email, so if you have any questions or something to ask me or whatever you want, just feel free to send me an email or whatever, okay? So the contents of this talk is, first I'm going to talk about how Python is being used currently in video games. And I'm going to show you my own game framework that I called it, Cobra. Okay, so how Python is currently used in video games? There are two ways of using Python in video games. The first one is using Python as a secondary language for scripting. And here the main game is programmed in another language like C++ for example. And the Python interpreter is in-vit inside the application. So the application calls that Python calls for certain actions such as what happens when two actors collide or so on. But this wastes Python potential. We don't use Python as much as we should. And Python is designed to be extended, not in-vit. So some examples of games that use Python this way as a secondary language for scripting is C++ for and MontanBlade. The other way of using Python for video games is using it as a primary language. Here the Python interpreter runs the main game loop, okay? And some buildings for improvement performance can be used such as C++ or so on. And we have a little slower than the previous example. So some examples of games that use Python as a primary language are E4line and Metin2. Next, these are the most used Python libraries. As a Python for primary language, we have Pygame that is very easy and very popular. You may know about it. And we have Coq2D that is a bit more complex than Pygame, but it has more features. For 3D, we have Blender3D that, as you may know, it's a 3D modeling tool. But it also has a game engine, okay? But the downside that Blender3D has is that it's GPL so we can make commercial games from it. And we have Panda3D that is a Disney's creation framework that was used for games such as Python. So, okay. Developing games with Python. Python is a very good language such as Csharp or JavaScript, okay? But the only problem is that there are very few game frameworks. Csharp is used for Unity that is very popular. JavaScript is used for games, but Python is now not currently used. Not because it's not good. It's because there aren't any game frameworks. And some of these frameworks are a bit limited or half an hour, so I'm going to explain Cobra. That is my game engine. It's an open source 3D game framework, okay? And it's three things. Dynamic, efficient, and easy. So why is it dynamic? It is dynamic because it uses an ECS architecture. I'm going to explain later what this ECS architecture is. So it's designed for real-world game development. As you may know, in the real-game game development teams, the game changes a lot because mainly of design. So this architecture helps us to change the game more easily. And this makes our game easily adaptable and extendable. So you may ask now, what is this entity component system? Well, this is a system for entity component system. So I'm going to show you an example. Here is, for example, this thing that we are working for EA, and we are developing the new Star Wars Battlefront. So we have our game entity, and from then we extend our ships, and that ship can be either enemy ships or player ships. And these enemy ships can either be enemy X-wing or Y-wing, and the same with the players. We are very happy our game runs very well, and suddenly our dear game designer comes and says, hey, what's up, Rambo? We have to change the game, so now we want a lead ship. And then we just have to rethink again all the architecture, we have to rewrite a lot of things. This is the new architecture that we have with an AESid and so on. And it's very probable that our game designer will come again and change our game over and over. So I wonder, is there a better way? Well, I have the pleasure to introduce you to the entity component system. It is used in modern game engines such as Unity 3D or Unreal Engine 4. And the basic concept is that the game entities that goes on our game are just a list containing components that make up that entity. Nothing more, it's just a list. So our previous example, our enemy X-wing, will be a game entity that has an enemy control because it's controlled by the enemy. It has a machine shooter because it shoots missiles, and it has a thruster movement because it moves forward. So these components communicate with each other via messages. So here, for example, our component enemy control wants to move forward. So the messages are sent to our entity components and then are consumed if needed. So the message shooter receives forward, but because he doesn't need to know if it's moving forward, then it ignores it. But our thruster movement says, oh, okay, so you want to move forward, I'll do it. And our X-wing will move forward. So our architecture of our previous example will look like this. Here we have our enemy X-wing that is a game entity that has a control, a message shooter, a thruster movement, as explained before. And the only difference from the Y-wing is that the message shooter is replaced with the message shooter. So making this with components, we can just reuse these components in different entities. And here, for example, the project X-wing and the enemy Y-wing, the only difference is that it's controlled with an AE, and the player is controlled with keyboard control, and the same goes with Y-wing. So what do we achieve with this entity component system? The enemy AE doesn't need to know anything about the entity weapon, or even if it has one. He just sends the message, and if there is something like a missile or blast or whatever, it will solve, for example. If there is nothing, it will do nothing. So we can quickly switch the blast with missile, with the changing anything of the code, making the code much more flexible to changes. And our friends, the designers, want to hurt us as much as they usually do. So I talk about entities, I talk about components, now I'm going to talk about the systems. So components don't do the hard work. Instead, they send commands to the systems. These commands are acute, and then later on, they are executed in different threads. This allows to use multiple cores using the full matching potential, and we don't have to, because each system does a separate thing, we don't have to worry about interlocks at all. So core is efficient, why is it efficient? Because it has a C++ core, and it's multi-core. So here's Cobra, uses Python, and Python communicates with C++ that has components and systems. You can also program your own components in Python or C++ and your systems on C++ or Python. And you may ask now, what about the GIL? Who doesn't know what the GIL is? Who doesn't? The GIL stands for Global Interpreter Lock. It's an implementation feature of C Python that doesn't allow us to run on more than one core. So how do we solve this? Well, Cobra is within C++ mostly. So this is how the main game loop goes. For each entity in Python, we update it, components we enqueue commands to the systems, and then when everything is done, we update the systems. And for the update, we will go to C++ and spawn a thread for each core of the system. After that, we update the system, and we join. And we can go back to Python without breaking anything of the GIL. This runs over and over on the game. So Cobra is easy. I tried to... I was inspired by Django to make this framework. So I'm going to explain now a little example of how to make a simple project with Cobra. So this is a simple Cobra project. We have five files. We have behaviors, controller, entities, scenes, and settings. In the entity file, we define each entity and the components you will have. Here, for example, I'm making a cube that it has three components. It has a Cartesian transform. This allows this entity to be placed in the world of the scene. And a renderer that allows us to draw a 3D model. In this example, I'm drawing a cube mesh and a behavior that makes the cube to rotate. So these are the most important components that Cobra has. Sorry, it doesn't read. Okay, for a space transform, we have a Cartesian transform that allows to, as I explained before, it allows that entity to be placed in our world. A polar transform that is just the same, but instead of using a Cartesian transform, it's used a polar transform from the region, or a screen transform that is a transformation in the screen space. And this can be used for user interface things. So these are the rendering. Okay, sorry for this. We have a mesh renderer that is for rendering meshes. We have a mesh animator that animates with the mesh. And we have a billboard renderer that will render a billboard. And a billboard is just a plane that always faces to the player. Okay, so for physics, we have read body, sub body, and box collider, for defining what kind of collisions we will have. And for audio, we have audio source that is from where the audio is being played, and the listener that in most cases, it will be our player. And there are much more components that you can use in Cobra. So, behaviors. In the behaviors file, we define the behaviors. Behaviors are components with refined messages. The most important ones are start, sorry, update, and input. In the start, it calls when the entity is created, update calls in the stick of the game, and input is when input in the game happens. So with Cobra's and inMessages, it's very easy. We just call a set of components, now the name of the message, and here we put which variables for output we want from that component. And here in the keyword arguments, we put the input variables. I'm going to explain now. Okay. Here's our last example, rotating behavior that just makes that entity to rotate. So here in the update that runs on the stick, I'm sending a message to get the current rotation. And as you can see here, I'm creating a big that will be the output for that message. And now I will have here my rotation. And with speed, with that speed, I'm going to update it. So, and now I send this message, set transform with our new rotation. It is also possible if you don't like to send messages to get a specific component and call a specific function on it. So you can just get the transform and call the rotation to get the current rotation. And here updated just the same way as before. Sorry. Okay. Controllers, we define mapping for different device inputs. Here I'm getting the keyword and XBOS controller. And here I define what will be our game controller. Our game controller will have a fire button and X axis. And here we make the mappings saying that this fire button is going to be a digital input that will map to the key from the keyboard and the e-button to the XBOS controller. So this way we will just check if the fire button is pressed and we don't have to worry if it was from the keyboard, from the XBOS controller, Kinect or whatever you want. And the same goes with X axis. This is an output. It was just the same. And the same file, we define the game scenes. Okay. So here I'm making... sorry. The scenes have two important methods. First they start that is called when the entity is created and then update that is called in each tick of the game. Here I'm creating the previous queue. Okay. And I'm calling each component to set the transform to the position 0, 0, 0. And in each update of the scene I'm displaying the current FPS of the game. And settings.p that is very important. It contains the game configuration and the system models that we want to use. I want to explain that later. Here we set the game name, in this case my name, and the available resolution that our game will run. Here for example, this resolution among other settings. The one thing about this file is that here we can set which systems will the entity component system use. Here I'm using OpenGL, OpenAL, sorry, bullet and my own input from Cobra. So if I don't want to use physics, I just comment this and my game won't have physics. I don't want to use bullet. I want to use another physics library. I change it for everything. The rendering, the same. If I want to use direct text, the same. So as you can see, Cobra is very extendable and you can just, it's very playable. And the other, okay. The current status of Cobra is still in development. The beta will be available on December. If you want to help me with this, I will be glad to work with you. My future goals are to implement a particle system, 3D send in editor, and work on support. Work on support. Do you know Vulkan? Anyone knows Vulkan? What is? Vulkan is the new graphics API that Chronos Group will release that will replace OpenGL in a short time. So I want to support it as well. And now I'm going to show you a simple Cobra demo. Okay. Don't blame me. Sorry. Why this hand? Okay. I love Python tools for Visual Studio. Don't blame me. Okay. I will just run it. See if it works. Well, it didn't work. Let's see. This is a game like Angry Birds. Okay. But in 3D. We have this link shot and our box. And in this example, it's very easy. We just hold this base bar. And when we release it, okay. And you can see here that I'm getting a score for the Xs that are going outside. And this is just the example. So let's check the code. The code is very easy. You can see here the entities that I have. I have the red box that has a partition to transform to be placed in the world. Mesh, rigid body, Corridor and a scorebox behavior. This behavior, the only thing that it does is to check if it has gone outside the... I don't know how to say it. Outside the... the area. Yes. Thank you. Thank you. Outside the area, we just add the score and we print our score. You can see it's very easy. Here we have the player that is just the same, but it has a player behavior. And this behavior, the only thing that it does is check if the speech key is being pressed. If so, it charges up and moves backwards. And then if it's released, we send a message to apply that in plus to forward. You can see it's very... and this is all. So... Thank you for your attention. I hope you enjoy it. Do you have Oculus Rift support or are you planning to add in it? I don't have Oculus support. Well, it would be a good idea. 2D, it works as well. You can just use a card transform, but only moving around the X or Y axis. You can use just the screen transform, if you want. Any more questions? Yeah. Sorry. So your message passing part is all in C, but then it's calling back into Python to execute the methods. So what is the... is it just for the message passing that the GIL is released? Or do you have a lot of components written in C? The components are written in C, yes, but you can also write components in Python if you want. The systems are... how do we say it? The systems and the components can be programmed in C or Python. If I understood your question, you're asking if all the messages are passed to all the entities. Sorry, to all the components. First of all, Manj. Yeah, so you're not getting any performance benefits if your components are written in Python. You only get them if you're writing them in C. Is that right? Well, the other performance is that the renderer is just running in C, and that's the open-gil calls that are the most expensive. Right, so it's basically the rendering that is threaded. Everything. You can just make a system in Python or in C just as you want. Each system, as I explained, you can just put it in the settings and it will just work. Any more questions? Yes? How does a framework approach the transport lag problem? For example, you have a delay between client and server if I fire a bullet from client and I have a lag. So is it possible with a framework to somehow approach that problem? Yes, well, the components are a bit bound to the systems. So you have also to write the components for your system if they are not compatible. Any more questions? I have one. Do you support Python 3? In the future. Make it a promise. Okay, do we have any more questions? Thank you for the talk. Where can I find the source code? The source code I will upload it into my GitHub. I'm going to also make a web page. It will be available on December if you want to or if you want, you can tell me your email and I will send to you if you want. But it's still in development. There are some things that need a bit of work. But if you're interested, I can send an email to you if the source is you want. Okay. How do you see the future of Python game development keeping in mind that there are a lot of high quality free engines out there? Well, I see it complicated. That's why I wanted to help Python to be a thing in video game with this framework. But yes, I see it complicated in the future. So, you saw code to render a mesh with what kind of shaders is that running? Is it like a deferred rendering engine or can you customize the shaders? It's rendering pipeline. It's the third rendering. Yes. And can you customize the rendering pipeline? Yes. Any more questions? Okay, so I'll ask what platforms do you support? Well, I didn't try but it's open source so it can compile to any platform. I'm using Boost, with Python for the bindings of C++. The system is used on DLL and so on but it's a bit complicated to port right now but it works for Windows for the moment. So, no Android for example? For mobile, I'm not planning for being able to make games for mobile for the moment. Okay, does anybody else have a question? Alright, thank you.
|
Alejandro Garcia - Python Gamedev MLG An overview of the currently available Python game development libraries and frameworks and how is Python currently being used in the videogame industry. Presentation of Kobra, a modern open source Python game development framework with ECS (Entity Component System) architecture and C++ bindings.
|
10.5446/20060 (DOI)
|
but making a few, making some claimers to begin with. First of all, it's my opinion, it's not bad at all. I'm really lazy to create slides that look okay. I found this, I like it, I took it. Pretty much all the stock was created by having multiple bad pizzas at multiple places that had like reviews that say it's the best pizza, that play at whatever. So that's what started. I'm a Brazilian Pythonist that moved to Scotland for the weather. I'm from Sao Paulo, a 20 million people city. So I moved from these into Dundee, 140,000 people. The fourth largest megalopolis in Scotland. Home of the Dundee cake and the birthplace of the marmalade. So again, from this into this, that's two minutes away walk from my place, so just round the corner. I'd like to define what is a good place to eat, but I'm not. We all know what it is. Stock is not about that. The stock is about, like, you can go to any first place in any restaurant site review, it's gonna be nice. But I'm worried, I wanna know about what, what is the 319th or the 314th site are. And that's what this site's about, it's how to find them. The first idea would be like stars and ratings. And if you get something like this, what does this tell us? Well, pretty much nothing. Is the first restaurant a McDonald's or a very good, bingus place? So we really don't have to judge that. Also, if it is a McDonald's, you get like five stars for a lot of people, it might be an interesting McDonald's to see if it might be the best one. So what does that tell us? It's just that restaurant A, Scarrow likes a little bit more restaurant A than Peter Sam likes restaurant B. Ratings just not so bad. And all the rating sites found that really early on. As any kid just from kindergarten will say that ratings is not gonna work. You have to use the lower bound of Wilson's core confidence interval for a Bernoulli parameter, of course. So another good metric that we might try is the number distribution of ratings. So it's got that. And for the first restaurant, you have like pretty much four reviews. It might not be good, but it's not nearly enough to know that. For the second restaurant, you got what's called a jig shaped curve. What do you call it in the agai? Where you have a lot of terrible reviews, a lot of excellent reviews. The thing is people tend to vote for the very good and very bad. And the average is just, you won't vote. That was all right. So that helps, but doesn't solve the problem. So now for something completely different. Linear algebra. When is that true? Yes, dirty clothes. So washing and drying and drying and washing are dirty clothes that are not commute. That is a for Wikipedia. They make me write that. But yes, matrix multiplication also between home, after home short. So a funny thing about matrix multiplications is that depending on the order that you do, you, that's pretty much all the math that's gonna need for this talk. You get different results with different dimensions. And that's pretty much based on the order that is. So the trick here is that in order to multiply two matrix, the middle values, that is the columns for the first matrix and the rows for the second one have to be the same number. And see, is when you multiply them, they pretty much vanish. You might as well use a pit emoji and that's gonna disappear. So that's something really nice. Back to Sao Paulo. Sao Paulo had a lot of immigrants. In the turn of the 19th century, it was the population of the city, 35% were Italians, 11% were Portuguese. That's almost 50% of immigrants in the center. When we got close to 1940, it was almost 700,000 Italians. A lot of models. At that time, it is for the state, the second one for the state, it couldn't find for the city. But, I didn't know my time around quite that. But the population at Sao Paulo at that time was less than one million people. So there's a lot of Italians there and a lot of immigrants there that created a food subculture. Pretty much close to what you have in Italy as well. In New York, like the New York style pizza. We do have a Sao Paulo style pizza. And that's important because it comes to my hypothesis that people that have the same background judge the things, judge food, use the same standards. If you have a culture of something, you're gonna judge, and all of you have the same culture, you're gonna judge it the same. If you don't have it, you're gonna judge differently. And here's me trying to prove it. That place, the first one, it's not even the top places I've never been there. But look how tight the Gaussian is. People agree quite well that it's a good pizza. If it was a bit pizza, it's gonna be lower, but it's gonna be as tight as well. The second one is the best pizza place in Dundee. And it's awful. But look how sports the Gaussian is. And Dundee has 10% of people from abroad. And each one has a different idea of what a good pizza is. And don't have a pizza that's defined by Dundee. The same thing happens here in Bilbao. The first one is just, it's not even, again, I like the middle ones. This is the first restaurant that I went here in Bilbao. And it's a solid four. But again, a very tight Gaussian. A lot of people know that's a four, and they vote for four. The tail end in Nanibra, it's a fish and ship place. And people there have a good fish and ship. And also very tight Gaussian that you have from people that would not. Thus, so many hypothesis. Proving my hypothesis. Well, it's not really a talk about diet science. It is just a yuck shave. I don't know when I'm going in the Gaussian rampage, analysis rampage. So this is just based on the hunch. And this is the hunch. That what can suit a good place to eat is based on individual background. And that's deeply based on the individual. Makes that if you agree for multiple reviews from someone, you tend to, the chance that you're gonna like another restaurant that they like will be much more likely. And then we go back to Lina Rausba. No food. I suppose that I have this, I got it somewhere for a friend. A huge list of users that gave an amount of stars for a restaurant. And I have a huge list of those. So you have for the 500 top restaurants in Bilbao, for example. What could I do with that? Of course, load in the matrix. And I create a matrix that is restaurants for each row, users for each column, and the value in the middle, value for each position is just how many stars they gave it. And we call it M, just M has dimensions, restaurants and users. So the same way that we could remove dimensions, we can also create them. So if I have a matrix, and I wanna create two new matrix out of there, if I can make that equality, I can generate any sizes that I want with this extra dimension that I'm creating. And that's quite useful because suppose that I create a matrix that's really an approximation of matrix C. There's an approximation of that matrix users that I got from real data. But this is theoretical, theoretical one. I created that. And that one is the result of users, or a matrix of restaurants, with some categories that I choose. And a matrix of users with the same number of categories. If I can multiply them and create a matrix C, that again, it's a good approximation of matrix M. I can pretty much classify the restaurants and the users in these categories. And that's pretty much what no negative matrix utilization is. It's a way to create weights for automatically generated categories. You don't know what each category is from beforehand. Sometimes they don't even make sense, but sometimes you can get pretty much a rough approximation of what they are. And our results is pretty much a matrix for restaurants that can tell me oh, this restaurant has these categories in different weights. And I can try to match those two other restaurants and try to find restaurants that I might like based on that one that I like. Oops. Okay. So the no negative part of it, and that's something I should have said, is just if you keep it all positive, or greater than one to zero, it's possible that it's a lot of the ways of generating C because one of the most usual ways of doing that, it's the least squares. And that's gonna be easier if you do just with positives ones. And okay, so we're good because stars are zero to five. So how do you generate C? This is taken from this book, it's programming the collective intelligence. It's a slightly older book. I guess it's 2010, probably. But it's one of those great books that when you return to them, from time to time you find different stuff. It's all done in Python. It's probably one of the most fun books you're gonna have in your bookshelf. So this is just pretty much his algorithm for that. And I just left the comments because that's pretty much all you have. All you need. So first you start R and U of just random values. And then you start interrating. And for each iteration you calculate how different C and M are. And then if they are the same, you just exit. But that doesn't happen to like realize data. But then you fixate one of the matrix and integrate the weights for the other one. Then you fixate the other ones and try to integrate the index for the weights for this one. Calculate again and try to improve it and see what it gets. And that's pretty much how I got here. That's the way to have to click here. This is the hard part because I can't, I have no view from here. So this is pretty much like a proof of concept what I did. This pretty much just loads a bunch of imports. My focus in the right place. And this loads the data that I got. No, it didn't. Okay. If I want to run the no negative matrix utilization just runs on here. But I've just run it. So I got an M matrix. That's, so it is 11,000 users long and 500 deep. From that they got the restaurant matrix. It is just the bottom of it. With the categories and weights that they have. And also the weights for the users. So if you look at here, you see that user zero voted for restaurant zero and so forth. There's a five here. This is a very sparse matrix. And the more connectivity that you have, the better results that you have. And make sense, like you try to find people that vote likely likewise for all the rest. And if you have like just someone just voted once, it does not gonna help your data that much. Here do a little bit of the pretty fine of that. So, okay. So, pretty matrix. So this is the restaurant one. So we have restaurants here. Each of the factors there are ruled all. The URL. And that goes on for 500 restaurants. Here have the users. I've just transposed the matrix. So it does make it more likely. How are we in time? Okay. Okay. Nice. I didn't start my timer. So here's again, users ratings. A lot of them is just like people that voted once. So they don't really influence stuff. Someone with an empty name that I found that just goes here. And here I do have some pandas magic to get people and show them. I can get a little bit back to this. But the thing is, I get the restaurant that I like. And this is how they look. Ah, no. Oh, come on. Let's go. I should have disabled controls, yes. Okay, so here's how they look. So it has a huge factor four, which we don't know what it is. A larger factor zero and factor three. And that's pretty much it. So we might want to find similar restaurants. So this is just the basic find similar that I found. It's very trivial. It's not even the best way. The best way is probably using some linear programming to find it, reducing the difference. But here it has an INC, which is a solid four. It has a little bit of zero here, but you can see it because the four is like 40. So it's huge four. And from that you can get an idea what that category might be. Because they sell code tapas, code pinches, and croquettes on top of toast. So that might be an indication of croquettes on top of toast kind of market. Costco also huge four and a large zero zone. Well, that might be an interesting restaurant to see. It doesn't have that much F3. The same restaurant, that's a good one. And La Deliciosa. And that's a weird one because I first looked at it, it didn't seem that interesting, but that's something that has a huge four. It has a good zero, some of them too. So that's something that this program, this way of thinking just suggests a new place to look at. El Vuevo Frito, that I didn't like it. This is just searching the similarity between the curves on the factoring, the same on the restaurant data. You can also do that through the users. You have factors here that are correlated some factors. So if I find who voted a lot for factor four might be people that likes this kind of factor for restaurants and they might be interested to find what they like. So I got to users this next team. This P63, P63 votes a lot for F0, F3, MNF4. That might be an interesting person to see what does they vote for. It's a little bit of a resonate and it's kind of tough. And then from them you can try to find familiar restaurants that people like. And again, it in Zini. Cafe Barbeau Bout is a different one. It's so covariate. Casco again, Vuevo. And that's it. Oh no, no. Okay. Focus. What? Sorry, it went back to the beginning. It's changing. It stopped following. Oh, thank you. Oh yeah, no, it's the right place. Sorry. And that's it. Thank you. Questions? Questions? Hi. You ended up with nine categories in your example. Was this by Jan? Jan, so what was this? The study? You ended up with nine categories. Oh, it's F0 to F9. So it's 10 categories. Oh, I chose 10. It's just a random value. I tried with 20. It got the results. The more categories you add, the closer they get to the more similar the matrix are. But also, it makes less sense of each category. I got a balance about this. I tried with five and I tried with 20. And pretty much 10 was like this with spot, and that's pretty much what I used. When you run it, I don't know if I showed that I was about to... Oh, sorry. Should we have the first one? Here. Come on. Wait. Let's see. So, five to Markdown. Oh, not Markdown. So, yeah, I can get the next... Go away! Yeah. So this is 13 directions. 15 directions, sorry. What was that? Oh, just 10. Yeah, not that. Totally not that. So, they are calculating it. Should be putting something in a while. So, it's a start. So, let it calculate. I'm going to show you how. So, more questions? Thank you. Oh, sorry. There's one question. So, having done all that, is it true that you can sort of tell how much you're going to like a Cesar restaurant just by looking at the shape of the... If you just go on Google reviews or something like that. I know that I've been walking with you and I've seen you do that. Yeah. You just say, oh, yeah, this one looks good. Just by quickly looking at it, you get a feel for... Yeah, yeah. The general shape of... It can't tell you if it is a good... It's a huge... From looking at the shape, you can clearly see if you got the J-shaped. And depending on how large is the terrible one. It could be an indication that there's a lot of average and poor ratings that you're not seeing. So, that's something that I try to avoid. Seeing as close as we can see to a tight Gaussian, it's usually a good sign. But even though a large one might mean just that the restaurant is not really stable, really doesn't always do stuff to the same standard. So, even though it might be an interesting or not place to see, but it does tell you some stuff. It is too early. Okay. Thank you. Okay. Just over here, I can just... So, here's how they do the results come from. This is the difference. And as it goes, it doesn't reduce anymore. And that's pretty much the amount of difference that I have between my C matrix, that is the theoretical one, to the actual one. So, with 5, that thing doesn't go that low. If 20, it does, but not that much and doesn't improve the results that much from what I saw. Again, not data science, it's just a guy with computer and trying to find a place to it. Thank you.
|
Adriano Petrich - Yak shaving a good place to eat using non negative matrix factorization Trying to find a good place to eat has become much easier and democratic with online reviews, but on the other hand, that creates new problems. Can you trust that 5 star review of fast food chain as much as the 1 star of a fancy restaurant because "Toast arrived far too early, and too thin"? We all like enjoy things differently. Starting of on the assumption that the "best pizza" is not the same for everyone. Can we group users into people that has similar tastes? Can we identify reviews and restaurants to make sense of it? Can that lead us to a better way to find restaurants that you like? Using some data handling techniques I walk you through my process and results that I've got from that idea. There are no requisites for this talk except basic python and math knowledge (matrices exist)
|
10.5446/20059 (DOI)
|
I work at a company called Braintree. We make it really easy for you to accept credit cards, PayPal, and other payment methods online and in mobile apps. This talk is not currently on GitHub, but it will be shortly, and you'll be able to find it there. So let's talk about the title of the talk. Python not recommended. It might sound like it's a joke, but I mean it quite literally. At the company where I work, we have something called the radar, and it contains information on the technologies we use, whether you should use them, whether you should never use them again, and sort of where to look for examples and other relevant information. In the radar, Python is specifically listed as currently in use, but not recommended. So this talk is about why we originally used it, why it's not recommended now, and what we still like it for and what we're going to use it for in the future. So first some background. The obvious question is what do we use instead? If we don't use Python, there's got to be something else we use as our general purpose language. And the answer is Ruby. Braintree loves Ruby. Braintree uses Ruby by default for pretty much everything. But I promise this isn't a Ruby talk. I personally am a Python person. I don't love Ruby, so we'll talk about it only as background. Braintree has bought into the Ruby ecosystem. We use Rails as our web framework, we use Capistrano for remote server management, we use Puppet for configuration management, and we use Rake for sort of scripting, builds, and automation. So you probably wonder what's it like doing Python at a Ruby shop, something I've been doing for three years now? There's a lot of jokes about Python because it doesn't have quote unquote real lambdas, which people just mean that function definitions are statements and not expressions, and people are like just to be able to define a function wherever they want, and even though functionally they're the same, people don't like the syntax difference. It's seen as an elegant compared to Ruby or Elixir or Closer. This is both because of the statement expression dichotomy, but also because of the generally structured syntax, and because in these languages magic is really easy with macros in the list or with everything, and Ruby is magic. And it's just a little harder in Python and people don't like that. It's sometimes gets dismissed as a dying language because of some of the negative publicity around the slow adoption of Python 3. I don't see it as a failure, I don't think the Python community sees it as a failure, but people outside the Python community sometimes see it that way. The languages are also similar in a lot of ways, almost too similar for their own good when you're a Ruby programmer. A lot of people will sort of apply judgments that you'd apply to Ruby code to Python code without taking into account the difference in idiom. And so if you were to apply to Braintree as a Python developer, it might hurt you a little because people are going to expect your Python code to be written like Ruby, and it also means that our Python code can be a little less than idiomatic. So despite using a lot of Ruby, first and foremost, we believe in using the right tool for the job. So you're not going to find a ton of this at Braintree, although it does happen once in a while. So what do I mean? When is Ruby not the right tool for the job? And we sort of found two main times. One is when you need the JVM ecosystem, you might say, Joe, just use JRuby, and you have access to everything. Well, we found that that gets a little messy and that it's not really a good long-term solution, although we have done it in the past when we needed Java in the short term. We'll need to use Java if some third party who maybe wrote their API in 1999 only has, doesn't have an open API spec, and they'll just give you a library and your choices are C++ or Java. Well, we're going to use Java. There's also a lot of great tools that work best with the JVM, like Apache Kafka, which is a message broker for handling high-volume data feeds, and Apache Cassandra, which is a distributed database for handling large amounts of data. So what do we use when we need the JVM? Well, like I said, we have tried JRuby historically along with Java. We haven't found it to be that great. And then more recently, over the past maybe year and a half, two years at this point, we have used Clojure pretty successfully, although we've still found that there are times that you really want to use Java directly. So the other main time that we find that Ruby is in the right tool is when you need to write a smart proxy. Now, our business is basically to sit between somebody who wants to sell something online and the banks and card networks. So we basically are a smart proxy. And so our big smart proxies build up a lot of little smart proxies. You can see this is the logo from a party that Braintree and PayPal threw at South by Southwest a couple years ago. It's smart. It has a brain, so it fits. We have really high uptime requirements. We need to be available when there are temporary networking problems, when we need to fail over services, when we have to run database migrations, when we get huge traffic spikes like Uber is, it's New Year's and Uber is running tons of rides and they want to charge people's credit cards. And we also have a big problem with the services behind us, the banks, the card networks going down. And we need, at the very worst case, requests to fail gracefully. And preferably if the outage times are short, like on the order of a few seconds, we want those requests to succeed even though the service behind us failed. So we use them to make our outgoing connections appear highly available to our internal services. We use them to pause incoming requests so that our internal services won't actually see any requests coming in and we can do whatever we want to them. But actually clients are still able to connect to us. We do custom rate limiting. We have pretty complex rules for how much capacity different clients and different types of requests can use. And we, because we integrate with a lot of legacy services, we often find that we need really weird SSL or persistent connection configurations. And this is something that we've often had to do ourselves. We also have very complex retry logic and a lot of other custom logic that we need. This is where we've historically used Python. It might seem a little odd since handling a large number of requests doing a lot of concurrency is not necessarily the first place you think of Python, but it was actually a pretty good fit for a long time, specifically because of Tornado. Tornado is a web server and framework for doing non-blocking I.O. And you get the benefits from Python of rapid development and easy to learn, and you still get the necessary I.O. concurrency to handle tens of thousands of requests. So back in 2013, Python was in pretty good shape at Braintree. It was in use for several of these internal proxies. It had served us well for a couple of years, and it had several internal advocates, not just me, but other people too. So what happened that I'm giving a talk called Python Not Recommended? The platform really did fail us as we started to scale. Concurrency in the framework really isn't enough. Nowadays, we sort of expect our languages to have concurrency built in, as Python 3 now does, and as Go and many other languages do. And you really expect that concurrency logic to not get in your way when you're writing business logic, which we found that the concurrency logic in Tornado really does. It was also too much work to keep up with changes in Tornado. We looked at new versions a couple of times, and it would have really been a complete rewrite to use the new APIs. And we really didn't want to spend the time for that, and we didn't trust that we would be able to make those changes without breaking anything. And so because we were using an outdated Tornado version, maintenance overhead could really be pretty high. You can't Google things easily because people have moved on to newer versions. It's hard to find the right docs, and you end up in callback hell because you're still using old, less elegant APIs. We found that logging actually has an unsettlingly high overhead. Every time we logged a line, we saw request pause, and eventually those pauses added up to enough to be really significant as our volume scaled. And then there's no SNI support in Python 2 historically. SNI lets you serve multiple SSL certificates in the same server port. A lot of our customers used it, and it was only very recently introduced into Python 2. The version of Tornado we're using is so old that even if the Python supports it, the Tornado version doesn't. The applications that we wrote also really failed us, and a lot of this ties in with those platform failures, but it's also to a large extent our fault. The smart proxies were really too smart. The logic was all mixed in with the concurrency boilerplate, making it hard to understand, and they were trying to do so many things that the code ended up that was meant to do one thing, was coupled with the code that was meant to do something else. So when we tried to rip out the code to do connection pausing, it completely broke rate limiting, and we had to put it back and leave this completely now unused code in the code base. Straightforward Python implementations, as we scaled, were not fast enough, so the rate limiting code started to add an unacceptable amount of overhead to every request that came in. We also found that in addition to being too smart, the proxies were not smart enough. We couldn't just write certain pieces of business logic in the proxy. The business logic had to be duplicated in the main application, and that's something that we really don't like. We don't like writing the same logic twice. And none of these applications were really built for horizontal scalability. They all assumed that a single or in some cases a small number of instances would be enough forever, and so they weren't designed for us to run 10 or 100 of these. They weren't designed for us to run two or three. So what solutions did we switch to? What made Python obsolete in these areas? So for our incoming request proxy, we've switched to a combination of nginx and haproxy, along with pgbouncer. So in nginx and haproxy, we're able to do approximately the same complicated rate limiting logic and load balancing that we were able to do previously with our proxy in a pull layer. And we've also moved pausing completely out of the proxy layer and into pgbouncer, which is basically another proxy that sits between the applications and PostgreSQL. We then wrote our main outgoing proxy as in Node.js, but this was actually a failed attempt. It had all the same types of problems as Python. It was still trying to write our own tool to do a job that we weren't experts at, and we had problems with failed persistent connections and with memory leaks in Node that led us to abandon it and move to nginx and haproxy. Now the key here was timing, is that haproxy 1.6 had features that we really needed, and so we've now moved to that, and it allows us to remove another custom piece of code from our system. Finally our sort of most complicated outgoing proxy, we've decided to rewrite to enclosure using Apache Kafka, and it allows us to centralize all the logic in a single application. We can build it pretty easily to horizontally scale almost linearly, and of course you get SNI support with the JVM right out of the box. And unfortunately we canceled that project because it wasn't a high priority, it was going to take a lot of time, and so instead I wrote a monkey patch of Tornado to support SNI even though it's a really, really old version of Tornado. So it's not great. We still have logic duplication between different applications, and we still have sort of lack of horizontal scalability. We basically run two of these proxies, but for now it's okay, but the problems are unsolved. So as of late 2014, all the smart proxies were on the way out, or in use, but not because we wanted to. Not recommended for new projects, it's official, it's in the documentation that we use at Braintree, and there were fewer internal advocates. This isn't because all the Python developers got mad and left Braintree, but a lot of the people who used to really like Python now have moved on and prefer languages like Go, or Closure, or Lixer, some of which have more in common with Ruby so you can understand it, and some of which just have better concurrency primitives than Python 2. Just to be clear, I don't fall into any of these groups. I still like Python and still use it outside of fork. So this is kind of sad, and it makes it really sound like the state of Python at Braintree was really sad, but that's because that's the point of the talk. There is something we will never use Python for again, but actually things are looking up overall. We're now using Python in areas where it really shows its strength instead of just sort of the Swiss Army Knife glue code of our code base. So the first place is data analysis. This is probably surprising to exactly no one. Our business analysts have really used it to replace Excel to write sort of one-off reports and do smaller monthly tasks that don't really need to be automated. Our data analysts have moved more and more from writing giant crazy blocks of SQL to putting more logic in Python so that the code is more maintainable and understandable to a larger group of people, and these are people who have maybe done a little programming before, but are really buying into Python very rapidly, which is cool. And finally, our data scientists are really using it to replace R. Part of that is because of the great modeling and analysis tools that Python now has, but primarily it's because it's a lot easier to deploy your solution, it's a lot easier to do the sort of ETL steps that happen before you do the modeling in Python than it ever was in R. Finally, well, the next thing is really infrastructure management, and this is somewhere that historically we bought into Puppet wholesale. We have huge repositories full of Puppet code. Puppet is a Ruby-based tool, but recently we found that Python has a really good niche here, and we use it to manage certain resources like IP address, physical ports, server locations. This is something we actually didn't do with Puppet, we did with a bunch of Google spreadsheets. So the centralized application with a lot of the, like, better views into the data is really helpful. We've also used it to manage cloud instances. Because we run our own physical hardware for a lot of things, we had a pretty unsophisticated setup for managing cloud infrastructures, namely we used the user interfaces, we basically log into the website and would start and stop instances. Now that we're starting to do more and more things automatically and need scaling, excuse me, use scaling, we had to have an automated tool, and Python is the right way to do that. And finally, we use our switches for slightly complicated setups. We do try and make sure that even if switches fail in our data center, everything keeps going, so everything is, all the networking is mesh, and the switch configurations are pretty complicated, and we didn't find a good way to manage them with Puppet. And we found that rather than writing a custom Puppet module, which can be pretty complicated, it was actually easier to emulate what Puppet does in Python and pull from the same data source that our Puppet repository does. And so all the code is in Python, and it's like 100 lines, rather than writing a very complicated Puppet module that we'd never be able to change because no one understood how it worked. The Python community is sort of the final thing. It's a big advantage of the language, in my opinion, and something that has been very beneficial to Braintree. We host a lot of Python meetups. We have a monthly project night. We host Chippy, the Chicago Python meetup, about twice a year. We've done a couple of events with PyLadiesNow and one with Django Girls, and we've also sponsored other events outside the office. And this has really helped with our hiring. We've hired several people now who first heard about Braintree through these events and came to Braintree, even though they don't get to write Python because they know we support the community. And recruiting is one of the hardest things we do, so this has been super helpful. We also find that giving talks is a great way to spread the word about Braintree, and I'm not the only one giving Python talks. One of my colleagues has given a couple in Chicago and has given one with me at Northwestern University. Our customers also use Python. So having us support the Python community makes them feel more connected to us. Some of the biggest startups in the world who are customers of ours have pretty large Python code bases, and it's much easier for them to keep using Python to connect to us. And that includes as they migrate to Python 3. Our Python library is single code base, Python 2 and Python 3, and a significant amount of Python 3 traffic comes through our API. So having that and supporting sort of the next, the future of Python has helped us gain and keep merchants. So now we're at 2015, so what's the state of Python now? Python 2 is definitely showing its age internally and in general, especially around concurrency. I don't think anybody really disagrees with that. And as the standard tools like HAProxy and Nginx improve more and more, we're sort of losing a use case for Python, being the jack of all trades, the language you can use to write whatever tool you need is a little less important as the standard tools tend to be there for scaling and for high availability. Data science is really important for Python's future, I think. It gets the foot in the door for Python at pretty much every company, which is a great way to keep people who like Python interested and to sort of keep it in your mind for when to choose it for other projects. The community is also really important. It's been great for us at Braintree and it's just one of the reasons that Python is as successful as it is. Thank you. That's all I have. I'm glad to take questions. Thanks for the great talk. Thanks a lot. I have a question about, like, if you see yourself as a tech hub, more or less, if you have some... I'm sorry, I'm having trouble understanding you. Sorry. I'm a little hard of hearing, so you're going to have to speak up. Thanks. The question is about if you have any interns, I mean, people who like students or such, walking into your company, are they keen on learning Python or something more? How does it act in your community? I'm sorry, could somebody else repeat the question? I just had trouble understanding that. Is it advice for students about choosing Python or something else? If they come to your company or your environment, are they keen on learning Python or something else like Ruby or stuff? What do they do? What do they choose? How do they act? Do they choose to learn Ruby or choose to learn Python? I think that we definitely look for people who don't want to do one specific thing. We look for people who want to learn and want to use whatever tool is best for the job at Braintree. We look for people who are open to learning anything. A lot of the people we hire and a lot of the students we talk to are people who have maybe done Python at school but are open to learning anything. I'm not sure if that answers your question. The non-professional programmers like the analysts you talked about, did they choose Python because they tried other options and like closure and didn't like them or because they already knew Python or because you suggested it or how does it find its way into their job? So the guy who started our business analytics team, he had been doing a lot of our reporting manually in Excel for years and he decided to learn to program. And so first he actually learned Ruby and he wrote a web application that we used in production for years.
|
Adam Forsyth - Python Not Recommended Braintree is a Ruby shop. By default, we use Ruby and Rails for projects. We also use Ruby-based projects for much of our tooling, including puppet, capistrano, and rake. However, we strongly believe in using the right tool for the job. What that means has evolved over me, and I'll discuss what solutions we chose in the past as well as our current choices. So what's it like doing Python at a Ruby shop? You get lots of jokes about language features Ruby has but Python lacks and lots of disbelief that Python will survive the 2/3 split. People also tend to apply the best practices and conventions of Ruby to Python code as if it hey were the same. Python's major inroad at Braintree has been, surprisingly enough, as a platform for high-concurrency situations. This is a direct result of the power of Tornado as a platform for asynchronous I/O. It also helps that many Python is very approachable and many developers have at least some experience with it. Braintree has three pieces of our infrastructure using Python and Tornado -- an incoming request proxy; an outgoing request proxy; and a webook delivery service. They've served us well for 3+ years but all suffer from a number of problems. The outdated concurrency feature s of CPython / Python 2 as well as our lack of experience with and commitment to Tornado have always been an issue. As the meat of the talk, I'll speak in depth about the other issues we've encountered with each of the three applications and our short- and long- term solutions to the problems. The state as of the end of 2014 appeared dire for Python at Braintree. All the old Python code in our stack is on the way out, and Python has been specifically recommended agaist for new projects. Our Python client library is used by some of our largest merchants, and is ready for the future by supporting Python 2.6+ and Python 3.3+ in a single codebase. We also have a vibrant Python community at Venmo, our sister company. Both Braintree and Venmo support Python by attending, hosting, sponsoring, and speaking at meetups, conferences, and other events in Chicago, New York, and elsewhere. At Braintree, our Data Science team uses Python almost exclusively and they're becoming a bigger part of our business every day. We also use custom tooling written in Python to manage our infrastructure.
|
10.5446/20054 (DOI)
|
Thank you. All right. It's four o'clock. It's time for lightning talks again. Thank you. Thank you. How was the rest of your Python? It was good. All right. All right. Thanks. I think all the people with the green shirts, everybody who's putting work into this, really appreciate that you like it. Okay. Let's get started. We don't have two setups again. We're going to stick with a quick switchover. So if you are giving a lightning talk today, please be prepared to quickly set up. We're going to start with Anton Turing about elliptics. And the person after that is going to be Austin Bingham on Being Super. Austin, are you there? Excellent. What position you could pick? We see it on that screen. All right. Let's go. Hello. My name is Anton. I work at Yandex. Sorry. Okay. Yandex is a Russian company. We provide services like searching and something wrong. Search, mail, video, music and so on. We have a lot of user-generated content services. And today I would like to introduce you to one of our open source projects named elliptics. Elliptics is distributed for tolerant key value storage with high availability, high scalability and a lot of other buzzwords. We use it in our production to store petabytes of data and billions of keys. And it works perfectly. It was founded to solve problem when one of our data centers goes down. Idea of Amazon Dynamics is in base of this storage. Of course, elliptics has DHT and provides replication by using mechanism of elliptics groups. Each node in one group takes responsibility for some range, that's you, for some range of keys and this DHT ring and when you store your data by key, hash is calculated from this key and according to that number, data is transferred for some node. If you need replication, you should write your data in three different groups, for example, in three different data centers. Elliptics is not simply a storage. It provides distributed cache, Russian words, and we use it in our content delivery network. We use this cache, for example, in engines. We use this cache in services which are related with some operations with sending files, for example, photos. If some photo became very popular, we could copy this photo to the cache, close it to the user so we could save a lot of IE and network. The other feature is that elliptics provides you an ability to start your own program on the same node when your data is stored. It's server-sized scripting. You could write your own program on Python, C++ and many other programs, languages, sorry, and you have a guarantee that your worker would launch on the proper node and you need not copy any data through network. Elliptics is easy to use and easy to enlarge. It has a rich and powerful Python API. It supports asynchronous operations so you could implement a really good scalable application using it. Of course, we provide C++ and go API and we have, of course, HTTP interface for our storage. It has buckets and keys like S3 interface, sorry, and it's not compatible, but full S3-compatible interface is being developed now. Oh, sorry. It's quite obvious that it's impossible to store billions of keys on ordinary file system, for example, on X3 and so on. So we implemented our own backend for elliptics storage. It named eBLOP. At the first glance, it looks like a simple large binary object and you just append that data to the end of this file and marked deleted keys as deleted. But there is a lot of rocket science in this part of elliptics and it's extremely fast. So the bottom line is if you need distributed storage in your own project and you could not rely on S3 because of big brother or something like this, I think I'm pretty sure that elliptics is a good choice for you. Thank you. Thank you, Anton. Hopefully this just works. The next one is Austin on Being Super and Miko Ottama is the next one. Miko, are you there? Okay. If it works, that is the big thing. I can see you, Miko. There we go. Okay. There you are. Thank you. All right. My name is Austin Bingham. I work for and part own a small software company in Norway called Sixty North and I want to talk about on Being Super. So who here uses Super? Just raise your hand. So, nearly everybody. That sounds about right. Who knows how Super actually works? Much smaller hand. Who knows about the super proxy objects and method resolution order and the C3 algorithm? Okay. All right. Basically, a lot of us use Super but don't really understand exactly what's going on inside of it. That's a position I was in a couple months ago and I had to develop some training for somebody and I wanted to know how does Super work. I use it for literally more than ten years knowing that Super is how I access the base class implementation. That was my model of it. I never had to really get any fancier than that. But as I looked into it, I learned that actually Super is really fascinating and it opens a door to a wide variety of really interesting design choices. So I thought it would be a great topic for something like a five minute lightning talk format. So here I am. So the first thing you need to understand to understand Super is method resolution order. Which a lot of you probably know already because you can use or you use method resolution order all the time implicitly even without Super. All it is is an ordering of an inheritance graph. So here I've created ABC and D, a standard diamond ‑‑ well, maybe not standard, maybe not a good idea. But it's a diamond inheritance graph and you can see the MRO for D in the end there is D, B, C, A and object. Of course, object is always in there. So it's just an order that Python has come up with for all of the classes in an inheritance graph. As I said, this is useful ‑‑ this is used for all method resolutions. So you call D dot foo, then the method resolution order is used by Python to figure out which of these objects ‑‑ which of these classes, I'm sorry, implements foo. And the first one that has foo is the one that gets used. That's fundamentally what MRO is. So how is MRO calculated? It can't just be randomly chosen. It must have some order. That's what the C3 algorithm is or C3 linearization is, the computer science term. This came out of the Dillon language, I think, I don't know, many, many years ago. Now it's used in Python, Perl, Dillon, probably some other languages. It's very popular for dynamic MRO calculation. It has three basic things that it guarantees when it calculates the MRO. One is that drive classes will come in the MRO before their base classes. So it guarantees that. It guarantees that whatever base class order you give in your class definition is also preserved. So the relative ordering is always preserved based on what you tell it, lexically. And finally, the first two constraints are conserved anywhere in inheritance graphs. So the relative ordering of classes is always the same no matter where you start. These are the rules that C3 provides. And that's how Python resolves functions, methods, I should say. C3 brings with it a little bit of baggage. One of the interesting side effects is that not all inheritance graphs are legal. So in this case, you see I've created another inheritance graph where D inherits from B, A, and C in that order. And because of the guarantees that C3 wants to make or is going to make, it's telling you it can't do that. It can't have A before C because A has to come after C because A is a, is that a bell I should be concerned about? Okay. I don't know what's going on with it. Really, I don't. So you can get yourself into a hole. You may have seen this. Maybe not. I'd never seen it, but I thought it was interesting. So that's C3. So finally, we come to super. What is super actually doing? And this is the most pithy terse explanation I can come up with that's not a poem, which may have been a good idea. But given a method resolution order, some MRO, calculated by C3, and a class C in that MRO somewhere, super gives you an object which will resolve function, resolve method calls using everything after C in that MRO. That is the definition of what super does. Does it make sense? Yeah. Read it a few times later if you want. When you call super, what you're actually doing is creating a proxy object, a super proxy, and that's the workhorse of super. That's the thing that embeds all this logic about looking in the MRO in the right place and using the tail and so forth. So I had this nice picture of not really horses. It's a horse and a donkey. One of them is named Mr. Henry, which I thought was a cool name for a donkey. So I put it up there. They're the workhorse of super. Super proxies are, like I said, just regular objects. You can take a reference to it. You can interrogate its type. You can see that it has a little dunder thing in it called this class, which is the class type you passed as the first argument. So you can examine super a little bit and get a better sense of how it's put together and what it's doing. So in summary, I think I'm well in time here. Python calculates an MRO for all classes. You might have known that already. C3 is the algorithm that does that. Super requires an MRO and a starting point in that MRO. That's how it knows what to trim off of the MRO. And finally, super proxies find the first class in the rest of the MRO that supports the function that you want to call or that you try to call, and that's how super works. Thank you. The next one after Miko will be Dimitri Milajev. I hope I pronounced that arbitrarily wrong. There you are. I can see you. Thank you. Also, what somebody might have shared with the ring was about, is that a nice message or please turn off your phones. We all will be happy with that. Hello Berlin. So who is having fun in Europe, Python, please raise your hands. What's that it? Again. Very good. My name is Miko and where I'm coming from, you are going to have more fun. So we are going to have a bike in Finland and I'm here to tell you why you should come go to her. And the first of all, we are having fun now, so let's compare Finland to Berlin. So first of all, we are a little bit smaller community chair, so we are still maturing, so we need a lot of guys to come to see us and to tell us that the buton is good. I'm especially proud of our little pilot is community started like one year ago with two members and now we have 60 members there. And then of course in Finland, it's not the Berlin, so you can use your laptop for the purpose it was created, which is forming up your laps. Still a lot of cool technology comes from Finland. So if you like Linux, SSA or IRC, you should come to pay us a visit. Looking for our freedom. And we are the only Python in the world where we have a sauna party. I need to tell the truth that I'm not actually one of the organizers and because I'm not the organizers, I can make any promises. So if there's no sauna party, you are invited to my home. I have a sauna for three people. And what I hope that you would do now that you go to a pi.pycon.org and what we lack in Finland is a good speaker. So we have had the problems to have these foreign superstars come to our conference. So if you have a talk in Python, please reduce the talk and come to tell the talk again in our Python in Finland. Thank you. Lynn, are you around? I can see you here next. Okay, cool. Okay, my name is Mitris Melaevs. I'm a PhD student at the University of London and I'm doing computational linguistics 24 hours, seven days a week. So this talk will be about distributional semantics. This is a study all about meaning and because I'm doing languages, so it's meaning of sentences and meaning of words. And like the main ideas of the whole field can be said in these two sentences. So you should know a word by the company it keeps or a little bit more formally. The semantically similar words tend to appear in similar contexts. So if you think about it, then you will see words like beer or wine occurring together with, I don't know, bar and parties. And probably such words as Python will occur with some different words. Like, right? So oh, sorry. This I Python is a bit confusing. So now we can, if we get a big text, we can just look to it and for every word we will look, what are the other words our word of occurs together? So for the boy, we will know, okay, we see it with A one time and we see it with Mitre one time and so on and so forth. And then we do it on many other sentences and do our counting and we get something like these. We got our boy and we know that it was together with time about 100 times and together with year about 102 times. And then we do the same with goal and what we notice that these numbers kind of similar is, right? At least they're not, the pattern is different than to notion and idea, right? So we can measure a way of similarity. So how can we do the similarity thing? What we're going to do is that we will see these words as vectors in a multi-dimensional space. And vectors, these are just some directions to points and between two points you can calculate the distance or you can calculate the cosine between these two points, right? So that's what we are going to do. So you seek it, learn and ask us, okay, given these vectors, please calculate us. The distance and what we see that indeed, boy is much closer to goal than to notion. And we would expect it. So can we go even further? Can we actually make it visible to us? So now our vectors, they're kind of in a so huge multi-dimensional space that we cannot even imagine how it looks like. So what this code is doing, it tries to get the same points on a two-dimensional space and tries to preserve the distances. So why it's two dimensions? Because then we can easily plot it. So if we plot, we get something like this. And so here you see that boy is close to the goal, mother is close to father. And all the family words, they cluster nicely in our field. And then kind of business-related words are here and colors are there and so on and so forth. So that's what distributional semantics is about. My main message is this. Sorry. So it was really few lines of code, but there were some intuition behind. So some science was behind. And if scientists tell you, they kind of know what to do, but developers know how to do things. And if you connect these two things together, you can achieve very great results. And it doesn't really apply to linguistics. It applies to any science. And I really encourage you to look for some cool scientific results if you're a programmer or learn programming if you are a scientist. Thank you. Thank you. Questions? Ok. All right. All right. All right. Oh, awesome. My turn. Okay. Hold on. I don't have slide notes, but I can kind of make shift. Hold on. Oh, damn it. It's not set up right. Totally unprepared. Okay. So first a little bit about me. I am Lynn Root. I actually live in San Francisco. I am a back-end developer for Spotify. I am also leader slash founder of the Pi Ladies of San Francisco and a board member on the Python software foundation. So earlier, Miko from PyCon Finland talked about them having Pi Ladies for about a year now, which is pretty awesome. Side note, you should all go to PyCon Finland. I spoke there last time. It was awesome. And you might be wondering what exactly is Pi Ladies. So Pi Ladies is sort of a mentorship group for women and friends in the Python slash open source community. We're there to support women in diversity. We've been around since the fall of 2011. And Pi Ladies in San Francisco started in April 2012 and we're now up to, I don't know, 1800 members. And we have about 50 locations all around the world except Antarctica, which is kind of the purpose of this talk. Maybe you can get a little Pi Ladies in Antarctica. Penguin Pi Ladies. So why would you want to start a Pi Ladies? I'm presenting about Pi Ladies. Why would you want that? It can be, it essentially is the motivation for you to learn or to better your Python knowledge. I started Pi Ladies by wanting to learn how to code in Python and got some friends with me and kind of blossomed from there. It also gives you leadership and organization and it creates sort of a networking kind of web for you so you can kind of, you know, find jobs or build your resume. And also we kind of want to take over the world. So that's kind of awesome. How can you get Pi Ladies? You can pip install it. I did, I created a Python package right before PyCon and it's up on PyPI. So you can go grab it. Yeah, there is literally a Pi Ladies on PyPI. So many people asked me about that. Yes, there's literally, not figuratively, literally a Python package. So pip install Pi Ladies and then run Pi Ladies Handbook. And what it does, what it gives you is it's a handbook and a checklist on how to start your own local chapter. And it has some assets and images so you can promote Pi Ladies. Some workshop materials so you don't have to like create your own workshop. A lot of them are beginner workshops. And I am still developing it and we'll have more scripts and stuff for like local organizers to work with Twitter and meet up and data mining about our meetup statistics. So if you don't want to pip install it, you can also go to kit.PiLadies.com to see what it's about. Let's just read the docs. And basically I wanted everyone to be able to have their own Pi Ladies. Thank you. And the next one would be Radu Mia Doppieralski. Radu Mia, are you here? All right, thank you. It's okay. I don't know. It's there. It's there. It's okay. If I try. It's okay. Sorry. I think it should be. I can't help with Russian. Okay, I just move you. Can you move it over there and there say, do you see that screen over there? There's another screen down there. Maybe you can use that for the presentation. I just move the presentation there and show it this way. Okay. My name is Harut. I wanted to present our forms like Django or WT forms, but our forms purpose to validate user input from form representation to Python internal representation, render the forms, and our key feature is that they work well with nested data. So we have a little bit of different abstraction layers. The main object is form. It contains nested structure of fields. Field represents a single atomic, single maybe atomic or maybe complex structure of data. Every field contains converter, has widgets, has permissions, and has data before converting and after converting. So here is the form. To create your own form, you should subclass form, class, and define a list of fields. Here you can see there are nested fields. To use form, you instantiate it with initial data. Then call accept with data you want to convert and either get data in form that Python data or get errors. Here you see a result of conversion of nested structure. So converter is object with two main methods. One is to Python, which accepts unicode string and returns object of type of whatever type you want. And the second method does reverse conversion. And you can define validating functions that just validate or do simple one-side conversions. So, and from these methods, you can throw validation error which is written in form.errors. So you put converter in the field and validators as arguments to converters. Converters can be required or not required. There is implementation of multi-dict features that allows us to add values under the same name, list of converter with nested converter. Here you see implementation of multiple selects by list of converter. You can easily tweak converters by copying them by call. And here is an example of implementation of a little bit complex converter that converts, not implementation, users that converts SQL, that converts dictionary to SQL Alchemy object. So widgets are a little bit simpler. They just take a field and render them to the template. You can render the entire form or single field. And where do we use? It is our CMS with some keywords presented. So here it is. Harris Butler, are you there? Hello, everybody. My name is Rana Nopiralski. There is a saying that Python is going either or. I thought I would work on the destroying part. So this is a project of mine. It is a killer robot that is going to run around and kill everybody. It is made with Raspberry Pi inside and programmed in Python, obviously. It started as a mechanized thing with some Arduino attached to it, but that didn't work too well. So I got a server controller for it and it is programmable kind. So I thought, oh, great, I can program it. It came with some language based loosely on FORT. I don't know if you know FORT. It wasn't as easy to program as I anticipated. Then I decided to upgrade the robot and give it another knee. So it has three degrees of freedom per leg. And thanks to that, I can make it move much smoother and I can make it tilt and do all sorts of cool acrobatics. But because of that, I had to do something called inverse kinematics, which involves a little bit of math, mostly trigonometrics and a little bit of linear algebra. And doing that in anything but Python was too painful for me. So I decided to remake it again and put Raspberry Pi inside. And also to put a bigger battery because all these extra servers were too much for the three AA batteries that I had in there. And it's growing. It's still growing. Right now, it has a gyro sensor so it can sense when you tilt it or when you pick it up or things like that. It has an audio. It has a speaker connected to it so it can talk, for example, the voices of turrets from FORT. Very useful. Like, hello, friend. I'm going to kill you. And it's coming on nicely. Unfortunately, I was planning on bringing it here and showing how it works. But unfortunately, just before the conference, it burned three servers in the legs. So I decided, no, I'm not going to risk that. But this is how it looked like recently. And that's all. Thank you. APPLAUSE. By the way, it's called Kubik. So you can Google that and see it. We are making very good progress. Everybody who's in the overflow slots has a very good chance to actually get the talks in here. Dimitri Semerov, are you around? Excellent. Get ready. Hi, everybody. So I got to practice this elevator pitch when I was at the open stack some in Atlanta a few weeks ago. So I've condensed that all for you to just a couple bullet points. So here we go. This is zero VM. And I work for Rackspace, by the way. Zero VM is not zero MQ. It's not Docker. I can't tell you how many people ask me, is that like Docker after I explained it to them in a few sentences? So it's not Docker. Docker is cool, but this is not Docker. It is not a drop in replacement for any other type of virtual machine. It's something completely different. It's not Nackle, but it's based on it. And if you want to know about Nackle, I'm not going to explain that to you. Just go read about it. It's pretty cool. It's open source. And Rackspace is sponsoring it, but that doesn't mean it's a Rackspace thing. Anyone who wants to get involved in it can. Just to give you a quick comparison of different types of virtualization technologies, something like KVM would be in the far left column. There's no kernel. There's no operating system. The overhead is extremely low. There's no interpretation. It starts up in about five milliseconds. And it's extremely secure. But of course, there are some limitations. A couple other key aspects. There's no place to get entropy from. Time and random functions behave completely deterministically. So it's just like a pure function. If you give it X for input, you will always, always, always get Y for outputs no matter what. And there's no persistent state. So you can't really write a demon with this thing. It's more of like a... Don't think of it so much like a program, but think of it as like a function. Your programs behave like functions. So you decompose your application to small, tiny, tiny little programs. So of course, you need to do something with this. So for I.O., you have to map all of your inputs and outputs beforehand. And we do that through an abstraction, what we call channels. A channel on the host on the outside of zero VM can be a file. It can be a pipe. It can be a socket or whatever. But inside, it's just treated as a file. Everything looks like a file. And you can read and write. And you can declare, okay, read only or read write or whatever. Like I said, it starts up in five milliseconds. No interpretation. The cool thing about it is that this gets really useful when you have an environment where multiple users are running arbitrary code. You don't want them to talk to each other, probably. So the worst thing a user can do with his or her code is to just crash itself. That's it. They can't break out. If you want to read about how that works, there's a thing called suffer fault isolation. That's the core concept in NACL, the native client. You can read about that. This means you can embed it in data stores like open stack Swift. We've done this today. It works already. We're still developing it to add some more cool features. So the cool thing about this is you can send code to the data due computations in place. And you have lots and lots of tiny little processes that live for just a few seconds. And then they're discarded and never used. If you have a program on this thing, you can write in C or C++. We've also ported C Python 2.7. We're working on Python 3. And we also support Lua for some reason. Why is this interesting to Python people? Well, most of our developer tools are written in Python. Testing tools and of course the thing that enables this, the glue between Swift and zero VM is this zero cloud thing. That's written in Python, of course. Everything is Apache 2. It's all open source. So use it, contribute to it. And if you want to find out more about zero VM, check out these websites. Come harass us on IRC or harass that Twitter handle right there. And that's all. Thank you very much. Another just gentle reminder, your phones have other settings than ring. Klaus Bremer, you're up next. Okay. So good afternoon, everyone. So my name is Dimitri Generov. I work for Google and I'm here to talk about zombies and application frameworks. So first of all, why zombies? Let's get this out of the way first. So some guys a few years ago wrote a scientific paper on modeling an outbreak of a zombie infection. And the math involved there involved solving a system of differential equations. And the sci-fi guys actually made a cookbook sample out of this. And what I did with this was deploy this to Google App Engine. So just to show that it actually works, so here's my little website. I can just change some values. Press the simulate button. And provided that my network connection is doing fine, it will show me the updated graph. I'm actually using my personal hotspot on the phone. Yep. So it works. But wait, this is an App Engine application. How could I deploy sci-fi and matplot to App Engine? This doesn't work. Actually, what I used here was a new feature of App Engine which is called manage VMs. And here's a link to some documentation of it. And what is essentially what this allows me to do is to run the App Engine runtime on the standard Google Compute Engine VMs. Which means that I still get to use most of the App Engine APIs that are familiar, like the data store, MAM cache, authentication. And in addition to that, I get to do a lot of stuff that wasn't previously possible, such as running background threads and processes, installing binary models, which I just created. I can have direct network and disk access, and I can even, with some caveats, direct the SSH into the machine running on the App Engine instance. And I also get the compute engine pricing. So unlike App Engine instances, compute engine VMs do not really start in like milliseconds. They start in minutes. So I actually have a long running VM, and I have to pay to keep it running. But still, it's relatively nice. So how can this actually be accomplished? So in my app.yaml file, in addition to the standard App Engine stuff that I usually have there, I add this key parameter which is called VM equals VM colon true. This means that I want to use this new manage VM stuff. When I do that, I also have to set up to tell which kind of instances I want and how many of them. So in this case, I'm telling that I want manual skating, I just want one instance. And here, I specify the type of the instance that I want to use. And one standard one is just some, it has one core, it has some amount of memory, just the default compute engine instance. And this wonderful App Get install line allows me to actually install binary packages onto my machine as it's installed. And I want to install NumPy, SciPy, and the Matplotlib package. And now, once I have done that, I can actually go into the Google Developers console and check the state of my instance. So this is my instance. I can see its state, I can see its IP address, and I can even press the SSH button to connect into it. And I think I have an SSH window already open. No, I don't. So let me open this again. And so this connects to the machine directly from my browser without any native, without any Chrome apps, without any native code, without any plugins, it's just an in-browser implementation of the SSH. And now that I have got there, I can actually, for example, get the list of processes and peek a little bit under the hood. So what's actually running there? Like there is like all kinds of interesting Google stuff running there. And for example, you can see Docker there. And you can see through that that my application was actually deployed as a Docker image. So this feature is actually now in limited preview. So if you want to use it for your own stuff, you have to sign up for the limited preview. So here's the link where you can do that. And if you have any more questions on this, then you can find me. I'll be around during the rest of the conference. Thanks, everyone. Thank you. Thanks, everyone. Yes, my name is Klaus Bremer, and I like to talk about the Alpha M Fritz box today. The box is a very popular Internet access router here, at least in Germany. And what you see here is a picture of an older model, but that doesn't matter. Some time ago, I have tried to access this box by means of Python. But unfortunately, I was not able to find any library that allowed that. And so I decided to write my own one. And I have named it Fritz Connection. And before you can talk to this box, you first have to know how this box talks to you. And this is based on UPNP and RIS style. And by the letter one, which is an XML-based dialect, the box tells you about their own API. And once you know this, you can start to exchange data by means of SOAP via HTTP. And the API itself is organized in services. Every service has a lot of action. And any action may get some parameters and may return a result. That depends on the action. Yes, to work with Fritz Connection, you first have to install it. Let's go by PIP install Fritz Connection. And then you may have to wait some time because it depends on requests. And LXML might have to compile, so this can take a few minutes. But afterwards, you are able to inspect the API. So this is just a two-liner. You say port Fritz Connection has FC. FC to make a long word short. And then you say FC print API. You get your offer. You send the address of your Fritz box, the IP address. But that's not, that's optional because Fritz Connection knows how to find your box. But you may have changed the IP or may have more than one Fritz box in your network. And then you give your password. And as a result, you get a very, very long list of all available services and corresponding action names. And the parameters for the actions as tuples here. The first item in this tuple is the name of the parameter. The second one, whether it's inbound or outbound. And the last one, the type of the parameter. Once you know this, you can start to use the API. This is done by the method call action. And call action needs the service name, the action name, and optional some parameters. And here is a very simple example. You can say call action one IP connection and false termination as action name. Then the box will reconnect and you may get a new external IP from your service provider. And because it's hard to remember all the service names and action names, you can write it and here is a shorter call. You can just call FC reconnect and it's done. There are more complicated examples for example this. There's a module named Fritz host, which lists you all active hosts which are connected to the box. This is a snapshot from my own home office, as you can see. And because this is a lightning talk, I see there is a repository for it. And there are a lot of links there to the documentation of the available service names and action names. And you can have a look at the code. And if you say, well, this code is quite ugly and I can do it better, so please feel free to improve it. In this sense, thank you very much. Okay, exercise. Pick up your phone. Unlock it. Check the icon that says whether it's on sound, ring, vibrate or silent. Make sure it's either vibration or silent. We are having currently one ring per lightning talk. I think we can get the ratio down a little bit. Thank you. It doesn't work if this is Apple. Is this an Apple connector? I don't see it. Okay. Ladies and gentlemen, my name is Larry Hastings. I was the release manager for Python 3.4 and I'm here to tell you a little bit about how it works internally and show you a problem that we were having and how we solved it with a new tool. And by we, I kind of mean me. I did most of this. So let's talk about the problem. Here we have a Python interpreter and it's running your wonderful code and you just happen to call OS.Doop2 and you pass in two handles, which are actually Python integers. So this goes into the Python interpreter and it comes out here. This is a C function that is the implementation of Doop2 for C Python. How does this work? And specifically, how do we turn those Python objects, which are H and H2, these Python integers, how do we turn them into native C integers, which is what the C code really wants to talk to? So here's the Doop2. There's the doc string, which is a C string, and there's the external interface for Doop2. Most of it is kind of interesting. It returns a py object star. Everything that is an object in C Python is a py object star. It takes in a module and here it takes the args and kwrs. And this is the interesting part, or at least for this talk. Args is appointed to a tuple and that contains the positional arguments. Kwrs is appointed to a dict and that contains the keyword arguments, see if there are any. But these are still Python values. How do we turn them into native C values, which is what we want to deal with? This is it. It's up to you to write this code. And so there's code that looks a lot like this all over the place in C Python. Every time that you want to deal with a parameter, you kind of have to write it in a bunch of different places. So in Python 3.4, we added this new inheritable parameter to Doop2. And we had to touch four different places. We had to add it to this keyword list so it knew its name for a keyword parameter. We had to declare the variable. We had to tell it what type it was. This little i means it's an integer and the pipe means it's optional. And this inheritable, that's how it actually writes it in. We actually had to touch five different places, but we forgot one. We forgot to touch the doc string. And the reason I point this out is because we're now up to five different places you have to touch when you add a new parameter. And it's kind of an error-prone process. So we're talking about adding a sixth one to get introspection information. Right now, if you call inspect.signature on os.doop2, you don't get a signature back. We wanted to fix that, but that would mean adding a sixth place. And now this seemed like it was going to be way too much work. It was going to be way too error-prone to manage all of these things. So I wrote a new tool. It's called argument clinic. The way this works is you write a comment inside of your file. It's literally a c comment with these extra funny strings at the beginning and the end. And inside of that, this is machine readable information formatted, sort of vaguely Python s kind of. It's not intended to look like Python. It's intended to be convenient for the person who's writing it. So you declare the name of your function. You declare your arguments, and you only have to write them once. Here you're declaring the arguments. You're declaring their default values. You're declaring their c-types. And you're declaring per argument doc strings, which is just a convention for encouraging people to document more, really. And then at the bottom, you have the actual doc string for the entire function. This is input to clinic. Clinic runs over the code, finds this, and then writes immediately afterwards in the c file its output. Its output is c code. This is similar to a tool, by the way, called the cog written by Ned Batchelder, which is a brilliant idea. So this is c code that's dealing with informing Python about dupetwo. So there's the doc string. And that's actually, by the way, where we hid the introspection information. It's that funny looking first line. Here is a method def, which is how we tell Python, here's a function called dupetwo. Here's the function you should call when you call it. This is the external implementation of dupetwo. And argument clinic writes that. And that has all of the argument parsing stuff. And it writes a new function for you. You write in the middle, called dupetwo. And you'll note, fd, fd2, and inheritable, it's done all the conversion for you. It's now in lovely native c types. And your code becomes much cleaner to read inside of c Python. All this upper code, this stuff, is actually hidden in a separate file. So you don't even see it anymore. That's about all I got. If you want to know more about it, you can read the pep 436, or you can look at the source code. It's shipped with Python, tools clinic.py, and it's only about 4,000 lines. Thanks very much. Really great. We got to know that. The next one after Sebastian will be Stefan Schwadzer. Are you there? All right, I can Let's try to begin. Wait for a second. Okay. Let me try to begin. Okay. What is the extent of that sweet? Okay. Okay. Okay. Okay. Hi, everyone. I'm Sebastian Kreff and today I'm going to talk about PEP 473, adding structured data to built-in exceptions. It's a draft PEP. So if you like the idea or you have any comments, please like contact me. I have started thinking about this when I was working on doing TDD on a huge code base and either because I was lacking some understanding of the code or typos, some test could fail and the error messages are not helpful at all. So the worst example is index error. So you don't get back either the offending index nor the size of the container. So it's really hard to get back to the test runner. So it's really hard to get back to the test runner. So it's really hard to get back to the test runner. So I started with a really hard solution. I instrumented the bytecode to temporarily store some additional information about the index and the receiving object and which is open source, by the way. And then I have a test runner that collects all this information and the output is nicely. So in this case, it's kind of much more direct to see that we have an off by one error. And so we can go and fix the code without having to divide it or add extra print statements or whatever method you like. And of course, the limitations of this is not portable. It only works with C Python 2 and also relies on the error messages. We are not standard within the standard library itself. So I decided to reach the community. I wrote to Python ideas and they were really supportive and they put me to some pre-existing issues related to this. Some of those are like older than 10 years old. And basically this is a summary of all the attempts of all these people, including Wido trying to have more useful exceptions. So for the case of index error we would add the target, the index, which is just an alias of the key. And for example for value error, we could have the unaccepted value. So then test runners could get this information and try to do some automated debugging for you. The same could be possible in interactive console like IPython or for diagnosing failing requests in a web application or running processes. And in the long term the idea is like with this information to provide a uniform and normalized error message for all the standard library. So if you find this interesting and want to see it implemented or have any comments go read the paper and send me an email..........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
|
Lightning Talks: Elliptics: Anton Tyurin, On Being Super: Austin Bingham, PYCON Finland: Mikko Ohtamaa, Regularities in language: Dmitrijs Milajevs, pip install pyladies: Lynn Root, Iktomi Forms: Harut Dagesyan, Killer Robots: Radomir Dopieralski, ZeroVM: Lars Butler, Zombies and Application Runtimes: Dmitry Jemerov, FritzConnection - Communicate with the AVM Fritz Box: Klaus Bremer, Argment Clinic: Larry Hastings, PEP 473: Adding structured data to builtin exceptions: Sebastian Kreft, Supporting Python 2 and Python 3 witth the same source code: Stefan Schwarzer, Birthday: Mark Shannon, nsupdate.info bepasty: Thomas Waldmann, Python Core Security: Christian Heimes, Hands On Unix: Rok Garbas, Deproulette: Joar Wandborg
|
10.5446/20053 (DOI)
|
. Nevertheless, we're going to have another 90 minutes of lightning talks, and it's going to be packed. I want you to stay after the lightning talks, though. There is something the organizers want you to be here for, and I'm not allowed to say anything more, except that having more people is better. Mysterious. Anyway, let's get started with the lightning talks. The first one is Stefan Nothausen, and he's going to talk about NetAdder. Hi, everybody. My name is Stefan Nothausen, and I will give you a quick introduction to NetAdder, which is a library that helps you to handle IP addresses, MAC addresses, and more. I'm one of the main contributors, and you can find the code on GitHub and also on PyPI. But first of all, I would like to show you a new build-in that was added in Python 3.3. You can now import IP address and IP network, and then you can do stuff with IP addresses and IP networks because there are now objects for that. So you can check if an IP address is in a certain network. And there's more goodies. You can do iterating, slicing, you can get the broadcast addresses, NetMass, you can even do IP arithmetic. NetAdder provides the same for you. We also have an IP address object, an IP network object, and you can do pretty much the same thing as with the IP address built-in of Python 3.3. So why would you install NetAdder? Well, first of all, it supports Python 2.4 and Python 3.0 because not everybody has Python 3.3 yet. And there's no further dependencies, no C-magic, nothing. Second of all, there's more than just IP addresses and IP networks. For example, there's IP ranges. NetAdder provides you with an IP range network object. And you can just say, I have an range that starts here, ends there, and then you can do operations like, is the IP address in the cluster? You can iterate over the hosts in your range, stuff like that. There's also an object for handling MAC addresses called the EOI because that's the official name of MAC addresses, extended unique identifier. And the funny thing about these is that there are so many different formats. Upper case, lower case, different kinds of separators. Some people leave out the leading zeros, but as you see in the example below, it's all the same for the EOI object. They all get parsed the same. That's the input. How about the output? Maybe you want your MAC address to look in a certain way. That's why NetAdder provides so-called dialects for you. One is the MAC Cisco dialect. And you can just instantiate your MAC address, set the dialect to Cisco, and then stringify it, and it will look like a MAC address that the Cisco systems are using. More such dialects are available. Customization is, of course, possible. There's OUI, organizationally unique identifiers, which are basically the first three bytes of a MAC address. And you can ask a MAC address for its OUI and the interesting part about the OUI is the registration information. So in the example you see on the screen, the registration says it's the Intel company from Malaysia. Always good to know stuff like this. And by the way, this works without Internet Access, the database is built in. Then there's sets of IP addresses. Wait a minute. Set is a Python built in. Why would you bother? Well, there's stuff like 10.0.0.0.8 and FE80.64, which are just really big networks. And you don't want to add two to the 64 objects to a Python set. That's not really fun. Here, IP set comes to the rescue. You import the IP set, you instantiate it, and then you can add IP networks, you can add IP ranges, you can add individual IP addresses, and you can mix IPv6 and IPv4 as you please. And then do the usual set operations. Check for membership, for example. Or in the second part, remove an address from the IP set. And even though that splits the IP range that was inserted in two pieces, let IP set take care of it. You don't have to. One disclaimer, though, your distribution may ship NetAdri already, but please use version 0.7.11 or higher, otherwise you will have a bad time. So to reiterate, the code is on GitHub, but also on PyPI. PyPI also has a link to the documentation and to a very nice tutorial. And that's it. Thanks for your attention. APPLAUSE The next speaker is Alex Apolsky about antifragile, and after that, Josie, you're next then. APPLAUSE OK. So I wonder how many of you have read any books or articles by Nesim Taleb, like the Black Swan or antifragile. OK, I see some hints in the audience. And I'm going to give a talk, a kind of more philosophical approach to Python. Well, Nesim Taleb is a kind of intellectual thinker who speaks about probability theory, about risk management, and how we can gain from instability from an unstable world. So the first book of his, of him, was Black Swan. It's about rare events that occur in our life and can change it completely, and that these events are not predictable, but we somehow should prepare for such events. And the second book of him is antifragile, is specifically about algorithm, how to deal with things that we cannot predict. There must be some way how to survive in unpredictable future. And in programming, we can speak about some problems that can occur to our code, to our system, and to our product in general. So three terms that he uses are fragile, antifragile, and robust. So fragile is a system that can break, and if it does not break now, it will break later and will cause a lot of, it will cause a lot of problems. So there must be a way to prepare, and we can think about what properties of such systems could be at the level of code, at the level of, say, Python module, at the level of application at all. So, for instance, ORM porting is fragile because if you, for example, change ORM, you have to redo all the application. If you do a product in a non-agile way, it's a problem too. Suddenly customer asks you to do something and you cannot deliver the product instantly. So, and if you write some code and do not keep in mind that there is some context of product stability, like you need login, you need error tracing, it's not fragile. So please use some login, et cetera. So what is robust? Robust is login, robust is transactions, robust is when your application falls back gracefully, for example, nice error displays, or if a part of your applications go away, still some front page will be working and your customer will be notified. And antifragile application is something that's not totally robust, but is kind of good enough. So what I'm speaking about is in Alex Martelis terms good enough, but I have ported into Nesim-Telib's concept of fragile, antifragile and robust. So he, one of the main strategies that he uses is the strategy of what's this heavyweight instrument is called, I forgot, barbell. Yes, the barbell strategy. Basically this means that you have to do very robust things and not to do some new things at all if they can break the system. But if something new and something cool does not break the system, you should exercise it and add it to your system. For instance, do not run for new things, do something that looks very simple, very stable and very understandable. But if new changes and cool changes do not cost too much, well, feel free to do them especially if they do not affect big parts of the system. So exercise new stuff in limited contexts and do very stable and simple things in general. So this metaphor is especially good for new programmers. As they come to a company, I used it to explain to middle and junior programmers how they can fit various ways of thinking about the system they build together in terms of fragile, anti-fragile and robust... All right. After Josie talked about Pixie Desk to us, there will be more coverage. All right. I see you. Do you want to use that microphone? If you look at that screen, you can see that everything is fine and they just need to switch it. Come on, guys, wake up. We are ready. Thank you. Hello, everybody. My name is Josie. I'm 13 years old. I come from Zagreb, Croatia. I got a scholarship from the Django Girls. So this is how I'm here. This is my group from the Django Girls project on the first day. And what is Pixie Desk? It is my idea to help girls from 13 to 18 start from being consumers to becoming producers of technology. So I think the big problem with technology for girls and teenagers my age is that we love it so much that we overuse it and we do not try to produce. So the consuming and producing would be 90, 10% and not 50, 50, which should be. I sometimes have this problem. I can be on multiple screens at one time and get lost in it for hours and then at the end of the day I feel so guilty that I didn't spend my weekend or day very well. My mom knows to call me a technology zombie because I can't get my eyes off of the screen. I'm just like, where am I? What am I doing here? So my solution would be bringing these girls from smaller countries because I'd like to start small than to big places because a lot of these companies have great ideas but start very big. They want to add big details, big, you know, get press and everything. And then they lose the money and go bankrupt, which I don't think anyone wants. So it's better to start small and grow, level up. So the resources that I need would be teachers, slash experts, sponsors and of course hosts. Yes, so this is my story and anyone can be a part of it. Thank you. Thank you. Dima Tijne, are you there? Excellent. For everybody who just joined, the organizers would like to stay you after the lightning talks. They want to do something and they need to be here. The more the better. Hi, everybody. I'm here to invite you to WebCamp. It's a conference in Zagreb, Croatia. It's organized by local user communities. And it's pretty awesome. So at this point, everybody asked me where I was. So it's in Europe. It's not far away and it's generally good enough connections to get there. So what's going on there? So in October, we will have for the third year in a row WebCamp conference. There is no specific topic, but all local user groups like Python, JavaScript, PHP, Ruby is involved. And I would like to suggest to you to submit your talk because it ends next week. And it's a great environment. Last year we had around 600 people. This year we are expecting around 800 people. So there are two parts of WebCamp conference. The first part is a week of WebCamp where local user groups have their meetups, smaller conferences and talks. And it's generally really relaxed. And I think everybody would enjoy it there. And the main event is WebCamp conference, which is something that everybody wants. You can expect talks about Elixir, Erlang, Go, Clojure, Python, JavaScript. Everybody can find something interesting. So once again, call for papers ends next week. So if you want to come, submit it. There are two places where we have this thing. First, smaller, for around 150 people is in the first picture. And the main event is like in a congress center similar to this one. And you can expect an audience of about 800 people this year. So just to confirm that it's a cool event, you can expect any kind of stuff to happen. And please consider coming. It's one of those small communities that are building a nonprofit conference and it's really cool. So see you in October. That's it. Thank you. APPLAUSE The next one after that will be Tosten Rehn. Are you around? Oh, there. OK. My name is Dimitri Snick and I hate testing. I know it's not a popular thing to say, but I'm here. We are pretty cool stage start-up. We are hiring in Bosnia. And I live in Bosnia now, which is really close to Berlin. We've got awesome developer Python club stuff. Please check us out. So now to the actual thing. So I had testing, but I'll leave that for later because I've got a lot to cover. First of all, let's see some code. So it's a real code. All names will change to protect the innocent. And as you can see, it has 100% code coverage. And first I have to say that coverage.py is awesome. It's at least five years old and it survived a lot. It is still being maintained. Now let's see what was the unit test that actually created this 100% code coverage. It works like this. In other words, only the function signature is tested at all. So what can we do about this? There is a formal thing called mutation testing. Unfortunately, it's not that great in Python, but it exists. There are two packages, PyMuteTester and MutePy. I hope I pronounce them right. PyMuteTester is dead. It's for Python 2 only and it doesn't work for me at all. MuteTester, on the other hand, is workable. It does work. It does work. It's only for Python 3. It produces results in very unintelligible format. And it requires unit test case which nobody should use anymore. So let's see how it works. First, let's have some simple test. Let's say this is code. We are testing it somewhere. We've got a decorator. We've got a context manager. We've got some logical operations. We've got a branch. And then let's run it. We give a magic command. Get a lot of output. What does it say? It says it tried 15 combinations of how to change my code and test it if my unit test still pass. 14 of them were killed by this unit test. So these changes were invalid. One of them survived, which means there is an alternative bit of code that still passes my test and that is a problem. Let's see how it looks like. So apparently, removing this decorator doesn't change anything, which I sort of expected. On the other hand, this lock is not tested anywhere either, but it didn't find it. So this tool is not perfect, but it exists. Moving on. In an ideal world, let's build a system that connects to a GitHub, takes your code, and if all your unit tests pass, or in fact all of your tests pass, it changes the code. It could do it in any way. It could be exhaustive search. It could be artificial intelligence, mechanical torque, whatever. But if your code is simple and the test pass, let's just issue a pull request right away and let the developer figure out whether the simplified code is actually better or maybe they need better tests. So I hate testing. Now for some fun. Amazonbee. You've seen this image at a keynote talk. I'm not a corporate drone. I'm a different kind of a zombie. I'm a zombie for a reason. I like brains. In fact, I believe that programs have brains and people that work with have brains and they're actually smart enough and we should not be constrained to some ridiculous systems. We chose a dynamic language for a reason. So while on the subject of keynote, both with polito's push for type annotation is completely misguided. Nobody said anything but whatever. Okay. If you want to get in touch with me, just remember my nickname. It's Jimo QQ. Jimo GitHub, whatever. It searches the Googles. That's it. Thank you. Thank you. Hi, everyone. So by now people get annoyed when you use tag lines that end in for humans. So there I fixed it. But I really want to talk about config management. There are a lot of config management tools out there. The most popular ones are probably Chef, Puppet and Ansible. And they're really great. They're really the workhorses of DevOps and have made config management in general somewhat popular. But I wanted a pony. A pony that is a tool rather than a solution, if that makes any sense. For example, I didn't want to maintain a server component. Also, having to install an agent on each node before you can start to bring that under management is kind of a pain, especially when dealing with legacy systems. And the same goes for SSH. I really want to use that as a transport and authentication mechanism because it's already there. Beyond that, I was also wishing for an interactive mode and a nicer command line UI that allows me to carefully review and pick each change that I make to an existing important critical system. And I really wish that PyPen was the only language I needed to know in order to use it. Last but not least, there's item level parallelism. This is a big one, actually. What that means is I wanted to continue working on small things like config files while it's installing, for example, a system scratch on the same node. I don't think any system currently does this. They all just work out a linear order and then go step by step. So about two years ago, I started hacking on this just for fun. And roughly 12 or 13 months ago, I threw it all away and started from scratch. So for the past year, a friend and I have been working on this and we're really trying to design a config management system that is extremely minimal and probably dangerously dynamic. Here goes. What we see here is just a file that has a dictionary in it that tells us there is a node called TARDIS and it has the MOTD bundle. Easy. Within that bundle, we have a file called Etsy MOTD with some inline content. You can also use templates, of course. Now, this is all just Python, which makes it very easy, perhaps too easy, to import and inject any data you like. It also makes for an interesting learning curve because you start out as a command line tool with a sort of JSON-ish configuration. But it quite fluidly transitions to feel more like a framework once you start to build these dictionaries dynamically. This is interactive mode. You get a nice color diff for this change in a file and can carefully view what you're doing. There's so much more features and bells and whistles to this that I can't show you right now. But just a few days ago, I tagged 1.0 of our tool called BundleWrap. After about 16, I think, alpha and beta releases. You can find out all about it on bundlewrap.org. There's more information, a quick start tutorial, and just generally everything you want to know. If any of this is even vaguely interesting to any of you, please come talk to me. If you just want to yell at me for reinventing some parts of Ansible, that's cool too. I'm here until Friday. You can find me as Tirain on GitHub and Twitter. Thank you very much. Simon Piazki, are you around? You ready? Yes. So I want to talk about a trick that helps us to keep code more clean. So sometimes we want to have to write such strange classes that if you see they are quite similar and the code is almost identical. So what we can do with that, actually this is a SQL alchemy class and SQL alchemy allows us to write this code by inheritance using the declared utter decorator. But I don't like this because it is too verbose. So let's revert and let's see what we can do with that. So as we know, Python classes created, its body is executed, and the result of that execution is passed to type. And the main difference between the function and class bodies is that function body can be executed many times. And so we can just move body from function, from class to function and call type in decorator. So here it is. This decorator do some trivial stuff. It just calls constructor and passes value to type. And returns to class. But there are some things that I don't like. For example, return locals. And I want just to remove it. Can I do this? Yes, I can. So. Revolved. Here is. And it's working. I just create two classes by decorator. These functions look like class. They behave like class. Probably they are class. So how I done this? I never met this before. It is a decorator that actually executes a function and extracts local variables at the end of execution. So I don't want to tell about how it works. So and we use it in real life. Like that. Here we have models. And we create similar models for different languages with little distinction. So here it is. Okay. No, I don't. Sorry. Simon Richard. It's your turn. Okay. So I would just like to tell you about some design flow that is with get at with the default argument and also has at. And why you shouldn't rely on these functions too much. We created a nice base class which is called character. And we created a function that is called get quote. It nicely formats some quote of this character. We also subclassed it. And we can see that it works here very well. We also think about ourselves. We are so smart because we can handle the situations where there is no quote. We know that Sir Ilympain isn't very talkative for some reasons. And here we handle it. Sir Ilympain won't say anything. So it's okay. Then we can add another characters. It also works very well. And we are very happy. But then we want to do something more. We just think that there are too many good quotes by Tyrion Lannister that we don't want to choose one of them. So we use the random module and we use random choice to select one of these quotes. We are doing this and what happened? Okay. Who saw this coming? Who saw that this would behave like this? Exactly. Not a lot of you. Okay. Who can tell me now what really happened here? Yes, there is a typo. But can we see from the behavior of this program that there was a typo? We can't. This typo here would raise attribute error. This attribute error would be silence. It would be called by this get-at. And we got our error silenced. This is something completely against Zen of Python. And this is something that makes it very hard to debug. This example here is super simple. But you could have this error raised some 10 frames deeper and you could spend two hours on debugging sometimes. It happened to me. So this is not a good way to do it. Now let's see if we change this just a little bit. Here we don't rely on get-at. We just set quote in our base class to none. And we check if it's still none or if it was changed in the inherited class. And now we can see, oh, snap. I made the typo. There is choice, not just. And if we... This is a thing that we can fix in two seconds and it works now. So if you don't have really, really good reasons, just don't rely on get-at and has-at because it can produce these kinds of very, very hard to debug problems. Thank you. APPLAUSE Antonio Ognio, are you around? All right. Richard, was it okay that I did not write your talk for you? Where is the... Oh, there it is. Okay. Where is it? Oh, there it is. Wow. Now I have to... Oh, wrong way. I think that screen is on your left side. Your down... Oh, got it. All right. So much for changing the plan at the last minute. Okay. That was a great idea. All right. Everyone can see that? Is that big enough? Yeah. All right. Good. So... Oh, that's the wrong one. So what I was going to start with was this... Oh, you want to... Cheers. So I was going to start with this golden rules of lightning talk thing, which was going to be really, like, ironic because, like, the first rule we all know, don't go over five minutes. The second rule, don't do live demos. But then everyone's been doing live demos. So... Well, I'm going to do my live demo now. Don't forget a piece of paper. That's rule three. Okay. So I'm going to show you a couple of... Just a few little things, however much I can fit in the five minutes. The first one is a module called E. Who's heard of the module called E? You don't count. You wrote part of it. Okay. So pip install E. All right. Off we go. So E is a rather funny little module. It's designed for the minus M command. So Python dash me. It's got some basic functionality. It can do evaluation of the pass... What the... Oh. This is why you don't do live demos, eh? All right. Let's try that again. Pip install E. All right. There we go. So it evaluates expressions you pass on the command line. It can do all sorts of stuff like import this and things like that. So that's really neat. One of the other things it can do which is kind of nice is say we're doing some Flask development and often when you're doing something like Flask development or whatever, you get a bit lost in the documentation. You have to kind of poke inside Flask to figure out how things are working. So one of the things that the E module will do is tell you where Flask is. So dash me and a module name and it gives you the location of that module. And so we ask for the Flask.config, for example. It will tell us where that is. You can also then say less and it will fire up less with the module so you can then have a look at it. Or if you say Vim, you know, it's in an editor which is kind of nice. It's just very handy if you then need to like insert and debugging into Flask to find out what the hell is going on to your program. So that's really neat. Now, who noticed some funniness when I was doing the pip installs there? Did anybody see how, I don't know, blazingly fast that went for a computer that's not even online? So the magic there is a thing called DevPy. Who went to Holger's talk today? For the rest of you, you should have gone to Holger's talk. DevPy is really cool. So you pip install DevPy and it's spelled pip install DevPy. It's already there. So it's going to say whatever. Okay. That installs, I've got plenty of time. Okay. So then you run DevPy server. Actually, I think it's already running. So, yeah, it's already running because I'm using it. So I'm going to stop it. Okay. When you start it up, it says cool, running, and there's a URL. You then run a command which is, oops, sorry, flip thing, DevPy use set config. This is all in the very basic startup help. You don't need to remember this. And that URL. Let's grab that. Okay. And that sets all of the configuration things that need to be set so that your pip commands and everything else will now start using your local DevPy server. What that means is it's a proxying cache which basically means every pip command you do, anything you install will be cached by DevPy locally. So subs will install commands, use the local version. Really, really, really fast. And it's trivial to use like I just showed. I've got 48 seconds. I'm going to show one more thing. All right. This, all right. So one of the other things that I do is I volunteer running the Python events calendar and Python user group calendar. And what I thought would be interesting yesterday was I grabbed all the calendar entries because I realized we were getting kind of busy lately. I grabbed the calendar entries and I grabbed the locations and I plotted them on a map. That's all the py commons. That's the, that are in the calendar. I chucked the user groups on there as well just for laugh. There's a few of us now. And that's my time.ику how enrollment. in the Yeah, what is it taking such a long time? Yeah, well I showed them. Do you have it? Yeah. Yeah, so that's what I asked. Okay. Okay. Okay. Yeah. Okay. Hello, my name is Antonio. I've come all the way from Peru, South America to this conference. I'm very happy to be here. And I want to talk about wrestle microservers with Python. So we need to ask first what is res anyway. Not everybody knows that answer very well. What is a microservice architecture? And what is it a good idea to write or build this kind of thing with Python? So what is res? Res stands for representational state traffic that doesn't tell us very much. In practice what res is is a way of taking full advantage of the existing web-related protocols and technologies. I mean something like web proxies, application firewall, client libraries and stuff like that. In order to manipulate objects remotely. That's what it is in practice. Taking consideration that res was not invented. It was reverse engineered after the existing weapon knowledge in 2000 by Roy Fielding, one of the original authors of the HTTP 101 specification. It requires for you to include links so your API is browsable. And most importantly, it requires you to respect the actual semantic of each verb. Like for example, you can never just get in order to make a change in the server side. If you are thinking that res is only mapping the crude actions to an HTTP web-ware, you are wrong. It's more than that. Please educate yourself a little bit. There's a lot of information on Google about this. So for example, there are more than 15 status codes that you should regularly be returning. Not just 200 or 404. In order to know if you're doing this right, for each resource, which is for each unique entity in your problem domain, you should be having only one URL. For example, you should be including a home document. A home document with links to all the other available resources. There are many rules like that, including hyperlinks. For example, if you make a search and you don't find the date in your search report, you should not return 404. You should return 200 with our resources explaining that there are no available results. So for learning more about these status codes, this is a website that really helps a lot, HTTP status.es. Please grab your copy of Roy Fielding's dissertation. It's a PDF. It's very easy to read, actually. I really like it because it compares this architectural style to others, and takes many chapters to justify why REST is a good idea. I also have a talk about this in English. REST for the rest of the fact. Just please Google it if you're interested. And these two books that other speakers have been mentioning during this conference are really good. So microservices. Basically, microservice is building a larger system or larger service, piecing together smaller services. Very much like doing a match up using APIs from Twitter, Facebook, and the like. The most important thing about this is that they have to be highly reusable, self-contained, conceptually simple, and easy to maintain. Instead of making a big monolithic system and having different people with different skills to agree on a system, you basically build cross-functional teams, like for each department. Each department builds its own web service, and you just use them together. And each one can have its own agile development cycle independently. So even when you deploy that, instead of having one huge process with all this functionality, you can have each functionality in a different process, which is far better for reliability issues. So what are the advantages of microservices? The most important thing is that once that you have defined an API, you can build the internals. You can rebuild the actual implementation without breaking anything. Other than that, only one developer can maintain the code because it's really simple. Martin Fowler, the very famous Martin Fowler, has been written about this. There's also a conference coming this November in London, which is very interesting about this. And why Python? Well, the most important thing about Python is productivity. You really have to prototype these web services. Camille, are you around? Excellent. You're next. And again, a reminder, at 5.30, there's a small surprise happening here. One note would be to keep some battery life in your computers. That would be nice. All right. I'm from Switzerland, and I'm actually not here to show you something, but to ask something from you. So we recently founded a new hacker space in my town. And one of our goals was to improve education and to, like, get children into technology. And one way to do that is in Switzerland, many primary and secondary schools offer something like a vacation program for children. So they get a vacation pass where they have many activities to choose from during the vacations. And we want to offer one of these programs, and I want to offer a programming workshop. So we thought about how to do this. So, um... Probably most of you know Minecraft, and probably even more children know Minecraft than you. There's actually something called Minecraft Pi Edition, and it's free. And if you have a Raspberry Pi, you can just download the binary and run it. And one of the good things is it's actually scriptable, so you can program the tiles. You can create a new block at a specific position. And the good thing is if you give one of the kids a Raspberry Pi and install that down there, you can actually go home, take the Raspberry Pi with him, and continue programming it. So my question is, if you have any experience with this kind of workshop, if you have good ideas, if you want to know more, then please contact me. And basically that's it. Thank you. Fernando Massanori, are you around? You'll be up. Good to have you after that. You ready? Yeah. All right. Hi. I would like to tell you about a tool, a tool for developers which I created. This tool records requests of web services. A basic idea of this tool is that all the time we call web services, we have to XMLs like request and reply. And to record them, to use them or record them, we can use XSLT transformation. There's web service. If we work with web service, there can be many times inaccessible. They're not, especially if you work with developer web services. So this can be really helpful if you use it in Dev mode. How it works. When you use web services, one, it's already working. It's record to the saved samples based on the config files. And when there is problem with web service, you can use service mode to use web service without noticing that it's not there. Right now, I will show you short demo. And here is three calls to the web services. This web service is really simple. It's just changing request to the reply. And I will call here. It creates this file which is recorded sample. Here is this response for this service. In the configuration for this message, we need to configure based on which parameters we want to create reply. Which parameters will be used in creating reply based on request. And all of them are used in this other sequence conditions. If you call more request, this file enlarges and there are other parts of it. If to prove that it's already serving from this. I hope. If we call web service again, it serves with changed files. As well, if you want, if web service has some errors, you can fix it in this sample files. Okay. Come back. I created a few configuration modes. Mostly, it's mix of recording and serving. And there is also transparent mode when it looks like there is no recording at all. So, what I'm already working on is implementing more, implementing XPAP2 functions which are not present in XML. But they can be very helpful if you create configuration file with this nodes. But it's just library only for the functions, not for them. And also, to not use whole blocks of XML, you can use nested nodes. So, it could be more complicated configuration file. In the future, it would be nice to create some support for rest and JSON. But unfortunately, JSON and Dress don't have such mature standards for XPAP or XSLT. Yeah. That's it. All right. Peter, are you here? Somebody is missing. Peter? Yeah. So, if Peter isn't here, then the next one would be Philip Kledzik. That's you. Excellent. Hi. I'm from Brazil. For the first time, by ComBrasil, it will be on the beach. Porto de Galinhas was voted by Magazine Best Beach in Brazil for 10 times consecutive. This is the venue of conference. The other thought. This is the Python community. Some speakers. And I invited you to submit call for papers are open. Thank you. Thank you. Jim Baker. Okay. I see you. I'm getting a little tired sitting and just for as a recommendation for you, just everybody, stand up once. Please stand up once. Stretch a little bit. We've been sitting for an hour consecutively. I don't want the other speakers to think we've all fallen asleep. All right. Some yawning. Okay. Thank you. Okay. Hi. I'm Philip Kłębczyk. I come from Poland. I wanted to talk about Python PL. Another Python. A little bit of history, but very fast because we don't have much time. Mostly photos. It started in my home city. We already had on the first edition over 100 attendees. Then we moved it into remote locations like mountains where weather in October is different depending on the year. We had a lot of attendees. Well-known people like from projects like PyPy. There's a very good atmosphere on the conference. We have three days of talks currently even four. We have different one-on-one speakers on the conference. We also have anniversaries. We do a barbecues in things like that. Yeah. Eyebrothers. And now we scaled up up to 300 attendees. We also have all the meals together because they are included in the price of conference. So let's talk about the current edition. So it has three to four days depending which option will you choose. It has three tracks, one full English track. We are ready for up to 500 attendees. We have parties and discussions all the time because all the attendees are in one place, in one hotel, which is awesome. And the prices are very low, I would say. And that includes accommodation and meals. So where the hell is Sturk? We always choose places that have difficult names because we love tongue twisters when people from US say, It's Sturk. So it's in the center of Europe. Look on this map. And near the border of Czech Republic and Slovakia. It's actually Besky mountains. It's southern Poland. How to reach? So there are nearest airports are Katowice, Krakow and Ostrava in Czech Republic. Ostrava is the closest one but has only flights to London and Paris. There are also Euro city, Euro night trains. You can also take a bus if you prefer buses with Wi-Fi on board. We will have over 50 talks and workshops. And at least half of them will be in English. Here you will have some of the speakers but the screen is too small to have them all on one slide. We don't have call for proposals because we've ended it two weeks ago but we have call for workshops proposals. So you can still send as a workshop proposal and if you will be accepted you will get a free ticket. There will be also communities from other countries on Pycom PL, some from Czech Republic, Slovakia, Ukraine and Belarus. I want to add that Pycom Belarus is the first Pycom in Belarus. So I recommend to support ladies that are organizing the Pycom Belarus. So we have an offer. If you are organizer of Pycom you can get one free ticket. It doesn't matter if it's Pycom Brazil, Finland or whatever or Pycom Germany. You can get a free ticket with accommodations and meals. So please contact us at Pycom PL at Pycom.org. And follow us on the web, so on Facebook, Google, Twitter, YouTube, so on the social services. That's all. I've made it. APPLAUSE I'm being totally naughty. If you are an organizer of a Pycom, please send the details of your conference to the Python Events calendar that can be in the calendar and everyone can know about it. APPLAUSE I got it. Are you around? Oh, well done. All right. APPLAUSE Can you see me? Hi, so I'm Jim Baker and I'm going to be talking about the state of Jethon. And I have 22 slides in less than five minutes. But I'm going to be talking about the state of Jethon. Because we've been doing so much work. It's under extremely active development right now as we're trying to close to get to final release by Q4. So I found in this development, along with all the other core devs and committers, that the language changes were pretty easy. It's the run time in libraries that we've definitely spent our work on. We can way more pages on the A and the script. The internet micro system is our current focus. Some stuff we're talking around GIB. You need to use Java 7. We have some interesting functionality. For instance mix Python and Java types. for interesting efficiencies, so on and so forth. There's this new work called Socket Reboot which re-implements the standard Python socket module as well as adding support for SSL and real support for select using NetE4, which is quite a really good event loop package for Java. And I would hope we would eventually use it for async.io support as well. So here's a good example of what we're using to actually provide socket support today. You can see it's actually just Python code binding to this NetE package. It actually is able in a very small function which is slightly cut off here to handle all differences between non-blocking and blocking with any timeouts for a given socket. This enables requests. So you can now use with the latest version of Jython, certainly what's in trunk, this popular client. And since it's used by PIP, this is really good. PIP used to work, but because of this change to using request and also the support for SSL that is now required when working with PYPI, this was important for us to do. We almost have this complete and it should be in the next release of PIP. Regular expressions used to have some problems and were slow. I heard someone talking about a potential cherry pie performance. I don't know about your specific case, but this might actually fix it for you. It was really cool to see this great commit by someone who contributed this, Indra Tullip. We are now also developing stuff outside the usual release schedule in Jython. So for instance, I'll talk a little bit about Clamp and Jiffy. There's also this new fireside, WISGI bridge to work with several containers. We've had that in the past, but this knows about site packages and is PIP compatible and is really nice in that fashion. Clamp allows you to directly import Python classes into Java. What does that look like? Let's say I had this Python class that extends some Java interfaces. Really simple. It's really implementing the callable interface. You add two lines here in terms of being able to say from Clamp import this Clamp base, Meta-class factory, use that Meta-class factory function to generate your bar base and then you add this bar base to your list of bases for that class. And now with just one more step in terms of in your set of tools using Clamp, you can go and this is now Java code. Sorry to show this to all of you here. That here you are directly importing a Python class into your Java code without any additional work. Just one line. How's that? It's pretty awesome. And Darius and I, he did a great job working out this with, you know, it was fun. So Jiffy were planning to provide CFFI support. There's this interesting patois. Jainai is a project that allows C extension API support. There's going to be a sprint on this in auction on Monday. We're also having a sprint on Saturday about Clamp. Again we're planning to get a beta four out. Release Canada is as needed. Mostly this is about going and there's obviously if you, for what we want to do, Java integration is key. Java nine is going to add some additional support. And what about Jython3a.exe? I'll let you guys read this as my time runs out. And some blomberg, are you here? Excellent. You next. Peter Koppatz, did you arrive here by accident or purpose in between? I'm going to pass it over to him, yeah. There you go. Okay, Paul. Okay. Can you give me the system settings? It should be. I mean, it doesn't have an html. It just has this one. Yeah, yeah, yeah, that's okay. Yeah. That's okay. That's okay. Yeah, yeah, yeah. Yeah, yeah. Yeah, yeah. Is it built in digital? No, it's not digital. This is HDMI. Do you have another adapter? I think I have one. It was just a matter of time. Yeah, yeah, yeah. Yeah, that's good. What do you mean? Yeah, analog. That's code of conduct. Okay, hello guys, I'm Agata, and I just want to have a short talk about how to make Python more sexy. The idea is pretty simple. We just have to improve Python to be more sexy. So let's start from the beginning. We can just simply improve packaging. Just guys use Python gems. And of course, improve parallelism, like replace the Jil with V8 and use more callbacks. What about the style of Python? Adopt an explicit Java-like naming convention because readability counts. And of course, guys, it's really simple. Just use more globals. No imports. Use file includes. And just build in carrying. And after all this stuff, I just want to say you one more thing. I was just kidding. Python is as sexy as it can be. And it cannot be more sexy, guys. And I want to thank you so much to inviting us for our Europe Python conference. And we just love you. And I hope to see you next year. And of course, I hope to see you in a shterk because Python Poland is waiting for you. Thank you very much. Thank you. Thank you. Thomas, you're after that. And again, there will be something at 5.30 where you need your computers. So save battery or plug in. I hope you've got the mobile phone and the tablet with you. We'll put those to some good use later. You're ready? Yep. All right. Hey, guys. My name is Anton Blomberg. I'm in my Swedish distributed systems gig. And I want to talk about parallel execution. Well, everybody knows there's a gel in Python. And therefore, as Pythonists, we're not very used to dealing with this concept. And, well, if you actually could do parallel execution, why would you? Well, of course, increase the speed of your program and reduce the latency. And maybe why would you not? Single-fared programs, they're easy. Multiple threads, it's really hard. And why is this? It's because you introduce concurrency into your programs. It leads to race conditions, deadlocks, and all sorts of nasty problems. They're really hard to debug. So what you do, you add locking, synchronizing, and then your code base is so hard to understand. But there is an alternative to program systems that's running concurrently. It's the actor model, which is a very old way, has been around for ages, where you do parallel concurrency without locks. So this is not your regular threaded model, where you acquire a lock and do your stuff, release the lock. Everything is done in independent share-nothing tasks. This is a bit similar to all the G event tasks with libraries, Node.js, blah, blah, blah. But this is more a way to organize your code to make it easier not to lead you into callback help. So what you do is you communicate by passing messages between these actors. And every actor is in itself like a process. It has a shared nothing with all the other actors. The only way to communicate is to send a message. And all messages are sent asynchronously and put on the queues for incoming messages for all the other actors. Since each actor can only execute one piece of code at once, you cannot go into race conditions because everything is in itself atomic. An actor can only receive one message at once and process it. So what you get with this is execution transparency, network transparency. As a coder, you don't need to know where your code is executing, in how many parallel threads it executes, or if it actually executes on another system, on another machine. So what you actually get is distributed systems for free. So if you want to, if you like Python, then maybe this is bad news for you. But the PyPy SDM, it actually has potential to bring this into Python. What it adds is parallel execution if there are no conflicts. And in an actor system, there is a shared nothing architecture, which means no conflicts. This means you can run Python code fully parallel on all cores. So what I did yesterday after having a couple of beers, was actually try to implement this in a Python only implementation to make use of an actor system in a fully distributed parallel environment on PyPy SDM. And well, it kind of worked. So it's not a library, it's a proof of concept. What they basically did was implement Actress, a multi-threaded, fully parallel event loop. A lot of buzzwords, but it's not your G event event loop that you execute in synchronicity, but sharing the time between tasks. You actually execute fully in parallel. Since actors share nothing, you can process the cues of the received messages in parallel and do all the actors on all cores. So what they did was basically give the application developer a generator statement and receive a new message with a yield and built into the language. You can now receive a message in a synchronous manner and execute fully in parallel. So it works in C Python. It's totally regular Python code and it shines on PyPy SDM, well, kind of. The overhead is still high on SDM, but it actually improves the more cores you add. So as a fun exercise, I also implemented pure Python-efficient immutable data structures like an immutable dictionary which gives copy and write semantics but doesn't actually copy anything. So you can find this at if you go to deadlock.se where there's a link to the GitHub so you don't need to remember it. And if you like distributed systems and see the potential for Python to actually run stuff in parallel in the future, donate huge amount of cash to the PyPy SDM guys and give these guys a round of applause. Christoph Neumann, are you around? All right. Can you make it in four minutes? All right, that's excellent. I have two announcements to make. It's both related to having private email. And to begin with, we'll make a GPG, PGP key signing. It's right after the lightning talks and surprise thing. So we meet at 1745 in the basement in the lounge area where there was so far and that stuff is. And just look for me and for Arnold Grille, we'll kind of lead this. And if you need another motivation after the first keynote, and if you still think you have nothing to hide, this is the slide for you. So what can we do if you don't use any GPG yet? Maybe in reality, you don't have private email yet because they are reading your email. It's not because they are especially interested in you, but just because they are reading all email. That's not encrypted. It's easier than you think. If you want to start right here at Europe, Python, we will have another event and it will be in the bar camp. And we will have a small crypto party. You just come there with your laptop, have Thunderbird installed, and we'll help you to get GPG going and to create your keys and so on. Just meet us there. We'll try to make a session about this. You can also use another email client. Just make sure there is some sort of GPG support for it. So for the key signing for the event today at 1745, what do you need? You need a PGP key, of course. It's a key signing. You need your name, email and key ID and fingerprint, for example, on such paper strips or on your business card. Alternatively, you can also have this data just on your laptop screen if you don't have any paper with you. Then you need a valid passport or ID card. So the procedure will be about like this. We will just row up in a kind of circle, but it will be a flattened circle. So everybody is facing somebody else. And then one guy will just verify the passport of the other guy. And we will compare the identities on the paper strip. And if everything is okay, you just take the paper strip and you can sign the keys later at home. You don't need to do it here. Then we just rotate right by one and the next verification step happens. And we repeat this until we all have checked everybody else. If you have no paper strips, you need this. Either you have a mobile client maybe on your Android phone or whatever. You can just show the key info page of the GPG implementation. Usually the info page shows your key fingerprint and then the other one can just make a photo of it so you don't need to exchange any paper. Also, if you have a laptop, you can just enter that command to print your key IDs, your email addresses, your name and also your fingerprint. Then just copy and paste it into a world processor and make it full screen so it's easy to photograph it. And then the other guy will just make a photo of it instead of taking the paper strip. What you need later is a tool called CAFF. If you run Debian or Ubuntu, it's in the package called Signing Party. And if you have a lot of keys to sign, that will save you a lot of work. You need a lot of work to set up that tool once because you need a local email server or at least some sort of relay agent. But it will save you a lot of work if you have done it once. Alternatively, you can also do it manually, but then you have more manual work. And the tool is rather easy. You just say CAFF and key ID, then it downloads the other guy's key from the server. You can also use the fingerprint so you can compare with the paper or the photo you have taken. Then you can sign each of his email. And the tool will automatically create multiple emails for all the email addresses you wanted to sign and sends each signature to the corresponding email. Yeah, this is the final slide for everybody who has maybe not seen it yet. It's from Germany. Thank you. So I'd like to get an overview of how many computers we have in here. Who has at least one computer with him? All right, that's a fair standard. You don't have to show me the computer. Who's got two? And computers, anything that has a CPU in it? Who has three? Okay, now that's the computer or smartphone in a tablet who's got four. So the fourth one, does that have network connectivity? Who has four with network? James, you are awesome. We need you. You're going to be a workhorse here in a minute. Ready? Ready. Let's go. Hi. My name is Christoph. I'm co-founder of Quantified Code. And what we do is we do data-driven code quality management. So we develop software, data-driven software that helps you monitor and improve your source code. And we are lucky we got a funding from the German government. And so we can, for the next year, spend our time without a lot of business pressure to build a great product. And I'm here today to ask you to help us to give us feedback and to check out our first alpha version because we want to build something that really matters and that really is going to be used by the community. So what we are striving for is basically we want you to check in your repository on our website. Then we will offer you continuous check and monitoring of your code quality. And finally, we want to not only check the code, of course, but we want to display results. We want to give you suggestions. And in the final state also offer you automatic improvements to the code. How do we do that? Basically, we apply machine learning algorithms to a lot of source code. Basically, to most of the source code you can find open source and GitHub. Where are we right now? So as I said, we introduced an alpha version. So we can already check the source code of open source, public Python repositories on GitHub. And you will already get commit by commit a report of what has changed. Did you add bugs or did you fix bugs? So let me show you the website. So this is basically our alpha version. So what you see here, if you visit our website, is basically you get a list of all the projects we already crawled on GitHub. So you can let's go into one of them. There is, by the way, no sign up needed. You don't need to do anything. It's all the projects can be used right away. And now you get an analysis of the mistakes and errors in your code. How did we do this? As I said, again, it's an alpha version. It is the first step. So what we did is we used common linters, pyflakes, PEP8, and so on. We checked if the messages, they give you a relevant. We categorized these messages. And here you can filter whether the message or the error is a critical one, a potential bug, et cetera. Let's zoom into one. So you click here and then you go directly to the piece of code where this issue occurred. So now what I really love to get from you or anyone who's interested in static code analysis, code management, code quality management, or any programaries, now this is a simple alpha version. It has functionality. It gives you a few of your code, but I would like to know what are the features you really need? What do you need in terms of workflows? Where do we need to integrate into Jenkins or any other systems? What are the common mistakes you get from other linters which you would like to see solved? Yeah. Just reach out to us. We want to develop the product with you. All this is here free and will ever stay free for open source. And also, as I said, we got a government funding. We have free slots there. So if you're interested in working with us for the next few months on the cool product, then get in touch with me. Thanks. All right. Was that a good set of lightning talks today? Was it? Thank you.
|
Lightning Talks: netaddr: Stefan Nordhausen, Pyxie Dust: Josie Zec, Tina Zec, Webcamp: Aljosa, Who watches the watchmen? Dima Tisnek, Config management for humans: Torsten Rehn, Some magic for class factories: Harut Dagesyan, The great Python design flaw: zefciu, Some cool stuff: Richard Jones, RESTful Microservices with Python: Antonio Ognio, Teaching Python to kids with Minecraft: Danilo Bargen, Record your Web Service: kkuj, PyCon Brazil: Fernando Masanori Ashikaga, PyMove3D translations: PeterK, PyCon Poland: Filip Kłębczyk, State of Jython: Jim Baker, How to make Python more sexy, Actors in Python: Anton Blomberg, Keysigning Intro: Thomas Waldmann, Data driven code quality management/automation: Christoph Neumann
|
10.5446/20050 (DOI)
|
You're welcome to the performance Python talk for the American algorithms. My name is Yves Hildpich. I'm the founder and managing director of the Python Quants. As the name suggests, we are mainly doing work in the financial industry. So my examples will be primarily financial, but I think they apply to many, many other areas. So if you're not from finance, you won't have trouble translating what I present today to your specific domain area. Before I come to the talk, a few words about me and us. As I said, I'm a founder and managing partner of the Python Quants. I'm also a lecturer for math finance at Sarlton University. I'm co-organizer of a couple of conferences and organizer of a couple of meetups. I actually all center around Python and Quant topics. I've written a book, Python for Finance, which will come out at Urali this autumn. It's already out as an e-book. I will show it later on. And another book, the Routes Analytics with Python, which will be published by Wiley Finance next year, probably. Apart from Python and Finance, I like to do martial arts, actually. This is the book. And actually, today's talk is based on chapter nine of the Python for Finance Urali book. As I said, it's already out as an early release, as an e-book, and the printed version will probably come out at, well, let's say mid-November is kind of the date. I'm finished with my editing. I hope Urali will come out with a printed or final version pretty soon. There's also a course right out now on Quants up, actually, which also covers the topics that I present today. It's completely online based one. Maybe you want to have a look when you come from the finance industry. I think then the benefits are the highest here in this area. We are doing otherwise is at the moment mainly working on what we call the Python-Quant platform. We want to provide a web-based infrastructure for Quants working with Python and applying all the nice things that I present today. I will show it quickly later on. Maybe with a couple of examples. We have integrated an iPad notebook there. We have an iPad and a shell, easy file management, web editing. Anything you want to need. In addition, we also provide our proprietary analytics suites, decision analytics on this platform. That's enough about us. Now about the talk. What is this talk about, actually? When it comes to performance, critical applications, two things should always be checked from my point of view. Are we using the right implementation paradigm? Sometimes this boils down to what is typically called idioms. Are we using the right performance libraries? I think many of you might have heard the prejudice that Python is slow. Of course, Python can be slow. But I think if you are doing it right, with Python, Python can be pretty fast. One of the major means in this regard are performance libraries that are available in the Python world. I can only briefly touch upon all of these that are listed here. I think there was yesterday the talk by Stefan about the Thyson. But any topic that you see here, you can have complete talks or even complete tutorials for day or even for week for some. So it's a complex topic, but my talk is more about showing what can be done. The main approach will be to say, this is what it was before. It was a little bit slow. Then we applied this and that, and afterwards we see these improvements. We don't go that much behind the scenes. We don't do any profiling during this talk. But you will see in many, many cases when it comes to numerics, Python and these libraries can help in improving performance substantially. Let me come to the first topic, Python paradigms and performance. As I said, what I call paradigm here in this context usually is called idioms, for example. This is just a function that you see here. Don't try to get the details. This is just a function that I will use regularly and have provided it here in the slides that you can use it afterwards as well. It's just a function to compare a little bit more systematically different implementation approaches and compare performance a little bit more riturously, but there's nothing special about that. Let me come to the first use case, a mathematical expression that doesn't make too much sense. We have a square root, we have absolute value, we have transcendental functions in there and a couple of things that are happening there. You might encounter these or similar expressions in many, many areas. As I mentioned before, in finance and math finance, you have these in physics and you name it in almost any science as of today, you find such or similar numerical expressions. We can implement this pretty easily as a Python function. As you see here, it's a single equation and we translate this mainly in a single line function. Nothing special about that. What we want to do, however, is to apply this particular function to a larger array, to a list object in the first case with 5,500,000 numbers actually. This is where usually the computation of burden comes in. When you have a huge data set and you want to apply these expressions to the huge data set, it's not that the single equation is complex, but it's in the end the mass of the computations that makes it typically slow. To start working with, we generate a list object using simply range with 500,000 numbers. What we then do is to implement another function which uses our regional function f, implementing the numerical expression, where we have a simple looping. This is the first implementation out of six that I present. This is a pretty simple straightforward function where there is a for loop in there. We have another list object and we just calculate the single values and append the results to the other list object. The function then returns our list with the results. A second one, second paradigm or idiom, if you like, is to use list comprehension. Actually the same thing is happening as before, but it's a much more compact way to write the same thing. We generate in a single line of code the list object by iterating over our list object a and collect the results given the values that the function f returns. A little bit more compact, maybe better readable, but if you're a Python coder, you might prefer this one. We can also do it. This is quite flexible. I wouldn't suggest to do it in that case. We will see this will be the slowest one, but it's very flexible. We are working, for example, with classes, objects, where we value derivatives and derivatives that have complex payoffs and so forth. You can describe these in a string format. It makes it pretty flexible to provide different payoffs for these classes. This is, for example, one area where we use it, but typically when we use it, it's only once that we have to evaluate the expression. In this case, you might notice that the expression is evaluated per single iteration of the list comprehension. As we will see, this is a very intense, a compute intense or interpreter intense way to do it, to like every time I iterate over the expression to evaluate it. This will make it pretty slow, as we will see. Of course, if you're working in numerics or science, you would be used to vectorization approach of NumPy. What we can do is implement the same thing, this time now on a NumPy and the array object, which is especially designed, of course, to handle such data sets and such data structures. With a single line of code and using vectorized expressions, we can accomplish the same thing. Now we were working on NumPy and the array objects and using NumPy universal functions to calculate what we're interested in. This is similar to the list comprehension syntax, but in the end, we would hope for speed improvement because this is especially designed to exactly do these kind of operations. Then we can also use a dedicated specific library, which is called NumExp for numerical expressions. Here, in this case, we again provide the whole expression as a string object. But in this case, actually what happens is that this string object, this expression, is compiled only once and then used afterwards. Here again, we are working on NumPy and the array object. NumExp is especially designed to work on NumPy and the array objects. In this case, we would also see hopefully some kind of improvement because it's kind of a dedicated specialized library to attack these kind of problems. You might have noticed that in the first example, I have set the number of threats to one to have kind of a benchmark value. We are only using one threat, one core in this case. Here I increase the number of threats to four. So if you have a four core machine, you would expect here kind of an improvement. But what kinds of improvement? Let us have a look. In summary, again, we have six different paradigms or items used with Python. In the end, this is kind of delivering in any case the same result. As is often the case, when you see people coming from other languages, coming to Python, being new to Python, not knowing all the idioms, they are using probably those that are used to from other languages like C or C++, you name it. Sometimes this can be pitfall in the sense that they are using maybe the wrong paradigm, the wrong idiom. But let us have a look at what the differences are. Now our comparison function comes into play. We have a clear winner, obviously. We have a multi-threaded version. The F6 was the last one. We are using multiple threats to evaluate the numerical expression on the array object. Then we have the single-threaded one, which is the second fastest. The third one is the NumPy version. And then the Python ones follow after that. So we see actually this kind of a given the list comprehension, for example. We have a 28 times increase in performance using the multi-threaded NumPy version. And as I mentioned already before, the F3, this was the one which uses the built-in evil function of Python. You see that we have a speedup in total here of 900 times. These can vary, of course, depending on number of threads they are using and so forth. But the message, I think, should be clear. We have in Python many, many ways to attack the same, the very same problem. And all the ways will yield the same results. But there might be considerable performance improvements when going the right route and avoiding pitfalls and especially avoiding implementations that are per se compute intensive. So this is, for example, where profiling would come into play. I don't present it here. I said my approach is more like this is before, then we do something, we compare it, and this is what is afterwards observed. But profiling, of course, would have revealed that evil is kind of a very time-consuming function and most time is spent, for example, with F3 in this type of implementation. Let me come just briefly to a rather subtle thing. When we've seen the numerical algorithms implemented based on NumPy and the array objects, be it directly by the use of NumPy universal functions or by using num expression, have been the fastest. But this kind of, in certain circumstances, and I encountered that quite a while ago, and it was first like a little bit like I didn't know what's going on in there. But later on it became pretty clear what was going on. So it's, from my point of view, worth mentioning, depending on the specific algorithm that you're using, that memory layout can play indeed a role when it comes to the performance. Let me start with a typical NumPy and the array object, which we instantiate by providing the D type, in this case, float64, and here the order or the memory layout comes into play. We have two choices with NumPy. There's C for C-like layout and F for Fortran-like layout. In this case, you don't see any difference. There's nothing special. You see just the numbers printed out. But don't get confused because this is just a graphical representation of what data is stored actually at the moment. But if we have an array like this, you can explain what memory layout is all about, actually. When we have to see like it has row-wise storage, we provide here the order C, this means that the ones, the twos, and the threes are stored next to each other. So this is how in memory, I mean memory is a one-dimensional thing. So we can store it given a unique location in memory. So we don't have kind of two-dimensional, n-dimensional things where we can store data into. It's kind of a linear thing. So we have to decide how to put multidimensional things into memory. This is how is it stored when you use the order of C. Using the order F, then in this case, we have a column-wise storage, which means that the one-through-three, the one-through-three, and the other one-through-threes store next to each other. Working with such small number in the array objects doesn't make that much of a difference. But when you come to larger ones, and in particular to asymmetric ones, like this one, we see we have three times 1.5 million elements in there, then we can expect some differences in performance. We essentially ate two different NERA objects here, the one with the order C, of course, and the other one with F to just compare it. But we now start calculating sums, for example, the C order. You see already kind of a difference when you're calculating the sums over the different axis. So an umpire is, of course, aware of the axis. List objects, when you construct like two-dimensional things with list objects, there is no awareness, or there's kind of no attribute for the axis. But in this case, we can calculate the sum row-wise or column-wise, if you like, and you see there's kind of a huge difference, here like a 50% difference when it comes to the two different axes. Only the performance of calculating sum. The one delivers back kind of 1.5 million, one-dimensional result, the other one returns a result which has only three elements in this case. But of course, the numerical operations are running differently over memory for both cases. For standard deviations, you observe the same thing. So according to results here, going along axis one, which means the second axis, of course, with the serial numbering is much, much faster than the other way around. So if you have these problems and you have to implement, it might be worth considering really doesn't make sense to have a three times 1.5 million array or 1.5 million times three array. So you will see considerable performance improvement going the one way or the other depending on what you're exactly interested in when it comes to the calculations. Third sums with the F order and the array object. You see, these operations are actually both slower. They're absolutely slower than the C order operations, but you see different relative performances. So in this case, doing the sum according along axis zero, which means the first axis, is faster relative to the other axis. The same actually holds true. Not really. This is pretty close for the standard deviations. And you see, this is also the absolute disadvantage might be due to the fact that C is the default and the C world in NumPy is a little bit more maintained or more important. But once you have to, for example, interact with the Fortron world and you are required, so to say, to work with the F order, then it might make sense again to consider the question is three times 1.5 million better or 1.5 million times three. So you will see in certain cases, huge differences. Let me come to another approach, which is usually, and I think all the approaches that I present today are like in a certain sense, low hanging fruits. There's typically not that much involved when it comes to, for example, the redesign of the algorithm itself. So I don't do any redesigns of algorithms. I'm always sticking to the same problem, to the same implementation and then showing what it can do. Sometimes you will see that, of course, using different libraries, you need to rewrite something, but it's not about the, for example, parallelization of a certain algorithm. Where we present now it's more like given a certain implementation of an algorithm, what can I do with parallel computing, actually? And SI already announced before, I'm used to use these financial examples and here is a Monte Carlo problem, which is about the simulation of the Plexigold's Merton Stochastic Differential Equation, which is kind of a standard geometric pruning motion, which also applied in many, many other areas in physics and so forth. What I want to do is kind of simulate this model and value a European call option. Don't worry about the details. I want you to say that this is usually kind of a very compute-intensive algorithm to work with and that might benefit usually from parallelization. But first the implementation of the algorithm. What I do here is kind of already a, I wouldn't say optimized implementation, but at least I'm using NumPy and using vectorization approaches to be faster than, for example, the typical Python looping that we have also seen as an alternative before. I could have done this also with Python, but this is the point here. I want to stick with this NumPy implementation and see what we can do when we parallelize the code. You see I have the import function here within the function because when we use iPython parallel, which I will do here, the whole thing will be pickled and we have to import within the function to get everything to the single workers. First as a benchmark, of course, is sequential calculation. This example is only about calling for a couple of times the same function and parameterizing the function by different strike prices, in this case. But again, you can replace this with any function you're aware of, which is similar from your specific area. And what we're doing here is kind of, indeed, just looping over the different strikes we interested in and collecting the results that we get back from the function. Nothing special in this. It's a simple loop collecting results and we're finished. So you see here we do it for 100 of option calculations and we get back the strikes, the list of strikes, and the results from our function. And in this case, the 100 calculations take 11.4 seconds. Just the results visualized that you get a feel. So going over the strikes, well, in European call option means the higher the strike, the lower the value. So this is what we would expect, so obviously, the function works pretty well. Now the parallel calculation. We use here, and there are many alternatives. I've seen already Celery and I know that there will be a couple of talks about alternatives, but IPython parallel usually, as I said, is kind of a low hanging fruit. Many people are working with IPython notebook these days and this is very well integrated. So we can just import from IPython parallel our client class, object here, and instantiate the client in the background or using, for example, the IPython notebook dashboard. I should have fired up already kind of either a local cluster or when working really in the cloud or in cloud-based services, you can have huge clients. So the largest ones I've heard of were about like 512 nodes. IPython parallel is known to be not that stable when it comes to like 1,000 nodes, for example. So it doesn't really scale beyond a certain point. But still, for example, people doing research or for smaller applications is kind of a pretty efficient way. What I'm doing here, once I have a client given my profile and my local cluster, for example, I generate a low balance view in this case. And the code that I need to do the same, what I've been doing before with the sequential calculation, it's just almost the same. They're kind of two difference, actually, worth mentioning. In this case, I don't directly call the function. I rather asynchronously apply my function given the parameterization to my view. I append the results and I have to wait until all the results have been returned, otherwise the whole thing will break down. So these are kind of only two lines, if you like, are attached in the code. And this is not even in the algorithm. This is just how I collect the results. So there's not that much overhead given the sequential implementation. We might have had three new lines and four new lines in total. And one line of code has been changed to implement the different approach, actually, in this case. And the parallel execution. I'm a little bit surprised. Why does it take 29 seconds? The wall time is not a white one. We have the, I've been looking at the wall time. But the total time for the execution was five seconds here, actually, in this case, because we have used multiple cores. So it speeds up by a factor here where we are. We have started, let me get back, like 11 seconds and a little bit. Yes, 11.4 seconds. And we end up here on this machine at five seconds, a total time. But to have a little bit more rigorous comparison, I come back to the performance comparison by again applying my performance comparison function. But here you might have noticed that implementing this approach leads to different results, actually. You don't get back kind of only the number. What we get back is the whole set of results in a metadata which the single jobs are returning. So for example, having a look at the metadata, you get much more information, like when it was completed, when it was submitted, and so forth. But we remain in the industry, of course, in the result. And this can be retrieved via this attribute. We have this results object. Here's the attribute. And in the end, I can hear why another loop collect all the single results from a parallel application of the algorithm. And just to compare here the sequential and the parallel calculation, of course, there are numerical differences because we're working with a numerical algorithm which implements simulations. So we would expect kind of numerical errors or differences even if you're doing the same thing, be it parallel or sequential. But what we are most interested in is the performance comparison. And to this end, we have used the function already. And you see here working on a machine with four cores in it leads to a speedup of roughly 3.1. So of course, you have an overhead for distributing the jobs and so forth, for collecting yourself. But in the end, you will see that applying this approach typically scales almost linearly in a number of cores. It's not in a number of threads. Hyperthreading for these kind of operations don't bring that much, but you would see usually, as I said, almost linear scaling in the number of workers. So for example, working with another server, we use these approaches with eight cores. There you see like kind of speedups of seven times points something. But again, not that much overhead involved. We haven't changed the algorithm at all. And by investing maybe an hour of work or whatever, you might improve your numerical computations considerably. If you're only working locally and are not interested in like spreading the parallelization over whole classes or whatever, then there's, of course, the built-in multiprocessing module. Again, I press on parallel scales over smaller medium-sized classes, but sometimes it is helpful even to parallelize code on local machines. And I mean, I know the percentage, but most machines as of today, even the smallest notebooks have multiple cores. And even using two cores already might lead to significant speedups. When you now think of a larger desktop machine, where you have four or eight cores, you will see also significant improvements. And again, the fruits are low-hanging in this case as well. So we import multiprocessing as MP. And our example algorithm here is, again, Monte Carlo simulation. This doesn't do the valuation, but this doesn't do actually the same thing. It does the simulation. So there's not that much of a difference. We have a different parametrization here that we apply. But in the end, it's kind of the same core algorithm that we use here to compare the performance. What this does is kind of gives us back simulated paths. In our case, it will be stock prices, but also many, many things in the real world, in physics and so forth, are simulated that way. I mean, probably motion was invented, so to say, in the first place for describing the movement of particles and water. So I mean, it comes from physics, but the finance guys have adopted all the approaches used in physics. So we are simulating paths over time, so to say. What we now do here is kind of a more, let's say, rigorous comparison or not rigorous. But what we do is kind of we change the number of threats that we use for the multiprocessing implementation. We have kind of a test series, and it's implemented on notebook with four cores, i7. And we use the following parameters. We have 10,000 paths that we simulate, and a number of times that's 10. And what we want to do is kind of 32 simulations, which translates to the number of tasks that have to be accomplished here in this case. So it's a simple looping over a list object starting from 1 and ending to 8. So we start with a single threat and end with 8 threats. And you see, there's not that much of code involved. It's actually pretty comparable to the iPython parallel example. You just have to define our pool of processes, our pool of workers that we use for the implementation. And then we map, here in this case, there are different approaches, I must say. But here we map our function to a set of input parameters, actually. It works pretty much the same than the map functional programming statement in Python. So we map our function to our set of parameters and say, well, please go ahead. And in the end, we wait for the finishing and append the times that it takes for the single runcy in this case. But as always, a picture says more than 1,000 words. And you see here, we start for the 32 processes with the time approaching almost 0.7 seconds. And we come down to, well, something about one point or 0.15 seconds. But you see, multi-threading doesn't bring that much here in this case, actually around 4, 5. Actually, here in this particular case, at 5, we have the lowest execution time. But you see, the benefits are pretty, pretty high here in this case. Again, almost scales linearly with a number of cores available, not with a number of threads, but with a number of cores available here for our 32 Monte Carlo simulations. And as you have seen, it's only mainly two lines of code that accomplish the whole trick. Let me come to another approach. We haven't really touched the code, the implementation. We just have taken an implementation for the last two examples and have parallelized the thing. But more often than not, you want to try first to optimize what is actually implemented. And one very efficient approach is dynamic compiling. There's available a library called number. This is an open source number where we optimize and compile for Python code, which is developed and maintained by Continuum Analytics. And it uses the LIVM compiler infrastructure. And this makes a couple of application areas for really efficient, yeah, getting the collecting of benefits and the low-hanging fruits that I've been mentioning so often right now, that it's almost like sometimes really surprising because we would see not that much effort, not that much overhead involved, but usually can expect, given a certain type of problem, really high speedups. First, an introductory example before I come to a more realistic, real-world example. And the example is only about counting the number of loops, but counting in a little bit like complex fashion, that we have to transcendental function of cosine here and then calculate the logarithm. But in the end, this nested loop structure doesn't do anything else but counting the number of loops. There's nothing about that. What we know, looping on a Python level typically is expensive in terms of performance and time spent. And we see it here when we parameterize this looping structure with 5,000 and 5,000, this takes about 10.4 seconds to execute. In the end, we have a looping, which shouldn't come as a surprise, over 25 million iterations here in this case. The benchmark again, 10.4 seconds to remember. We can, of course, do a NumPy vectorized approach to accomplish the same result. It actually wouldn't make sense to count only loops, but there are typical numerical and financial algorithms that are based on nested loops that easily vectorize with NumPy. So this kind of very general and very powerful approach. But we would see what the negative consequences are here in this case. Again, the function is pretty compact in a sense that we just instantiate here our array object, which is symmetric in this particular case. And we just do the calculation. We just do the summing over our resulting array object where we have applied before the logarithm and the cosine function and then do the summing over the results. In this case, I mean, it's always the same, always coming up with the one. But nevertheless, it's compute intensive. We see there's already a huge speedup. The execution time is below one second here in this case by using the vectorized approach. So NumPy, as we know, is mainly implemented in C where we are doing this kind of here. We like delegating the costly looping on the Python level to NumPy, and NumPy does it at the speed of C code, which is a little bit faster as we see here. Actually, we have a speedup of more than 10 times. But there's one drawback. Instantiating such a huge array leads to memory requirements. Of course, here we see we need an array object, which in the end consumes 200 megabytes of memory. I mean, it's not kind of a nice feature. You have an algorithm which doesn't consume any memory. And here in this case, using NumPy, vectorization, leads to memory burn of 200 megabytes. And now think of kind of larger problems. And you will certainly find some where memory doesn't suffice in the end. So this is kind of nice because it's faster, but in the end, it consumes lots of memory. It's memory not an issue. You might go that route, but there is an alternative, actually. And this is number that I mentioned before. And again, the overhead is kind of minimal in this case. I just import the library, usually abbreviated by NB, and call the jit function for just and time compiling. I mean, it's not just in time. It's not compiled in runtime. It's compiled at call time, actually. But it's called jit here in this case. And I don't do anything with the Python function at all. So I just let the Python function as it is, the f underscore pi. And I generate a compiled version of it by using number. So now executing this, we see that when I call it for the first time, it's still not that fast because for the first time, I said it's compiled at call time. There is kind of a huge overhead involved, but when I call it for the second time, you see this is then ridiculously fast given the Python implementation. So here we get to speed. So we say, well, now we can compare to C implementations to optimize C implementations because number uses the LLVM infrastructure. And on the LLVM level, there are kind of all these optimized compilers that compile it optimally to the given hardware at hand. So this works as well as I will show later on with a different example also on the GPU, actually. So here we see huge improvements in speedup. And again, I can only stress the point, there's not that much effort involved. It's just the calling of the JIT to the original Python function. And here you see kind of huge, huge, huge, huge speedups given this implementation. So it might be worth considering using number when you have a similar problem with your own asset loops and do this and that and so forth. And the beauty is, which comes on top, is that the number implementation is as memory-efficient as the original one. There's no need to kind of instantiate an NDRA object with 200 megabytes or even larger. So the beauty of the memory efficiency remains and you get these huge improvements but just compiling it with number. You know me, option pricing is kind of a very popular, very important in America, rather than in the financial world. So let's see if it works with that as well. Don't worry about the details. Again, it's just like a parameterization of the model. What we have to do here is kind of simulate something, then we calculate some inner values of an option and we do a discounting. So we have kind of a three-step procedure, if you like. And the three steps are like illustrated here. I can do maybe a little bit smaller. Again the code is not that important but there are two points worth noting. The one is that I do the whole thing based on NumPy arrays. So I do, if you like, Python looping but based on NumPy arrays. So I'm not working on lists with Python loops. I have my Python NumPy in the array objects and I do Python looping here over my arrays. And you see we have three nested loops to implement this when I go the looping route. This is not that I say you should do that by no means but I will show the effect of going different routes afterwards. So just remember looping over NumPy and the array objects and we have three nested loops. And by now we should know that looping on the Python level should be costly. What does it mean costly in this case? The execution for a given number of time steps takes 3.07 seconds. Actually this benign option pricing algorithm solves the same problem as we have been attacking before with the Monte Carlo simulation. So we can compare the result and you see here the Monte Carlo simulation which is usually considered to be the most expensive one when it comes to computational power that is needed. It's even faster in that case. It's not that exact. I must say there are numerical errors in there. But three seconds for the benign option pricing model here compared to the 82 milliseconds given our Monte Carlo simulation from before. But you see there are similar results that we get from the two numerical methods. But this is actually the point. Just to say these two algorithms solve actually the same problem in a sense. The first improvement again we can go the NumPy vectorization route. I said well I don't touch the algorithms themselves. I wouldn't say that touch the algorithm here. This is just kind of using different idioms using different paradigms to do to implement the same algorithm in Python. And here we can make use of course again of NumPy vectorization. Is it two minutes left? Oh, okay. We can do the NumPy vectorization actually. And what you see from the vectorization process is that it's again already much, much faster. But we now apply the JIT from number and get back a compiled, a machine code compiled version of our Python one. Then you see that we again get a speedup of three times. Comparing this more rigorously you see here well we have the number version is 54 times faster in this case and three times faster than the number version. Let me skip through a couple of slides. There's a study compiling some with Syson as well. At this point before I forget it, if you go to my Twitter feed, dygh, I have tweeted the links to two presentations. Actually this one I have also a side presentation so you can read all the things to it. I might do it right now going to twitter.com and it is dygh. And I have tweeted links to it so I'm not able to present everything but here you find the links to the presentations on my Twitter feed. Actually study compiling with Syson works similarly. Here we have examples where you also get kind of huge improvements. I skipped through that in order to have a couple of minutes left for questions. But again if we do a performance comparison in this regard for example here I'm working with floats so if you have a look at it there's no need to work with floats but still having this kind of rigorous performance comparison when you go to the algorithm back you see I have an implementation using Syson and another one with number and here in this case they are actually pretty similar when it comes to performance. So with Syson you usually have to touch the code and you have to do kind of static declarations and so forth but with number sometimes I don't say that always don't get me wrong but sometimes you can get the same speed ups but just using the just-in-time compiling approach of JIT. Actually last topic is generation of random numbers on GPUs. I want to spend the last minute on that because this might be useful in many, many circumstances and usually it's considered kind of a very hard thing to get the power of GPUs included in your work. What I'm using here is number pro which is a commercial library of continuum analytics which is kind of the sister or brother library of number and what I use is kind of the native libraries that are provided in the Q-der lip in order to generate random numbers. There's not that much specialties included. We just generate random numbers which are stored in a two-dimensional array in that sense. Here's the code for the Q-der function. Q-der only gives back like a one-dimensional array so we have to reshape it afterwards but I mean this is straightforward. What I do here is kind of compare the performance for different sizes of arrays where we want to get standard normally distributed random numbers back and I skipped the first slide because I have kind of implemented a rigorous implementation. What we see here in this one chart and this almost says it already that if you just want to generate a few random numbers so to say then you see that the CPU might be the better one because you have overhead involved when you're moving data, when you're moving code from the CPU to the GPU. This overhead of course but once you reach a certain size of the random number set you see that the CPU is not a linear scaling is that what you see the increase in time needed and you see that there's hardly any increase in the time needed on the Q-der device here to generate the random numbers. Here again the message if you have only a small set of random numbers don't go to the GPU there's too much overhead involved. Remain on the CPU but again if you're working with really large random number sets and here the largest one that I'm generating is 400 megabytes in size per random number set then you see that of course the Q-der approach pretty much well yeah of course outperforms the NumPy in-memory version with the CPU based on the Q-der approach here in this case. So again only couple of lines of code it's a single library that you call and you get all the benefits from that and you see there's a huge speed advantage over the Q-der device over the number one. The last thing I just want to mention is hardware bound I.O. Python is not only when it comes to numerical operations but I had it included in my abstract. Python is also pretty good when you want to harness the power of today's I.O. hardware and usually it's pretty hard to get to the speed limit of the hardware but with Python and here working with an example array object 800 megabytes and just natively save that. You can also use your PyTables and actually a five format and a couple of other things but it's already built in into NumPy that you can save your arrays to disk. You see this almost at the speed of the hardware that is allowed here writing on the MacBook with SST drive you see for the 800 megabytes. This is much much faster to save and to load as you see as it is to generate a memory. For the memory generation of this 800 megabyte array with the memory allocation and the calculation of the random numbers takes 5.3 seconds but on this machine it only takes two seconds to write it and two seconds to read it. So you see how fast it can be with Python and there's no kind of performance trickery involved. This is just like batteries included and Python typically makes it pretty efficient and pretty easy to harness the power of the available hardware as of today. This brings me to the end and thank you for your attention. I'll play it here. It's already open time. We still have some five minutes. Also wants to wait for the lunch a little longer. Sorry to stay between lunch and. You just asked a question if you I think I can hear you I can repeat it. Of course of course. What I was saying is kind of for this particular algorithm with data parallelism and code parallelism but this is kind of the most single simple scenario you can think of. Of course I'm pretty aware of that. Yeah of course it's kind of one of the center protocols that are used within. We haven't used it actually but of course there are kind of bindings and many application examples and pretty good use cases for that. Okay so hi thanks for the talk. Well done. Small question so suppose you were doing a huge time series analysis. It's not in this scope but obviously that's something it's kind of hard to do in parallel. There are algorithms that work very nice in parallel and there's algorithms that don't work very nice in parallel. What's your gist on doing things with algorithms that don't work nice in parallel? What's the besides compiling what are some tricks you maybe use? This is a question out of curiosity. So actually time series analysis is of course one thing we are like concerned with all the day in finance so it's one of my major things so to say but I didn't get the question in the end. So there are algorithms that you can really easily approach something parallel and there are algorithms where this is not so easy. I can imagine for certain like very advanced time series models say ARIMA type of models those don't parallelize that nicely. What's your gist on if the problem is hard to parallelize what's the best tactic to approach those problems? I mean of course not everything is kind of well suited to be parallelized. We're using for example working heavily with kind of least squares Monte Carlo where you need kind of the cross section. Every Monte Carlo you would say 100,000 paths I can parallelize into two times 50,000 paths. It's the same with the time series analysis. You could do like I have 100,000 observations I can implement my algorithm on the first 50,000 and second 50,000. There's one approach to do it but not every algorithm is like well suited because you need kind of the whole history or whatever built up so you need the cross section of the information in order to have your algorithm produce the results that it's supposed to produce. Usually I mean the approach that I presented is kind of using parallelization for an unoptimized algorithm is kind of usually not the way to go. What you would do in this case when you say well I don't have the algorithm that can be pretty easily parallelized. Of course you would in any case go for the optimization of your algorithm by using the Thyson and everything. But I think all the libraries there's not only Thyson. What I haven't mentioned is kind of theano. For example if you have a look at pi MC3 this makes heavy use of theano just in time compiling where kind of your objects in classes are like on the fly dynamically optimized for the problem given at hand using just in time compiling or call time compiling. It's kind of a slight difference. But this is typically I think the approach that you would take. They say well let's first optimize with any given means that we have available the algorithms. But I agree not every algorithm is kind of well suited to be parallelized. But again if you have two times serious to analyze then you can start thinking about that again. This was my point actually. Many similar tasks and this is of course the trivial case for parallelization. So if you're starting off with serial Python code I think it's pretty obvious that using these parallel tools will make it go faster. But I think lots of people believe that Python is not what you should be using if you want to have efficient parallelization because of the additional overhead because you have to use multiprocessing because of the gil and obviously Python is a higher level language. So I'm not talking about benchmarks but just in real world applications like actual applications that you have written what do you say to people who would be tempted to stick with C++ to squeeze out that last bit of performance. Do you find that Python is sufficient for kind of everyday needs for these kinds of simulations? So far we haven't ever gone the route of going to C++ or C for all our client projects nor for the things that we've been implementing. So we're using for example the multiprocessing tool for Dx Analytics or library where we have simple, easy scaling and parallelization on the machine at least. This is typically where the things are run on larger machines with multiple cores with huge memory. This kind of the scenario that I have. I mean many, many, I can understand people that have issued like with the scaling or like classes and so forth but of course I mean you have this route and with Thyson is kind of the very good example where you can say well even using Thyson I can decide whether I have 90% Python and 10% C so to say or I have something that looks like 90% C and a little bit of Python you can still notice in between. So this is the beauty of Python that you don't have usually these either or things. You can even go for like after profiling where you say well this is the real bottleneck of the whole thing. Let's go for the C route for that. A recently met doing our Python for Quant Finance meetup in London, a guy he was saying well I again did something in Assembler because we thought this would be, I don't know if this is the right thing but still you have the flexibility and there's a beauty of Python that is not either or that integrates pretty well and of course C and C++ are the two worlds where Python interacts natively with. So whenever I say well I have this approach and I can use this better with C++ why not going there. So many people are doing that and the financial industry is kind of a standard to do that. But we for the things we have been doing can only say it again for our stuff and for stuff that we have been implementing for our clients we have always stuck to the Python world. Of course using performance libraries and all that I mean under the hood this comes down to C and other stuff but not on the top level where we have been implementing things. People will only be applicants for Python and let them go off all the lists and it's not the question for
|
Yves - Performance Python for Numerical Algorithms This talk is about several approaches to implement high performing numerical algorithms and applications in Python. It introduces into approaches like vectorization, multi-threading, parallelization (CPU/GPU), dynamic compiling, high throughput IO operations. The approach is a practical one in that every approach is illustrated by specific Python examples. The talk uses, among others, the following libraries: * NumPy * numexpr * IPython.Parallel * Numba * NumbaPro * PyTables
|
10.5446/20041 (DOI)
|
Welcome back, everybody. I hope you had a good lunch. Well, won't waste any time. Tom Christie is going to talk about documenting your project because it's MK docs. Hi there. Thanks for coming today. My name is Tom Christie. I'm going to give you a quick run down today of a little project that I've been working on, which is a documentation builder for Markdown called MK docs or make docs or muck docs. I'm not really sure which yet. So here we go. Let's have a little look through. First of all, I want to apologize today if there's any bits of this that are a little bit patchy. I've been a little bit busy lately in the last few days. I launched a Kickstarter for a project in mind called Django Res Framework, which I've been like outrageously, outrageously successful. So I know there are probably some of you who've donated here today. So thank you all of you. Wonderful. Anyway, on with the show. Why, why, why am I building my docs? Because it's something that I needed. So for when I started working on the release of Django Res Framework 2, I had some very specific ideas about how I wanted the documentation to look and how I wanted it to work. And at the time, Sphinx didn't quite fit in with that. The themes are much, much nicer now, but at the time I wasn't happy with any of that. And also because prior to using my docs, I was writing my documentation in RST, in Ziblum Text. And I'm quite a visual person. And for me, being able to write my documentation in Markdown gives me a better flow in some of the nice Markdown editors, a better feel of the flow of the documentation. For me, I feel like being able to write my documentation with these kinds of tools helps me write better documentation because I can get more of a feel for how they're going to present to the end users. So I really wanted something that was nice and simple and used Markdown to generate the documentation. So I started hacking away on a little script. And at some point, decided that I ought to take this hacky little Python script that just lives in the REST framework repository and turn that into something a bit more reusable and hopefully be able to use that for some future projects as well and open it up to everybody else to be able to use. So this was the end result of how the documentation looked with this hacky little script. And my users were happy with it as well. So that's nice. Right. I want to spend most of this talk just giving you a very brief demo of using McDocs just so you can get an idea of what the documentation layout looks like when you're working with it and how simple it is. So only a couple of prerequisites, Python and PIP. I would like, perhaps at some time in the future, to package this up in a way that is invisible that it uses Python so that we can deliver it to a wider community. But that's something on the long-term roadmap. So what do we do to get started? Install McDocs from PIPI. Create a new project which will populate the directory with a couple of initial files that we'll look at in a minute. And then we're going to start serving our site. So let's do that. So I already have McDocs installed here. So McDocs new demo. What's the matter? Okay. So if we take a look down here, we can see that it's written this demo folder. And if we go in and have a look, and I hope that's just about where I'll un-follow. So there's inside the directory that it's created, we've got two things. We've got a single configuration file, which is a YAML configuration file, and we've got a docs folder with a single markdown file in it, which is our first page of documentation. So let's see into the directory, McDocs serve. There we go. And there we go. This is the live server that you can run. And there we have our documentation being served locally. Fine. Okay. So what can we do with this? One of the nice things that's built into McDocs is a nice little live reload feature. This means anytime you alter anything in the configuration or any of the documentation, the site that's being served will be automatically rebuilt. And all you have to do is go in and hit refresh in the browser, and you can see the effects of your changes. So if we go into here, let's just open up the index page, and let's change this to McDocs, ROPs, there's a Ryan's. There you go. Nice and easy. Ah, yeah, sure. I'll do a bit of that as well. Okay. So quick rundown of how the documentation source files are organized. Everything goes into a folder called docs. It has to have an index.md file, which will be the home page. You can sling other media in there. You can also sling in CSS and JavaScript. Any CSS and JavaScript that you sling in will automatically get included into your theme without you doing anything else. So you can make nice little tweaks to say how the hero headers get displayed on your leading page and nice little, like, things like that without having to change the theme wholesale. So have a quick look at that. Here's a folder where we've got a few more pages of documentation. If we just go back to the example that we're working on here, what I'm going to do is, add a couple more pages. So let's create a... You can see I've just added an about. And now I've added a new folder called user guide, which also has a couple more pages in it. We go and reload the documentation again. You can see we've got a couple more pages in it. And then we've got a new documentation again. You can see we've now got an app bar at the top that has included some extra pages. So here's our about page that we've just added and a couple of other pages. And we can page back and forth between those. And all we've had to do was add the new markdown files into the folder in order to add them. Similarly with images, just throw them into the docs directory, then you can hyperlink to them from the markdown files exactly as you normally would, and they'll be included in the output. Again, with CSS. And you can also put in other useful things, such as, for instance, this CNAME file, which is used by the GitHub pages if you want to provide a custom domain, and you're hosting your documentation on GitHub pages, you can include this little CNAME file where you just put in the domain name that GitHub should understand the documentation to be served under. So one of the things that restructured text is very strong at, which markdown isn't so strong at, is interlinking. And obviously interlinking in your documentation is very important. So how do you do that with MacDocs? Well, the simplest way to link between pages is just to use standard hyperlinks to the documentation, to the other pages. What you do is you include the hyperlinks as if they're hyperlinks to the markdown source files. MacDocs will automatically translate that into the equivalent URL when it's building the documentation or when it's serving the documentation. This has quite a nice effect in that when you're working with your documentation in the editors, you're able to click on the links and it will automatically end up bringing you up the next page that you're working on, which is quite a nice way to work on it. And the other thing that I'm in the process of adding to MacDocs, which isn't quite there, is a syntax for a slightly more intelligent interlinking that allows you both to interlink to particular pages but also to particular sections of particular pages. So there's a simple syntax for doing that, which allows you either to just put in a ref without adding exactly any text that is referenced against and the top link here would link to any section it can find in the documents called Project License. If you don't want the text of the link to match the section that you're linking against, you can instead be explicit about the name of the section that you're linking against, which is the second one down there. That's still in process, but that's the idea. So configuration, everything goes in the one YAML configuration file. It must exist and it has one required setting and everything else is optional. So it's nice and simple. There's an example there. And you can use the... So with the example that I've shown you at the moment, we've added some markdown files, but we haven't specified anything in the configuration for how to order them, so it's just automatically decided on an ordering for those. You can... I won't bother doing it now, but you can set up the ordering for your pages by using the configuration file. You add a key called pages and then you just list the order of the source files that you want them to appear in. There you go. Oh, yeah, getting ahead of myself. So themes, okay. There's two different ways that you can theme your documentation. There's a whole bunch of built-in themes that you can use, and you can also provide custom themes. So of the built-in themes that are available, we've got this fairly kind of bootstrappy style. There's also a style based on the really fucking cool Read the Docs theme that... I can't remember who it was who developed it, but it's really quite nice. And because the default style is based on bootstrap as well, there's also a whole stack of boot swatch-based themes available as well, which is really nice and easy to use. So for example, if we go into our configuration file here and we say theme, United, perhaps... There we go. Brand new theme, lovely. Similarly, if you want to create a completely custom theme of your own, you can do that. Nice and simple. The only thing that you need is a base.html file in your theme directory. You can then include any other files that you need, and all of the context that gets passed into that template is... Well, some of it is documented. Some of it's in the process of being documented. What we don't do is have anything like particular pages that only get pulled in, particular HTML pages that only get pulled in for particular markdown source files or anything complicated like that, or anything like partially overriding a theme. If you want a new theme, either you're using some CSS in your project directory, or you just create a brand new theme directory with everything from scratch. It's just simpler that way around rather than dealing with, okay, I've got this base theme, but I want to override this bit and that bit. Oh, yeah. And here's an example of what a theme directory might look like. So we've got a bunch of HTML files which will get translated using Ginger 2, passing in the context, a couple of images, some fonts, styles, JavaScript. There you go. Okay. So let's have a look at building the documentation. I'm going to go and do that now. There we go. So you can see building the documentation ends up creating a folder in there called site which has all of the final HTML files in there and all of the other media. There we go. And it builds completely static sites, so you can just host them from anywhere. Most of my documentation, I happen to host from GitHub pages because it's really good and it's really simple. Amazon S3 would work equally well. And maybe one day there might be integrated support with Read the Docs, perhaps. Eric has talked about that and said that he liked that, but I just need to find some more time to work on the project first to make that happen. So one of the nice things that it also has built in is a nice easy way of deploying your site to GitHub pages. So in case you don't already know, GitHub pages is a way for GitHub to serve up static pages. And what it does is you have to host your site on a branch of the main repo called ghpages. And then GitHub will expose that on a particular domain. I can't remember exactly what it looks like, but we'll find out in a minute. Oh, yeah, it will look like this. So let's go and have a look at the built-in GitHub pages integration. All we have to do is mcdocs ghdeploy. Hang on a minute. I haven't added this site to GitHub yet. That would be a good plan. So here's my empty repository. Tom Christie slash demo hasn't got anything in it yet. Let's just push our documentation up to GitHub. Okay, that's done. So there's our documentation source up there. And now all we need to do is ghdeploy. So it builds the documentation and pushes that up to the ghpages branch and tells you the URL that that should now be available on. And our documentation is live in the Internet. Yay. Okay. So I'm trying to keep this super simple. I'm not interested in exposing, say, a programmatic API to allow developers to override this with Python. Really, I just want to keep the overrides based on you can change the theme, and that's it. This isn't a semantic about being a semantic markup tool in the same way that RST can give you lots of extra information about, well, this word is a class, and it means this, that, or the other. This is just about taking simple markdown files and rendering them into HTML. So it won't easily support going out into lots of other different formats, but I don't care. And, yeah, this is a nice thing as well. The wonderful people at Docker have started using this for their documentation, so I better get my act together. And what else? Yeah, stuff is happening. It's fun. And, yeah, that's about all I've got to say. Thank you. Thank you. That's my boy, two questions. Would you consider adding the audience to the upload with the documentation to the cheese shop documentation, Christine? So the question was about uploading the documentation to cheese shop documentation hosting. I didn't know there was such a thing. What, they host static sites or? I still have.py upload docs. Oh, okay. New to me. Setup.py upload docs. I'll have a look. Yeah. That sounds like a sensible option. Is there a feature or essentially how to generate the API docs? No. The question was about generating API documentation. So I assume you mean inspecting doc strings in Python and automatic. No, deliberately not, actually. Really, I'm interested in aiming at pro-style documentation. I enjoy reading and writing pro-style documentation more than I do automatically generated API documentation. I'm not a big fan of that style. So, yeah, this is just about textual stuff. Yeah. Okay. Come and say hi. Okay, thanks, everybody. Thank you.
|
Tom Christie - Documenting your project with MkDocs. MkDocs is a new tool for creating documentation from Markdown. The talk will cover: How to write, theme and publish your documentation. The background and motivation for MkDocs. Choosing between MkDocs or Sphinx. ----- This talk will be a practical introduction to MkDocs, a new tool for creating documentation from Markdown: * The background behind MkDocs and the motivation for creating a new documentation tool. * Comparing against Sphinx - what benefits each tool provides. * Getting starting with MkDocs - how to write, theme and publish your documentation. * Under the covers - how MkDocs works, and some asides on a couple of the neat Python libraries that it uses.
|
10.5446/20040 (DOI)
|
Thanks so much for coming out to hear me talk about my favorite topic, probabilistic programming. I thought I'd introduce myself real quick. I recently relocated back to Germany after studying at Brown where I did my PhD on Bayesian modeling and we studied decision making. And for a couple of years, I've also been working with Quantopian, which is a Boston based startup. And as a quantitative research, and there we are building the world's first with McTrading platform in the web browser. The talk will be tangentially related. So I'm going to say I'm going to just kind of show you that screenshot of what it looks like. So this is essentially what you see when you go on the website. It's a web based IDE where you can code Python and code up your trading strategy. And then we provide historical data so that you can test how you would have done if it was 2012. And then on the right here, you see how well it did in often. And that's what I'll refer back to. You are interested in whether you beat the market or lose against the market. And also I should add it's completely free and everyone can use it. OK, so I think at every talk, that should be the main question as well. Why should you care about probabilistic programming? And it's not really an easy talk just because talking about probabilistic programming, you need to have at least a basic understanding of some concepts of probability theory and statistics. So the first 20 minutes, I will just give a very quick primer focusing on an intuitive level of understanding. Can I get a quick show of hands like who sort of understands on an intuitive level how Bayes formula works? OK, so most of you, so maybe you won't even need that primer, but it might still be interesting. And towards the end, then we have a simple example and then a more advanced example that should be interested, even if you know already quite a bit about Bayes and statistics. So to motivate this further, I really like this contrast that Olivier gave at his talk about machine learning. And that is chances are you are a data scientist, maybe, and you use scikit-learn to train your machine learning classifiers. So what this looks like is on the left you have data that you used to train your algorithm and then that algorithm makes predictions. And if those predictions are all you care about, then that might be fine, right? But one central problem that most of these algorithms have is that they're very bad at conveying what they have learned. So it's very difficult to inquire what goes on in this black box right here. So on the other hand, probabilistic programming is inherently open box. And I think the best way to think about this is that it is a statistical toolbox for you to create these very rich models that are really tailored to the specific data that you're working with. And then you can inquire that model and really see what's going on and what was learned so that you can learn something about your data rather than just making predictions, right? And the other big benefit, I think, and we'll see that later, is that these type of models work with a Black Box inference engine, which are sampling algorithms that work across a huge variety of models. So you don't really have to worry about the inference set. All you have to do is basically build that model and then hit the inference button. And in most cases, you'll just get the answers that you're looking for. So there's not really much in terms of solving equations, which is always nice. So throughout this talk, I want to use a very simple example that most of you will be familiar with, and that is AB testing. As you know, when you have two websites and you want to know which one works better in some measure that you're interested in, maybe the conversion rate or how many users click on an ad, what do you do to test that? So you split your users into two groups and give group one website A and give group two website B, and then you want to look which had the higher measure, right? That problem is, of course, much more general. And since I'm coming more from a finance background, I'm going to sort of switch back and forth between the statistically speaking identical problem, where you have two trading algorithms and you want to know which one has a higher chance of beating the market on each day. So here, I'm just going to generate some data to basically see what the trivial answers, trivial answers that you might come up with yield and how we can approve upon that. And you might be surprised that I'm not using real data, but I think that is actually a critical step is before you apply your model on real data, you should always use simulated data where you really know what's going on in the parameters that you want to recover so that you know that your model works correctly, and only then you can be sure that you'll get correct answers by applying it to real data, right? So the data that we're going to work with will be binary, so just Boolean events. And that type of statistics, statistical process called Bernoulli, and that is essentially just a coin flip, right? The probabilities of coin flips. And I can use that from sci-pi stats. And then I call it Bernoulli. And here I pass in the probability of that coin of coming up heads or that algorithm of beating the market on a particular day or that website of converting the user. And here I'm sampling 10 trials. So this will be the result, right? Just a bunch of binary zeros and ones. So I'm generating two rhythms, one with 50%, one with 60%. So you want to know which one is better. The easiest thing that you might want to that you might come up with is just, well, let's take the mean, right? And actually, statistically speaking, that's not a terrible idea, and it's called the maximum likelihood estimate. And if you ask an applied mathematician what you should do, then that might be the answer. And I took a cause and applied math. And the proofs always work in a very similar way. You basically have this problem, and then you say, well, OK, let's have our data go to infinity, and then you solve. And then you get the estimator works correctly in that case. And that's great. But what do you do if you don't have an infinite amount of data, right? And that's the much more likely case that you be in. And that, I think, is where basal statistics really work well. So what happens in our case now, where I just take the mean of the data I just generated? As you can see, in this case, we estimate that the chance of this algorithm beating the market is 10% and 40% for the other one. So obviously, that's completely wrong. It's 50% and 60% to generate it. And the obvious answer of why this goes wrong is just I was unlucky. And the observant members in the audience will have noticed that I used a particular random seed here. So I took that random seed to produce this very weird sequence of events that basically produced this pattern. But certainly that can happen with real later, right? You can just be unlucky and the first 10 visitors of your website just don't click. And the central thing that I think is missing here is the uncertainty in that estimate, right? 10% 40%? That's just a number. But we're missing how confident we are in that number. So for the remainder of the talk, that will be a recurring topic is really trying to quantify that uncertainty. Then you might say, well, there is this huge branch of statistics, frequent statistics, which designed these statistical tests to decide which one of those two is better or whether there is a significant difference. Then you might run a t test and that returns a probability value that indicates how likely are you to observe that data if it was generated by chance. And that's certainly the correct thing to do. But one of the central problems with frequent statistics is that it's incredibly easy to misuse it. For example, you might collect some data and the test doesn't turn out anything. And then on the next day, you get more data. So what do you do? Well, you just run another test with all the data you have now, right? You have more data, so the test should be more accurate. Unfortunately, that's not the case. And you can see that here, just create a very simple example where do that procedure. I generate 50 random binaries with 50% probability both. So there is no difference between them. And then I start with just two events. I run a t test. If that is not significant, then I do three around the t test, right? So just that process of continuously adding data and testing whether there's a difference. And if there's a difference of smaller than 0.05, then I return true. If it isn't, then I return false. And then I return that a thousand times. And I look at what the probability is that even though there is no statistical, there is no difference at all between those two, they're both 0.5. What is the chance of this test yielding an answer that there is a significant difference? And it's 36.6% in that case, which is also absurdly high, right? So this procedure really fails if you use it in that way. And granted, I misused the test, right? It's not designed to work in that specific scenario, but it's extremely common that people do that. And for me, one of the central problems is that frequency statistics really are dependent on your intentions of collecting the data. So if you use a different procedure of collecting the data, for example, say what I just did, I just add data every day, then you need a different statistical test. And if you think about this more, it's actually pretty crazy, right? If you're a data scientist and you just get data from a database, you have no idea what the intentions were of gathering that data, right? So and you want to be very free in exploring the data set and running all kinds of statistical tests to see what's going on. So I think while frequency statistics is certainly not wrong, it's often very constricting in what it allows you to do. And if you don't do things correctly, you might shoot yourself in the foot. And I think that's really a good setup for Bayeson statistics. And I'm just going to introduce that very quickly. So at the core, we have Bayes formula. And if you don't know what that is, essentially it's just a formula that tells us how to update our beliefs when we observe data. That implies that we have prior beliefs about the world, that we have to formalize. And then here we apply, then we see data and we apply Bayes formula to update all beliefs in light of that new data to give us our posterior. And in general, these beliefs are represented as random variables. And I'm also going to very quickly talk about what those are and what intuitive ways of thinking about those. So the decisions like to call their parameters, their random variables theta. So that's what I'm going to use here. And let's define a prior for our random variable theta. And theta will be the random variable about the algorithm beating the market or the single algorithm beating the market or the website converting the user. So what is the chance that that happens? Oops. So I didn't want to show that. I just wanted to show that. So the best way to think about that random variable is as opposed to a variable that you might know from Python programming, which just has a single value, say, i equals 5, is here we don't know the value, right? We want a reason about that value. We have some idea, some rough idea about that value. So rather than just having one, we allow for multiple values and assign each possible value to a probability. And this is what that shows here. So on the x-axis, we have the possible states that the system can be in. For example, the algorithm can have a chance of 50% of beating the market. And I'm going to assume that that is the most likely case. That's my personal prior belief without having seen anything. I'm going to assume that, on average, 50% is probably a good estimate. But I wouldn't be terribly surprised to see something with 60%, even though it's less likely. 80% considerably less likely, but still possible. 100% that beats the market on every day, that I think would be next to impossible, right? So I'm going to assign a very low probability to that. So I think that's a very intuitive way of thinking about that. So now let's see what happens if I observe data. And for that, I created this widget here. And where I can add data when I use this slider, and then it will update that probability distribution down here. And so that will be our posterior, right? Currently there's no data available, so our posterior will just be our prior. So that is just the belief we have without having seen anything. And now I'm going to add a single data point, a single success. So we just ran the algorithm for a single day and it beat the market. So now as you might have seen, the distribution here shifted a little bit to the right side, right? And that represents, I'll update it, believe that it's a little bit more likely now that the algorithm is generating positive returns. So now let's reproduce that example from before where we had one success and nine failures, right? There was algorithm A and there we estimated it has a 10% chance of beating the market. So that was ridiculous, right? With that amount of data, no way we could say that. And also with our prior knowledge, no way we would assume that 10% is actually the probability. So now I created that and is updating this probability distribution down here, which is now our updated belief that certainly with nine failures, we're going to assume that there is a lower chance of success of that algorithm, which is represented by this distribution moving to the left. But still note that 10% is still extremely unlikely under this condition, right? And that is the influence of the prior. We said 10% is unlikely, so that will influence our estimates away from these very low values. The other thing to note is that the distribution is still pretty wide. So here now we have our uncertainty measure in the width of the distribution, right? The wider it is, the less certain I am about that particular value. So now I want you to, in your head, just imagine what the distribution look like if I move this up to 90 and the success up to 10, right? So basically now we're observing data that is in line with a hypothesis that it has 90% failure probability. So as you can see, the main thing that happens is it moves to the left, but also it gets much narrower and that represents our increased confidence. With having seen more data, we have more confidence in that estimate. That's exactly what we want. By the way, how cool is it that I can use these widgets in a live notebook? Okay, so where's the catch with all of that, right? This sounds a little bit too good to be true. You just create that model and you update your beliefs and you're done, right? Unfortunately, it's not always that easy. One of the main difficulties is that this formula in the middle here in most cases cannot be solved. The case that I just showed you is extremely simple. You just apply Bayes formula and you can solve it and then you can compute your posterior analytically, but even with just tiny bit more complex models, you get multi-dimensional integrals over infinity that will make your eyes bleed and no sane human would be able to solve. And I think historically that's one of the main reasons why Bayes, which has been around since the 16th century, has not been used up until recently now where it's kind of having a renaissance is just people weren't able to solve for it. And the central idea of probabilistic programming is that while we can't solve something, then we approximate it. And luckily for us, there's this class of algorithms that are most commonly used called Markov chain Monte Carlo. And instead of computing the posterior analytically, that curve that we've seen, it draws samples from it. That's about the next best thing we can do. So just due to time constraints, I won't go into the details of MCMC. So we're just going to assume that it's pure black magic and works. And it's sort of, it's a very simple algorithm, but the fact that it works in such general cases is still mind blowing to me. And the big benefit is that, yeah, it can be applied very widely. So often you just define your model, you say go, and then it'll give you your posterior. So what does MCMC sampling look like? As we've seen before, this is the posterior that we want, right? This neat closed form solution, which we can't get in reality. So instead we're going to draw samples from that distribution. And if we have enough samples, we can do histogram, and then it'll start resembling what it is. Okay, so let's get to PIMC3. PIMC3 is a probabilistic program framework written in Python and for Python. And it allows for construction of probabilistic models using intuitive syntax. And one of the reasons for doing PIMC3, rather than two, maybe some of you use PIMC2, PIMC3 is actually a complete rewrite. It uses no code from PIMC2. There were a couple of reasons. There was just technological debt that the code base of PIMC2 is pretty complex. It requires you to compile Fortran code, which always causes huge headaches for users to get working. So PIMC3 is actually very simple code. And one of the reasons is that we're using Theano for all things for the whole compute engine. So basically we're just creating that compute graph and then shifting everything off to Theano. The other benefit we get from Theano is that it compiles, that it can give us the gradient information of the model. And there's this new class of algorithms called Hamiltonian Monte Carlo that work, that are advanced samplers. And those work even better in very complex models. So they're much more powerful, but they require that extra step. And that's not easy to get. Luckily for us, Theano provides that out of the box. So we don't really have to do anything. The other point I want to stress is that PIMC3 is very extensible. And also it allows you to interact with the model much more freely. So maybe you have used JAGs or WinBugs or Stan, which is a very interesting recent probabilistic programming framework. And while those are really cool, one problem I personally have with them is that they require you to write your probabilistic program in this specific language. And then you compile that, you have some wrapper code to get the data into Stan, and then you have some wrapper code to get it out of Stan, the results. And for me, there's always very cumbersome. So you can't really see what's going on in the model. You can't debug it. So PIMC3, you can write your model in Python code and then really interact with it freely. So you never have to leave, essentially, Python. And that, for me, is very, very powerful. And so you can think of it much more as a library, and we'll see that in a second. Just the authors. So John Salvatier is the main guy who came up with it, and Chris von Especk also programmed quite a bit. Currently, we're still in alpha. It still works. It works fairly well already. The main reason why it's alpha is mainly that we're missing good documentation. And we're currently writing those, but if you are up for it and would like to help out with that, that's certainly more than appreciated. Okay, so let's look at that model from our early example that we wanted to and see how we can solve it now in PIMC3. And for that, I'm just going to write down that model, how you would write it in statistical terms. So we have these two random variables that we want to reason about, theta A and theta B, and that will represent the chance of the algorithm beating the market. And here we say this tilt means it's distributed as, so we're not working with numbers but with distributions. So this is a beta distribution, and that is the distribution that we have been looking at at the beginning, just from zero to one. If you're working with probabilities, the beta distribution is the one to use. So this is the thing that we want to reason, that we want to learn about, given data. And then how do we learn about it? Well, we observe data, and the data that are simulated was binary, so that came from a Bernoulli distribution. So we have to assume that the data is distributed according to a Bernoulli distribution, so they are zeroes and ones, and the probability of that Bernoulli distribution before just fixed point five, right? Here now we actually want to infer that, so since we don't know that value, we replace it with a random variable, and that is the random variable theta A that we had above here. So that is how commonly these models look like. And the other point I want to make here is that here you really see how you're basically creating a generative model, right? So you might wonder, like, how can I construct my own model? And I think a good path for that is to just think of how the data would have been generated, right? Here I know, well, there's this probability and a generated Bernoulli data, so that's the model I'm going to create. But you can get arbitrarily complex and then say, well, I have all these hidden causes that somehow relate in complex ways to the data, and then you're going to invert that model using Bayes formula to infer these hidden causes. So here I'm just going to, again, generate data a little bit more now, so again, 50 and 60 percent probably of beating the market or conversion rate and 300 values. And this is what the model looks like in PIMC3. So first we just import PIMC as PM, and we instantiate the model object which will have all the random variables and whatnot. And the other improvement over PIMC2 is that everything you specify about your model, you do under this with context. And basically what that does is that everything you do underneath here will be included in that model object so that you don't have to pass it in all the time. So underneath here now, this should look pretty familiar from before where I just had these random variables, right? Theta a distributed as a beta distribution. So here I now write the same, but in Python code where I say, well, theta a is a beta distribution, we give it a name, and we give it the two parameters, and alpha and beta are the two parameters that this distribution takes, the number of successes and failures. So this is the prior that I showed you before that was centered around 50%. And I do the same thing for theta b. And now I'm going to relate those random variables to my data. And as I said before, that's a Bernoulli, which I'm going to instantiate, I've got a name, and instead of the fixed p value now, I give it the random variable, right, that we want to link together. And since this is an observed node, we give it that array of 300 binary numbers that are generated as like before, right? So this links it to the data and links it up to the random variable. And the same for b. So up until here, nothing happened. We just basically pluck together a few probability distributions that make up how I think my data is structured. Now it's often a good idea to start the sampler from a good position. And for that, we're going to just optimize the log probability of the model using find map for find the maximum upper stereo value. And then I'm going to instantiate the sampler I want to use. There are various you can choose from. I'm using a slice sampler, which works quite well for these simple models. And now I actually want to draw the samples from the posterior, right? And for that, I call the sample function. And I tell it how many samples I want, 10,000. I provide it with the step method. And I give it the starting value. And when I do this call, it'll take a couple of seconds to run the sampling algorithm, and then it really would return the structure to, which I call trace here. And that is essentially a dictionary for each random variable that I have assigned. I will get the samples that were drawn. And now that I ran that, I can inquire about my posterior, right? So here I'm using Seaborn, which just as an aside is an awesome plotting library on top of Medplotlib. You should definitely check it out. Creates very nice statistical plots. For example, it has that nice this plot function that basically just creates a histogram, but one that looks much nicer and has, for example, this nice shaped line. And I give it the samples that I drew, that my MCMC sampling algorithm drew, of theta A and theta B. And then it will plot the posterior now that I created. And that is, again, the combination of my prior belief updated by the data that I've seen, and now I can reason about that. And the first thing to see is, well, the theta B, the probability of the chance of that algorithm beating the market is 60%. And that's what I used to generate the data. So that's good that we get that back. And again, that's why we use simulated data to know, actually, that we're doing the right thing. And the other one is around 50% or 49%. The other thing to note is that here now, instead of just having that single number that seemingly fell from the sky that we would get if we just take the mean, we have our confidence estimate, right? We know how white that distribution is. We can answer many questions about it, like how likely is it that the chance of success for that algorithm is 65%. And then we get a specific number out that represents our level of certainty. And we can do other interesting things, like hypothesis testing, to answer our initial question, which of the two actually does better? And for this, we can just compare the number of samples that were drawn for theta A to the samples of theta B. So we're just going to ask, well, how many of those are larger than the other one? And that will tell us, well, with the probability of 99.11%, algorithm B is better than A. And that is exactly what we want, right? So by consistently having our confidence estimate carried through from the beginning to the end gives us that benefit of everything we say has that confidence and probability estimate associated with it. OK, so that was super boring up until now. Hopefully it gets a little bit more interesting now. So consider the case where instead of just two algorithms, we might have 20. And that is what we have on Quantopian. Many users have these algorithms. And maybe we want to know not only each individual algorithm's the chance of success, but also the algorithms overall. The group average. Are they also doing, are they also consistently beating the market or not? So the easiest model you can probably build is just the one we did before. But instead of two theta A and B, we have 20 theta's, right? And while that's fair, and this is called an unpooled model, it's somehow unsatisfying, right, because we probably assume that these are not completely separate, right? If the algorithms work in the same market environment, some of them will have similar properties, some similar algorithms that they're using. So they will be related somehow, right? They will have differences, but they will also have similarities. And this model does not incorporate that, right? There's no way of what I learned about theta one, I would apply to theta two. The other extreme alternative would be to have a fully pooled model, where instead of assuming each one has its own random variable, I just assume one random variable for all of them. And that's also unsatisfying because we know that there is that structure in our data, and which we're not exploiting. And also, even though we might get group estimates, we could not say anything about a particular algorithm, how well that one did, right? So the solution which I think is really elegant is called a partially pooled or hierarchical model. And for that, we add another layer on top of the individual random variables, right? Up until here, we only have the model we had before with all these independently, but what we can do is, instead of placing a fixed prior on that, we can actually learn that prior for each of them and have a group distribution that will apply to all of them. And those models are really powerful and have very many nice properties. One of them is, well, what I learned about theta one from the data will shape my group distribution, and that in turn will shape the estimate of theta two. So everything I learned about the individuals, I learned about the group, and what I learned about the group, I can apply to constrain the individuals. And another example where we do this quite frequently is from my research on, say, psychology, where we have a behavioral task that we test 20 subjects on, and often we don't have enough time to collect lots of data. So each subject by itself, the estimates we would get if we fit a model to that guy, it will be very noisy. And that is a way to build a hierarchical model to basically learn from the group and apply that back to the group so we will get much more accurate estimates for each individual. That's a very nice property of these hierarchical models. So here I'm just going to generate, again, some data. And essentially the data will be just an array, 20 times 300. 20 subjects, 300 trials, and it will just be each row is the binaries of each individual, right? And then for convenience, I also create this indexing mask that I will use in a second that might not make sense right now. But just keep at the back of your mind that basically I'm indexing the first row will be just an index for the first subject, and indexing into that random variable. But this is the data that we're going to work with. Okay, so how does that model look like in PMC3? So here I'm going to first create my group variables, the mean and group scale. So what's the average rate, the average chance of feeding the market of all algorithms and how variable are they? That's going to be the scale parameter. And this is a choice you make in modeling, which price you want to use here. I use the gamma distribution, and for the variance, I use, sorry, I use a beta distribution for the group mean, and I use a gamma distribution because variance can only be positive with certain parameters. But the details of that are not that critical. Then unfortunately, the beta distribution is parameters in terms of an alpha and a beta parameter and not in terms of a mean and variance. Fortunately, there's this very simple transformation we can do to these mean and variance parameters to convert them to alpha and beta values that I'm doing here. And while the specifics of that are not important, I just wanted to show how easy it is. And if you use some other languages, that's not a given that you can just really very freely combine these random variables and transform them and still have it work out. And the reason is that these are just these theano expression graphs that once I multiply them, it will actually take the probability distributions or the formula and combine that and actually do the math in the background of combining that. So then I need to hook that up with the Thetas with my random variables for each algorithm. And instead of having a for loop and just generating 20 of them, I can pass in the shape argument and that will generate a vector of 20 random variables that will be Theta. So this is not a single one, but actually 21. And before, we will note that I had just my hard coded prior of five and five here, right, in the previous model. But now I'm replacing that with the group estimates that I also gonna learn about. And now, again, my data is gonna be Bernoulli distributed. And for the probability now, I'm gonna use that index that I showed you before. And essentially that will index into this vector in a way so that it will turn that into a two dimensional array of the same shape as my data. And then it matches it one to one and it just does the right thing. And then I pass in the 2D array of the rows of binary variables for each algorithm. And again, I'm running, I'm finding Google's starting point and note here that I'm using now this called nuts sampler, which is this state of the art sampler that uses the gradient formation and works much better in complex models. Specifically these hierarchical models are very difficult to estimate, but this type of sampler does a much better job. And that was one of the reasons actually to develop I'm C3. And then with the trace plot command, we can just create a plot. So don't mind about the right side. But here now we get our estimates of the group mean. And again, we have not a single value, but rather the confidence. So on average, we think it's about 46%. We have the scale parameter and we have 20 individual algorithms. So that would be theta one, two, theta 20. And all of them constraining each other in that model. So that's pretty cool. So to wrap up, I hope I convinced you that probabilistic programming is pretty cool. And that allows you to tell a generative story about your data. And if you listen to any tutorial on how to be good data-signed, it is telling stories about your data. So how can you tell stories if all you have is that black box inference algorithm? So I think that's where probabilistic programming is really quite an improvement. You don't have to worry about inference. These black box algorithms work pretty well. You have to know what it looks like if they fail. And it can be tricky then to get it going. So it's not super trivial, but still, they often work out of the box. And lastly, PIMC3 gives you these advanced samples. I'm going to skip that and go to further reading. So check out Quantopia and everyone on design. Everyone that have, hopefully, a higher chance than 50% of beating the market. For some content on PIMC3, actually, I have written a couple of blog posts on that. And currently, that's probably the best resource for getting started. And mainly, that's just because there is not that much else written about PIMC3 in terms of documentation. And down here, these are also some really good resources that I recommend to learning about that. So thanks a lot. Yes, please. All these uh tool methods for the way of attempting a products from SWAT service, members at SWAT is broaden space and they use all these tools to put bowels on the variables how do you compare it to an FTC? Yeah, so the question is, Stan provides a lot of tools for assessing convergence and many diagnostics, but also a very nice feature of transforming the variables and placing bounds on that. And so, PMC3 has the most common statistics that you want to look at, like the government Rubin or hat statistic and all of that. And you can sample in parallel and then compare. And we do have support for transformed variables. It's not as polished as Stan, just because it's still an alpha. But yeah, it's there, and you can bound your parameters. So yeah, that works. But it's not quite as streamlined yet. More questions? Sure. So I'm not saying I have a real life model that I can do Hamiltonian Monte Carlo, because it's expensive, but I don't like it. Or I can't mount sampler. And how hard is it to include this into the PMC given I want to use only infrastructure? Great question. So the question was, I can't use this sampler that we provide here, Hamiltonian Monte Carlo, because it's too expensive to use. So how difficult would it be to use my own sampler? And that, I think, is a big benefit of PMC3, is that you just basically inherit from the sampler base class, and then you overwrite the step method, and then you can do your own proposals, and acceptance, and rejections. So that's very easy. And if you look at Stan, for example, I hadn't done it. But I imagine that it's quite difficult just when I look at the code. It's really hardcore C++, and all the templates make my head hurt. The other question, if I understood you correctly, was if you can't evaluate the likelihood or? Yeah, just if I can, I just need to think about it. Oh, yeah. So I would put up this code that it sends that in our own base class. And if I write a step method for going something that we do, and this will be a Python method, you can call it in on the number of other new solutions of how this is. So the question is, how does it compare speedwise, I guess? Or if you write your own sample, and Python won't that be slow? And so I think most of the time is actually not spanned in the sampler, but rather in evaluating the log likelihood of the model. And also the grading computation is very difficult. And it's true that Stan is fast, but it's fast once it gets started, but it takes quite a while to compile the model actually. So in that sense, I haven't really done the speed comparison. And we recently have noticed some areas where PyMC3 is not fast. And we need to fix those and speed it up. And certainly the Stan guys have done a lot to really make a try. And that's the benefit of having C++. But on the other hand, one benefit, I think, to Theano is that it does all these simplifications to the compute graph and does clever caching. And you can even run it on the GPU. So we haven't really explored that to the fullest extent yet, but I think there's lots of potential speedups that just Theano could give us. And another answer to your question as well, if you, for example, you really spend that much time in your sampler of just proposing drums, you could also use Scython, for example, and code your sample in that. Yes? What are you considered to do parallelism? The question is about parallel sampling. And that is possible. So there is just a P sample function instead of the sample function. And that will distribute the model. It doesn't quite work in every instance yet. But yeah, it uses multiprocessing. So you get true parallelization. And just as an aside, there's this really cool project that someone on the mailing list just wrote about that is about PIMC2. But the same trick could be applied to PIMC3. And he uses Spark to basically do the sampling in parallel on big data like if you have data that doesn't fit on a single machine. You can run individual samplers on subsets of the data in parallel and then aggregate them. And Spark lets you do that very nicely. And he basically hooked up PIMC and Spark. So that's really, really exciting. Unfortunately, we got to thank you. You can ask questions at times. Dolly. So let's give him a big round of applause. Thanks so much. Thank you.
|
Thomas Wiecki - Probabilistic Programming in Python Probabilistic Programming allows flexible specification of statistical models to gain insight from data. Estimation of best fitting parameter values, as well as uncertainty in these estimations, can be automated by sampling algorithms like Markov chain Monte Carlo (MCMC). The high interpretability and flexibility of this approach has lead to a huge paradigm shift in scientific fields ranging from Cognitive Science to Data Science and Quantitative Finance. PyMC3 is a new Python module that features next generation sampling algorithms and an intuitive model specification syntax. The whole code base is written in pure Python and Just-in-time compiled via Theano for speed. In this talk I will provide an intuitive introduction to Bayesian statistics and how probabilistic models can be specified and estimated using PyMC3.
|
10.5446/20038 (DOI)
|
I hope that won't sound too sponsoredish. My intent is actually to talk about some of the technologies we're working on that are open source. I'll give you a little brief insight into what we do as a company, but mostly I'm going to talk about the open source tools we're doing that really drive from my experience as NumPy and SciPy communities. We are basically a team of scientists, engineers, and data scientists trying to build tools for others, scientists, engineers, data scientists. We feel like in the wider ecosystem of computer science and computer technology, that category of people, the domain experts, the scientists, the domain scientists tend to get left behind as people focus on developer tools only. We tend to be developers that focus on the scientific tools. There's a lot of need for this in the real essence of the big data movement is really getting insight from those data. That insight requires scientific models typically. I'm Travis Olyphant. My background is in NumPy, SciPy. I'm actually on a PSF. I'm a PSF director currently as of June. I started the NumFocus Foundation. We'll talk a little about that beyond. I've been a professor of UIU, been a scientist myself. My roots are as a scientist, but we created a company really to allow other people to build open source software. We love open source software. Peter Wang is my co-founder. Two and a half years ago, we built Continuum. Our whole purpose is really to allow other people to help us build open source and deliver it to the enterprise and really make it a part of everybody's enterprise experience. That's what we're about. We love open source. It's part of our DNA. I've been contributing to open source since 1998 when I first found Python. I've been a Linux user. A lot of us do a lot with open source. Now we've got 50 people worldwide. We have remote developers. Depending on the project, remote developers work really well. Sometimes it can be difficult. So we try to find those projects where remote developers can work really well, but they are available. We have major contributors to NumPy, SciPy, PANDA, Sympy, Ipython, and we love more. We love newer, new open source products as well. We think that open source can be more than just a hobby. Our desire is to grow the community. That's why we started the NumFocus Foundation two and a half years ago as well. This foundation, its whole purpose is to promote accessible computing and the sciences and to back NumPy, SciPy, PANDA, Sympy. A lot of these projects, they are emergent open source projects with just kind of loosely affiliated community and not much money to help them. NumFocus' purpose is to gather money from enterprises and drive it towards sprint development, towards open source scholarships, for diversity training, diversity events. NumFocus also sponsors and promotes and actually receives any residual income from the PYDATA ecosystem, the PYDATA conference series. We're having one as an affiliated event to this event, so please come to the PYDATA conference. You'll hear all about the great scientific tools, the great data analysis tools that are emerging. Now, as a company, what we sell is enterprise consulting and solutions, optimizing performance to managing DevOps and a big data pipeline to building native applications in the web or on the desktop. We also provide training, Python for Science, Python for Finance, as well as practical Python through our partners Dave Beasley and Ray Van Henninger. And then we are building the Continuum Platform, which is a product for kind of taking the desktop to the data center and back that allows people to deploy data analysis applications and dashboards. So, our products are all centered around that platform that kind of take the appearance of Anaconda Add-ons and Anaconda Server, Wacari Enterprise. I'll show you briefly just those. The key behind these products is to really give experts and scientists what they really are asking for. I've been spending a lot of time myself as a scientist. I kind of understand what the workflows they desire are, and we're trying to bring that to large organizations, large companies. So, this is a picture I show of the Continuum Platform. You can see that it rests on an open source base. And an open source base that we contribute to greatly. We continue to contribute to it. The IPython, Sympi, SciPy, NumPy, Pandas, that basic baseline, and we have additional open source products that we're writing and growing. Numba, Bokeh, Blaze, Dye and Conda, LVMPy, PyParallel, all these things are trying to bring high level scientific applications, make them easier to write, make them faster, make them take advantage of the hardware that's changing today, GPUs, multicore. I wrote NumPy six years ago. I still know all the bad places where it does not optimize. There's many, many places, and it's not optimized because it can't take advantage of multiple cores or can't take advantage of multiple GPUs. On top of that, we deliver Anaconda, and then above that are some of the proprietary applications that we provide, all about creating applications that can deploy in the enterprise very, very quickly, and really empower the domain experts that exist in every organization. Why Python? We love it because it provides a spectrum. What you'll see here in the Python community is kind of different categories of people. We have some people that are web developers, and they love that. Some people are DevOps folks and system administrators, and they love that. Then I kind of in the camp of data scientists, scientists, and sometimes it can be challenging because we don't all speak the same language, and so we kind of talk and use different words and different terms, different libraries. One thing that's great about the Python community is it is a community, and people, for the most part, listen to each other, try to work forward on solutions that help everybody. In particular, some of those people that are in the Python community aren't even developers. They're what I call an occasional developer. They're the cut-and-paste programmers. I have an idea. I kind of want to put a few things together, and Python, it fits my brain. It's partially leveraging my English language center so I can kind of understand what it's saying and not have to be a developer to use it, and I can build things very quickly. Python does that. It's very unique, actually, among all programming languages. Now, NumPy, it plays a central role in the kinds of tools that we build. It's the center of a large stack of data analytics libraries. There's a lot of users of NumPy, actually. I think about three and a half million. It's hard to tell because they don't ever tell me. They don't write home and send me a postcard. Sometimes it would be nice if you could actually get a sense of who did and who used it. As a company, so that's kind of what we build on, but as a company, we ship Anaconda. Anaconda is a free, easy-to-install distribution of Python plus 100 libraries. One thing that's challenging about the NumPy stack is it uses extension modules. It uses C, sometimes Fortran, for SciPy. How do you get that installed? It's not enough to just have a source-install solution. We have to have a binary-install solution. We invented Conda. Conda is, and we work with the Python packaging authority to try to promote Conda, help understand how it fits in to the overall packaging story in Python. Essentially, it's like Yum and Apgit for Linux, except it's for all platforms, Linux, Mac, and Windows. It's a fantastic distribution that people rave about. They love it when they use it. Why do they love it? I think Conda is a big reason. Conda is a cross-platform package manager. It helps you manage a package and all its binary dependencies. It's an easy-to-install distribution that supports both Python 2.7 and Python 3.3. You can actually install Anaconda for 2.7, then create environments. Just had a talk by Red Hat. They call these software collections in the next space. We call them environments. They're system-level environments that let you, they're more than just Python. They support anything. So you can run Python 3.3 in a separate environment on a Python 2.7 base. You can also do the reverse. Get a Python 3.3 base. You can run Python 2.2.7 as a compatibility test development environment separately. It's a fantastic solution for bridging the gap between Python 3 and Python 2. Then there's over 200 packages available. Scikit-learn, Scikit-image, iPython notebook, just at your fingertips. Conda install gets them and you're off and running. You know we're compiling dependencies and we're trying to figure out how to install it. And this is all for free, completely free. You can even redistribute the binaries we make. So that's Anaconda. The purpose is to make Python ubiquitous in data science and have, there should be no excuse for anybody in the world using Python to solve their data analysis needs. And that's why we make Anaconda. Get it at Continuum Iow Downloads. It's free for downloading, free for distributing as well. And we do sell some things on top of that. As a company, we have to stay in business. We have to sell something. And part of that is an Anaconda server. It's a commercial supported Anaconda. Provides support, provides identification licensing. It also provides a package mirror and kind of a management tool. And if you're interested in that, I can talk more about that to others. Come see me later. Binstar.org is, you can see kind of what Anaconda server might look like on a, on premise installation. But going to binstar.org, signing up, get a free account and you can upload there any package you like. There's a three gigabyte limit, so don't just show up all your movies and content packages. But you can put any binary package you like and share that with somebody else so that they can easily install your solution. And as long as it's public, as long as anybody can download it, it's completely free. Mokkari is our hosted analytics environment solution. It's a fantastic way to quickly and easily get running with the IPython notebook. You can sign up and instantly you're in an IPython notebook running code. Now the free version gives you a node with only a little bit of memory and only a tiny bit of computational power. But it's a great for teaching, for showing, for demonstrating. If you want more power, you can easily upgrade to get as powerful a node as you like. Then Mokkari Enterprise is the on premise version of that cloud story. It's adapted for, the UI has changed to allow LDAP support integration to, it installs to internal servers, it has a notion of projects and teams, and that's people instantly collaborate on a large scale project and then share the results of their workflows with others very easily. So from desktop to data center is kind of our platform story. It helps you anaconda on the desktop, Mokkari on the data center, and a seamless connection between the two so you can go from writing code on your desktop to deployed applications that are on the cloud or on the data center on premise. So that's our solution. That's the thing we are building together as a company that helps all enterprises everywhere. But the part I like the best is the open source tools that we're actually building as a part of this. We feel it's critically important to continue to build open source technology. So we have key open source technology that builds on top of NumPy, SciPy, Pandas, and the rest. So Blaze, Bokeh, Numbaconda, I don't really have time to explain all of these in the brief time I have. Tomorrow, my keynote, I'll be talking not about all these technologies. A little bit, I'll mention NumBa, probably mostly about Blaze and kind of how I see it as part of the story for the future of big data analytics. We do have some add-ons. I talked about those before. So I'm going to briefly talk about kind of these technologies. Get you excited about it. We're looking for help. We're looking for developers who can help us with each of these. These are paid positions. So one is NumBa. NumBa is really a technology about taking the CPython stack and providing compiled technology to it. So PyPy is a fantastic project, but it doesn't integrate with the NumPy stack very well. NumPy, Matplotlib, SciPy, Sympy. So we took and took the LLVM technology stack and with decorators, we can take a function, compile it to machine code, and integrate it in with the rest of the NumPy stack very easily. It takes advantage of the LLVM tool stack. So the kind of work that we're doing is to basically translate a function that looks like this with a decorator, create a general assembly kind of code via the translator, and then LLVM takes that code and runs on your platform. It can do amazing things. I think it changes the game. It lets Python essentially be like a compiled language. It's a subset of Python, and we can go to the details if you like later. But a subset of Python can now, you can write it in the Python syntax and get compiled performance just as if you'd written C or C++. And we have numerous examples of that. Very, very easy way to get optimized performance out of your Python code. Here's a simple example of a manual brought generation. Got to have your manual brought generation example. It illustrates the ability to call functions and have that bypass the Python runtime and essentially be a low-level machine code. So this is one way to bypass the GIL. Use Numba to add a JIT, and now you have a Python, now you have a, it's not in the Python runtime anymore. It's actually compiled code, and you can release the GIL and execute that. So that's Numba. And Blaze is our data to code seamlessly. It's about taking the fundamental problem Blaze tries to solve is when you have data in, let's say it's an HDFS or somebody else in your teams as well, I think we should have it in Postgres with Green Plum, or maybe we should have it in Netiza, or maybe we should have just a bunch of HDF5 files. That decision of how you store your data ends up determining how you write code, how you write your queries, how you write your solution in Python. It shouldn't be that way. There ought to be a way to write an expressive table-oriented code that then you just plug in whatever data you have, and even allow you to cross different tables and have the same expression work across all those tables. So Blaze is a foundation for large-scale array computing that leverages the technologies that are out there already. So with data, this is describing some of the pain involved in data. There's many, many kinds of formats. The data pipelines are constantly changing. It can be difficult to reuse code in that environment. The Blaze architecture has an API, it has some fundamental pieces, a deferred expression, and a compute, a pluggable compute infrastructure, and a pluggable data infrastructure. So it's a flexible architecture that it can scale across multiple use cases. So data, for example, it can be stored as CSV files, or a collection of JSON files, or HDFS, or HDF5, or just in SQL. You can add your own custom data type. So a simple API lets you add it, but then your Python level expression is common. It's more numpy-like. You can slice it. You can grab pieces of it. And then you can write in a compute graph that refers to part of that data. So this is a compute abstraction that basically can sit on the top of multiple back-end libraries. Things like pandas. Dined is a next-generation numpy equivalent. It's a C++ library that does the same things as numpy, but allows other... It's more general, allows things like variable link strings, ragged arrays, and categorical data types, which are missing from numpy. You can also sit on top of Spark, which is part of the Hadoop ecosystem, PyTables, which from our friend, Francesc, who's sitting in the back. Then this blaze expression graph, you can write a single expression and have it attached to multiple data sources and pull it all together in a single application. Here's a simple example. We have a generalized data declaration format called DataShape, which generalizes numpy's D-type. And this DataShape allows you to describe data universally in a way that can sit on top of multiple data formats. So here I'm creating a symbolic table. And this symbolic table, then I can write an expression involving that symbolic table, including joining, group buys, aggregations. Now that creates a deferred expression. And then the load data, there's different implementations of load data, depending if my data is in SQL or if it's in Spark. And then I simply map the elements of what I've loaded to basically a dictionary representation of the namespace that that compute is going to evaluate in. And then the compute maps the expression graph to the actual back end calculations that are needed. So whether it be pandas in memory or Spark on a 100 node cluster, the same code can be executed. So this is the load data showing the difference between the Spark and pandas. I'll talk more about this tomorrow because I think it really sets the stage for reusable computing and reusable expressions and helping people make sense of the diverse and changing world of big data and large scale array-oriented computing. So the last technology, I didn't tell you a lot about Kanda because I've got a lot of videos out there. If you want to hear about Kanda, there's actually some jokes about me constantly talking about Kanda because I love it so much. You can find videos about Kanda on the web. I'm going to talk about Bokeh, which is our visualization library. I'm really excited about the visualization library. A lot of people are as well. It basically allows you to do interactive plotting in the web without writing JavaScript. So as a Python developer, you can write interactive visualizations in the same spirit as D3 but using Python. Now, it's still in development, but quite a bit can be done already. You can have novel graphics. Actually, the violin plot came from a Seaborn library using the map plot lib compatibility of Bokeh. So you could have a map plot lib plot and then essentially render it with Bokeh to provide the interactivity and the JavaScript rendering. Lots of different kinds of graphics can be built. There's even streaming and dynamic data that can be built. I have a simple demo here. I'd like to show basically it's running in the background. So if I go to... This is just my basic computer and it was been running for a while and it's the microphone. What I'm doing is using the NumPy stack to do a Fourier transform on the audio coming from the microphone and show that spectrogram in a couple of different ways. So I can see the time series. I can see the frequency. Here's the time series. Here's the frequency spectrum. And then here's a image map of the frequency spectrum. Take this line and rotate it and stick it in an image and then it kind of moves across so we get a spectrogram image over time. And then here's just a radio plot just for fun. So you can see that this is sampling the microphone. I can't whistle that high. Anyway, there's things to do with the game. Thank you. So this is a JavaScript library and you can actually do this from Python. Currently this demo is currently written in... It's taken advantage of the Bokeh.js back end but it's being written in Python so you can show just how to do this sort of thing in Python and create these kind of visual apps. It illustrates many things about the platform that I think is the new platform for visualization which is the web browser. So this is what we're doing. It's what it's about. So one of the things you can do if you come with us, if you come work with us, let's go back to my presentation. Not the Twitter feed. Although that may be some of you tweeted. The other aspect of dynamic interaction is that because it runs... There's a web socket communication and an object model. Bokeh creates a scene graph and the web browser and an object model that can be reflected in the Python side. So you can write objects, an object model in Python that gets reflected to the browser and you can have server side control. You can also just have all that logic in the browser and have kind of a static web page that has all the interactive logic in the browser. So this is just an example of essentially a web... The web service updating the plot and then the backend server updating the plot in Python and having the web display change. So it's a great way to handle streaming data and all kinds of different interactions. You can also do big data kinds of analysis pretty easily with this kind of setup. I can go to... This is running actually in the US. So these are time series that are stored on a server and I have just a ability to zoom in. So you can see it's actually updates. It zooms in initially with the data it has and then it goes back to the server and updates a higher resolution version. And these are all links, these different plots. So it's just a simple example of resampling. Then I can reset the view and then it expands out after it grabs the data. This is actually back in the US so there's a little bit of latency between them. Here I have an example. I'm actually looking at the whole world. You can see I've zoomed into a particular slice. This is a worldview. It's a three-dimensional time series, about four gigabytes of data we got from the JPL from NASA. And it shows the ocean view and it's in time. So I'm seeing a 2D vision, 2D projection of the world, but this slider changes the time view and takes a little bit to bring back all that data. But if I zoom in to a particularly interesting area of the world, I can see that it updates from the server. It gives me back this resolution view and then I have projections that show the period per time and I can change which slice it shows. You can see it's updating down here. So that's just an example of an application built with the visualization and the kind of things you can do very quickly and then deploy in a web browser across your organization. You have a little widgets you can provide. This is just an example of a simple widget and some dummy data about downloads and I can adjust as I slide through it. These are the kinds of things you can do from Python without running JavaScript using bokeh and its application technology. So that's the gist of what we're trying to do on the platform, basically from data to visualization and beyond and make it easy for people to do it at a high level so they don't have to be expert developers. They don't have to change and know everything about SQL and about JavaScript and about development operations in order to get solutions that take advantage of multiple kinds of hardware, multiple kinds of data sets and high level ideas. So no JavaScript just a little more example of the kind of plots you can do. With bokeh there's actually going to be a tutorial at PyData. I invite you all to come to PyData in the tutorial given by one of the, by the principal author of bokeh, Brian VanDeven will be here. There's also a great website you can go to that will explain bokeh, bokeh.py.py.org. It's got a gallery. You can go in and look at the code. It's still a work in progress. It's 0.5 just came out. The widgets just came out. It's making rapid progress. But it's usable today. But if you find something that you want and it's not there, let us know. And I'm sure it's either on the roadmap or it will be added if you let us know about your particular needs. Okay. So that's a quick run through of the technologies we build and the kinds of things we do. And basically I'll end by talking about the openings that we have. There's many openings for the number team, for the blaze team, the bokeh team, embedded consultants. If you want to live in New York, come talk to me. I have great opportunities for you in New York City. And these are opportunities that not only work with a client but work with the rest of our team in helping us build this platform based on open search technology that can benefit large and small scale organizations around the world. We're really excited what we're doing. We think we have ideas that can really help and transform the way people write code and write code for high level data analysis and we'd love to have you join us. So I think with that I'll ask for questions or anything else you want to know about. So we got any questions? Thanks for the talk. I have two questions regarding the Python part of the bokeh. So first of all, I remember that at the beginning the bokeh was something which was trying to implement the grammar of graphics for Python. But recently I saw that there's no mention in the documentation about the grammar of graphics. Are you still using the same kind of interface or just? I would say it's not the grammar of graphics. Well, I know that some other developers see the grammar of graphics as a good direction but not necessarily complete. And so bokeh.js itself uses concepts of the grammar of graphics and its architecture. The interface is something that can be added on top. So for example, ggplot, which is a, it currently has a back end in Matplotlib, can easily be retarget the bokeh.js. In fact, we have examples using Matplotlib's interface of doing that. So I would say the grammar of graphics discussion is higher level than where, then, kind of bokeh and bokeh.js. And the second question is regarding the widgets and interactivity of the plots. So I understand widgets is something that you can play with to, for example, select some data points, get some all information, both particular data. Is it something that you have to implement in JavaScript or you can just use Python code to define the widgets? For which? For example, if I want to select some data points and print maybe some tool. Right. Selecting data points and pretty them. I believe that's on the roadmap to be done from Python. I think currently to do that, you have a nascent Python interface to that. And so if it works for you, it might be enough, but it's possible that it's still not quite complete that API. So the idea is you won't have to use JavaScript. I'm not sure if it were completely finished with that API on the selection of points side. Any more questions? And Brian will be here later today and he can give you a lot more explanation of Bokeh. Anyone else? No? Okay. Thank you very much. All right. Thank you. Thank you so much.
|
Travis Oliphant - The Continuum Platform: Advanced Analytics and Web-based Interactive Visualization for Enterprises The people at Continuum have been involved in the Python community for decades. As a company our mission is to empower domain experts inside enterprises with the best tools for producing software solutions that deal with large and quickly-changing data. The Continuum Platform brings the world of open source together into one complete, easy-to-manage analytics and visualization platform. In this talk, Dr. Oliphant will review the open source libraries that Continuum is building and contributing to the community as part of this effort, including Numba, Bokeh, Blaze, conda, llvmpy, PyParallel, and DyND, as well as describe the freely available components of the Continuum Platform that anyone can benefit from today: Anaconda, wakari.io, and binstar.org.
|
10.5446/20036 (DOI)
|
Okay. Ready for the next talk, then? This is Federico Fraenguieli, who's a C++ in Python developer at Evernove, and he's going to talk about how to make a full-fledged REST API with Django and Oth Toolkit. So, over to you. Hi, everyone. So, the goal of this talk is to show you how to create a REST API protected with Oth Tool. But first, I want to tell you how you should know how to do it. And I want to tell you a story. So, let me introduce you to this small and simple application, the Time Tracker, which is, of course, a web application that tracks the... allows the users to track the time they spend on their activities. And at the beginning, at first, I had... I had to choose to pick one tool, which was Django, and they had one single big project, and I deployed it once, and everything was fine, more or less. But as he used to see, the times they are changing. And what has changed? Actually, what Frontend development has changed a lot. Sorry. Today, we have a lot of web frontend frameworks, and that allows you to create amazing frontend applications. And they have their own development tools to build, to test, and to run the application. And so they are completely separate applications. And I also had to support multiple devices, which means to support different browsers and different platforms. But I should also take care of the native applications. So, I ended up with a lot of projects. I had a Time Tracker backend, a Time Tracker web, which is the frontend application, and the Android project, the IIS project, and the desktop application for the old desktops. And, moreover, there are third-party services that want to connect with my Time Tracker application. They want to send me data. They want to read data from my Time Tracker applications. So, what happens in the backend application? What's in the backend application? There is a service that exposed an amazing and reliable REST API. And this is the recipe. Django, Django REST framework, and Django of Toolkit. These are the models. They are really, really simple. There is the activity and the time entry, which is the model that allows us to track the user, the time that a user spent on an activity or a task. And these are the endpoints that I want to create. Yeah, on the mouse left column, the URL, the HTTP method supported, and the semantic meaning of each method. So, for example, the first row, you have API slash activities. If you send a request, you get back a list of the available activities. To create the endpoint, I need to show how Django REST framework works in, I hope, in less than five minutes. So, the first thing you have to do is to serialize your data. And this is really, really straightforward in Django REST framework. You can use this base class, serializer. And this works just like Django Forms. So, you just define the fields you need and add some code to restore or create the instance of the model from its serialized representation. And then you can use, easily use the serializer and you'll get back a dictionary representation of your object. Of course, this is boilerplate code, which should be repeated for every model of your application. But you can avoid to write that code using model serializer, which allows you to just specify which is the model you are serializing. You want to serialize. Then we have to create the views, the endpoints. And what do we need? We need, of course, to respect the semantic meaning. And we should take care of the user authentication. And also we should take care of permission checks. Also, sometimes, object level permissions. Sometimes you need to paginate your endpoints because you get a lot of results. And also you want, maybe, to handle response and request formatting, to support, for example, JSON, XML, YAML. So, this is a lot of stuff. But you just keep calm and use Django as framework because Django's framework is really, has a lot of settings that allows you to customize its behavior, its default behavior. These are just small examples. The first one allows you to define which is the class that takes care of the user authentication. And then we have the default permission class. So, if you're not authenticated, you won't get anything just for 01, for 03, sorry. And the default renderer and the default parser for the formatting, for the format. Well, when you, to create the endpoints, you can use the API view-based class provided by Django as framework, which allows you to add some code to the handler meters. And this class will use, this base class will use the settings. We can see here to create, to realize, to create an endpoint with the correct behavior, with the behavior you want. And, well, the code, it is really, really easy to understand. Here we have the query set to retrieve all the activities. We serialize the query set and return the serialized response. And, of course, this is boilerplate code. So, you can, you have to repeat this code, 44 endpoints, but you can avoid to write this code using the generic class-based view provided by Django as framework. Here you just need to specify which is the base query set and the serializer class you want to use. There is also a built-in browserable API provided by Django as framework, which we are going to see at the end of the talk. So, the next step is how do you authorize client applications? I mean, your applications like the Time Tracker Android and the Time Tracker iOS, they need to be authorized to talk to the Time Tracker backend API. And also, there are third-party apps that want to access your user's data. So, you need an authorization, you need some authorization engine behind. If you don't have this authorization engine, these are the problems with your going to face. First of all, you have to store the first solution without the authorization framework as to store the user password in the application, which is not good, of course, because the application gets a full access to the user account. And if the user wants to revoke his password, it wants to revoke the access, sorry, to the application. He needs to change his password. So, also compromised apps can expose the user password and username. This is the solution, the ought to authorization framework. So, how does it work? And I want to explain how it works using this simple use case. So, imagine there is a songify, music streaming service that wants to connect with the Time Tracker application. So, their users can track their listening activities on the Time Tracker application. These are the actors. This terminology is the same used in the RFC, in the ought to RFC. And I'm just trying to translate these terms to this use case. So, the resource owner is, of course, the unit, the resource server is the Time Tracker API. And the authorization server, in this case, is the same as the resource server. And the client is the songify application. I want to explain you what the ought to authorization framework defines for flows, for authorization flows. I want to show you how one of these flows work. This is the most popular one, the authorization code flow. So, the first step is when the client registers with the authorization server, and the authorization server provides a client, provides a client ID and a client secret. The client, of course, is the songify application. So, there is someone at songify that, for example, goes to developer.timetracker.com, add developer application, and it gets back a client ID and a client secret key. The second step is when the songify application redirects the user to the Time Tracker application via its user agent, via its browser, for instance. And next, the Time Tracker application authenticates the user and obtains the authorization to communicate with the songify application from the user. Now, the Time Tracker application redirects the user back to the songify application with an authorization code, which is later exchanged for a token. And the token can be used by the client to authenticate requests. How to do that in Django? With Django, our toolkit, of course, which supports Django from 1.4 to 1.7, Python 2 and Python 3, and it is built on top of Outlib, which actually is a really great library. It takes care of the compliance with the RFC. We just wrote some glue code. And its integration with Django is really, really easy. You add the auto provider application to the installed apps, add our URLs to your patterns, and you can create a protected endpoint using our generic protected resource view. And here you have an API endpoint, which is protected with OAuth 2. Now, it comes with batteries included. So we have built in views to register developer apps and a form view for the user authorization. It is integrated with Django as framework. You just need to switch the default authentication classes with the one provided by Django, our toolkit package. And now I want to show you how it works. So these are the steps. We are going to the authorization. First, we are going to create a developer application that we are going to simulate the step when the user is redirected to the authorization endpoint. So here you can see this is one of the built-in views to register new developer applications. So you create new application. You add the name. Here you have your client ID and client secret. You can choose these are details from the OAuth 2 framework. Anyway, I got my Sangify application ready. Here it is. So we can just use this one to... Step one, the Sangify application is redirecting the user agent of the user to the authorization form. But first the user has to authenticate. And now the application is asking for my authorization. And we authorize, of course. Come on. Okay. Of course, this should be the URL of the client application. So Sangify.com, for example. Now we can take this authorization code. Just substitute the code here. And we are going to exchange the code to obtain a token. Here is the response. And this is the token. And the token can be used to create a proper request with this header, authorization bearer with token. So I just want to show you that I'm not lying. If I try to get the list of the activities, just tell me that I'm missing the authentication credential. If I use my new token, I can get back the list of the activities. And that's all. The future plans for Django, toolkit art support, alt1. Maybe alt support for the open AD connector, RFC. We really don't know. I still have to read the paper. And add no SQL storage support for the applications storage. So we need some help. And thank you. That's all. Any questions, please? Anyone? Thank you for your talk. My question is, can we use the same framework to post tweets on Twitter? Django toolkit actually is the implementation of the server side part of the alt2 specification. So if you want to send tweets to Twitter, you need a client side implementation of the alt2 RFC, the authorization framework. In your examples, the authorization server was always the same as the actual service. With your toolkit, can you make two services? Can you separate out the authorization part? We have to work to keep the resource server and the authorization server separated. But maybe you can. Actually, you have to write some more code to keep the authorization server and the resource server separated. Hi. One quick question. If you want to expose your API differently, resources differently from the actual model definition, can you handle that in the serializer class so you don't have to use the Django or the RAM? Yes, of course. If your data is, for example, in MongoDB, you just write your, you just use the serializer base class to write your own serializer. And it just works. You have to write some more code, but it works. It is okay. Okay. Thanks. You're welcome. Hello. Thanks for your talk. What are you using for object ownership? Sorry? Object ownership. How do you say this object belongs to this user and I only showed this request coming from this user? Well, this is object level permission. You are using any component for that? We are just filtering the query set looking at the, well, actually, okay. We sometimes you, we can get back, which is the user bound to the token. Okay. And with the user instance, you can filter the query set. This is a solution, really simple solution. Yeah. No problem. Any more questions? Okay. Thank you very much. Thank you. Yeah.
|
synasius - How to make a full fledged REST API with Django OAuth Toolkit World is going mobile and the need of a backend talking with your apps is getting more and more important. What if I told you writing REST APIs in Python is so easy you don’t need to be a backend expert? Take generous tablespoons of Django, mix thoroughly with Django REST Framework and dust with Django OAuth Toolkit to bake the perfect API in minutes.
|
10.5446/20034 (DOI)
|
So, good morning everybody. Stefan Schwarzer is a Pythonist for 15 years. He has written articles and a book on Python. So he's a regular speaker at Linux and Python conferences. And also he maintains the FTP Util library, which is quite handy. Thank you. Today he's going to discuss problems and back practices for maintaining code for Python 2 and 3. It's going to be about like a 35-minute talk and we have five minutes for Q&A afterwards. So please welcome Stefan. Thank you very much. Wonderful. Good morning. Okay. Yeah, I want to give the talk about supporting Python 2 and 3 with the same code. Maybe for the interaction, something about me. Okay. Some things were already mentioned. But okay, I have a degree in chemical engineering, but in 2000 I kind of switched sides. I mean, I've been programming since I was 15 before and in 2000 I became a full-time software developer since 2005. I'm self-employed. I maintain this FTP Util client library and the starting point for this talk was, yeah, that I was myself in the situation when users asked for Python 3 support and FTP Util. There was one ticket and then at some point there was a question on the mailing list and yeah, but I had shied away a bit from this and yeah, so I went through. Okay. Yeah, last year we had this, or I had this FTP Util 3 zero release with Python 3 in addition to 6 and to 7. Okay. Yeah, Python 2 or Python 3, yeah, I think most in the audience will actually have read this. This is from the Python website. When you go to the download section, there's a link, Python 2 or Python 3 and this is the wiki page you get when you click on that link. Yeah, so it says Python 2 is legacy. I mean, it doesn't really say you shouldn't use Python 2 anymore, but yeah, it's strongly in the direction. Python 3 is the present and future of the language. Okay, so it's Python 2 obsolete. So in a way it is, yes. On the other hand, it isn't because it's very widely used. There's lots and lots of legacy code. And yeah, Python 2 is also the pre-installed version. When you say, yeah, I'm installed Python or aptitude installed Python, you get Python 2 or it's even pre-installed usually on Ubuntu and also Fedora and Red Hat. Okay, and Python 3 is optional. Red Hat Enterprise is only yet getting there. Yeah, I mean, they have a Python 3 package now, but yeah. Okay, also if you want to host, if you buy hosting and don't do the hosting yourself, there are more hosting offers available for Python 2, I think. And many libraries don't have a Python 3 version yet and I guess that's a reason why some of you are here because you want to change that. Okay, so my recommendation is use Python 3 if you can. And one thing I find very important is also if you migrate or adapt your library, you use the transition for others who say, okay, I still need your library to migrate myself or adopt myself, my software for Python 3. Okay, there are different approaches. The first one or the, yeah, in the beginning when Python 3 came out, the recommendation from the Python development team was use 2 to 3, writes Python 2 code, converts this to 2 to 3 with this command line tool. Then later there was some 3 to 2 tool. I think it's rarely used, but the idea is that Python 3 code is usually looks a bit cleaner and it's nicer to be able to maintain Python 3 code than to maintain Python 2 code. Okay, but what it seems the Python world ended up doing or mostly doing is developing or having the same code for Python 2 and 3. So you don't have to run 2 to 3. Neither the user needs to run this nor do you during development if you want to test for Python 2 and 3. So it's just always the same source code. Okay, yeah, one major problem. Yeah, okay, I mean, there are many things if you look at what's new in Python 3.0, there's lots of stuff, but I think the most important thing when it comes to this adoption of the code for Python 3 is the spites versus unicode topic. So I want to just say maybe refresh something on this. Okay, we have, and yeah, in both Python 2 and 3 we have bytes or byte type or byte strings. I think Python 2 terminology is mostly byte strings and in Python 3 usually when you read documentation it says bytes, it's not really byte strings. So because the intention is you don't use it for strings, not for corrector data. Before encoded corrector data. Okay, so these are the kind of raw bytes that you store on a disk or send over a socket when you need to make a decision now. I have unicode, but it needs to go somewhere and I need to encode it. On the other hand, unicode text represents corrector data where characters have number code points. And this is unrelated to how the characters are stored. So it's not just another corrector encoding like let in one or so or some code page, I don't know, yeah. But the characters are numbered these code points and at this point it's unrelated to how they, how this is later represented in bytes. Okay, so this unicode text can be encoded to bytes. You can choose an encoding, yeah. Of course the encoding should support the characters you have in your string, in your unicode string. So if you have, for example, Chinese characters, you can't convert this, encode this to let in one. Okay, and this here at the bottom of the page, this is just one example. These would be the unicode code points for this German word hören to here. And this is, and the bytes, if you encode this unicode string to UTF-8. So this one byte becomes these two bytes, otherwise it's unchanged, but this depends completely on the decoding, on the encoding you apply. Okay, yeah. Something that might be confusing while you are working on this adoption for Python 3 is that both Python 2 and 3 call their default string type. So if you write a byte literal in Python 3, this is the byte string type, the binary type here, and in Python 3 it's actually the unicode type, the text type. Yeah, you also have, yeah, and the byte string type is named bytes in Python 3 and unicode is the unicode type in Python 2. Okay. Yeah, one major change, and this is, yeah, again, one of the difficulties for Python 3 support often is that in Python 2 you can just, yeah, add a byte string and a unicode string with this prefix. And mostly it works unless this byte string contains anything non-esky, then you get a unicode decode error dependent on the process data. So one day you run the code and run it's fine, and the other day you read the text file which has an umlaut or something or some special character, and you get an exception. And Python 3 this has changed, and you can't, you don't have these implicit conversions anymore. Get a type error every time you want to add a byte string and a unicode string, which I think mostly is a good thing, but yeah, but it makes it this migration harder, this adoption for Python 3. In Python 3 you have to be explicit, and for example if you have a byte string which has this prefix, you need to decode it explicitly, then this becomes a unicode string, and this is in Python 3, yeah, and this is the unicode literal as well. Okay, but in code, in actual code you shouldn't usually do it like this, but I will talk about this later. Okay, also in Python 3, yeah, I mean, which I find, I think that I find logical, that makes sense is that everything, or almost everything which took strings or byte strings in Python 2 now requires unicode strings. So in Python 3, this is again by default, this is a unicode string, this works, but if you pass the constructor of decimal, a byte string, it complains and gives you a type error. There's also a new file API in Python 3, yeah, here the thing is that if you, yeah, in Python 2, if you open a text file, it still gives you byte strings, yeah, if you open a text file for reading, it always gives you byte strings, yeah, whereas in Python 3 it gives you unicode strings, really, and Python 2 just does, whether you choose text or binary, only changes the line ending conversions. And there's also no more file objects, the return value of open depends on the arguments of the open call. The good thing for this migration, for this adoption for Python 3 is that this open function, which is the build in open in Python 3 is also available in the IO module in Python 2.6 and 2.7. One thing that tripped me up in another small tool is that standard in and standard out and argument values here, this is a unicode strings, yeah, or I mean, system, SCD in, read gives you unicode strings and write requires a unicode argument. But you can work around by using the buffer attributes, which gives you the file object, which works with the raw bytes with the binary data. Okay, steps, yeah. So now the first part was introduction now, yeah, but how you actually do this migration or some tips or steps here. Okay, you should have automated unit tests, possible unit tests. I like this keynote from Emily Bach on, I think, Tuesday, where she also mentioned approval testing, yeah, which so, so if you don't have unit tests. But I recommend you have them if you don't have unit tests. So you should at least have some automated tests that you can run on your code. The code, the unit test should pass 100% on a Python 2. And since you are just starting out supporting Python 3, you can't expect that they work on Python 3. Actually, lots of these tests will fail if you just do the experiment and you have unit tests for a Python 2 version and you run the same test with Python 3, you probably get lots of failures and errors, yeah. Okay, that's completely normal. Also, sometimes I notice that, for example, in FTP that most of these failures come from string literals in the test code in the unit tests. So and sometimes you, another thing is that sometimes you only need to change a few functions which do the conversions in your code and maybe even then lots of your tests pass again, even under Python 3. You can also do this test or try running even the Python 2 version with the option minus 3, this is supported in Python, I think from Python 2.6 up, will give you some information, will print some warnings or information on things that need to change for Python 3. While you are changing your code to adapt to Python 3, make sure then, yeah, you have to change the actual code, the production code and the tests and keep them in sync. And so you don't have, yeah, if you cannot, that many failed tests. Okay, one tip, I found this nice tool problem, I can imagine many of you know this already, this is TOX, nice tool because you can easily run Python 2 and 3, the test both on Python 2 and 3. So you can check if they still work on Python 2 and if they already work or to which extent they work for Python 3. I also like that implicitly checks whether the packaging works for your library or for your code. Okay, since we are going to support Python 2 and 3 with the same code, you run, I mean 2 and 3 is still useful, but I recommend running it once. So for example, print became a function and some, the exception syntax, yeah, changed in Python 3 or the required exception syntax. And 2 and 3 does many of these, straightforward conversions. So in this way, it's nice. You should have a look at the documentation on 4.2, 2.3, about the fixers. There's also, yeah, these are different steps, different conversions that can be applied. So it's not just everything or nothing, but you can, yeah, individually turn on certain conversions. You should exclude the future fixer because this removes all from future import statements. What you don't want, you want to keep from future imports for your Python 2 code. Many of the changes will be, you don't want to keep them literally. I mean, the print conversions you probably want to keep as they are. And some changes, for example, changed module names, for example, config parser in Python 2, which was spelled with a capital C and capital P is now our lowercase in Python 3. So 2 to 3, it expects that Python 3 is your final target destination. It removes, it just changes the import to our lowercase config parser. But if you look at the changes that 2 to 3 did, you see, yeah, you do a diff with a version control system and you see what it changes. So, for example, you see the change import statement and you need to take care of this. Have different code or switch or something later for Python 3. Yeah, yeah, really check all the changes. One, someone I talked with, I think yesterday also, yeah, mentioned or suggested running the fixers individually. So with each diff, you see the changes only from this fixer. Of course, this requires that you know all the fixers. I mean, you can get this information, but yeah, you have to make sure or should make sure that you don't forget any of them. Okay, and after you ran 2 to 3 and made your changes, everything at this point should run again under Python 2. Okay, and if you're lucky, you already have an API that you can keep for Python 3 or doesn't make you jump through hoops if you want to support Python 2 and 3. I was not so lucky with FTPute, so I had, yeah, I would say I had to change APIs to support Python 2 and 3. So, the new FTPute 3 is not backwards compatible with FTPute 2.8, which was the previous version. Okay, and the standard library in Python 2, yeah, almost everything that accepts a string accepts either Unicode or bytes, and in Python 3, yeah, with rare exceptions, you have to use Unicode for strings. And my suggestion, therefore, is use Unicode for text data. If you have to, if you want to decide and should, you should decide on Unicode for text data and not keep the bytes interface that you may have for Python 2. And you need to know or even define what data is text data, sometimes corner cases where you really have to think or maybe even define whether something is supposed to be bytes or Unicode. Yeah, one recommendation is encode and decode text at system boundaries. Yeah, so everywhere, this example is reading a file from the file system, decoded, and later, if, for example, you want to send this over a socket to some other host, you encode this, but only then, yeah, so you should try to have most of your code deal with Unicode strings. So, if you, yeah, I mean, as far as strings are concerned, I mean, of course, the other types are not affected. Yeah, so if you look at code, so you don't have to think, yeah, in Python 2, this will do that thing. On this case, I will have byte strings, and when I run this on Python 3, these strings will, yeah, I think they will be Unicode strings or something like that. Yeah, so you should avoid this. You should try to, if you have strings in your code, yeah, try to get to the point that, yeah, you can confidently say that, or know the locations where it's Unicode and where it is not. Okay. Sometimes it can get, yeah, a bit hard. For example, in FTPo2, yeah, maybe I should show something. This is code you can write with FTPo2. It's mostly sitting on top of, yeah, encapsulates FTP lib, FTP, but it's more high level. You can write, the intention is that you can write code as if you were using the OS module or SHU till module, OS part. You can even, you can do, you can use walk and list on these host objects. You support the with statement is supported. You can use this file, or you can end up with these convenience methods like download, for example. And you can also open remote files, yeah, and read from remote files, for example, or write to remote files. Okay. And for, yeah, since the, yeah, one difficulty with FTPo2 was that I don't only have to convert between bytes and Unicodes for file contents, for the remote files, but the harder part was dealing with the encoding of file names or directory names when I sent them over socket. And I also checked a bit, yeah, these are some sections. So what you see on the next slide is this upper right corner. And on the slide after that, you see this part. So for example, I made this diagram and attached nodes and, yeah, how these interfaces behave on both Python 2 and 3 to, to, to wrap my mind around this. And, yeah, I get this straight how I should deal with this. This is the other part where I'm, since I'm using FTPlib, how does FTPlib handle this in Python 2 and Python 3? Where are the conversions? Because the string arguments for files that FTPlib uses, it also requires Unicode. So I was wondering how, what, how do they encode this or how do they know the encoding or something, yeah, when I want to finally send this to the FTP server, this file name. So it really depends on your project how complicated that is. Sometimes maybe it's straightforward, then you are kind of lucky. I mean, you are even more lucky if you can keep your API and only make internal changes. But especially if you need to change the API, if you can't come up with a clean API with that is, yeah, expected by the user or, yeah, that a user can easily work with in Python 2 and 3, now then you need to think harder. Okay. And that's really the part that can't be automated. Some more tips, don't let functions or methods accept both Unicode or byte strings. So something like, yeah, if I get a byte string, I convert this to Unicode or if I get a Unicode string, I convert this to a byte string or something because this makes the API confusing and you always have to think about, yeah, is this now, under which conditions does a byte string or Unicode string and it also makes the tests more complicated because you always have to check both, this have to check both for Unicode and byte strings for the arguments and imagine you have something that takes two or three strings and you want to test all combinations maybe even, yeah. So you should try to avoid this. Yeah, special case, file-like objects or strings for paths because in both in Python 2 and 3, you can use all these APIs that accept file names or directory names with either byte strings or Unicode strings and it will, yeah, under the hood will call different APIs on the operating system. So and this also gave me some headaches with FTP because I do this as well and so in this case I have to accept both Unicode or bytes and do different things because I try to mimic these APIs from, for example, the OS module. Okay, if you, yeah, but if you don't accept the strings but let, accept file-like objects, so this is handled before so it's not part of your library anymore. So if you can get away with this, you probably should use file-like objects instead of accepting file name strings. I mean, you can of course still have this for user convenience but this makes it a bit harder. So, yeah, also avoid different APIs for Python 2 and 3. For example, for FTPutile, I was thinking at some point for backwards compatibility when you run on Python 3 and I open a text file for reading, it should give you, as before, to be backwards compatible, give you byte strings and Python 3 give you Unicode strings because you expect this under Python 3. But this would really be a mess, yeah. I made up my mind and wrote summaries on the advantages and disadvantages and posted this to Comlang Python. What do you think? How should I deal with this? And I got two answers, both saying explicitly go for the unified API even if it breaks backwards compatibility. Okay, and I think this was a very good decision to make. Yeah. Also make a list of changes before actually changing the API because, yeah, this will, yeah, hopefully make sure that you don't forget to change, you know, that you still maybe need to change different parts of your code and, yeah, don't want to forget something. And it also helps you write some release notes, for example, for FTPutile, I wrote what's new in FTPutile 3.0. So I could check this list and make sure I don't forget anything. You could also use commit messages, but commit messages are usually more fine grained and maybe not so useful for this purpose. Yeah, and another tip. Yeah, if you need to change the API, increase the major version number the first part and, yeah, then, yeah, I mean, it's after all adding Python 3.0 support is a major change. And so you can get away with the API changes. So I think it's justified. I mean, it's not just a trick, but I think it's fine. Okay, some other tips in general on this Python 3.0 adaption, read what's new in Python 3.0. At least I actually recommend you also, I mean, not instead, but also read this porting to Python 3. I have links at the end of the slides and I will put the slides online. But even if you just search on the net for porting to Python 3.0, you will probably get this website. This is practically an online book, which is really nice. Yeah, if possible, support only Python 2.6 and up because Python 2.0 and Python, Python 2.6 and Python 2.7 have some very useful Python 3.0 features backported. For example, you can say from future import print function and you will have print as a function with the same behavior as in Python 3. Also, the exception syntax, so starting from Python 2.6, you can write accept exception class as exception object. You can't do this in Python 2.5. And if you need the exception object for exception handling and want to support Python 2.5 and lower versions, it gets really messy and it's not really nice. It isn't fun, I guess. Also Python 2.6 and 2.7 have this I.O. module. So you can just, you can say if you want to from I.O. import open and then you have the like the build in open function from Python 3. But have it available in Python 2. If you need to support Python 2.5, you can use the 6 library and so, yeah, in summary, anything below 2.6 will probably be out what to support. Okay, thanks. Yeah, okay. There will still be some things that need to be different for Python 2.3. And I recommend, I mean, I'm not the only person recommending this. I saw on the net to use a compact module, for example, for Python, for FTP, it looks like this. You just say if Python 2, if the Python version is Python 2, by the way, use this indexing if you want to run your code on Python 2.6 because in Python 2.7, this is a name tuple and you can write dot major, but this doesn't work on Python 2.6. Okay, and here I have int types, a tuple of int types, unicode type, bytes type. This is much easier than reading the code and trying to remember, reiterate, yeah, I'm now on Python 2 and the SDR I'm seeing here is the bytes type or something. This is, yeah, error prone. Yeah, if you have a larger project, also have a look at the future or six libraries. I think the future library looks a bit more modern. It's actually newer than the six library. For FTP, I decided to not use it because FTP doesn't have any dependencies apart from the standard library. I didn't want to introduce dependency just for these few things you saw in my com.py. Okay, and to, yeah, for every Python file, I mean, that's at least what I suggest. These are some changes which make your Python 2 code behave more like Python 3 code. For example, the absolute imports are required. The float division or the integer division rather has changed from Python 2 to 3. For example, if you use from future import division, you get this Python 3 behavior for integer division. I already mentioned the print function. Unicode literals, when you do from future import, all the literal strings in your code will, even on a Python 2, will become Unicode string literals. Alternatively you can use, yeah, alternative to Unicode literals in Python 3.3 and up is the U prefix. This was removed in Python 3.0 but was reintroduced in 3.3. But you still have to know what string type your literals are. And, yeah, I mean, it's maybe a better of taste whether you generally use Unicode literals for the import or if you use the U prefix explicitly. Okay, then summary. So Python 2 is still in wider use but, yeah, I recommend using or developing for Python 3 if you can. Using the same source code to support Python 2 and 3 is feasible, makes sense. Also larger projects are doing this. Django, for example, have gone this way and have the same source code for Python 2 and 3. You need to know the concepts of Unicode bytes and encodings and the changes from Python 2 to 3. So, yeah, again, read what's new in Python 3.0 and porting to Python 3. You should have tests for adapting to Python 3. Otherwise, yeah, it's much more difficult. You should at least have some tests even if you have, yeah, kind of acceptance tests or these approval tests that Emily mentioned. Yeah, you should prefer APIs in Python 3 style, so write modern Python, plan and implement necessary API changes carefully. I mean, as you would design, I mean, if you designed your API for your library from scratch, maybe even, yeah, do what makes sense for Python 2 and 3. I already mentioned reading what's new in Python 3.0 and, yeah, if you can, if you can actually use, require at least Python 2.6 because this will give you several of these future imports and you don't, if you want to support Python 2.5, for example, yeah, you have to write some convoluted code, maybe. Yeah? Okay, that's what I'm talking about. Thank you very much, Stefan Schwadze. So, if you have a question, please raise your hand and now we'll come by with a microphone. Well this is half question, half remark. I have a project where I also maintain both Python 2 and 3 at the same time with the same code base. But I noticed that while it is feasible and cool to be able to do that, you give something up. So you give some of the features in Python 3.0. So, for instance, I spent a long time struggling to find out if there's equivalence in Python 3.0 of the, admittedly, weird construction in Python 2 where you can re-raise so you catch an exception and you want to raise another exception with the same trace back as the original one. And you don't like to raise exception, comma. Can you get you the microphone? Okay, in Python 3, what you probably mean is you already have these changed exceptions and in Python 2, I think you can use an additional comma-separated argument to give a trace back object or something. But I feel in the end you have to do it kind of manually and setting these extra attributes. That's also one thing I thought about. Yeah. This one. Thank you for the talk. It was really interesting tips. But what about porting C extensions? C extensions like for Python. Okay, this is more complicated. Sorry, sorry. Okay. I haven't migrated or adapted any C extensions so far. So I can't, okay. From what I've heard, changing C extensions is much more complicated. It is. But there is a document that ships with Python. I think it's in the how-to's directory that gives you guidance on porting C extensions from Python 2 to Python 3. Any more questions? Let me check. Last question. Thank you for the talk. I think right now you're one of the most competent persons for the topic in the world. If you would re-convert your FTP utility library again starting from scratch with all the knowledge you have now. What do you measure it would take by means of person-taste to do it again? With my knowledge now, I can really, maybe a week or maybe less. I really don't know. I mean this was stretched out over several weeks because this is a free time project. To find all these things out you give me this nice presentation. I have this habit reading a lot of stuff before I start. So I did some research before. For example, this compact module is also mentioned by Armin Rohnacher in some blog post or something, this recommendation. But I think it makes sense anyway. So if I would say it takes several person-taste if you just mechanically apply all these. Some of the mechanical changes can be done by 2 to 3. This at least helps. I find it very useful to run 2 to 3. You should read what's new in Python 3 but running 2 to 3 on your code will also give you some insights or maybe some things you forgot when you read the what's new document. So it will change some things and you might wonder what did it change there? This seems to be something different in Python 3 in comparison to Python 2. Thank you. Okay. Thanks a lot, Stefan. Thank you.
|
Stefan Schwarzer - Support Python 2 and 3 with the same code Your library supports only Python 2, - but your users keep nagging you about Python 3 support? As Python 3 gets adopted more and more, users ask for Python 3 support in existing libraries for Python 2. This talk mentions some approaches for giving users a Python 3 version, but will quickly focus on using the very same code for a Python 2 and a Python 3 version. This is much easier if you require Python 2.6 and up, and yet a bit easier if you require Python 3.3 as the minimum Python 3 version. The talk discusses main problems when supporting Python 3 (some are easily solved): * `print` is a function. * More Python APIs return iterators that used to return lists. * There's now a clear distinction between bytes and unicode (text) strings. * Files are opened as text by default, requiring an encoding to apply on reading and writing. The talk also explains some best practices: * Start with a good automatic test coverage. * Deal with many automatic conversions with a one-time 2to3 run. * Think about how your library should handle bytes and unicode strings. (Rule of thumb: Decode bytes as early as possible; encode unicode text as late as possible.) * Should you break compatibility with your existing Python 2 API? (Yes, if there's no other way to design a sane API for Python 2 and 3. If you do it, raise the first part of the version number.) * Try to keep code that's different for Python 2 and 3 minimal. Put code that needs to be different for Python 2 and 3 into a `compat` module. Or use third-party libraries like `six` or `future`. Finally, the talk will mention some helpful resources on the web.
|
10.5446/20031 (DOI)
|
Hi, everyone. So thanks for the introduction. So this talk is called Red Hat, this Unicode character and Python. I guess you now know what that Unicode character is. Do you all know what that Unicode character is? Can you check? Is this? So this talk is called Red Hat Lost Python. And so during this talk, I'd like to basically tell you about two important things about Red Hat and Python. I'd like to tell you where Red Hat uses Python, until you learn that we are really heavy users of Python for all kinds of tasks. And what is really tightly connected to that is how you, as Python developers, can use the upstream projects that Red Hat contributes to, and how you can use Red Hat-supported products for your work, for your development, deployment, and so forth. So just before I start talking about the communities and Red Hat products, let me just briefly explain how Red Hat works in case you don't know that. So we have this kind of motto that says upstream first. That means that we just collaborate with communities, all the features, all the bug fixes, go to upstreams first. So we send all these patches. We do planning with the communities. We propose new features to communities. We make an agreement with them, and then we send the patches, basically. If we find some bugs or some security issues, then, again, we first send them to upstreams. And so we like to make the world a better place like this. And so at certain points at time, we just take upstream project and basically productize them downstream. What that means, that we do some additional quality assurance. We do integration testing. We integrate different products or projects together so that we make sure that they really work well. And if we find a bug or we want to add a new feature, then it all repeats. We go to upstream. We send them patches. We discuss with them, and so on and so forth. So starting with what actually made Red Hat what it is, like everyone knows Red Hat for Red Hat Enterprise Linux. So who doesn't know Red Hat Enterprise Linux? OK, so you all know Red Hat Enterprise Linux. That's great. Who knows Fedora? OK, so that's almost everyone. So Fedora is sort of abstract. So Fedora is sort of upstream for Red Hat Enterprise Linux. That means that all the development takes place in Fedora. And at certain point at time, we just take Fedora, branch it downstream, and we do some additional Q&A, and et cetera, et cetera. And we create Red Hat Enterprise Linux out of it. So as I've said, we are heavy users of Python. So in Fedora, for example, we have two parallel Python stacks. We have Python 2 stack and Python 3 stack. Currently in the supported releases of Fedora, that's 2, 7, and 3, 3. And so let me just briefly skim through what is written in Python in Fedora. So we have Anaconda, which is the system installer. We have Yam written in Python, which is the package manager. We have two build system, Koji and Koper, both written entirely in Python. And as they're building back end for RPMs, they use Mock, which is also written in Python. This is basically, Mock is basically a sort of change route in its own way. And the whole Fedora community, like it uses Python for pretty much the whole infrastructure. So like if Python disappears tomorrow, there is basically no Fedora. You can't install it. You can't install packages. We can't build packages. We can't live without Python, really. And I'm proud that I'm a Pythonist, and that I can say that my distribution can't live without Python. It's so important. And so one of our plans, obviously, is to make Python 3 a default. Not obviously, but it is a plan. I obviously honestly think that Python 3 is better as a language, but we can keep that to some corridor discussion, something. So hopefully we'll be switching to Python 3 as a default in Fedora 22. So Fedora is a rapidly moving Linux distributions. We basically make releases every half a year. It goes forward. It's really, everything is really rapid. And basically, Red Hat Enterprise Linux that is made out of Fedora is quite the opposite. It's really slow. It has very long support cycle. It's very stable. It's very secure, which can be a good thing if you have an application that you need to run for, I don't know, 10 or 15 years. But it's not that optimal if you want to move forward. And we've also come up with a mechanism to allow people to run this super stable system, but to also go forward, to follow upstream, provide new versions of Python and databases, I will be talking about it. So we currently have three supported releases of Red Hat Enterprise Linux, shortly REL. It's five, six, and seven. They have these Python versions. And maybe sometime in the future we will also release REL 8. And so I guess you can extrapolate from these years like when that might be. But you can't really extrapolate the Python version. Right? I personally am sincerely, honestly, hoping that it will be Python 3. But there are lots of stakeholders, big companies, big players. We'll see about that. Who would like to see Python 3 in REL 8? A applause for you. I like that. Thank you. So a minute ago I said that REL is a really slowly moving target. It's very stable. It basically doesn't change versions. Once we place Python 2.6 in REL 6, we will never update that. We'll keep with Python 2.6 forever, basically, as long as REL 6 lives. So what we came up with is a technology that's called Software Collections. And we're building a few of our products on top of this technology. Software Collections are basically an RPM way of providing multiple versions of basically any type of software on top of RPM-based distribution. That includes Fedora. That includes Red Hat Enterprise Linux and CentOS. I'll be talking about CentOS in a few minutes if you don't know what that is. So Software Collections are a general RPM-based technology. We have upstream for them. That is called softwarecollections.org. And usually if I talk to Pythonists and they ask me, so what software collections are, I usually say it's a system-wide virtual NF based on RPM. That's pretty much it. These are just packages that install somewhere under slash opt. And they have new versions of not only Python extension packages, but actually the interpreter itself. So we can provide, in this way, you can have Python 3.3 on top of Red Hat Enterprise Linux 5 or 6 or something like that. And the way you use Software Collections is basically that you run sl-enable-python, just the name of the collection bash, instead of source, then activate. So that's pretty much it about Software Collections. So for Red Hat Enterprise Linux, we have a product that is called Red Hat Software Collections. We first released it last year. And it basically brings fresh versions of various useful developer and system-in-tools on top of RL6 and now also on top of RL7. So what's most interesting here is that we have Python 2.7 and Python 3.3. So because I'm a RL Python maintainer and people from community have been coming to me and saying Red Hat is the only thing that is really preventing us from moving to Python 3. And I was always like, sorry, I can't do anything about it. Well, now I can and I did it and it works great. So Red Hat Software Collections are basically a product that is installable on top of your system. It doesn't replace your base system versions. Like if you have RL6, you will still have your Python 2.6 that is in the system. And on top of that, you can get Python 2.7 or Python 3.3. And you can also have one of these other components that are listed on the slide. And the good thing about it is that it just works. It just does. So Red Hat Software Collections are obviously a product for Red Hat Enterprise Linux. And they are designed in a way that they move faster than the system itself. Now, if you, for example, take Fedora, which itself is moving very fast, then for Fedora, it perhaps makes sense to create communities of third collections that actually move slower than the system so that you can create more stable software on top of Fedora. So the community does build some software collections. Like I know they've been building Ruby collections for Fedora and stuff like that. So this is really general technology. You can build your own collections. And it just works. So another thing that Red Hat does is Cloud. I've been advised to say the word Cloud at least 10 times during the presentation. Because in the past, like two years ago, when I was giving presentations, I just said Cloud once and everyone was like, what? Did he say Cloud? Great. Now, these days, you have to say it at least 10 times so that people would actually listen. So it's not easy at all. But we have a lot of Cloud. We have OpenStack. Who knows OpenStack? OK, who doesn't know what OpenStack is? OK, so there are some people who don't know what OpenStack is. OpenStack is an infrastructure as a service type of Cloud. It's basically a huge upstream project or more likely a set of APIs that happen to have an implementation as a huge upstream open source project. And so as some people talk about it as the next Linux, there's so much contributors to this project. Red Hat is there, HP is there, Dell is there. Like lots of huge companies contributing to OpenStack. And all of them think that this is really the future of Cloud computing, of the infrastructure as a service type of Cloud computing. So Red Hat has been the number one contributors to the last two releases of OpenStack. And so we contribute a lot to that. And of course, we also take that downstream and productize that. And we have what we call Red Hat Enterprise Linux OpenStack platform, which is basically Red Hat Enterprise Linux with OpenStack packages. So basically, you can type yum install OpenStack, and it just installs OpenStack. And that's it. So this type of Cloud is not really useful for programmers. It's just like it provides virtual machines. OK, virtual machines are nice. But what the programmers really want is platform as a service cloud. We want, if you're creating a web-based application, we just want the cloud to set up the environment for us so that we wouldn't need to care. We don't want to care about databases. We don't want to care about deployment. We just want to code. And then do git push and let the cloud take care of everything else. So for that, there is another open source project that is called OpenShift. It's written in Ruby, sorry. So platform as a service cloud is really what I just explained. It basically provides some sort of environment for your application so that you can just push your source code there, and it would just run. OpenShift works in a way that there are like two important terms that one needs to understand to understand OpenShift. There are gears and there are cartridges. Gears are basically containers, not necessarily in the Docker sense of containers. But these are just isolated runtime units that provide you some processor cycles, some memory, some storage. And then you have cartridges, which are basically languages or services. So for example, in OpenShift Online, which is like an open red-hatch provided instance of OpenShift, you have cartridges for Python 2.6, Python 2.7, and Python 3.3. So the way it works is that you register your application. You can even do that for free on openshift.com. So you register your application and you'll get two or three gears. And you say, I want this gear to contain Python 3.3 and I want this second gear to contain MongoDB. So the cloud will create this for you, and then you just push your application there, and it just works. So OpenShift really does this in a very good way, and it also has a prepared environment for a Django-based application. You can actually use any web framework in OpenShift, but the Django type of applications is supported out of the box. You don't have to do any additional setup by yourself. It just works. So we have what we call OpenShift Online, which is what you can see when you go to OpenShift.com. And if you actually want to run this in your data center, you can also get OpenShift by Red Hat, or it's also called OpenShift Enterprise. So you can basically deploy this in your data center and have your own platform as a service cloud. So how cool is that? Now, you can really combine all Red Hat technologies in basically any ways you can think of. So you can have OpenShift that uses Red Hat software collections for some of its cartridges. You can run it on OpenStack with transcend of the frail. And I guess what I want to say here is that nothing of this would be doable without Python. It's basically Python is throughout the whole stack, no matter what you want to do. It's just there. It's everywhere. So I just said that OpenShift Online uses Red Hat software collections for some of its cartridges. So what that means is that you can reproduce the same environment that you have in cloud on your REL6 or REL7 system. Which is quite cool for developers because you can get REL6 or REL7 on your development machine. You can get Red Hat software collections, which you basically get for free with Red Hat Enterprise Linux. And you can just start coding, then push your code to the cloud, and everything just works. So doing this, you can get your applications running for free in cloud in a matter of a few hours, basically. And of course, that's not all. Like Red Hat uses internally and externally lots of other upstream Python-based projects or projects that rely heavily on Python. So for example, what we use, we use Beaker, which is used for hardware integration testing. We use SPALP, which is software repository management. This is like kind of you can create a server that basically will serve all your machines in your data center with YAM updates. So if you want to do security updates of all of your machines in your data center, you don't need to download the package 100 times. You just download it once to your PULP server, and it distributes the updates to all the machines. All written in Python. We are also a heavy contributor to Gluster FS, which is a distributed file system, which itself is not written in Python, but really uses Python for lots of its like utilities around the core itself. Fedora infrastructure, for example, is a heavy user of Ansible, which is an automation tool for CIS admins. Basically, you can say like, what's the CIS admins, basically, you can say like, I have a basically a recipe that describes how a system should be created and booted up, and you just pass it to Ansible, and it just does the stuff for you. And we also have quite a fresh project called Dev Assistant, which is like my pet project, so I just had to write it there, although it's like a bit smaller than the others. But it's supposed to be what Ansible is for CIS admins. Dev Assistant is supposed to be the same for developers. Also entirely written in Python, so basically you write some recipes, how, let's say, a project should be created, and you give it to someone else, and he can just create the project the way you created the recipe. So kind of like this. And I promised I will also speak about CentOS. So there has been some confusion about what CentOS actually is. Do people know what CentOS does or how it like comes to being, or do you know anything about CentOS? Who knows? Who doesn't? OK, so there are some people who don't. So CentOS basically is a community rebuilt of red data enterprise Linux sources. What that means is we have Fedora, which is moving fast forward, we have RHEL, which is very stable. And some people thought, OK, we need something that people don't have to pay for, but it's also stable, and it's like RHEL. So people from CentOS community just basically take RHEL sources and rebuild them and provide them for free. Just like, OK, so you can get RHEL for free. You don't get redhead support with that, but it can be good for testing or something like that. So and the way I like to talk about it is that it's a community platform to run community projects. So you can get it for free, and it's not moving forward as fast as Fedora. So people from this big projects like RDO, which is basically OpenStack packaging projects so that you can install OpenStack easily on downstream distributions, or GlastroFS like to use CentOS for their development because it's very stable and their environment is not changing so rapidly. So this is pretty much everything for me. I guess that the whole message that I'm trying to send here is that redhead is really grateful to people in communities that make all of this possible. So really, if there is an applause at the end of this presentation, and I hope it is, it really goes to you people who work in communities and make this all possible. So thank you. APPLAUSE I think we've got time just for maybe one question. If anyone wants to ask a question? No? OK, we'll start. OK, so if you think of anything, just approach me somewhere, just ask away. Thanks. APPLAUSE
|
Slavek Kabrda - Red Hat Loves Python Come learn about what Red Hat is doing with Python and the Python community, and how you can benefit from these efforts. Whether it is the new Python versions in Red Hat Enterprise Linux via the new Red Hat Software Collections, compatible Python cartridges in OpenShift Platform-as-a-Service (PaaS), or being the leading contributor to OpenStack, there's a lot going on at Red Hat. We're Pythonistas, too!
|
10.5446/20030 (DOI)
|
Thank you, and morning everyone. I'm going to be speaking about, well, I'm speaking on conversing with people living in poverty. And before we really get into the talk, I want to kind of bring your attention to some of the words used in the title of the talk. So I really just want to highlight a few things because words are important. And especially if you're trying to do good, you really have to guard against sort of preconceived notions or slightly lazy thinking. If you're trying to do good, you have to be careful not to kind of do bad. We also all come with sort of preconceived ideas and our own sort of cultural baggage. And it's really, there is no one else to own that and to accept it and work with it. It's only human. So conversing, this is really, if you're going to reach out to people living in poverty, did we just lose the mic? Maybe I just need to lean a bit closer. If you're going to be reaching out to people, you really want to make sure you're thinking of it as a conversation. This is not a one-way, we're educating people. It's not a one-way, we know what's right and we're telling poor people what to do. We're really learning from them as much as they're learning from us and it really needs to be a dialogue and we'll get to a bit more on that later. The next word I really want to highlight is people. It's very easy to become kind of obsessed with the differences between oneself and the people one is helping. And those differences are important. But at the same time, we sort of share a common humanity and it's important not to lose sight of that. I mean, these people that you're helping, they have lives, dreams, families, aspirations. And it's tempting to kind of think as sort of ourselves, as sort of middle-class people who are already kind of reaching out as in some sense being better. I mean, certainly we have more money, usually we're better educated, but I think it would be really a mistake to think of ourselves as better humans. And lastly, just the phrase living in poverty. I try to avoid speaking of people as poor, which makes it sound like being poor is some sort of innate condition. Living in poverty really highlights that poverty is a circumstance that people happen to define themselves in rather than something that they are. Cool. So just a quick introduction to me. And I work for a nonprofit called Precelt. We operate throughout Africa. We also have one employee in India and one in London. There are about 50 of us in total. I'm the lead engineer on Vumi, which is our messaging platform. There are currently three developers and recently we've hired one more, so there are four, I haven't actually met him yet because his first day was on Monday and I was here. So we're not a big team, but we're already trying to make a difference. So Vumi is a text messaging system, just so you kind of know what it is. So it's an engine for moving text messages around. Just like to tell people that we write IRC bots to help people. And I think really we're kind of on the cusp of seeing chat bots become a really major thing. Instant messaging networks have really taken off. If you see someone using their phone, there's a good chance that they're chatting to people over text and we haven't really seen chat bots taking off yet. So Vumi is a text messaging system. It's really designed to reach out to those living in poverty and really we're aiming to reach massive scale like whole countries with people because we really want to have a big impact. I like to think that we're trying to build infrastructure. So non-profits do cool projects, but which are in the end kind of really small. And at some point if you're really going to transform society, what you're doing has to become a kind of infrastructure. Our Python is our primary language that we write Vumi in and we use Twisted for pretty much everything. So this is the UN definition of poverty. So poverty is a lack of basic capacity to participate effectively in society. And I really like this definition because rather than highlighting a lack of money, it really highlights people's isolation. It's pretty hard for us today to imagine quite how disconnected people can be both from each other and from the society that they're meant to be a part of. We'll get to that a bit more in a minute. So this is Africa. This is where we do most of our work. The dots are places where we've done projects and where we have connectivity to mobile network operators. So places like Libya, Nigeria, Kenya, South Africa, where I'm from. So just some things about Africa. It's pretty big. It's 30 million square kilometers physically. That's roughly three times the size of Europe and three quarters of the size of Asia. There's a billion people. So about a third more than they are in Europe. Obviously they're a bit more spread out. But obviously a lot smaller than Asia, which is about four billion. And there's more than a thousand languages. So really if you want to work across Africa, I guess it's the same in Europe. You really have to take care of localization and internationalization. Even just in South Africa, we have 11 official languages. In practice, we only have two languages of record. But in theory, any citizen can ask the government to interact with them in their home language if it's one of the 11 official ones. So I'd like to kind of go back in time a bit to 1994, which wasn't all that long ago, 20 years. I was just finishing school. So I guess I was already an adult in many ways. In 1994, Nigeria had 100 million people. But only 100,000 landlines. So if you think about that, that's one telephone per thousand people. And probably most of those telephones were owned by kind of richer people. So you can imagine just 20 years ago, people living in rural Nigeria might never have made a phone call or probably never had made a phone call. And you can probably, if you try, imagine how isolating that is. And if you have, say, a service delivery problem, say your water isn't working, who do you tell? You don't have a phone. You probably don't even know. Even if you did have a phone, you wouldn't necessarily know who to call. The nearest government official might be, say, 500 kilometers away. And it's not like there's great public transport to get you there. So when things are going wrong, it's very hard to reach out and tell someone. And this is really, I mean, it really drives home. The problem is more isolation than lack of money. And the problem goes the other way, too. Imagine if you're an elected official in a country where your constituents have no way of getting hold of you and we're collecting information is hard. How can you serve the people that you've been elected to serve if just finding out anything about them is hard? And no surprise, there was no internet access in Nigeria in 1994. In fact, even in South Africa, it was, well, really pretty much unheard of. Let's, well, so one other thing. So 1994 was a very exciting year for one reason. The first mobile phone network launched in Africa. I should say that the most popular phone in 1994 weighed half a kilogram and cost about $2,000. Let's fast forward a bit some 18 years to 2012, so two years ago. Suddenly there's 65% mobile penetration on average in Africa. So roughly 65% of people suddenly have a telephone, which is quite an improvement from one in a thousand. In places like Uganda, we see incredible statistics like there are more mobile phones in Uganda than light bulbs. And they really actually are getting good use out of their mobile phones. UNICEF Uganda ran quite a famous project called UReport, which is, it is a citizen kind of sort of liquid democracy program where you can sign up as a member of UReport and you can submit feedback kind of directly to your government and they can ask you questions. And really what we're seeing now in Africa is a kind of mobile generation. So UNI might have laptops that we kind of carry around with us everywhere and that we run our lives on. And probably we also have a smartphone as a second device. The people, young people in Africa really are, their mobile phones are their laptops, they are their offices, they're how they socialize, they're how they do business. And by contrast to 1994, we're now seeing $20 phones which have internet access, they have SMS, they have USSD, they have instant messaging. Just to just economically, Africa is growing, the continent's averaging 5% GDP growth, which is pretty good growing. The population of Africa is still very young, more than 50% of Africans are less than 20. So at that point Precelt had been mostly a consulting company and the rise of kind of mobile phones was, well really we saw this as an opportunity to reach out to people. So we started doing small projects in the social space and one of the earliest of those was Text Alert. So Text Alert is really just a simple system to remind people of their clinic visits. So I'm sure you know about HIV, I'm not sure how many people know about tuberculosis TB. Well, so typically in Africa if you, or certainly in South Africa if you die from HIV what actually kills you is tuberculosis. And tuberculosis is a kind of a very bad chest infection is how it manifests. And tuberculosis is very curable, you need to take a course of antibiotics. But unfortunately for various reasons people don't finish their antibiotic courses, forget to go follow, to take follow-up visits or if they're being treated with anti-retrovirals for HIV they kind of forget to go to the clinics for checkups and things. And Text Alert was just a system which sent people a reminder kind of a few days before their visits saying hey you have an appointment at the clinic, if you can't make it just reply and let us know. And that dropped the number of people who were missing their clinic visits from about 35% to about 15%. So that was really a big success for us and made us want to do more. However, it wasn't all plain sailing and we learned a lot from projects like Text Alert. And one of the things that we learned was that if you're going to be doing a lot of projects it's important to have reusable software because otherwise you're rewriting things every time. And it's easy to accidentally make things not reusable. So with something like Text Alert you need to integrate with a mobile network operator or an aggregator who connects you to the cell phone network. And they all have special interfaces which are unique snowflakes. And it's easy to accidentally tie your application to something which is really network specific so they might give you a unique identifier and you start relying on that. And then you change to a different network operator who doesn't give you this identifier and suddenly you have to rearchitect your application. So we're struggling with reusability. The other issue was scaling. I'm sure those of you who have worked on small systems as they've grown have noticed this, you always get something wrong no matter how good your intentions are. And to really be able to kind of process things quickly you need to try. And if you get something wrong then you have to kind of go and rewrite something. And we didn't want to be kind of trying to scale every single project that we were involved with. And the last thing is tooling, kind of if you write a quick prototype you usually leave out things which become really important later like monitoring, good error reporting, good failure handling. I should maybe say some things about failure. So one of the exciting things about operating in Africa is that you have failure conditions often aren't the exception, they're often the rule. So for example we run some projects in Ethiopia. There's one ISP in Ethiopia, it's state owned, there's no competition. It occasionally goes down for a week and you can't connect to the country. So as you can imagine that makes designing systems to handle that, well designing systems to handle that can be tricky. So in response to these challenges that we encountered during text alert we created Vumi. And Vumi is a messaging engine which attempts to provide a reusable framework to separate kind of the social applications that we're trying to build from the connectivity to mobile operators. And to kind of give us the tooling and kind of production readiness that on all of our projects without having to rewrite it. Vumi is a Swahili word, it means something like distant roar or buzz or hum or roll of thunder, it's kind of a distant noise. So architecturally this is kind of what Vumi looks like. If you look on the left you can see a cell phone, so you can imagine someone kind of holding a cell phone in their hand. If they send us an SMS that goes to a cell phone tower, eventually that comes to one of our servers. You can see that labels transport, so transport is what we refer to, it's a twisted process that sends and receives messages to a network operator or to an instant messaging service. The next column is another set of processes which we call dispatchers. They're routers, they decide where messages go, usually based on the telephone number they're being sent or received from. And then lastly on the right we have really the important part, the left hand side is mostly plumbing. On the right we have applications, so these are ideas that these are reusable things that can be plugged into different transports and different dispatchers. So the way that things are architected we use kind of horizontally scalable workers, so we write workers using Twisted, the asynchronous event framework, and then if we need to handle more messages we fire up more processes. And I really want to say thank you to all of the people who've worked on software like Twisted that we use and Python. Using the infrastructure to build things on really helps. All of our messaging between these horizontally scalable processes happens over RabbitMQ, which is a messaging bus, so it just sends messages backwards and forwards and workers can subscribe to receive messages which they process. If workers fail to process messages they go back onto the queue and get reprocessed later. For data storage we use React, which is a distributed key value store. We chose React rather than Postgres because we do want to reach massive scale. So at the moment we have kind of say, as I said earlier about, I think about seven, we have about 7 million people who we've interacted with so far and Postgres would probably still have been fine for that, but we really would like to be able to reach the point where as I said we can speak to billions of people. We maintain Reactisaurus, which is the Twisted React client. It's pretty much a direct port of Bashos React clients. So if anyone wants something fun to work on we'd really appreciate some help maintaining that. So some of the things that we built with Vumi, in Kenya we did a project called CCNMani, which was an attempt to curb election violence in Kenya. In Kenya, in not the most recent election but election before, there was a lot of violence in townships and there was the impression that this violence was triggered by mobile phones were enabling this violence. So what would happen was someone's house would get burnt down. A few local people would decide that it was say the members of another political party's fault. And then they would estimate that they would be angry obviously. Someone's house had just been burnt down, I think they'd all be angry. And so then they would SMS their friends and say it was these people's fault. And then more people would be angry and violence would break out. So CCNMani was an attempt to counteract this by introducing sort of peace officers, also with mobile phones. And the job of these peace officers would be that if they received such a message or they had the impression that violence was flaring up, they would describe the situation to people at the NGO's head office who would then attempt to carefully craft some sort of response, which would also then be disseminated by SMS through the kind of peace officer process well, via the peace officers. So as you can imagine, you need to respond quickly. I mean, you're talking about kind of violence breaking out on a time scale of hours. And ideally, you want people to kind of calm down and think a bit about things more on the scale of minutes. So CCNMani was a system we built to allow that feedback to reach the NGO and for messages to be sent back out again afterwards. And the last Kenyan election had less violence. It's a bit difficult to have a control to measure against, but we also do, we're part of the Wikipedia zero project. So Wikipedia zero is zero rated, well, the main Wikipedia zero project is zero rated access. So that's free access to Wikipedia over internet on mobile phones. We do Wikipedia text, which is accessing Wikipedia over SMS and instant messaging. And that's to just kind of make things, well, lower the barrier to entry given further. So speaking of lowering barriers to entry, after we had created Vumi, there were still some problems. And the one was that just the biggest problem was just that there were too many projects. And setting up, and we really hoped initially that Vumi would be a tool that other nonprofits could use. But it turns out that there are difficulties. So one, connecting to network operators is expensive. Running Amazon EC2 instances is expensive. And really just we don't have enough people to solve all of the world problems ourselves. Not a surprise, I guess. So this led us to make Vumi Go, which is a hosted instance of Vumi. And the idea is that by providing Vumi Go, we can provide a way for people to help themselves. So we run Amazon EC2 instances. We deal with connecting to the network operators. And we provide people with a hosted service where they can come along and build their own applications to fulfill their own needs. And usually they know better than we do what those needs are. So for example, Vumi Go has a simple survey builder, which we call a dialogue application. Again to remind people that you're not asking a bunch of questions and getting some anonymous feedback. We want people to really think about this as a dialogue between themselves and the people that are kind of speaking to them over their mobile phones. And then we also wrote a JavaScript sandbox so that a lot of young Africans can code or very excited about coding are technically savvy. And we really wanted to provide a way for an excited young African who's one of the many African innovation hubs to be able to come to Vumi Go, write their application in JavaScript and run it. I'm sorry it's JavaScript, but that's what most people know. I'm hoping to build a Python one when I get a moment. So we're kind of reaching the end of time, but just a quick kind of where we are now. So Vumi Go has now interacted in the last year with a total of 7 million people, up from 1 million at the start of the year. So that's seven times as many people as we had before in the last six months, which is good. We sent and received 14 million messages, that's up from 12 million previously. We also registered voters in Libya. We registered 16% of the Libyan population of SMS and USSD to vote. So that's 16% of the total population, which I was very proud that we could be involved with that. In South Africa we ran election monitoring. What is interesting about the South African election monitoring campaign is that we really, the first time we were really running a big project over lots and lots of different channels of interaction. So that used SMS, USSD, Twitter and Mixit. For those of you who don't know what Mixit is, it's a big internet messaging network in South Africa. I think it has about 50 million users. In Nigeria we ran an agricultural awareness campaign, just making people aware of the importance of agriculture. People could download a ringtone from a famous Nigerian musician and that reached out to 1 million people. So what next? We're adding lots of APIs, again lowering barriers to entry, making the system easier for people to use. Things like, obviously the technology space isn't static, so things are constantly changing. We're seeing instant messaging move to multimedia messaging and if you're using WhatsApp, you can send photos and videos, send voice recordings. We're also actually moving into voice itself. There are many literate people in Africa and if you're going to reach out to them, you need to do so by voice. And lastly, we're trying to get better dashboarding and analysis so that we can be sure we're actually having the impact we intended to. We have some bigger plans, which we really like help with. We'd like to build an African content distribution network because there currently isn't one and 350 millisecond minimum latency sucks. Next we'd really like to build a federated instant messaging protocol. So think WhatsApp but structured like email so that we aren't all tied into a single provider. Just in closing, something that Constance said during her keynote, which really, she was speaking about it in a security context, but I think it applies equally to the kind of social space. Really show that we care, don't accept that things have to be the way they are and work to change them. Thank you. Thank you.
|
Simon Cross - Conversing with people living in poverty Vumi is a text messaging system designed to reach out to those in poverty on a massive scale via their mobile phones. It's written in Python using Twisted. This talk is about how and why we built it and how you can join us in making the world a better place. ----- 43% of the world's population live on less than €1.5 per day. The United Nations defines poverty as a "lack of basic capacity to participate effectively in society". While we often think of the poor as lacking primarily food and shelter, the UN definition highlights their isolation. They have the least access to society's knowledge and services and the most difficulty making themselves and their needs heard in our democracies. While smart phones and an exploding ability to collect and process information are transforming our access to knowledge and the way we organize and participate in our societies, those living in poverty have largely been left out. This has to change. Basic mobile phones present an opportunity to effect this change. Only three countries in the world have fewer than 65 mobile phones per 100 people. The majority of these phones are not Android or iPhones, but they do nevertheless provide a means of communication -- via voice calls, SMSes, USSD and instant messaging. By comparison, 25 countries have less than 5% internet penetration. Vumi is an open source text messaging system designed to reach out to those in poverty on a massive scale via their mobile phones. It's written in Python using Twisted. Vumi is already used to: This talk will cover: * a brief overview of mobile networking and cellphone use in Africa * why we built Vumi * the challenges of operating in unreliable environments * an overview of Vumi's features and architecture * how you can help! Vumi features some cutting edge design choices: * horizontally scalable Twisted processes communicating using RabbitMQ. * declarative data models backed by Riak. * sharing common data models between Django and Twisted. * sandboxing hosted Javascript code from Python.
|
10.5446/20029 (DOI)
|
As a lightning replacement, you will now hear Shlomo Shapiro on open source. He is a systems architect and open source enthusiast working at M.O. Scout. So give him a warm welcome please. Oh, sorry. I'll fix the screen mirroring. Please excuse the wrong format. I did not prepare this. Okay. And the slides are a bit in German because I gave this talk last year at the Linux talk, Linux tag in Berlin. But it's about the content not about the slides. I'm an open source evangelist at M.O. Scout 24, which is a German real estate listing portal. And I'm there for now about five years. And I think it's a bit interesting to tell how to introduce open source in a company also on the enterprise level. And also how to really make a company benefit from open source projects and from investing into open source. And my personal point of view has always been that open source, it's not just about enthusiasm. It's actually about solid business decisions. And my mission is to combine business, mentality, business decisions with open source and open knowledge. Few words about M.O.M.E. in Scout. We have two data centers, about 1600 virtual machines. The company is more than 15 years in business, so we are not a startup. And our entire technology stack is based on open source solutions. Mostly Linux, of course, but also a lot of Java stuff, which is almost exclusively open source. We have a lot of people. There are about 200 people working in IT in 30 cross-functional teams. So we have a lot of changes going on. And we are actually big enough that we have internal open source projects, which is also an interesting thing how to take open source methodology into your company to get things done internally. Well, open source sometimes has gurus or rabbis or, you know, the big people of open source with long beards. But in truth, open source is a lot about money. And if you look at the open source companies today, a lot of them are actually about money. For example, Red Hat. Red Hat is one of the largest open source companies, and they take a big pride in the fact that they make money based on open source. And I know a lot of other companies, a lot of them also attending this conference, who are also actually making money out of open source. And I think this is not shameful. I think this is a good thing, because in the end somebody has to pay the salaries of the people doing the work. And to pay salaries, you need money. And, you know, enthusiasm alone doesn't feed your family. So that's why I think in the long term it's a worthwhile thing to do. So why Linux and why open source? I guess you know this idea. Linux is just a big toolkit. And the open source world is the bigger toolkit surrounding your little toolkit. And everybody who likes to tinker or play Lego or build stuff loves toolkits and tools and components. And the reason for people to invest into open source is actually to invest into your own tools and into creating your own perfect, optimized, optimum toolkit which drives your business much better than anything else you could buy. And that's the reason why your company should be investing into open source. And that's how you can easily convince any boss who thinks that's a silly idea. Look for example at a house. When you build a house, when you want a house, you want something like that. A nice house, a castle, whatever. Something with a few windows which looks good and so on. But then you go to the commercial companies, to the proprietary solutions and that's what you get. Looks nice on the outside like that. But then a night comes and that's what you get. You wake up at night and you realize well, there's a monster in the basement. Because there's no operability. They didn't think about updates. They never test updates. With each new feature, the old features stop to work and so on. That's what happens with so many commercial applications that we run at our site that a lot of managers started to think, okay, what would be the alternatives? And the alternatives, just the way how to deal with that is a bit different. Because open source usually looks like that. It's a beautiful house. It has a great design. It's just not finished. And the nice thing about open source is that you can finish it yourself or you can pay somebody to finish it. And that's the fundamental difference between proprietary software and open source. And the main challenge when you talk about open source in companies is actually making this connection from let's invest into that and let's gain some business value from that investment. And that's the biggest challenge because then they're no like huge companies like Oracle who will send you sales droids on Mars and be happy to take your millions of dollars or euros. They're just individual people there and you have to deal with them. And that's something which takes work and it pays to understand how the open source system works really differently than the commercial software system. We've done this many times and we also try to talk with commercial vendors, for example Red Hat, about turning bugs into features or not seeing bugs as features. Because what happens a lot is that we submit a bug request and they say, well, that's a feature. And if you look at the Red Hat bug tracker, you see a few of our bug requests which, well, they are features. So the thing is how do you get a commercial company, a commercial partner to take you serious? And the sad truth is only with money and even more sad truth is with much more money than you could ever dream about paying. Just to give you a few figures, I know from personal experience that Linux distributions start to build customer solutions, meaning stuff which the customer actually needs, only after you are ready to put up several million euro per year. So if you're big enough to pay that much, then you can use proprietary software and adjust it to your own needs. If your budget is not that large, then you better look into open source sponsoring. And why is that interesting? Because open source sponsoring means you pay money into your own organization. You give money away to other people, but you actually invest into your own organization. And the reason is that with open source, you invest into knowledge and into features. And you don't invest into property of licenses or into having some license paper which you can paste on the wall. So any euro you spend on open source sponsoring goes either into consulting or into fixing errors. By the way, there are two types of open source companies. One, they sell you an open source core which is nice and useless. And then they take license fees for the extra features. And the others, they give you everything for free and they take money for consulting and for fixing problems. I personally prefer the latter one because they don't feel so much pressured into buying their ad on products. But I can also understand the companies who choose to refinance the open source development through selling licenses. A famous example for that model, by the way, is Opsi, an open source deployment and automation tool for Windows desktops. It has an open source core, it has commercial ad on features. And as soon as they refinance the development of these commercial ad on features, they turn them into open source. So I mentioned it, people. Here are a lot of people. This is an open source conference, please partially. And it's all about people. And the main thing which you need to understand when dealing with open source is you're dealing not with projects, you're not dealing with companies, you're dealing with individual people. And individual people are individual and they need individual treatment. And you have to really understand the people behind the projects if you want to interact with them or if you want to influence how the project will be developed. And as soon as you manage to have somebody in your company who sees it as their job to deal with the people doing open source, that's the moment you will be successful as a company in utilizing open source towards your own ends. So what do you have to deal with? Mailing lists. Avoid the flame wars. They're just a waste of time. A much bigger problem is some of the people doing open source actually do it really just for fun. Hobbyists. And that's actually a really tough problem for people like me because I've had it already. There's a great project which I would like to use and which just lets you know this little tiny bit of polishing or extra feature or whatever. I write an email to this person and I don't get a reply. I write more emails, eventually I get a reply and the reply is, I'm not interested. Like here I'm waving money but the people are not interested. That's a big problem. It sounds crazy maybe but it exists and then if that happens you are a little bit out of options because then you can't use the main competence on that project, the author. You have to find somebody else internally or externally. You can take a freelancer, ask them to work on the code and so on. It makes it a bit more difficult but thanks to open source you can actually do that because you can take the code. You can fork the project. You can take a freelancer, put him on it, pay some money, get the thing solved. And if you're a business what you care about is solving problems through money. That's the core of all business decisions. And with open source that works quite well especially if you have open source orientated companies. And I want to show you a few examples of successful corporations which we as Immobilian Scout had been doing recently. Just to show you that it is possible to find open source projects which are really good, really important, which are backed by individuals or by companies who are also willing to work with us as a company and to support their product exactly as we needed. One thing you have to care about is the legally stuff and the legal pitfalls. Because especially in Germany there are different kind of contracts which you can make. And depending on the type of contract you pay either for the time of the person or you pay for the actual artifact being produced. And when you pay for the artifact being produced then there's a big pitfall of warranty. And that's why many open source developers they don't want to be paid for a feature, they want to be paid for a time. And as a company you then have to go with them and say okay I'll pay you two days worth of development for doing support. You know on the contract you write support. And then you actually ask them to write a feature and then it's all okay from the legal side. And I also was doing this a lot as a consultant and I always told my customers I'm supporting you and as part of that support there will be magically a new release of my product on GitHub which you can then download following the disclaimers in the GPL so that there's no personal warranty involved. It's a bit Germany legal stuff you have to know about it. Then of course your boss will be afraid. People are afraid. People are afraid of the unknown. And how do you overcome fear? You make a small step. Let's say you find a small project where you can take 100 euro or 500 euro and get a big improvement. Something simple. Hey there's no spec file or no Debian directory for packaging. You pay the developer 100 euro he'll add one. To make a first step to show your boss that okay we paid money we got a result worked out well was great. And then you go from small steps to bigger steps. Don't start open source like don't start open source sponsoring from a big project. It will fail because you as the sponsor in your company will not have the experience to do it right. And then there will be problems and then the whole idea of open source sponsoring will be spoiled from starting on a big thing. Start small grow with the experience. So what do you need trust and that's again back to the people thing. I actually met with a few of the open source people who work on projects which we're using just in order to get to know them that they won't see my face that when we exchange emails. It's not just somebody anonymous but it's somebody they've met before. And trust builds bridges and when you have a bridge of trust then you can actually go further and do also bigger projects. That's the stuff which is actually a very good point for open source sponsoring because nobody likes to do it. No developer likes to write documentation. Nobody likes to work on tests which don't produce features unless you're a test driven friend like me. So when we do open source sponsoring all our sponsoring contracts have these things written into them develop write tests documented and create an upstream release. And we pay actually for the upstream release we don't pay for custom code. Here's your target that which you can install. That's not worth anything because custom code will not be maintained what you want is an upstream release that will be maintained by the original author. As he continues to develop his project. And these are actually valid points which you can use to sell open source sponsoring within your company because obviously if you have an upstream release that has been tested through test automation and is well documented. Then you will be spending less internal effort in integrating that in your platform. So you save money by spending money at the development side. You spend internal effort or maybe consulting money which you would have to spend. So that's a worthwhile investment. A few examples from my personal past open VPN gateway builder was the first open source project which I launched commercially. And there was a customer who came to me and said I need some handmade custom VPN configuration between two computers. And I said well no problem. How will you maintain that. He said I don't know we just switch it on and it runs forever. Which as we know is not a good plan. So I came up with an idea to create a build system that would generate bootable CDs that you can just pop into the VPN endpoints and then maintaining it is just generating a new CD. He thought the idea is cool sponsored the project and if you Google for it you can still find it even though it's not maintained anymore. Relax and recover was another open source project which is very successful. It is now the de facto standard for Linux disaster recovery automated. And it started also small as a small project where I offered my customer well don't buy 30,000 euro commercial disaster recovery tools. Hire me for about half of that and I'll write you an open source tool that does the same job more automated than the commercial tool. So I could do the job cheaper and better than the proprietary alternative. And again an open source project which is still alive today. And the project has since been extended and advanced through many many more consulting projects at many customers. And they're all very happy because they pay like a few days of development and they get a fully featured tool supporting their personal proprietary backup solution. So if you care about disaster recovery check it out. The prices are really the interesting thing here because all of these were cheaper than any other alternative that the customer had. And that's the strength of open source that the initial cost of development can be sometimes even cheaper than alternatives. And then of course the cost of further development even if it's very special for one customer is usually still cheaper than other alternatives. And the way how to market that to a customer or to your boss if you're working in a company is basically the trick which you need to do to get open source sponsoring on the market. A few things we've done at Immobilien Scout. Icinga probably everybody knows. Hands up who knows Icinga. Okay. Well it's a clone of Nageos the standard monitoring tool. Icinga has been forked already several years ago and is maintained by many people around the world. But a German company holds a significant part of Icinga developers which made it really convenient for us because we could just talk to that German company called Netways. Some people know them and ask them to implement a few features to fix a few bucks. For example, reloading the service took ages. So we paid them some money and they redesigned the reloading code internally and then it's now done in a few minutes. So in the subversion probably everybody knows. Again, there's a company actually here in Berlin who hold a couple of subversion developers. They also hold a yearly subversion conference here in Berlin. And we had a few compatibility problems with subversion 1.7. So we asked them to fix it. They fixed it. It cost us a thousand euro and everybody was happy. And that actually was a major blocker in rolling out subversion 1.7. And we could have spent days and days internally to try to get over that, but it was much cheaper for us to hire that company to fix the code upstream and create an upstream release than to deal with it ourselves. And another example is X2Go. X2Go is a Linux terminal server solution which allows you to create a terminal server and then access that through various clients running on Linux Windows and Mac OS. And we're using that in our data center to create a Bastion host. So you have to first go on this Bastion host and then you can work on all the platform. And again, a cool product developed by a lot of people around the globe, three developers in Germany and they're doing little fixing, little bug improvements, little features for us already for several years. Because we're using that on a daily base. And it's great if we can spend a little bit money and fix our big problems, which make this thing work much better. We also have our own open source projects. Our biggest open source project is yet an augmented deployment tool. It's our deployment chain, which is completely open sourced and it manages everything in our data center, how we roll out software, how we do configuration management. It's package based, so it's a little bit different from what you maybe know from other tools. And why did we do that? Because as a company, we learned that open source pays. As a company, we learned that investing into open source pays and that sharing what you're doing is actually a benefit because you get feedback, you get patches, you get bug reports, and that's the way how you can simply extend your internal development force with external help. And of course reviews and code reviews and questions and so on. So yes, if you manage to establish open source in a company, then the next step is to start internal open source projects and to take your own code and show it to the world and be part of the open source community. I'm at the end. I hope... applause I hope I managed to convince you a little bit more to use open source not only for fun, but also for business. I hope I was able to give you a few arguments to take home, and I hope we still have time for a few questions. Yes, thank you for stepping in for Kenneth, and yes, we will have some questions. I'm just... So as a developer, how would you try to fit some kind of custom requirement from a customer in your open source project if it's not really what it was designed for? Well, this is a good question. If you look at Relax and Recover, this project has been faced with that question many times, and our philosophy is anything that doesn't harm other people is most welcome. And the code is highly modularized so that it's really easy to implement something which will be run only in a very specific scenario. We say, please write it so that other people don't suffer from your peculiarities, and then you're welcome to put your code into our project so that when you deploy our software in your environment, you get out of the box a working solution and don't have to extend it with some local stuff. Okay, nice. Another question. If you are paying instead a developer who's not the author to work on a project, how do you still kind of try to make it go upstream if the original author is not interested in the patch that your guy is doing? Well, we actually, in the contracts, we pay for upstream releases. So in some cases, we split the contract into the functional part and into an extra added money for getting an upstream release so that we really pay for the work of the communication with the author and of convincing the author to accept this code. Okay. Because that is actually work. Yeah. All right. And there's very last thing. If you need something very quickly, it takes time to get something in upstream release, how would you do that? Well, if you take subversion as an example, it took less than a week from initial contact, contract, fixing the thing and creating an upstream release. But yes, the people working at this company, can I have my slides back, please? The people working there actually are part of the core team so they can just create a release if they want. Initially, they told us, oh, we have to talk with all the team and all the project has to agree that we make a release. In the end, it was no big deal. Okay. Yeah. Thank you. Okay. We have to finish. And there's a general announcement following. Thanks a lot. Thank you.
|
Schlomo Shapiro - Sponsoring Open Source und damit den Chef überzeugen
|
10.5446/20028 (DOI)
|
Hi, my name is Shlomo Shapiro and I'm working at Möbinska 24, which is Germany's leading real estate listing portal. So if you live in Germany, you probably already found a new home through our website. If not, come and check it out. We have lots to offer. But here I'm more to talk about DevOps and especially what happens if you already do DevOps for quite some time like we do and how to deal with the risks that probably are now different than in the times before you were doing DevOps. Let's start from a question. Who is doing DevOps here? Okay, very interesting. Not everybody. So I hope that this talk will help those who don't do DevOps to maybe get yet another argument for their bosses why it would be interesting to check out how to do things the DevOps way. Well, this is probably common wisdom. If you take software from planning through development testing and into production, then of course errors happen and need to be fixed. And the cost of fixing these errors, of course, changes. Fixing an error here is much cheaper than actually fixing it there. That's why it pays off to try to fix all errors early on in design. And those of you who run old software, old meaning older than 12 months, probably already thought about a redesign or were upset about the initial design. And an older company like ours, Immoblian Scout is now more than 15 years old, is running code that is partially also 15 years old. So a lot of the design decisions which we made early on are not valid anymore. So we suffer from that. One of the learnings we have is that we try to fix errors as early and not as late as possible. And DevOps doesn't make this easier because if you look at the development cycle of software in software development, that's how it looks at least in our company. You have a rather long time of planning and designing and user experience and wire frames and what else. After that you have a shorter development time. And after that you have test and in even shorter production time. This works quite well. And I think that this helps the developers actually to reduce design errors. In operations we also do software development and everybody who is doing operations is actually doing software development even if you don't call it that. The difference is that usually we have an idea over coffee and then we start hacking. And then we put it in production and call it testing. And then we run it and then we're often afraid to touch it because we know that if you touch it it's probably going to break. And I have a long history in operations so I know what I'm talking about. If you look at these two things in comparison, the first thing I would notice is that actually operations seem to be more risky than development because in operations we spend less time on planning and designing and we spend less time on the cheap fixing area so to say. And in production we much faster go into production and then fixing errors, especially fixing design errors is really costly. Why does that matter? Well, as a company you obviously don't earn money on broken stuff. So if you look at typical outages, at least typical in our kind of business which is running a website, you can ask yourself, okay, who's guilty for those outages? Who did the initial error? Who should have done something different? And we all know that the blame game doesn't help but in the end it helps to understand what to do different the next time you go there. And if you look at these, I'm not going to go into detail. I'm sure everybody can find themselves somewhere here. Me too included. I did almost all of them already. So I'm not ashamed to show them. So what about DevOps? DevOps in the nutshell is respect and learning in my opinion. And it goes both ways, both sides, developers and admins have to respect each other and have to learn from each other. And the devs, the developers can learn from ops a lot about operability, how to optimize software so that it's not only nice to develop but also nice to run it. And if you look at software development cycles, in some cases you develop it for let's say half a year and then you run it for 10 years. So why not optimize a little bit better for how to run it for 10 years and not only for how to program it? The admins of course also can learn a lot from the developers. For example, incremental improvement. Start small, do a minimum viable product and see how it develops. Improve it further on. Or coding instead of hacking. Let's not go into that. Test driven is a big thing from development which is already really established in development and in operations we're slowly learning how to do test driven. And actually this talk is also about how to do test driven in infrastructure development. A very nice thing is code quality. Who's a developer and cares about code quality? Who's an admin and cares about code quality? Okay. Why the difference? Why is code quality in operations different from code quality in development? It's all about craftsmanship, about writing code to be read later, write for reading, not write for it works. Don't do comments, do readable code, all this stuff is code quality. So that stuff which really works well in development and it works even better in operations. Actually I believe that the reason is that the stuff which we develop in operations is more complex than the stuff being developed in development. Because we instrument systems and landscapes of systems and very complex things that need to play together and that are often very difficult to test in a sandbox. That's why I think that the challenge of development in infrastructure is actually at least as high as the challenge of development in pure software development. And my favorite of course test automation and yes, test automation is the only way how to solve this problem. Because this is a big truth, untested means broken. And another big truth is no tests means legacy. Because if there are no tests, you don't know how to touch this code. You have to be afraid of touching this code. And the only way how to fight this fear is by having tests and test automation. And in our world this is true, untested means broken and there is a very nice example for which I can tell you. We recently did a complete rewrite of our system authentication layer. Like how the Linux systems authenticate users when they log in. And of course we did that test driven. And of course we forgot one use case. And when the original servers held up, we switched off. And only the new servers, active directory, state available, of course nothing worked anymore. Because we forgot about this little use case which was necessary there. And we looked into the code and said, well, where is the test? There's no test. So obviously it won't work. Right? It's simple. It's something a developer would always do. But in operations we also do that. No test, no work. So then we wrote a test, we fixed the code and it worked again. The problem was of course that our PAM patching code didn't expect that the PAM LDAP module was missing. Which just happens if you set up a server without PAM LDAP and then our hook for patching the file was missing. So no patching happens. So no login was possible. Actually simple but again no test, no work. So what is this thing about tests? There are a lot of books about tests and what you can do with tests and how to write tests and I think for us guys in operations the simple version is enough. The simple version is that there are two types of tests. Number one is unit tests. And the unit test tests the smallest possible component in an artificial environment. So try to think how to cut down everything that is not needed to test a single feature, a single aspect, a single function. In development this is much more complex and you do unit testing on all kind of levels. But in operations that is okay. Try to think how to cut it down, how to strip it down. And if stripping down means setting up a server and running something that can be a unit test. In development you say unit test doesn't have any external dependencies and so on but in operations you have to see what fits the problem. The other test you need is the opposite. It's the system test. The system test has the job of testing the entire application in a realistic environment and also testing it with other applications together because you need to test the cooperation, the inter-operation between different applications before you roll a change into production. That's all you need to get into test-driven development in infrastructure. So a little overview. Typical things for unit tests are they're part of the build process early on. And they have quick feedback cycles. A unit test should give you an answer with mere seconds so you can run it after every save of a file after every code change, a single line change, you run the test, it tells you yes or no. That's a unit test. Also, very important, syntax checks. Sounds stupid, sounds silly but hey, how many failures that you have due to a missing semicolon or other stuff happens to everyone so easy to fix by the test. And the other side, system tests usually start from installing something on a test server because you want to test your code in a realistic environment. You have to install it on a realistic test server that behaves like the real thing. And very important, you run tests from outside because usually you use servers from outside, you use their services so you're also in the test case, you have to do the same. And you of course also can run tests from inside which is especially useful to simulate error conditions. Like you remove the network, what happens? Of course, you remove the network from inside and RSH is a very useful tool for that because in that scenario you don't need the super duper security, you need the super duper automation. And SSH and automation suck. And don't forget, a reboot is also a test. The last thing I did in my consulting years was reboot before leaving the customer because I didn't want to go back to the customer the next morning after they rebooted the server. So yes, rebooting is a test, it doesn't cost much, you do it from inside, so do reboot, you wait a moment, you run the standard test from outside and you know if it's good. And if it doesn't work, then you know you have to fix it. And that actually will save you once getting up at night or save you buying your admin colleagues a crate of beer. Few examples from the real world, those who know us, they already heard that we use RPMs for everything. Software configuration doesn't matter what, everything that goes on our servers has to be packaged in an RPM package. And RPM packages have spec files as their master plan. This is a typical spec file, test like some preparation, installation, installing into some fake chain shoot environment and some files that are then shipped as part of the package. And the most simple thing you can do as a unit test is a syntax check as part of the build phase of your package or whatever other tool you use to ship stuff to servers. And if you use Zudo and you ship Zudoers, please syntax check them like that. Because if you don't, you'll cut off the tree you're sitting on. Because if you have a syntax error in the Zudoers file, Zudo will refuse cooperation. Even if the rule you would be using is in a different file, the complete Zudo stops to work. Another typical example you can find in hundreds of packages in our source repository is this. Syntax check, batch syntax check, Python syntax check, YAML. Very important. If you have configuration, test it before deployment. Because configuration is also code and configuration can break your server just the same like code can break your server. So the more you test configuration at least for obvious errors like syntax errors, the more robust your world will be and the more resilient your deployments will be. Because that means if it doesn't work, it won't build. And if it won't build, it can't harm my system. There are lots of more examples. If you look on my home page, you'll see another talk about this topic which has a few more. The more interesting part, of course, are system tests. Like in this example, a system test tests the entire system in a realistic environment. Same when you have a car checked every two years. They don't take off the wheels to check them. They check the wheels on the car as it runs on the street. And they put a fake street under it so that it's easier to handle the then stationary car. And that's the important thing about system tests. The important thing is how to mock away. The things that are irrelevant for the test. And this is a perfect example for mocking. The car feels like driving on the street. It behaves like driving on the street. But it's actually stationary in the garage where the test is being run on these wheels. So everything is real and everything that's irrelevant for the test is mocked away with these little two wheels here. And that allows this test to run anywhere, anytime under stable conditions. And actually we have trailers with the setup driving around the countryside so that people can check their brakes. And the same about unit tests, about system tests in IT. You want the system test to be exactly so that your code runs as expected without depending too much on external environments which you cannot provide together with the test. A little bit about build automation. Of course nobody runs these tests manually because then you would be busy testing instead of coding. So in our world the build automation looks like that. We have a source repository like everybody else. We have a central build automation tool. In our case Team City. Many people use Jenkins but there are others. Even a sophisticated bash script could be enough for that purpose. If a change happens it gets checked out on a build server which runs unit tests, creates an RPM package and uploads it into a dev yum repository. So far so easy. The next step is deploying that package onto a test server and running system tests. Now this takes maybe 20 seconds. This can take several minutes. But if the unit tests fail I don't need to run the system test because it's irrelevant. That's the thing about quick feedback, slower feedback. Small test, big test. If the test was successful the same RPM package is moved to a production yum repository and from there it's deployed by the same build automation to our production servers. And that way we basically instrument our entire platform. And any change goes this way from source code to production. And that's actually how DevOps works. Because DevOps in our case means devs and ops can provide commits into these source repositories. And it doesn't matter if the source code is turned into our whatever billing application or into our OS provisioning. Everybody can contribute to both of them. If they're there, if they know, if they ask their colleagues for a code review and so on, but they can. And that's the big change in DevOps. They can. They want to just go do it and we have test automation. So if you break it, it won't build. Don't be afraid. Just try it out. And that's the important thing to learn here. It's not enough to allow people to change code. You have to help people to overcome their fears, to overcome their kind of natural resistance to work in fields where they're not really proficient. Because many improvements are small improvement. I don't like that provisioning takes five minutes. But look, I see there is a simple solution to fix it. Okay, one minute saved. But maybe nobody in production had time for that. Maybe the developer who was testing the provisioning, including the setup of his software, if it runs in this initial kind of border case condition, he was sitting there and waiting for machines to boot and install and boot and install and got annoyed. So he fixed it. A few more examples from the system testing world. And yes, they are ops related. Who uses persistent storage? Okay. Everybody else? What do you do? I mean, you have to store stuff somewhere. So in our world, each virtual machine has one or two hard disks. One hard disk is the system disk. And we always wipe it and format it and install it from fresh. And if you store stuff on the system disk, you know that it will be gone eventually. If you need persistent storage in a virtual machine, you have to add a persistent storage disk. For those who use AWS, EBS is the keyword here. The idea is the same. You have a system and you have a persistent storage. Now, how does the persistent storage get configured into the system? Where to mount it? Where to format it and so on. In our case, we wrote a service for that, the Xan mount service, which uses certain algorithms to determine what to do. Like, oh, I have one extra disk. Oh, it has a file system label, persist something. Let's mount that. Actually not difficult, about 200 lines of bash. But if that service fails, then in our platform, the persistent storage will be gone. So how do we protect ourselves against this risk? We write a test. In this case, we write a test that runs through all possible permutations of actions that this service could do, including error scenarios. And we use mocking so that we don't have to connect real storage, but we use a low setup to provide an image file as a persistent disk. This is very convenient because I can also simulate different scenarios. I can add two or three disks and see what happens and so on. And I test, of course, service start-stop that the service mounts and unmounts my persistent storage as required. And now I have a delivery chain with a source code, sign mount service. And as part of that, a virtual machine gets provisioned and set up with a little bit of fake storage and all the tests run. And now I can tell everybody, you don't like the persistent storage handling, fix it. And they can fix it. And I can be sure that if the tests run, it will work in production. The important thing is always what to mock away and what not. And Linux provides you such a huge basket with little tools and tricks how to mock stuff. You can use routing or firewalling to fake network problems. You might actually, you should be doing both because if you set a route to DevNull, it behaves different than if you drop packages. And you might test your software against those two scenarios. Another example, who's using a proxy for their servers to access the internet? Okay. Who's allowing direct connection from web servers to the internet? Okay. Impressive. Because we don't trust our web servers. Web servers can be hacked and hacked web servers can download additional stuff. So we use a web proxy, squid in this case, as an application layer firewall for outgoing HTTP traffic. Again, if a configuration change in the proxy service would go wrong, then all our platform would not be able to talk with the internet. And then a lot of additional value services on our platform would stop to work. So we wanted to cover the entire proxy configuration with tests, not the proxy code. The proxy code is squid and it's from upstream and we never touch it. But for us, the configuration of the proxy service is also code that can break the platform. And we wanted to cover that with tests. And the way how we do that is we run each configuration change through a big set of system tests. For that, we set up a test squid server, load the configuration there, and then for each function group, which is in our world kind of a role, we do at least one test to make sure that the most important HTTP call for that function group goes through this configuration set. And we use, for example, X forwarded four headers to spoof the source address so that on our build server, we can set, let's, this request comes from function group five. And the rule set will think it's function group five. And then we check for access denied messages because, of course, the test server doesn't have internet access. You don't want your test calls to go against the production servers of your partners. It might make them upset and that might cost you money if it's a billable service. So obviously, the test server doesn't have any internet access, which leads to a very funny result. If I have a request, and this is the server that should be allowed to do the request to an external URL, in the good case, I get a bad gateway error because the squid on the test server allowed the request, but it can't go to the internet. And if the rule is wrong, or if I have an error there, then I get a forbidden from the rule set in the squid and error access denied. And this, in the test case, would mean test failed. So this is completely upside down. Like, bad gateway is good, forbidden is bad, but the test was successful. And this is, again, about mocking. You need to know what you're mocking. And then when you know what to mock, you know how to write the test that reacts to the right trigger from your mocking environment. Last example, VM provisioning. We also have service in our data center, which we need to provision. And every morning, we have a test set up running for 15 minutes and setting up various virtual machines, some of them broken, some of them good. And we check that the broken ones are not allowed to work, and the good ones are allowed to work, and that on the good ones, actually, the automated environment will set up a working Linux operating system. All that happens every night so that we know, okay, we can still provision new systems. And that's a very huge system test, but it's also very valuable. Because there are about, I don't know, 20 software packages that go into this automated provisioning set up, which we have for virtual machines and for hardware. This is actually open source, and you can go there and find all the code for the system test and so on. If you do that for your platform, you get different release cycles for different software packages, and each release software package goes its own way from development to production. And in the end, they all meet somewhere in production. And you know that they work together because of your tests. We call it continuous life deployment. That's our way how to maintain, stably maintain, always changing platform. And the general rule is we deploy applications when they're ready, and we automate the delivery chains from source to production. The end result is low risk, lots of fun. And that's the whole thing about DevOps risk mitigation. You have all the fun from DevOps, from doing stuff together, but you have a low risk, and you're not afraid to do stuff together. You'll find the slides here, plus a few more links and other talks about this topic. I'm at the end of my talk, needless to say, we're hiring. So if you have a passion for automation and for keeping things simple, please talk to the people who have our logo on the back. Thank you very much, and we have 15 minutes for questions. So any questions from anyone? So first thanks for a really interesting talk. In one of your examples, actually the last one, testing the proxy, you kind of showed how you would use error conditions because you don't want to rely on external services. And I wonder if you could compare this to actually mocking it by recording and replaying responses or stuff like this. Why did you not choose it, or what do you think is in this? Okay, it's a good question. And I think that's exactly a question about DevOps. So as a dev person, I would think, how can I mock away the Internet? As an ops person, I say, I don't need to mock the Internet, I just need to deal with it. And this code has been written by an ops development team. And so the solution is, let's take a server, let's let it do what it usually does, because setting up a server with a proxy and the configuration, it's just standard. You say, new server type proxy, go, go, go, done. So there's nothing needed to do for that. The only thing that we had to do for the mocking actually was to set up the test server in our dev environment, which in general doesn't have Internet access. So in this case, I would say the answer is it was the easiest thing to do. Like the cheapest in terms of effort and in terms of changing the system. The only other change which we did to the proxy configuration was to allow the build servers that run the tests to use the X forwarded for header to simulate the actual originating IP. That's the only real change to the configuration which we did. We use load balances internally. So normally only the load balances are allowed to use X forwarded for and nobody else. And for the system test to work, of course, the build agent that runs the script that runs for all that needs also to be allowed to use X forwarded for. But that's all. Except everything else is the original production configuration. And maybe I didn't say that. In this case, the entire proxy configuration resides in a software RPM package. So we have a source repository and there we have all the proxy configuration and the test cases. And each time somebody needs to change the proxy configuration, we just do a new release of the proxy configuration RPM. And that RPM goes first on the test server. It runs for all the tests. And then the same RPM is installed on the production proxy servers. And then we know for sure that that set of configuration works. And the end result is that now developers add proxy configuration to their function groups. If you're a developer and you set up a new software, then you go into this software package that contains the proxy configuration and add your own proxy configuration for the calls you need to make. Plus one few test cases and you're done. You just wait 10 minutes and then it's live in production. And that's how we play DevOps and that's how we bring DevOps together in improving our platform and reducing turnover cycles, development cycles, and so on. So you had this chart about automation. Automation is great because we are all lazy, right? But there you even just repeated that, that if I commit something, that automatically goes live. But you also had the slide with the release cycles. So what's your politics there? Everything goes live as soon as I commit it and it passes the tests or is there anything else? Well, it depends on the team and on the software product. If more teams put more trust in their tests than in the ability of the product manager to push the release button. But the ideal situation is you trust your tests because the tests are documented knowledge about your platform. And the manager pushing the release button is just believe. I believe this will be good. Yes? So in the setup that you've described here, how would you typically deal with replicating say your production database onto the test server? Would you fully replicate it or try to do something partial? Because one of the things that we find most dangerous when we're deploying is things like schema changes that are very difficult to test against fully? Well in our world, everything has to be a package and everything that acts has to be a service. Any acting part in our platform has to be a service that can be started, stopped and that has a status. So database changes also have to be a service. And we have services that do database changes as needed when they detect a new database schema. And that also happens, for example, here. Because together with that package comes a new database schema, the services here say, oh, new database schema, let's update the database. And the task of reducing data from production to test belongs to the developer who's creating the data or whose software is creating the data. And for each software, one of the tasks in the checklist for production ready is, did you write something that will create a test database from production? And then the reduction of the data, anonymization, removal of personal information, of people information, it's all their problem because they created the database. And that's the only way how you can scale to hundreds of roles or function groups. Otherwise, you have a team which just runs behind the others and always needs to adapt their changes into the conversion process. I wouldn't want to work in that team. So you described how you cover your ops code with tests. Do you also do it in test-driven way, like in a narrow sense, writing tests first, doing baby steps, refactoring, does it make sense in this scenario? Yes, we do that. We have a lot of Python code in operations. What I mentioned initially, the example I mentioned about the authentication code, it's managed by a Python script that does all the patching on the Linux configuration file level. And that has full test coverage with Python unit tests. And yes, we write, in this case, we first wrote the test and then the code that does the patches. And that's why in the end, the feature for which we didn't write a test also didn't have the code. Yes, test first. It doesn't mean that it's easy test first, by the way, especially in operations where sometimes testing means setting up a lot of stuff, but yes, we do that. You describe a bootland automation workflow that actually tests lots of RPMs. What about when you are testing things like configuration management, Puppet, ChF, all that kind of systems? How do you have those kind of changes also included in your workflow? Well, as I said, all configuration is in packages. We don't have Puppet. You don't use that. We don't need Puppet because Puppet solves problems which we don't have. Okay, fine. So, sorry. No, no, it's good to know. But if you would go to a Puppet conference, you would find out that making Puppet recipes testable is a really big problem. And the reason it's a big problem is that Puppet combines code and configuration in a very nasty way. I mean, okay, it's fine. But leaving Puppet aside, if you would try to do testing some kinds of things like duplications or stuff like that, you need to set up a really, really complex networking setup. And that's go way beyond just getting a VM. It's like getting two or three VMs that you have to get networking with between them and try to check stuff like that is not so easy like just running a test there. Actually, here I mentioned one package, but our automation can easily handle an arbitrary number of packages which are involved in this change because the system tests running here trigger the propagation of the packages. And we, of course, have a hook that actually propagates the packages that were installed on the test server. So in the test, in the job here, I say five packages are relevant for this feature. And then it will propagate all these five RPMs if they're here and installed on the test server to the production repository. Which is fine for deployment, but what happens when you have to test like my SQL that is not responding and your slave is way beyond whatever it should be. How do you test that your, the rest of your system is coping with that? Well, we, I'm just looking for what's the framework used to automate that kind of tasks. We use the framework called keep it simple. Keep it simple means that the average developer doesn't take too long to write his first piece of useful code. In many cases, that means we have some part on the build server which runs the job that will RSH into the test server and do some nasty manipulation before running a test. And we use RSH because in RSH you can just say, well, this IP range is allowed and you don't need to bother about faking away SSH keys. So basically you have to write the logic into the spec files, the logic which is mostly included in top it. Now the spec files provide a simplified way of doing the same thing. We separate configuration files depending on how they change. And that's why we get away with a lot less patching than you usually do in a puppet world. Puppet is good at patching stuff, but we lay out configuration so that it doesn't need to be patched. That's why I said puppet solves a problem which we don't have. I just wanted to add that the puppet situation about testing and testable code is much better than it used to be. There is beaker and there are some other tools to deal with that. I know. The community has been active because there was a big problem. But the end, I think what I'm saying is that you need unit tests and system tests regardless of the tooling you use. And even if you use puppet, whatever tooling, you still need unit tests and system tests. And the unit tests will still test something small. And the system tests will test something big. The question is always how do you express it? How do you abstract it? And so on. And how do you make sure that the stuff you tested goes unchanged into production? In our case, it's simple because the RPM is created once. And we never create RPMs again after successful tests. That's evil. You deploy in production that what you tested and not something else. And if you have a world where you test something and then you create something else to deploy in production, you'll always have this little gap that can go wrong. And you'll for sure find a smart hacker who will be happy to use that gap for his own purposes. Okay, I have to ask for a detail because it just sounds too great what you're saying. We were discussing the migrations and the continuous deployment. Do you manage to do migrations without service interruptions? Because when we do migrations, we have to at least partially shut down services. And we can't do that just because somebody said push to the repository. Well, there are several layers on which this question needs to be answered. The first layer is can your application handle the situation where the old and the new version runs alongside? Because if the application can't handle that, then the deployment won't help. So first, make your application so that version five and version six can work together on the same database. Make it so that the database upgrade from version five to six will not harm the version five code using that same database. That's the first step on the application level. Then you can go to the operations level and say, okay, I have 20 web servers and I want to have a rolling upgrade going through these web servers so that there's no external impact. And yes, we're doing that. In our world, it's very simple. Any server is allowed to install the RPEM packages presented to it through the various YAM repositories attached. And we have a tooling called Yatchel, which does the rolling upgrade, including load balancer on off monitoring on off services down packages upgrade services up monitoring on load balancer on checking next server. Yes, we do that. We do it actually in an automated fashion. Team city in our case does trigger this kind of waves. And if there's any problem, the wave just stops and we go and check what happened. I think we've got time for just one more question. Hi. So a lot of teams, when they're deploying to a large set of servers, do the so-called canarying. So first, you deploy to a server handling 1% of traffic or 1.5% of traffic, let it work with this small person for a while, see if there are memory leaks, et cetera, then to 1%, 10% and finally, fully, do you do such a thing? We do that in a few cases. In most cases, we don't do it so far. We are on the process of getting there. Our yet tooling can, for example, do exponential deployments that can deploy one, then deploy five, and then can deploy the remaining. And if we, but the thing is here, yum repositories represent the target state in our world. So this time of deployment, it's kind of a fuzzy gray zone. And in our world, it's always okay to deploy the latest updates. It's never wrong. Never, nobody can be punished for doing yum upgrade. So how do you, because if I understand correctly, your commit to deploy cycle is around minutes, maybe hour. So how do you deal with bugs that show up after a few hours of work or a few millions of requests, like I'm leaking four bytes of memory per request? Well, as a developer, it's your responsibility to think about the potential danger of your change. And if you have stuff that could go wrong in that way, then you need to deal with it already on the development side and not expect the deployment to solve your problems. And if you want to have a longer lasting state of different versions, then in our world, you create yum repositories for that. You create, for example, our big core application creates a new yum repository for each build. And in that yum repository are a few hundred RPM packages with a few gigabytes of stuff. And in that case, we can always take a few servers, hook them up to the yum repository of the next version, upgrade them all, and wait a little bit. And even if you would reinstall one of the cannery servers, it would automatically get the N plus one version. It's supposed to be running. And we do a lot of state management, like you mentioned, with the help of yum repositories, by just creating special yum repositories and putting packages there. Okay. I think that's time. So thank you very much.
|
Schlomo Schapiro - DevOps Risk Mitigation: Test Driven Infrastructure The (perceived) risk of the DevOps is that too many people get the right to "break" the platform. Test Driven Infrastructure is about adapting proven ideas from our developer colleagues to the development and operations of Infrastructure services like virtualization, OS provisioning, postfix configuration, httpd configuration, ssh tuning, SAN LUN mounting and others. This talk shows how ImmobilienScout24 utilizes more and more test driven development in IT operations to increase quality and to mitigate the risk of opening up the infrastructure developmen to all developers. ----- Common wisdom has it that the test effort should be related to the risk of a change. However, the reality is different: Developers build elaborate automated test chains to test every single commit of their application. Admins regularly “test” changes on the live platform in production. But which change carries a higher risk of taking the live platform down? What about the software that runs at the “lower levels” of your platform, e.g. systems automation, provisioning, proxy configuration, mail server configuration, database systems etc. An outage of any of those systems can have a financial impact that is as severe as a bug in the “main” software! One of the biggest learnings that any Ops person can learn from a Dev person is Test Driven Development. Easy to say - difficult to apply is my personal experience with the TDD challenge. This talk throws some light on recent developments at ImmobilienScout24 that help us to develop the core of our infrastructure services with a test driven approach: * How to do unit tests, integration tests and systems tests for infrastructure services? * How to automatically verify Proxy, DNS, Postfix configurations before deploying them on live servers? * How to test “dangerous” services like our PXE boot environment or the automated SAN mounting scripts? * How to add a little bit of test coverage to everything we do. * Test Driven: First write a failing test and then the code that fixes it.
|
10.5446/20027 (DOI)
|
dweud hynny a driran yng Nghymru, lydych ynthersent ar gyfladau a siarad symudol a compromise wychordeb wat என perbywm allfryd col程id fel gan y tu�� dros ynghyd y introduced, dwi gyd, os yw di modern wedi sylch ein cynll rerwch oeny viewch. Mynd gwybod y gallful y gwuddedigIG gan sicrhau remainrft o unf tongueLide committee mae galwch this ynsea'u cy patriarch at Testing I phonwys ominonu'i iechwynser a风os cyfaint twydon Merci Arla drwy ad vegetables Werang i…right, pan conseguirau y cyzasen Citizen Daицwn i rhan o cael g bonus ar ymarfambbi a dwy beth neu'r bys ymwelch yn siiHer, dwi'n credu rlyız. Mae newidyn arall arnor mewn Nghadir Coryn F onto12. Lleid likwad, ar g observau a'r cyscêl ond mae'r gwaith cyflog yn diemnell yng â ich пwyd idyn a pheth bethau otorth yn y glas a'r чerddonже pa'r Swt ninety cof… …nog ddim yn amliodwyd rhywbeth mwy o rang homi mawr in arweidences. Mae gennym plants iddyn nhw i'w gwahod iawn yn y sgwrs gwahodio'r ond wrth ei gall Police Diwyddi… …RO Job o b dysguant gwnaeth gyda, thetha f Aboriginal Atholbeithloedd yn y galw II. So rydym ni yw blaen d weakness, Ond myrt o gwen redo genna Siates? Er umr unterwegseth a Dwi ichan ei gyllian y t summonau. Gyftwo dyw'r Boxing T我說 powered complex yw'rhal leads Cy Wir, byddai castwyn a med depois institution gyda rebellion o Bou arifer hunting Underwater dan oermi ñ mae hynny'r cy mentally ir corog Dorothy i wneud ymateb i'i beth ddau beth wnaed llawer a'r吧tywater olabynch Lord Shawsea sy'n teimlo'n amlod ni ar y Llund,搞xonol sorted o'r restaurant yw'r ll 밝혔 CDs petith Notes roeddy arrestwyn hanesonwyr cymdeithiau'r 65 corps. Rydyn ni'n hollwch... ac ddim yn het podcasts ymateb ac lle. Ismrwydd yma'r gŵl. Tyw'r bodial i ddiwyllgor waith fairio cy cavalryha Kam pan mewn. Ewch i yn ymddwn sydd yn ei ddwefydwyrמש hynny... yn digwydd Fine is Wel i ddim leol-, jweuswil o behalfen sy'n falch hwn yn ddeficio ADP, mae'r llumiaeth fe'w goht gan fframeog. Has i ddim leol-ig暖 chi posibol flennig neu gonellidaeth caslu? F dobrze go i gymryd wedi hwnneud! Felly bydda'r gafodacedaidd i taethradau wedi ddwy gingh a meddwl worarr na aldog. Alo'r facilitategiad f基llin Wyr y mod ydy'r idea wedi bod ydy'r hyn yn ei wneud. Mae'r transbutau er mwyn i'r cwmwysgol, ac mae'r cyfnodd yn ychydig, ac mae'r cyfnodd yn ymddangos. Mae'r cwmwysgol yn ymddangos y cwmwysgol. Mae'r gweithio'r llai'r ysgolwydd yn ymddangos. Mae'r CSP yn cyfnoddol yn cyfnoddol. Mae'r cwmwysgol yn ymddangos, mae'r cyfnoddol yn cyfnoddol. Mae'r cyfnoddol yn ymddangos, a enghreithu drwg ein bodun, Sydney neither. Mae'r cwmwysgol yn lle'u gweld mewn ymddangos. Environilt y mae fydd i bod yn cael ei amfel oedd ynumb mewn hreit ip moddol yn ymddangos, ac ymddangos gyda sy'n astud. Mae nid i'r meddwl fyd關 ar hyn. Mae hynny'n nhw'n chi. Spedı onto rwy'r pas y mynd culwyd dros o'r ysgol yma, a ydy'n gwybod bod yn fawr. Yn y prosesalgwyb ar gyfer yw'r ysgol, ymgyrch yn ei wneud o'r cymaint cymaint sydd yn cyfaintol, ac mae'n ddweud o'r ddweud, ac mae'n ddweud o'r prosesau sydd yn cymaint. Mae'n ddweud o'r ddweud o'r ddweud. Mae'n ddweud o'r prosesau sydd yn cymaint sydd yn cymaint ymgyrch yn cymaint. Mae ysgol yn rhaglenol, onf o'r grwpastech me'n comparisongyngor ac nergyrchteil, mae gweithio addressesol yn Afryd, ac yn chwysr symud o'r bars accommodation ac mae'n j ni ddim yn�痴 iawn. Mae gwaith eu ga teknol gyda llunw. O'r swinchildren yw cyfnod o wigel a gynnaldiolaeth a'u bod i chi ceblifnu ond, mae th eighteen yma o han corona wrth gwrs weld bydd gweithf swwydd mor ffordd palwy neu i wnaetta ni'n unfabn iddyll. A hyn cyffer delw o'm aptrach yn Du o MoIwn o carefullyb Nash Algżado De Anioedd y Ac yn gwelafodd gyntaf justyn nhun bant yn hyvau hynny yn ymddangos. felly byw'r arni a'r gwyffrŵ을 yn digwyddol. economy Queen yn wircaf iawn. A gallwn ei osionum o'r anit am f experi penteig言iau Tamodol sy'n ni g door. Llwyddo i weithio cael gwlad Meet, wedi sicrhau nhw'n gweld rawyddion honno yng ngry autobiol fel mae'n ddesmyt ch widelyll ar ac yn llei ochreitchau ch coffin.ceddiolol wedi clynyddu wasname. Rwy'r rhwyap bellachol, probicswyl yr ur messages, ysgol sydd i gyntrдо affythio, arweinyw ID 저eras, yr sechczyn achosol, yn y merthw i s子'r pes yw'r s slappeden diwrnod, achos. Fyg mermaid rondd Zwg Femaf yw rhaniaeth a all you Gardens mor gaer mwyafol o gaeliai rhagor mewn rhes hubau gyfyrdd iaith y nhw, chi'n mynd ar gyrwch symud graeniant cy súdd, menocal i dod ei fod yn ardu i'r f更azei amweiliadau. Rost fair mewn hwn. Rwy'n nad y gaer hwnno iawn. peninsula o chi i eni'w das i gwael? Fe ychydig", obl 말이 hwo ddwy o'i defnyddio'r unrefr? Manchester-Popoda Edgl Strath breathing room ffaltau dweud hyn. Felly am Сп Authority, dysparwydd gan cyfr分u......ad am y Ghazyd detector hynny. Y ddu bach, maen nhw yn holl i gyf strategau sí pond Broaddpan. ac mae fydd gwaith hir速au hyperbwyntol i'r arcain biogel trefenn i'r googol. Felly, all y twл engh которуюw — mae'r hyn yn gweld icreinio. Efallai efallai o'n g령oli i~? Mae'n creduetus y daethau, rapeaceien i butthiau meun incluso. Felly, efallai efallais eu rhailedd â raihe zenain yn meddwl i'r meetydd yng Nghydri o gwerth interviewwyr yllюсь yn ddefnyddio'r languages. Rhaid eu xeniadon mwy wneud een unig mewn byteb. Pwault yn oTrwm iaw rwy'r bl 이 lui fyddbwyll Pi便. Yn meddwl eich gwyrdd cyfryd, a on iddo i genlwyddnog chi a appeal'r myl iawn i hynny credw yer ers maysbidd i chi dif againstwaith yr oedd yn gade. Rydych chi'n cyff公fer mae'n rhywbeth hyn yardsin ni'n myl flat ar ôl yn demp explode. Wal e'i hidingb tracing iddyn am fort hyn yn oed droi'r plo, mae'n rwy'n gwneud yn fag ar-redig yn rhanhu g gref таeth yn meddwl y r pimp hwn yn tu, ac cela提 o gy tight. ond lle'r chwarae ruman er sonnwys. Rydych yn erbyn y subавливau, bydd sefydlu gweithio eu cydureddayım y g stretched a'n deall yn achon ap ddiwedd Newbeth sydd wedi Spinwermu COVID-19. Sefyd vulyn cyylem flashgell fwy fydd Welnodd完了 cawn er Cult Front ac bydd y modd tyfo'r unedÏ y byddug yn AUY yn drws ond lyso a'r ysgilriaeth than an operating system thread. A then what we're doing on this line with this sort of funny syntax is we're sending the string, hello world, down this channel, my channel. And then we're going to run this Go routine straight away. So it's going to be running in the background of our program. So these sorts of funny bits of syntax, I'm afraid, really pervade these ideas in languages. And you'll see a lot of weird syntax. The CSP syntax for doing this sort of thing is a pling, an exclamation mark, or to receive something down a channel, a question mark. So it kind of hasn't got better through the ages, in my view, I'm afraid. So on this line here what we're doing is we're receiving something down this channel, from this channel, and whatever we receive we're just printing out. So this is a simple sort of hello world in that language. Rust is another new language, does similar things. So we've got the same sort of idea here. We've got a channel being created here. We've got a background process running here where we're doing some sending. Then we're receiving this value and printing it out. What's slightly different with Rust is that if you are familiar with working with UNIX Pipes in C, you'll know that when you create a UNIX pipe in C, what the operating system gives you is two ends of the pipe, a sending end and a receiving end. That's what's happening here with Rust. We've got a sending end of this channel and we've got a receiving end of this channel. The idea there is to prevent you from doing silly things like sending down the receiving end or receiving down the sending end. So you, the programmer in Rust, have to decide ahead of time where in my programme am I going to want to send down this channel and where am I going to want to receive down this channel. It's something you need to think about at compile time. So Scala, being a JVM language, is taking up the whole screen with its verbose braces. But this is exactly the same sort of thing. So we have an actor which is similar to the sort of background processes that we were talking about and co-routines and so on. We've got this slightly backwards here. So here we're sending. So this is really using CSP syntax with the pling. And in here we're receiving something and printing it out. And Python CSP. So Python CSP is my own library. And some of you in this room have been really generous and contributing to it, particularly Stefan over there in previous year of pythons. So this is an attempt at doing something like this in a Pythonic way for Python, but built as an add-on to the language as a library. So here we've got two CSP processes. So we're not saying in the code here how those processes are sort of reified, whether they're co-routines or threads or operating system processes. But we've got two processes here that can run in parallel with the decorators. We've got channels that can be shared between them, and we can read and write with those channels. And again, we're just sending Hello World, printing it out. And then on this line, so this is a much more sort of CSP-ish way of doing things than perhaps the other examples. We're saying, well, we're going to take these two processes and run them in parallel and start them off. And if we had a huge program with many processes, we might decide, well, we'll run them all in parallel, or we might run a few, then run a few more in sequence, whatever we wish. So we've got quite a lot of flexibility there about how our program is sort of put together. So this last example is by a student of mine, Sam Giles. And what he was looking at was really interesting, which was can we build a language like this that has the sort of go or rust style channels and concurrences on the R-Python toolshane? So can we use the technology that the PYPY team would develop to do this? So this is exactly the same Hello World example there. We've got channels here. We're going to send down that channel a Hello World. And then we've got Sam's sort of unusual receive syntax there to receive something from the channel and print it out. And this function here is being run in the background as a sort of asynchronous coroutine. So that's a really nice project. And it's a really nice way of working. And I think obviously Sam was a very, very good student. But I think it's a testament to the good engineering of the PYPY team that an undergraduate student can produce a working language like that in the small amount of time for a final year project. So I'm not going to talk in great detail about optimization and speed and efficiency and those sorts of issues. But I just wanted to show you quickly one of Sam's benchmarks, which shows quite nicely that with a tracing jit, Nalang can perform well compared to other languages of this sort. So go here and OCamPy, which is a descendant of OCam, both compiled languages. And Nalang compares pretty reasonably well to them. This is only a small benchmark, so we perhaps shouldn't take it as gospel. But it's a good indication that this sort of way of working might be positive. On the other hand, I haven't got for you here the same benchmark with Python CSP. But we looked at similar things with Python CSP. And that was engineered very differently. So I'll talk a little bit more about the design decisions that message passing concurrency implementers might take in the next section of the talk. But what we found with Python CSP is that our implementations of channels were very, very slow. Very, very slow. So you wouldn't expect, or I wouldn't expect, a Python implementation of this to be as fast as something like Go OCamPy that's compiled and has a lot of engineering going into these features. But I perhaps wouldn't expect, or I would hope, that message passing would not be the bottleneck in any program. And what we actually found was that OCamPy is incredibly fast. It's designed exactly for this. But compared to other sorts of interpreted languages, we looked at JCSP for the JVM, thinking maybe because JCSP is built on Java threads. Java threads are OK, but they're operating system threads. Maybe we could get something like that performance. And we actually got 100 or so times worse and didn't do very well at all. So there are some lessons learned there. And there's some interesting stories. But part of the takeaway of this is that actually it's very difficult to engineer that kind of performance if you're starting from an interpreted language that hasn't been built with this sort of concurrency in mind. So this next section of the talk is all about the sorts of varieties of message passing concurrency that can be created and the different decisions that an implementer would have to make if they were going to implement something like this in Python. So one choice is synchronous channels versus asynchronous channels. So in the CSP way of thinking and in the process algebra way of thinking, that sort of very mathematical formalism, the idea is that all channels block on reading and writing. And you don't move forward with the computation until your read or your write is finished. And some people, including me, think that the nice thing about this is it's then very easy to understand what your program's doing and reason about it because you know exactly in what order everything's going to happen. You know, this piece of code will not move forward until it's finished this read and then it'll do this and then all the other things that are waiting on it will be able to move forward as well. Asynchronous channels though are quite common as well in different languages and some people suggest that they're a bit faster and sometimes that seems to be true. And certainly the benchmark I showed you before showed that in that particular benchmark, SAMS asynchronous channels for now were a little bit faster than synchronous ones. If you do have synchronous channels though, you need to think a little bit about avoiding some of the common problems that people have with concurrency like starvation and you don't necessarily want a process to block for a very long time if it doesn't have to. Sometimes it might have to, sometimes it might be waiting on a long computation, but if you don't have to block, you would probably prefer not to. So a common feature of message passing languages and libraries is some way of selecting the next ready event to process. So if we're thinking in terms of events being channels or some message passing down channels, if you have a lot of channels that you're waiting on and you want to read from, so for example if you've got a map-regeased type problem or a worker-pharma type problem, then you might say, well, give me the one that's ready first and that's called alternating in sort of OCam, old-fashioned language or selection more generally. So you can say select for me the channel that's ready to read. And usually if you're implementing that selection, you do that rather carefully, because although you might want to select the next ready channel to read from, if you've got a channel that's always ready to read from and some that are taking a little while, you don't want those other channels to not be processed. So usually there's a little bit of work goes into that to avoid starvation and do some good load balancing. So that's one issue. If you're synchronous or asynchronous channels, or you might say buffered or unbuffered channels. Another issue is, are your channels bi-directional or are they unidirectional? So we saw in Rust you get what you get in Unix C, which is a read-end and a write-end of a channel. And that's quite a common way of working with channels to avoid some mistakes in your code. If you look at the JCSP library, which is a very nice library because it's been engineered very well with a lot of thought going into its crackness, the JCSP library is very Java-like. And Java people don't mind having thousands of classes to choose from and large amounts of documentation. And they don't mind pressing control space in their ID and getting a long, long, long list of things. And so JCSP sort of works with that paradigm. And it has lots of different channel types that are all classes. I haven't listed them all because the slides are small. But so you can have things like you can have a one-to-one channel that has one reader and one writer process attached to it at any one time. You can have it any-to-any channel that has any number attached to them at any time, and so on. And then you always have the read-end and the write-end of that channel, wherever you are. And the idea here is to use the type checker to design out a lot of potential faults that might creep into your code. So that's nice for Java because it fits well with the sort of Java way of doing things. It's what Java people would expect. So when I wrote Python CSP and designed that, I made all the channels any-to-any channel. And I didn't give people the read-end and the write-end. I let them shoot themselves in the foot because it seems to me to be a bit more of a sort of dynamic way of doing things and a bit more pythonic. But not faultless, not foolproof. So those are a couple of different design choices. Another is mobile or immobile channels. So this is something that wasn't built into CSP originally, but it was built into a different process algebra called the Py Calculus by Robin Milner. And then the Kent, the team at Kent University, who sort of took over the development of Occam, created Occam Py, which sort of fused together the Py Calculus and the CSP way of doing things. So a mobile channel is a channel that can be sent down another channel to a different process. And the idea of doing that is that you can think of your message passing programme as being like a graph where the nodes of the graph are your processes and the arcs between processes are your channels that link those processes together. At runtime, you may wish to change the topology of that graph and change its shape. So two good reasons why you might do this. One might be a bit to do with load balancing. If you have a computation that's split among a lot of processes, you might find some of them are more active than others and you might decide to change the load balance between them, which might also mean changing the topology of the graph and who's reporting their data back to who and who's aggregating the data and so forth. That's one reason. Another reason might be that you might be running these processes across a network. So you might not only be working with one machine, you might have some processes farmed out to another machine on your network and then you might have issues like latency or you might have issues like network failure or whatever. That might make you think, well, during the running of my computation, I'd like to change the topology to make the most efficient use of that network of machines. So that's one reason. So this leads to two issues. Mobile channels can be great if you can use them really well and you've got a good use case for them. If you're in a situation where you need to shut down this network and graph of running concurrent processes, then you need to notify each node in your graph that it needs to shut down. And so doing that safely is quite an important thing to do. So in the message passing world, one way to do this is called poisoning, which means that you tell a channel, or the node that decides to shut everything down or shut a few things down, tells a channel or all of its channels that it knows about that they need to start shutting down and they need to propagate the message that this program is going to halt. And this is called poisoning. So you poison a channel and the idea is that it poisons the well of the whole program and each process shuts itself down safely. And that's something that takes a little bit of care and a little bit of good engineering because you need to think, well, if I'm a process and you're all processes, and I say I want you to all die and then I'm going to die, then that needs to happen in the right order. If I kill myself first, you won't know what to do. LAUGHTER Or not. Who knows? OK. So the other...we talked about channels, we talked about mobility, different sorts of channels. The other thing is how to represent the processes, and there are a lot of different choices there too. So in some languages, one CSP process is one co-routine and that makes sense in some paradigms, so I think this is kind of how Node.js works. And that leads to very fast message passing because in the runtime system, all those processes share memory. They're all really in the same operating system thread, so they can do a lot of things very, very fast and they can pass messages down channels very, very fast, but then it's hard to take advantage of multicore if you're all in one thread. You could have a one-to-one mapping where one CSP process is one OS thread. That gives you much slower message passing because whoever implements that does have to deal with locking and all those low-level issues, but then you can start taking advantage of the features that your operating system has. You could make one CSP process one OS process, and that's a really good choice if you're thinking about migrating processes around a network and running your code on more than one computer at once. So, sort of MPI style if you're into MPI. Or you can have some sort of multiplexed version of all of those options. So, you can have some CSP process that are co-routines, but live inside an OS thread, and there are other CSP processes that live inside another OS thread, but are really co-routines themselves and all sorts of combinations therein. This is really where why Python CSP was not as fast as we'd hoped because we were looking at taking advantage of multicore in the network, so we were using these sorts of one-to-one mappings, which are not the best in terms of speed. So, I'm not going to talk for a huge amount longer because hopefully we can have a good discussion, but I wanted to say a little bit about message passing in Python. So, there are lots of... Although Python is not a message passing language in the way that Go is and Rust is and Acro are, all those other things, Python does sort of have a lot of these ideas built into its ecosystem. Sometimes in libraries, sometimes in different implementations of the interpreter, sometimes in all sorts of other ways. So, I was really pleased looking through the EuroPython schedule to find that actually there are a lot of different talks in this conference that in some way have quite a lot to do with the ideas that I've been talking about today. So, not necessarily straightforward implementations of message passing in the way that Python CSP was, but they take on some of those ideas either by implementing co-routines or using co-routines or using channels and so on. So, in a sense, message passing for Python is already here, in a sense. And also, of course, in Python 3.4, we have co-routines built in. So, there's perhaps a big opportunity there to think about building these things in to the core of the language. So, if you're interested in this stuff, then I'd certainly be interested in talking to you. My next steps for this, Python CSP has been sort of in abeyance for the last few years while my day job has taken me to do different things. But Python for the Parallela board will be coming out this summer. So, I've got a project working on that this summer starting sort of mid-August. And we'll be looking at nice and hopefully efficient ways of using Python for the Parallela that ideally would use message passing in some way, but we'll see how that works out. The jury's rather out on that one. And Python CSP is certainly moving back into regular development. Sam's language now will be continuing, so I'll be at the Pi Pi Sprint this Saturday, doing a little bit more on that. And if you are interested in this stuff, then please do come and catch me sometime. Thank you very much. APPLAUSE So, do you still have time for questions? If you have questions, can you please lie down on the mic on the other side of the floor? Hello. I have seen that we are always relying on the Operative System layer for the threads or the core routines, etc. Has been any enhancement on how processors can pass from one to another information besides the caches and all those things? Yes, so there was an interesting development in the Open MPI library a few years ago when they found that their message passing was a little bit slower than they would like. Open MPI people tend to work on Linux, so the Linux kernel brought in a new way of doing that, which is called Cross Memory Attach. The idea of Cross Memory Attach, and I think it is only a Linux thing now, but the idea of it is that you've got two different operating system processes. Rather than saying, rather than doing what you would do in a pipe, which is that you copy the memory, you keep the memory in one place and then you pass around a sort of handle to that memory between processes. So that's a much quicker way of doing it that was built specifically for MPI, but would possibly be a really good way forward for any other implementation like an implementation on Python. So yes, there's definitely some interesting work there. Hello, I wanted to ask what the Python CSP library provides that Sagee event doesn't, apart from the Simpler API. Well, it's a good question. It's a different API. I don't know if it's simpler or not. Python CSP sort of started because I wanted something like this in Python, but the only things that were available were direct ports of the Java JCSP language. So the idea of this is that it's much more similar to a CSP way of doing things than anything else. So what does it provide? It provides processes which can be various sorts of processes. It provides channels, it provides selection or alternation, it provides a small library of built-in processes that might be useful. So the reason for that is that the way that CSP people tend to think about this is that the more concurrency you have, the better. So rather than saying, well, I've got my nice sequential program, how do I split it up to make it efficient or concurrent or sensible or whatever it is? They say, well, you know, make everything you possibly can concurrent. They tend to have libraries of processes that do things like have two channels, read down those two, read two numbers from those channels, add them together, and send them out down a third channel. So a process that just does addition, and then a process that just does all the other arithmetic things. So there's support for that way of working. If that way of working is something that's interesting to you, I suspect, though, that that way of working is probably only interesting to people who are interested in CSP for its own sake, because it's not a terribly pragmatic way of working. So, I mean, the answer to your question really is that Python CSP implements all the sort of basic things that you expect of a message passing library. It's just a matter of how it implements them and how well. And I think we probably score about five out of 10 for that at the moment, but hopefully it'll get better. So a quick question on Python CSP and multiple processes. Is it currently implemented with multi? Does it use something like multiple processing? Multi-processing. How are you doing the message passing across processes? Is it using pickle to serialize them? So we've got two different ways of doing it. One with threads and one with processes, but I didn't use multi-processing. I just use OS.fork and that sort of thing. So Windows is out of the question. Yeah, so the idea of that was that multi-processing is really built for a particular way of working, and it has a lot of internal code that supports that way of working, but isn't so useful if you want to do things the CSP way. So, for example, I think when you spawn a process in multi-processing, that process also spawns a watchdog thread for that process. But in a CSP library, you don't need that. So the idea was to be just a little tiny bit more efficient by not having those multi-processing internals. I think in reality, if you compared a version of Python CSP using multi-processing on one without, you probably wouldn't find a vast amount of difference. So you could easily do most of these things using multi-processing because you've got pipes in the MP library. So in that sense, Python has some of these things built in already. So you use the OES-Fort memory copying to do the message passing? No, to do the processes, and then pickle. So I think I might know why your message passing is the bottom line. Yes, I think that's a very good, absolutely. So shared memory is a problem for object passing because you still have the question of then who owns the reference count. So some kind of library where you could have immutable data structures and you have a convention that the receiving channel owns the message that's been passed. So it's responsible for the destruction. And then you could do reliable message passing between channels. Yes, I think that makes a lot of sense. So the version of Python CSP that uses OES processes uses sort of Unix shared memory type things, and that's still quite slow, partly because I think shared memory is more efficient when you're copying a large amount of data or copying data many times through the shared memory. It's not really intended for sort of one-off sends and receives, which is kind of what you're doing when you do a message pass in CSP. So it's not really the right tool for the job, whereas something like cross memory attach might be. Interesting. So where do you see the future of message passing in Python? Would it be something like PySTM or rather something like async.io? So async.io sort of does this kind of thing already, but for IO processes that you want to run in the background and for that particular use case, which is great. But for more general computation, I think it would be interesting to see message passing used together with Python 3.4 co-routines and see how that goes. I think that would be a really interesting experiment to do, and really interesting to benchmark that and see if it could get really fast and usable for the sort of, as it were, the ordinary programmer, rather than someone who's got a particular use case like background.io. So the PyPySTM, would that be any hope? So PyPySTM is a fantastic piece of work. As I understand it, the purpose of PyPySTM is to make the core interpreter concurrent in a sense, which means that you can then build these high-level sorts of concurrency that the programmer would see on top of that. So I would expect that, I hope PyPySTM is really successful. I wouldn't expect that that would mean that ordinary Python programmers rather use STM in their own applications. I think that's kind of the wrong level of abstraction for the programmer. So I think building message parsing on top of PyPySTM would be really interesting. I have another question. You already mentioned Staglars-Pyzen. Did you look into the co-routines in Staglars-Pyzen, the called tasklets and the channels provided by Staglars-Pyzen? So I didn't quite catch that. So Staglars-Pyzen is an alternative implementation of the Python interpreter. It already provides co-routines and channels and message parsing over these channels. Did you look into it? Yes, so I think my understanding is that Staglars has a different implementation of the Python interpreter. So it's not quite CPython, which is why it's Staglars. Actually, it is CPython with some additions. Yes, okay, we have some changes. So it's fully binary compatible with CPython. Yes, so I did see that that's also a really interesting piece of work. Yeah. Yeah.
|
Sarah Mount - Message-passing concurrency for Python Concurrency and parallelism in Python are always hot topics. This talk will look the variety of forms of concurrency and parallelism. In particular this talk will give an overview of various forms of message-passing concurrency which have become popular in languages like Scala and Go. A Python library called python-csp which implements similar ideas in a Pythonic way will be introduced and we will look at how this style of programming can be used to avoid deadlocks, race hazards and "callback hell". ----- Early Python versions had a threading library to perform concurrency over operating system threads, Python version 2.6 introduced the multiprocessing library and Python 3.2 has introduced a futures library for asynchronous tasks. In addition to the modules in the standard library a number of packages such as gevent exist on PyPI to implement concurrency with "green threads". This talk will look the variety of forms of concurrency and parallelism. When are the different libraries useful and how does their performance compare? Why do programmers want to "remove the GIL" and why is it so hard to do? In particular this talk will give an overview of various forms of message-passing concurrency which have become popular in languages like Scala and Go. A Python library called python-csp which implements similar ideas in a Pythonic way will be introduced and we will look at how this style of programming can be used to avoid deadlocks, race hazards and "callback hell".
|
10.5446/20024 (DOI)
|
My name is John Pinner. I am by team organized Picon UK and Euro Python 2009, 2010 we did. One of our team members is Richard, who is going to tell us all about DNS and Twisted. Thanks John. Thank you, thank you all for coming. Yes, so my name is Richard Wall. I'll start by telling you a little bit about me. I don't want to spend too long. I'm a Python programmer from the UK and as John says, I'm involved in the Picon UK conference as an occasional speaker and my most important job there is to organize the conference dinner table plan. I run that, I maintain the table plan software. I'm also an enthusiast, a real enthusiast for Twisted, which is a framework which hopefully you've all heard of. It's one of the oldest Python frameworks. And I have become over the last year a contributor and a core contributor, I suppose I could sort of call myself a core contributor and I guess the de facto maintainer of Twisted names which is the components of Twisted which I'm going to be talking to you today about. And I'm currently working in Bristol in the UK for a company called Cluster HQ where we're working on some software to manage the deployment and the state of Docker containers. It's a really interesting open source project that we're working on called Flocker. So you should look that up if you're interested in any of those technologies. I'm working on that with some of the Twisted founders which is super exciting for me. Okay. Well, I haven't got long this talk and I, when I did this in the UK last year, I over underestimated how long it was going to take and so I've cut out a lot of stuff. This is going to be a much shorter version if anyone saw the talk in the UK. I'm going to talk less about myself, less about the history of Twisted and more about the technology in Twisted names and also the project that I've been working on recently to implement EDNS and DNS support in Twisted. I'm going to start with an overview of DNS. Well, very short overview. I'll explain why in a minute. I'll give you a tour of the components in Twisted names. I'll hopefully give you some interesting examples, some quite interesting examples. I'll give you a status report on this project and then I hope to have some time at the end to answer any questions. I hope there will be some questions. So I had planned to give an overview of the domain name system but I don't think I'm going to have time and I don't think I need to anyway because you've probably already been to a talk on Wednesday by Lin Root who explained really clearly about the domain name system and its structure, its operation and the terminology and I think some of the software that you may be familiar with for serving and sending DNS requests. So she did a much better job than I probably can at explaining it and so what I'm going to say is you should just go and watch her talk on YouTube. It's a great talk. I didn't make it to the talk. I wish I had been able to but I watched it last night and I'm glad she did it because it means I can get to the interesting bits, the Twisted bits that I want to talk about. So I'm going to skip this, skip this, skip this, skip this, skip this, skip this and somewhere. I might talk briefly about the software that you may be familiar with just as a contrast to the Twisted name system which I'll explain in a minute. So probably you're all familiar with Bind which is the original DNS server. I think that's true to say it's the original DNS server. It's an authoritative DNS server and a recursive DNS server and a forwarding DNS server and all sorts of other gubbins that are mixed in with it and that's part of its problem. It tries to do too much. It's feature packed but it's over complicated and it's full of vulnerabilities. This idea of one binary to satisfy all the different DNS server requirements is a mistake which has been learned and implemented better in another piece of software called PowerDNS. So if you're using Bind then I recommend you go and look at PowerDNS which is a much more modern, much better designed, a much more secure DNS server. It's actually more powerful than Bind in a way because it has a much cleaner way of interfacing with a database back end for example and it also splits the duty of authoritative server from the duty of recursive DNS server which is important to avoid cache poisoning attacks. Other servers you may have come across are Unbound and NSD. I mentioned those because they are written by an organisation that I've been involved with this project that I'm going to tell you about, NLNet Labs. And again they are much more modern, much more secure DNS servers dedicated to, Unbound is dedicated to answering recursive requests, NSD is dedicated to answering authoritative requests. So let's now get to the subject of this talk, Twisted Names. So Twisted Names is kind of as old as Twisted itself. It's celebrating its 13th birthday this year. It probably started life as you may or may not be able to see from this check-in. This is the first change set where Twisted Names was first landed. It was probably introduced as a sort of demo of what was then new UDP transport facility in Twisted. I did a little bit of digging and found the commits from the beginning of Twisted Names life and some of the newer commits that I've been working on and Julian's worked on a bit, I see. So you can see it was originally written by a guy called Moshi Zadka back in 2001. And that was in the good old days when Twisted had a kind of wild west development process. Everyone was just committing randomly to trunk. And I guess they hadn't yet implemented what is now called the ultimate quality development system, which is a talk in its own right. It's a way of, it's a way, it's the way we develop in Twisted. It's the way of developing in branches and ensuring that every change that gets merged to trunk has been code reviewed, that it's fully tested, that the code is fully covered and that there's an audit trail showing between the ticket and the code that lands in trunk. But you should go and read about the ultimate quality development system if you haven't already. So Twisted Names was, it was kind of actively maintained to start with. And over the years, it kind of, I think it's true to say it's been neglected a bit. And then I started getting involved about two years ago and I had a background in DNS, so I thought that's part of Twisted where I could help out. And I've been busily updating the documentation, adding test coverage, adding some new examples to demonstrating how to use Twisted Names and you'll find all of those on the new Read the Docs documentation website. I'll link to that in a minute. So yeah, like the rest of Twisted, the Twisted Names package is really well tested. It's got comprehensive unit tests which are run using a tool called trial, which is a great test runner. The unit tests in Twisted Names, if you care to read the code, and probably not the most sophisticated unit tests in the world, but that's a reflection of the way these testing techniques improve over time. Twisted is, as I said earlier, it's over 13 years old. So if you do go and start hacking on it, you'll find that there are bits of it which are kind of hard to read, hard to look at without bursting into tears. But then again, you have to recognize this history of the project and it's actually quite interesting to see how particular developers who have been with Twisted from the very start have changed their ideas and their approaches to things like testing. I think that would be another interesting talk in its own right. It's interesting from the point of view of a new contributor who has to deal with this old style code and the new style code and understand what's the current best way of doing this, developing the Twisted. So we've got plenty of unit tests and we've got reasonable coverage of the code. Some of the modules, I don't know whether you can see it on the slide, are not very well covered, but those are areas which I'm working on now. Modes such as the authoritative DNS server and the secondary DNS server are not particularly well covered. And in fact, that's sort of really its ugly head lately in a bug that's been discovered in the secondary name server in the latest release of Twisted. It's partly down to a lack of test coverage. This bug wasn't noticed earlier and it's partly down to the old style design of that part of the package. Hopefully I and others are going to improve that over time. Twisted names wasn't particularly well documented, but that's improving. As I said earlier, I've been working quite hard on improving the documentation for Twisted names and Twisted as a whole is better documented these days. You can go like most of the projects these days and read the docs on Read the Docs. And it's nicely presented and nicely indexed and easily searchable. So I recommend if you're interested that you go and read the documentation for Twisted names and for the rest of Twisted because it's much easier these days to navigate the documentation. So how am I doing for time? I'm running out of time. I'm going over time. I'm running out of time. So let's crack on. We'll have a look now at the different modules in Twisted names. I'm going to start at the lowest level and work up. Like everywhere in Twisted, there are layers of abstraction, layers upon layers and on layers. And we'll start at the bottom and give some examples of how you can use these low level APIs and then later we'll see some of the higher level abstractions. So let's start with the Twisted Names DNS module. Now this contains protocol level APIs, representations of the DNS records, representations of DNS messages, routines for serializing and deserializing these messages from the wire. And it's also in this module that you'll find the protocol implementations both for UDP and TCP because DNS operates over both of those transports. So we've got a little example here which I'll try and talk you through. Can you see that? Let me try and zoom in a bit. This is Reveal.js and I'm not sure whether it's going to... Oh, yeah. Yeah. Can everyone see that? Great. Okay. So what we're looking at here, there's a couple of things I need to explain. We have got, first of all, let's start at the bottom and look at the last line which is task.react. Now if you've used Twisted before, you may not have come across this API, but this is a new way for you to start the reactor for a short-lived Twisted program. And what it does, it supplies a method, you supply it with a function that you want to run, a function which must return a deferred. And TaskReact will take that function and supply it with the reactor, run your function and then wait for the deferred that it returns to fire. And upon firing, TaskReact will then tear down the reactor, take care of stopping the services in the right order, and it will log any errors that haven't been handled on that deferred. So if we then move up to the main method, we can see that having supplied the reactor, we're instantiating a DNS datagram protocol. So in this example, we're only going to be, this DNS client example, we're only going to be dealing with UDP. And we instantiate the protocol and then we pass that to React to listen UDP on port zero, which means any high ephemeral port. And because this is UDP, we don't have any connections. So we have to, so it may look odd to be in a client calling a method called listen, but we are going to send a UDP datagram and we have to be listening for the response, whereas in TCP, the operating system would set up the connection for you and you wouldn't have to choose the ephemeral port that the response comes in on. So we listen on UDP port XYZ and then we send a query using the protocol, using, we call protocquery to send our DNS query to, in this case, the Google DNS servers. And then when that query has been answered, we're going to print the result and that's by way of adding a callback to the deferred returned by protocall.query. Again, I sometimes think we should start every talk with an introduction to deferreds, but I guess everyone's heard it and more people these days are familiar with the idea because it's now part of JavaScript, I think. So yeah, when the answer comes back, we simply, we take the result and we take it, the result is a message, a DNS message, which I'll explain in a little bit more detail later, but a message has three attributes, has an answer's attribute, an authority attribute and an additional attribute. And those represent the three categories of records that might be returned by a DNS server. And in this case, all we're doing is printing out the answers returned by the DNS server and in particular, we're returning, we're going to print out the payload, which is the, either the, in this case, the quad A record, it might be the A record or the MX record. We're not interested in printing out the header information which wraps around that payload. And so I'll show you the, show you the output. Oh, I'm running out of time rapidly. Okay, so we've got an answer from the server, a quad A, one single quad A record. Now let's quickly move on to the next example. So that was a client. The next example is a server. And in this case, it's quite similar, but we instantiate the datagram protocol this time with a controller which takes care of handling the, the, the, the, the query which comes into our server. And when a query is received by the protocol on port 10, 10,053, our protocol then calls out to the controller and calls its message received method. And it's the message received method on the controller, which is responsible for, for constructing an answer to that message. So this is how we write low level servers in low level DNS servers in Twisted. And in this case, we're just going to respond with a count A record with a, with a fixed IP address. So hopefully that makes sense. I haven't got time. I'd like to go into it in more detail, but I haven't got time. There's the, there's the server running and there's us issuing a request to it using dig. Okay. So now, those are low level APIs. If we move up now to twistednames.client, this is a much higher level API, a much more friendly way of interacting with Twisted names. And in this example, we're going to, you, we're going to introduce a couple of new concepts. We're going to use Twisted names to look up concurrently the, the reverse DNS records for a whole class C network. And so you can see in our main method that we are constructing a list of all the IP addresses in a, in a, in a slash 24 network using a really useful module called net adder, which does all the, the construction of those reverse DNS names for us. I think there was a, I think there was a lightning talk on it yesterday. So I recommend that module. And for each of those reverse DNS names, we're going to call client.lookup pointer. And client has a, has a series of these lookup methods, one for each type of DNS record that you can receive from a DNS server. It doesn't have all of them that we're working on implementing some of the missing ones, but it has a lookup method for almost every common DNS type. And so we construct a series of, a list of deferreds, all of them in flight. And all of them are then added to what's called a deferred list. Now a deferred list is a, a really useful API for collecting the responses to a list of deferreds. And then it fires its callback when all of the deferreds have themselves fired or, or failed. So when we handle the results in this example, we are looping through the results, checking whether the result was a success or a failure. And if it was a, if it was a success, we print the, we again print the payload. And we also print a summary, summarizing how many of the requests were answered successfully and how many of them were not answered, either because the record didn't exist or perhaps the, the, the query timed out. So the, the results to that are as follows. So you can see it all happens rather quickly. And because everything is happening, happening concurrently. So that's a real advantage of using Twisted for this sort of work. I might skip now to a better example of that. One which follows on from Lynn's talk on Wednesday. So let me, let me quickly summarize some of the other modules. We have, in Twisted, we have the modules for, for creating DNS servers. I had an example of that. And it's really easy to use because there's a Twisted DNS plugin for the Twisted command which comes with Twisted. And so you should explore that and explore all the options that you have using that command. Twisted itself runs its, TwistedMatrix.com, that domain is actually served from, from a Twisted DNS server. So it's, it's, it's pretty, it's pretty stable. And it's, it's not, it's not a, it's not a fully featured DNS server, but it's good enough for some cases. You can see that when you start the server, it logs to standard out and that we can query that server once it started up. We also have an authoritative server and it's interesting, but I haven't got time to go into it, that you can load a DNS zone based by defining it as a Python module. And so here we have a Python module with describing the zone, but the, the, the, and the, the objects that you see there are all globals which are imported at the time that this module is evaluated by Twisted names. It's quite a clever mechanism, but you should look into that too. It's an interesting piece of code. And again, there's the example of how it runs. And again, these examples are all on the documentation, the Twisted documentation site. So I recommend you go and read those. There's a bunch of other modules which I have to skip through. Common contains some helpers and some, some APIs common to all of the Twisted clients and servers. Resolve, I won't go into cache, is about caching the responses to, to queries. Root is about doing recursive DNS resolution, which Lynn talked about in her talk. C is about transferring zones and serving them authoritatively from another authoritative server, which some of you may be familiar with. And the point I wanted to make by describing all of these is that these, all of these building blocks can be put together in interesting ways. And I've done a couple of examples of this on the website. For example, you could create very easily using the low-level APIs, a module or a script for test, for compliance testing of DNS servers or clients, because you have complete control over the, the, the flags and the, and the payloads that you put into messages. So it's easy to construct non-compliant messages to see how DNS servers respond to those. Or it's easy to, to, to, it's easy to see how clients respond to non-compliant responses from servers. So that's a, that's a good use of these building blocks. We use the, we use Twisted at Work for, Twisted Names at Work for functional testing. So we have a bunch of code which does DNS lookups. And we want to, in our tests, supply canned responses to those DNS lookups. And it's very easy using Twisted to set up a, set up a lightweight DNS server and then tear it down at the end of the test. It would be really easy to set up a database-backed DNS server or a DNS server which looked up its data from a REST API, for example. Or it'd be easy you to combine these with other parts, other components in Twisted, like the Web module or the LDAP module, to look up DNS records from an LDAP database or to, to control and manage the DNS records in your server using a REST API. So now let me see if I can quickly show you, finish off with a quick example, a more complicated example of using Twisted Names. So in Lin's talk, it was really interesting, there's a tool called DNS Map which is a tool for, for brute forcing a zone, for guessing which names may be in a zone, a DNS zone. And it does that using a dictionary of words or common sub-domains. And as, as you said in your talk, it does it in series. It's quite, quite a slow, it's quite slow to complete because there's about a thousand words in its dictionary. So I thought it'd be interesting to write a tool, write the same tool in Twisted because it can do all of these lookups concurrently. So this is, this is how DNS, the DNS, the original DNS map is documented. You pass it a domain name and you pass it, you can pass it a list of words which it will then look up each of those words as a sub-domain of the supplied parent domain. But as you can see, it is quite slow. So I started this going against Spotify.com and 48 seconds later it had only reached G. So it was going to take forever. Now I want to compare that to another example, the example I wrote which, I don't know whether I've actually given you the link, but I've put all this code on my GitHub page. I've got the link at the end of my talk. And in this example, we are, we're actually sending all of our requests concurrently, but in, not, not, we're not sending a thousand requests at a time. We're using some, another interesting part of Twisted called the co-operator, the co-operator API in the task module. And what we can do using that is to limit the concurrency so that we can say there's only ever a hundred in-flight DNS requests at a time. So we're not going to overwhelm the DNS server that we're querying. I haven't implemented in this the random time-outs that DNS map actually puts in between its requests. So it's not quite the same. But if I show you the results, in this case, it's looked up all, it's about a thousand sub-domains in two and a half seconds. So that's a good demonstration of the power of Twisted and the power of the APIs and the way that it can efficiently process, efficiently send out requests and process the responses. All this code is on my GitHub page. Now I think I've run out of time, so I wanted to talk about my project. I'd love to talk to you about it. I'm going to be sprinting at the end, tomorrow at least, on Twisted. So if any of you are interested in helping out or learning about the development process, if any of you are DNS experts and want to help me with my project, then I'd love to hear from you. It's all about EDNS, DNSSEC, I've got funding, so you might get paid for it. I've made modest progress and I think that's the summary of this talk. It's the summary of what it's how I wanted this talk to be, but I haven't had time to cover it all. So those are the links, those are the links to the documentation, that's the GitHub link to the examples in my talk if you want to investigate those. If you, I'll put these on GitHub as well, and I'll link to these from Twitter or somewhere. I'll make them available on the conference website. Have I got any time for questions? No. If there's any questions, catch up with me afterwards and I'd be delighted to talk to you about it. Thank you. Thank you.
|
Richard Wall - Twisted Names: DNS Building Blocks for Python Programmers In this talk I will report on my efforts to update the DNS components of Twisted and discuss some of the things I've learned along the way. I'll demonstrate the EDNS0, DNSSEC and DANE client support which I have been working on and show how these new Twisted Names components can be glued together to build novel DNS servers and clients. Twisted is an event-driven networking engine written in Python and licensed under the open source MIT license. It is a platform for developing internet applications.
|
10.5446/20021 (DOI)
|
Hello, and this year we have another contest winner and it's a child 13 years old and because it's a contest driven in Germany for German speaking, some slide here is mostly in German, but we will explain what it is and maybe we get some more languages or countries doing this also. My name is Reimar Bauer, I'm one of the initiators of this project, I'm also a board member of the Python software for Band and with me there is also someone who has a marvelous idea how to do this that school year students are interested in Python. This is Peter Coppards and he says also some words about himself. Hello, my name is Peter Coppards and I have developed large parts of the content from the course material and I have held some courses and I'm very proud to see some boys and girls again who has this course, participated in this course. Okay, so there will be some more actors today, then after some little bit explanation then we have an interview with Mika Greif, this is a winner, I just show him to you. Okay, my name is Mika Greif, I'm 13 years old and I come from Berlin, I go to the Bettina of Arnhem and my hobby is actually to do you too and I used to play tennis in a club. Okay, so that's I was in German, so that makes me it easy, we have someone who will be maybe visit tomorrow at the next other midday, at midday there will be some training for us and this is by Thomas Fabula, he makes us with some sports happy here and also he's very good in translations and so on and I know he speaks a lot of languages and he helps us today that Mika just had not so much English lessons yet that he helps with translations, so he gives for now a short summary. Thank you Reima, thank you Mika. So Mika is 13 years old as you understand and he is a fan of Jujutsu, this is martial art, if you are interested in martial arts we will have the opportunity to have tomorrow and the next three days some activation of your cognitive and your mental and physical health but now back to Mika, I asked him he was also doing six years tennis and he's going to the Bettina of Arnhem school here in Berlin and now yeah is a question how did you get to programming and what was nice there and what is the next things? How did you get to programming? I got to programming because my father is a professional programmer and because I used to have a robot, a meinstorm Lego that you could program that he would go forward or sideways and I thought that was pretty cool and I wanted to learn a little more from my father in javascript or c-sharp or whatever. Okay did you understand? You have to speak a little louder, okay, again very briefly, last three sentences. My father is a professional programmer and in the past I had a robot, a Lego robot in meinstorm and he could program that for example the motors can go forward or backward or sideways or can shoot and yes of course I wanted to learn a little more about programming, my father asked and then he brought a little more about javascript, python and c-sharp, I started to program him a little bit, for example 1 plus 1 is 2. Okay very good, so his father was the man who was the mother who motivated him, he's a professional programmer, I heard about something like embedded systems or things like this yeah because he mentioned Raspberry Pi, I think we have to look for Raspberry Pi for meca because Raspberry Pi and python is a very good combination, he used to program with Lego Mindstorm with a little orbiter to not for surveillance but for remote control and a little bit with javascript I think and also he told me besides with Unity game engine, okay, but how did you notice of this competition and this PiMove 3D 3D, how did you learn about it and how did you actually get to it? I got to the competition because my father and I were in Chemnitz once in the Linux days, so we had a few workshops in Stilbert and then we met Peter Koppatz at the Blender tutorial about low poly and he told us at the end that there is a PiMove Fed-bewerb and I wanted to try it out and see what it's all about. So it was also in Berlin at the Linux days, I think it was some months ago, we could have met there, I was at the Python and Plown of Ebruz also and Peter Koppatz, you already have heard him, they had contact to this competition and he wanted to move something and he was very interested to get something done and he knows a little bit Blender I think and this is the question, how did you get to Blender and Python and what exactly did you do? Maybe you say two words about the topic, what did your program and why do you use Python and Blender? I came to Python and Blender because I wanted to develop a game and then I visited a game engine called Unity, I developed a few games with it, but of course I wanted to bring in my own object, such as a cylinder or something, then I came to Blender and a few months later I realized that Blender also has its own game engine, I then tried it out and realized that it was powered by Python and then I learned a bit Python, my father sometimes asked me how it works with the classes or with the threads and so it was. So he mentioned that he wanted to move something and he was used with Blender, I think with 2.69, but he had to switch to 2.7.0a this version because only this was running with his Python program, but he was very enthusiastic in getting some things moved and I think he's a dynamic boy, so applause already now I think for him. Okay. So last question, maybe something to the technical background. What have you done technically in the background and what can you improve and what does the future look like, what can you imagine, what should you do, how is the future, maybe you can give an outlook. So I could improve the drops, the drops of the ice drops, also fall down realistically and not just the same speeds all the time, but at the beginning they are slower, then they get faster and faster, or for example that the ice drops have a realistic shape and I improved in the last update that such sounds, as soon as a drop falls down, then a sound comes, and when it comes down again, a sound like a record, and yes. Okay, so you will see what he has done, his creative work, he wants to improve it, of course with sound, with more dynamic moving icicles or falling drops, he wants to get more physics into this whole stuff, I understand, and he wants to make it mostly realistic. But I asked him how long did you program and how did you have time to solve this problem, it's not a problem, it's a challenge, it's only three weeks, three weeks, three weeks, three weeks, is that right? Great, thank you Mika, great. Please stay here, there is still the free contact. I know, thank you very much, nice interview, and please take a seat, and you also, thank you. Yeah, now you have heard how much interest such approach can be, and maybe we just wanted to tell a bit more about what it is, and maybe you have childs or you know childs, and next year maybe we can have an international contest or again a German contest. The idea is we need some platform or some program, and we want to use Python on a computer, yeah, then we have several platforms and it's not so easy to get the same version of Python on each platform easily, but if you do this with Blender, there's no problem. And Blender has a very good documentation also, and it has a Python API. So the contest is about to use now Blender, but don't click or use any of these buttons tools, make everything in Python, and then just do by your Python program a movie, and the movie you have to submit, and also the Python program, and also screencast and some other things, and so after the submission, your Python program must make the movie, so that has was done by Mika also. And this is how this wonderful Blender IDE looks like, and Peter wants to explain a little bit more about this. And you see also, yeah, there's a movie in it. At the first, I have two interesting news. The first one has nothing to do with Blender and this contest. We have found in a library an old paper, here it is, the origin, and you know we have the anniversary of William Shakespeare this year, and you know the famous words to be or not to be, and that isn't true anymore. On this paper, I can read, to use Python or not to use Python. This is the question. And this is a maxim in our contest. And the second interesting news is that we have to put Python on every operating system, and this is not true for all operating systems by default, and we have to put a Trojan horse on the operating system to get Python to work. This little effort. And so we decided to use Blender. And if you didn't already know, in Blender is integrated a Python, it is the newest Python version 3.4. And it's not visible in the first place, you have to switch some windows and then you have an IDE. As you can see here, as an example, you have an editor, you don't need to install Eclipse or such some. You have a Python console, you can use directly, and in the end, you can construct your own world, not only textual, hello world, you can this hello world print in 3D and other nice stuff. And this is the goal, to have fun and to move in the end all the objects. But you can do really serious things in science and in all, in all, in all, in all, in math, in chemistry, in art, at school, you can learn Python and combine the programming experience with other themes at school. Here you can see the eyes drops from Mika. And can we run this again? I have a longer one. I'll show you this later. So, and also, you maybe want to read about a bit and use Google translate to English, because the course material, we worked on it to make it in English, and we want also your help. Maybe you have time for a sprint at the weekend and make a translation or documentation sprint to get all this material we have into English and maybe then later to other languages. Also, we have this duplicate because we like many places where this course material is on the net. Also the Blender documentation with this Python API is very, very good and also a lot of people are working there. And so what you get is we offer a platform for newbies to learn Python. And they directly learn object-oriented programming because in Blender, everything is a visible object. You can just rotate it, scale it, and just, but this explanation of methods is very easy and you don't have to explain first a hello world example, you just make things flying. And you just produce movies as a result. And you have a very different way to present also your results. You have seen, I have a method, a video in my talks. And just to make you more or get more of your interest, there is a Blender training today after lunch. And if you like to be a trainer for PyMove 3D, it's just the only thing you need is basics of Blender. So if you take this training, you maybe are able to give a course like we did. And to find new interesting people doing so or some students learning, starting to learn programming. And it's a nice task to see what you can get from this. And also we have a much longer talk about this course material and how we want to proceed tomorrow at 2.30. And last year we had also two winners for the contest. And you can meet and talk to them at the poster session today. This was just on the beginning. And this is a submission of Mika-Kraef. It has also some, you see, icicles building and then falling tops. And so we want to know to give him also his prizes. So this will be done by Armin Strostacinski. Hi, Mika, come with me. We have some prizes for you, some placeholder prizes. We have more downstairs in a box. Okay. First of all, you get a certificate of winning the contest. And there are signatures from the members of the board of the Python software for bond. And we give you a second one, which is a little bit stronger for your hall of fame of all your great programs you write in the future or hidden in the past. And this is a small book about Python, which you can keep handy if you are on the road. And if you want to look for some modules or something like that. This is a bigger one. And this is a book about Raspberry Pi and programming with Python. We have a Raspberry Pi complete with additional hardware downstairs in a box. I think it's so much you cannot carry on the stage. Do you want to say something? Thank you. Thank you. It's a quite simple statement. And I'm seeing the next generation contest winners already outside in the room. I don't look in the direction of Torben, for example. And I hope you enjoyed the stuff. Do you want to say something? Thank you for the audience. We hope to have some more winners next year. And of course, Mika will continue Python and Raspberry Pi. Thank you all. Thank you.
|
pymove3D Winner Announcement
|
10.5446/36879 (DOI)
|
Ladies and gentlemen, good morning. Welcome to Europe-Ice in 2014. I hope you all had a great morning and like us, can't wait to start with Europe-Ice in 2014. First I want to introduce us. This is Tony. My name is Karina. We will lead a little bit through Europe-Ice and by giving some organizational updates every morning so that you're always up to date what's happening. And so we just start with this today. All right. So the first news are kind of sad because you didn't get your guidebooks yet. The guidebooks are going to be around tomorrow and you can pick them up probably around noon from the desk. If you need some offline material to look at the conference schedule, there are displays around in the foyer, in the entrance and in the basement. If you need wireless, if you need Wi-Fi, the SSID that you should be using is EP14 and the password is EuroPython2014 with a capital E, capital P. This is going to be in the guidebook and it's going to be printed around all the place but if you're missing it right now, EuroPython2014, capital E, capital P. So I see some of you already got their bags with the t-shirts. If you now decide that you perhaps choose the wrong t-shirt size, it is possible to switch the t-shirt but just tomorrow after the lunch break before it's not possible. Also you can't take a bag before with you then because yeah, we have to make sure that everybody gets a fresh t-shirt and we can't guarantee then if somebody brings it back that it's not already worn so we don't want to have that. So if you want to change your mind about your t-shirt size, just come back tomorrow afternoon to the t-shirt desk which then will be at the info desk. All right, that's already from us. I want to introduce you to the chair of EuroPython 2014, Mike Miller. Thank you very much for the introduction. Good morning everybody and I also would like you to welcome to EuroPython 2014 whether European-Python community meets. First to remind you, we are in Berlin. Welcome to Berlin and it's the year 2014. And obviously you found a place that I don't need to introduce the place to you. First I want to start a little bit looking back. The EuroPython since 2002 and then I discovered some data about this conference at the EuroPython Society website and I used this data. And I like to program and this is why I put a little bit of programming in this talk and actually this whole talk is written in an iPython notebook and so there is this data there. And this is the last EuroPython over the last years. This year is starting 2002 location in 10Ds. Unfortunately we don't have data for all the times but we have a few. But this is just the numbers but now you use a little bit of Python, we re-answer data and put them in dictionaries and after we do this we can plot them very easily. It looks like this after we put in a dictionary and then you have a nice plot with Matplot live and that's how the development over the end of the year went over the last years. See we are rising. A lot of people are coming to the conference and this conference has a whopping 1200 participants. Okay, so much about the past but we want to look at the future and look at the next week. And first I would like to invite Fabio Klichor, the chairman of the EuroPython Society to say a few words and say hello also. Hi everybody. I hope you are doing well and I would like just two words to say that I am really happy to see all the people here. You can see but it's a lot of people. And actually I will take a picture and tweet that. Okay. Hands up. That's nice. And I'm the current chairman of the EuroPython Society and the EuroPython Society is basically a society founded 10 years ago to help to maintain the carrier of the EuroPython conference during the years to maintain the way the EuroPython conference is set to keep it a community conference, an open conference. And our job now is to try and do this as the conference scale. And for us to do this we really need your help and we cannot do that by ourselves and we would like the community, the EuroPython community to be as open as it could, the more open as it can and for that come, become a member, be part of the conference and help us to keep scaling and to keep the same kind of conference. The next days on Thursday we will be presenting the PSF and presenting the new ideas, the EuroPython Society, sorry, and the new idea we have in mind to keep the conference serious. So come and join us. Thank you. Thank you very much. So do become a member of the EuroPython Society. And that's what we're going to do the next days. That's a program that doesn't even fit on the slide, there's even more, but I want to go through quickly what you can expect. Of course the main part or big part of the conference are the talks. So we have five prior tracks of talks. We had about 300 submissions, 300 people want to speak. Unfortunately we have only room for 100, so that means we had a pretty strong review process and you can expect very high quality talks. So you saw it already in the program. It would be very interesting talks. I would like to call all of them. But they all will be on video so you can watch them later. Also about the talks I have a call. So we need people to help us to run this conference. It's a very big conference and we need session chairs. So if you are interested and would like to be session chairs, session chairs are the person who attend, helps the three speakers or how many are in the session. And please sign up, go to this website or if you cannot write it down, ask at the info desk about becoming a session chair. So it would be very helpful because we need people to help us with the conference. Also if you are a speaker just a short announcement, there's a speaker preparation room in CO3 that's up here. It sees this level up here. So if you need to prepare for your talk, you need some quiet room, then you have to speak over. But please become a session chair. That would help us a lot. It's not a lot of work, it's just one session and then you are done. It's a help for the conference. Training sessions. So we have training sessions. There are two parallel training tracks. So very interesting trainings. They are about three hours. So if you want to get a bit deeper on the topic, then a talk, then the training is the way to go. But you need to sign up. So you had a chance to do online sign up before you might have seen it. If you haven't done so, please go down to the info desk and sign up. If you sign up and you are on a list, you are guaranteed to have a seat. If you don't sign up, you might not get a seat in the training. I might not see you in space, you might be lucky, but if you really want to see a training, please sign up. That helps you and helps us also to plan a bit about the rooms and stuff like this. So please take advantage of those interesting trainings and sign up for the training session, for your favorite training sessions. So we have a lot of things going on. One of the things going on right now today, the Jungle Girls workshop. Yes. Right now at the base level, there is about 40 women learning programming with Python and Django. We had about more than 300 applications for this, but we have limited room obviously because we have a two-tour student range of one to three. So we have limited room and limited two-tours for sure. And therefore, we could only select a fraction of them, but it's a small sort of event, so it's free for the participants. It's just once there's a whole application, they don't have to pay for it, which is a great thing. Keynotes. There will be a lot of keynotes. We start with two keynotes today, and I just put like a hashtag here. So today, we will hear keynotes about Snowden. I just put Snowden there. Everybody knows what it means. Haskell, test-driven development, something about decentralized systems and big data. These are the keynotes we want to enjoy over the week. And you see the keynote speakers. I'm not going to introduce them right now. That will be done right after me when the next keynote comes. Lightning talks are very interesting, though if you don't know lightning talks, it's very short, five minutes per talk, and it's very short. There's no answer question thing, question-answer thing. And if you're not done by five minutes, you'll be able to turn on the microphone and the next person's talk. You have to sign up. So if you wanted to give a keynote about anything that should be mostly Python related, just go downstairs to the input desk. There's a flip chart. Put your name in. And it's a first-come, first-served basis, so you put your name in. You have a chance to get in. Don't miss the lightning talks for many people. That's one of the favorite parts of the conference, and very often they're very interesting because they're very short and precise. And I enjoy them very much. Postal session. So we decided this year I have a postal session, so there will be a postal session. It will be today and after noon, so after lunch there will be a postal session. And this is a chance to talk to an author. It's a bit different than a talk. A talk, you have a few question-answered talk. A poster is much more personal because usually it's just you and the author of a few people. And it's a very good chance to get deeper in the one topic and a discussion. And also, if you're a poster presenter, you will have a chance to introduce your poster very quickly with a one-sentence thing. You just say your name and the title of your poster and what's about a few seconds. And this will be right after this, in this recruiting session. So if you look in the schedule, you'll find the recruiting session. We have the recruiters in the two-sentence and you have a chance to do this. So if your poster presenter take advantage, of course this potentially attracts more people to your poster. More. So I mentioned recruitment sessions. Though I just put the three main sponsors here. We have 44 sponsors, and I didn't want the pluses to go with the sponsors. You will see they're all over the place. So thank you very much to our sponsors. They are very important for this conference. I can say this out of the sponsors, the conference wouldn't be possible as it is now. They would have been much more expensive or whatever. Or wouldn't have happened at all. So sponsors are very important. And they are a recruitment session. So if you're looking for a pysonger, that's a good chance. Or you want to know what those sponsors are doing. There will be this recruitment session in the afternoon. And they should come. And the every sponsor has a three-minute short intro with a slide or two about themselves. And then after this intro, that was just mentioned, poster introduction. Then there will be a poster session. And this poster session, the poster session will be in parallel to the recruitment session. So all the recruiters will be there with a small booth. And you can talk to them in person if you're interested. And please take advantage of this opportunity because these are companies that are interested in good pysong programs. And if you're not interested in recruitment, you can talk to the poster people. And the poster session is running longer. So you still have a chance after this recruitment session to go to the posters. Next thing, sponsor booth. There will be booth over the place. You see when you come in, in the entrance area, there's a big booth. A one-level lower. There is a nice place to sit and relax. There's a night light, a lot of nice things, and a lot of other sponsor booths. So I encourage you to go by the booths, but the sponsors, because as I said, the sponsors help us to run this conference. And you should speak to them. And they have some nice specs. They have pencils. And they have a lot of other nice gimmicks. So you might take something home for yourself or for your kids or whatever. Okay. That's the sponsor booth. There's a social event. On Wednesday night, there will be a social event. So there will be very good food. I heard they have very good food here. And also, after the social event, there will be a club downstairs. There's a club. And for people that still have energy left, they can come and join as a club. The great food, I say, that's a very nice culture program. I just put a picture of the guys that are going to run the program here. So looking forward for a very interesting program. And also, it's a very relaxed atmosphere. And pretty much everybody will be here. Also partners are invited. So there can be different, not technical conference, but relaxed. And they're still possibly as far as to buy extra tickets. So if you like to, somebody just wants to come for this event, there's still some tickets left if I'm not mistaken here. The partner program, very important. So if you didn't come alone, there's a chance if you have your partner with you or your family to explore Berlin. So it's pretty inexpensive there because it's a big package of things you can do, including a cruise with a boat here and a picnic, museums and other things. Everything packaged for the whole week, something to do. And in a group to experience Berlin. So I encourage you, if you have somebody with you who might be interested to enroll them in the program, you can still buy tickets downstairs at the infodesk. Pie ladies, lunch and barbecue. So the pie ladies are involved in the organization of the conference and the conference in general. And there's two events I would like to pinpoint here. It's a lunch on Tuesday, which is tomorrow and the barbecue on Thursday. And you need to sign up if you want to go there. So ladies, pie ladies or potential pie ladies, whoever would, are interested in this pie ladies, go there, sign up and meet like-minded people at these events. And there's some food also, good food. Sprints. So the conference is five days as we see this talks and trainings. On the weekends we have sprints. So Saturday and Sunday, please sign up. That helps us to organize things because the sprints will be catered, but we need to know how many people are there. So if you're interested either in just taking part of the program or even just on session. So if you want to have your topic covered, go to the website that will be wiki. You can put your topic in. Just register as they put your topic in and then hopefully a few people come to work with you on your topic. So join, there's quite a few topics up there already. Join this session that's up and help them develop the Python software on this two days. There's also bar camp. So if you don't like this kind of very traditional format of a conference or if in addition you like something else and they have a bar camp which is sometimes called an unconference. So it's a different type of conference. You have one hour sessions and there's no pre-planned schedule or talks or anything. People meet in the morning and they decide what they're going to talk about. That can be something like more traditional talks. That can be very open discussions. Sometimes you want to know about the topic and people teach you. So it's a very interesting format and I encourage you to go. This is bar camp. It was very interesting for me. I attended a few and all this very enjoyable. So please come. There's a bar camp. There's also a bar camp tool so you can also please sign up there so that you know that you're coming. Okay. Your Python is not alone. We have a satellite conference. It's called Pi Data. So if you're interested in Python and big data, then Pi Data is for you. It's an event running, has been running many times in North America and once in Europe in February but now it's in Berlin. So if you're interested in this very hot topic, big data and Python, go to the Pi Data. We can meet all the folks that are instrumental in this field and talk to them. It's also on the weekend, so on Saturday and Sunday. Good. That's not nearly it. I just put together a few more things. So about the name tag, I just think so if you get a name tag, they have versions with thicker paper and I heard they also have some kind of enforcement papers. Of course, it's a little bit thin. There was a small glitch when they printed it. At the back, we said already, so it's kind of double here. We're streaming. Everything will be streamed. If you're interested in things after the conference, please go to the website and look for this buckling thing that's some sort of a heads-up to organize. If you want to go to some pub and you want to meet some other Pythonistas, just go there, sign up and then the chance is higher that you meet other people that are interested in Python and just hang out in the evening somewhere. And those of you have a bunch of sponsor events, so please go to the schedule. There's a few sponsor events that you can go to. They have some very interesting things. Go there and some of them you need to sign up. Some of them you can show up, but please go there and encourage you to go there also. Okay. That's what I just said. Quick run through. And this is what I would like to do. I would like to have a nice conference and enjoyable in. Thank you. Thank you.
|
Get first hand information on future research directions and learn about innovation and its transfer into industry from leading researchers at ETH Zurich. The Industry Day is an annually recurring event which showcases the research activities of ETH Zurich and offers a platform for industry to engage with ETH researchers. We aim to cover a broad range of research topics in order to provide a comprehensive overview of the activities at ETH Zurich.
|
10.5446/20014 (DOI)
|
Hello, welcome to my talk, log everything with LogStash and Elasticsearch to begin with. Just raise your hands, who uses logging in your applications? Yes, that's great, who uses a central log server? That's okay, I hope we'll be some more after this talk. Little bit about me, you can follow me on Twitter, you can get the slides on GitHub, I'll post them afterwards and of course you can visit my blog at peterminusoffmann.com. I'm a software developer at Blueyanda, we are a sponsor of this event. Blueyanda is the leading software as a service provider for predictive analytics in the European market. We have our headquarters in Karlsruhe and offices in Hamburg and London and about 120 employees. We use the full Python stack for development, we use Flask for web front end, SQL alchemy for database access and the pandas numpy cycle learn stack for our machine learning tools. Most of our core algorithms are written in C++ and executed on a custom parallel execution engine and of course we are hiring. So log everything. When your application grows beyond one machine, you need a central space to log, monitor and analyze what's going on. LogStash and Elasticsearch store your logs in a structured way and Kibana is a great web front end to search and aggregate your logs. Just a little disclaimer, I'll talk a lot about LogStash but I think the same accounts for Greylog. Greylog is also a great tool to collect your logs and I think they have similar strengths and some differences. So what do you need if you want to have a central logging for your applications? Of course your log producers can be your front end, might even be a JavaScript single page application which uses a custom API to ship the logs to the back end. It might be some API or back end service, might be an authentication service, might be even a database system or the operation system. You have to transport your logs to a central station. I think everybody knows Syslog. I'll talk a little bit about GALF, that's the Greylog extended logging format but you can also ship your logs via RedisQs or via the RabbitMQ system. You could even log to log files and pass them back with the regular expressions but I think you have more benefits if you log your messages in a structured way. Then you have to root and filter your logs. You can do this with LogStash or with Greylog2 server and of course you need some storage where you can store your log files. I think Elasticsearch is one of the great open source tools. It not only allows you to search your logs but do all kinds of analysis based on your log files. To do the analysis you need a front end to access your logs. I'll talk a little bit about Kibana. It's a JavaScript only framework. The Greylog2 server has bundled a web interface but you can also use the plain Elasticsearch head, it's a JavaScript application tool or even use Python with the Pius library to build custom queries and reports against your log files. What I'm going to talk today is the logging chain to transport your logs with GALF to a LogStash server. The LogStash server pipes them into an Elasticsearch engine and you access the logs with the Kibana web framework. It's the pattern transport, root, store and analyze. If you need to grow further you can scale each part of the system. You can add more nodes to Elasticsearch. You can use multiple LogStash instances or even add a message broker in front of your LogStash. The message broker collects the logs and then ships them to LogStash to handle the load better. What is GALF? GALF is the Greylog extended logging format. It's basically JSON over UDP. That means it's non-blocking. But it avoids some shortcomings that you have from plain Swiss log which is also text over UDP. It's not limited to one kilobyte. I know Swiss log and GE can handle more but the plain Swiss log can handle one kilobyte. Often one kilobyte is not enough especially when you do application, monetary logging because of backtraces and you just have more data. GALF also adds structure to your logs. You have a key value relation in JSON and it has compression built in. The possibility to add chunking, one log message can be chunked to I think about 120 messages. Also Swiss log per default has no support for additional fields and metadata. In GALF you can add arbitrary fields and arbitrary metadata to your log messages. I think GALF is a great choice for logging with applications. Of course there is the GreyPy Python handler and clients for all kinds of messages too. One thing you have to consider when you want to log with GALF because it's sent by UDP. It's not reliable. If your network is flaky or if the server is under high load, messages could get lost. If you really want to get sure that your log message arrives at the server, you have to consider different transport formats. Like I said earlier, RADIS or the RabbitMQ system. What does log message in GALF look like? You have a mandatory version field. You have the host field where the log message comes from. You have a short message. You have a timestamp. You have the log level. Then you'll have an arbitrary number of custom fields like a facility or some request ID. I'll talk about this later. How to use this with Python? It's pretty straightforward. Works with the standard Python logging. You just add a GALF handler host and port. You can just log as normal. The handler will push it into your GALF-aware service. In our case, to log-stache. What is log-stache? Log-stache is a tool for receiving, processing and outputting logs. It's written in JRuby and runs in the Java virtual machine. It's based on the pipes and filter pattern. So you have incoming pipes. You transform the messages. You filter the messages. You may even add fields or delete fields. And you have a pipe where you output it to elastic search. Jordan Sizzle, the creator of log-stache is now employed by elastic search. The Kibana web analysis toolkit is also under the hood of the elastic search company. So how do you run log-stache? Of course, you just download it, unpack it, and you need some simple configuration. As I said earlier, you have to define inputs, filters and outputs. The filters are optional. Here I'll just drop all messages with the log-level debug. For our system, we define a GALF-import input filter. But log-stache can also provide input types like syslog or radius or other tools like I said earlier. The output is to elastic search. You can also output to a file, but it's, of course, you get only the benefit if you put your structured logs into a log-stache. What's Kibana? Kibana is a single-page JavaScript application. You need no install. Just unzip it in your NGNX root folder or Apache root folder. And it's a tool to search and analyze time-based data in elastic search. It has a rich set of visualizations and provides the access to the full, powerful search syntax from elastic search. And you can create and share dashboards for yourself or within the company. A big advantage in using Kibana is it is possible for non-programmers or not-so-skilled people to query and analyze logs. And I think a really important point is you don't have to have access to your servers to analyze your logs. But you have to consider, Kibana has no authentication built in. So it directly talks to an elastic search service. And who can read from elastic search can also write to elastic search. So if you need extra security, you have to put a proxy in between and do some authentication. The next slides are some possibilities to visualize search queries from elastic search with Kibana. Bettermap uses geographic coordinates to create clusters on a map. You can zoom in. You can do this based on country codes in your log messages. And yes, if you want to drill in, you can click on the clusters and have a better view on it. You can build panels with histograms. Histograms display time charts. It displays counters, mean, minimum, maximum, and a total number of numeric fields. You can build spark lines. Spark lines are a great tool to get an overview of your system what's going on. It's based on tiny time charts. And you don't get the exact numbers. But if you look at a spark line, normally you can really fast access what's going on and if there's something wrong with the system. Then Kibana provides some visualization for the facet calculation from elastic search. Facet calculation means based on a set of filters, you can see how one term is distributed. You can see, I think, that there are logs from a web server, what kind of files you have delivered, mostly HTML, some PHP, and some images. So it's also nice to get a quick overview of your system. After talking a little bit about the technology, I'd like to present some logging patterns that are useful when you want to add structured logging to your application. They are all based on adding context to your log messages. So the easiest way to add context to a log message is just use the extra field from your log message. It just takes a dict where you can add arbitrary key value pairs and the gray log Gulf handler just pushes them into log stage. A little bit more advanced is using a filter. With a filter, you can add context to all of your logging afterwards. So if we have a web application with a user logged in, we can add a filter which adds the logged in username to all the logging messages just afterwards. The request ID, let's you call it all log message from a request together. So if you generate a request ID at the beginning of a web request and just add them with a context to all the following log messages, it's easy to identify messages from the same request and it makes debugging much more easier. How does this work? Okay, you get a request application, you set the request ID and all the logging messages have the request ID applied. How could you implement this? This is an example for Flask. Flask provides a before request handler. It's always called when a new request starts. We are generating a UUID and we are adding a filter to the logging so that every log message has this request ID applied. The correlation ID, let's you call log messages from different applications and systems. If you have a front-end server and some backend AP servers, you want to call your log message over all these servers. So what do you do? At the beginning of the request on the front-end server, you generate a correlation ID and when you make requests to the backend servers, you add the correlation ID to all your requests and the backend servers just read the X correlation ID header field and add this correlation ID to the log messages. Same ID here. All the log messages have the same correlation ID and you can follow a web request across different applications. Implementation for Flask, pretty straightforward. You just get the header field if it's set and again you add a filter. I started the talk with the claim log everything. That's not always true. If you have really big systems, maybe you don't want to log every debug message and there's a really cool handler. It's not yet available in the Python locking but it's available in the logbook from Rwannacher. What is the thing across handler? The handler wraps another handler and buffers all the log messages until an action level is triggered. That means you can buffer all the debug messages and if there's afterwards an error message, then it outputs all the debug messages if there's no error messages, they are done. In the error case, you have all your debug messages in your system and if everything works okay, you're just they are done the way. It's the implementation. Pretty clear, I think. I really like logbook. I think it's a worthy alternative to the standard logging but you have always to wait the benefits of using an extra library to the benefits of using the standard what is in Python. So that's my talk. I'm finished. One minute left. Thank you very much. And... Thank you. You
|
Peter Hoffmann - log everything with logstash and elasticsearch When your application grows beyond one machine you need a central space to log, monitor and analyze what is going on. Logstash and elasticsearch let you store your logs in a structured way. Kibana is a web fronted to search and aggregate your logs. ----- The talk will give an overview on how to add centralized, structured logging to a python application running on multiple servers. It will focus on useful patterns and show the benefits from structured logging.
|
10.5446/20013 (DOI)
|
Please welcome Tom. Hi, so good afternoon and thank you very much for coming. My name is Tom Ron and I will present a joint work with the friend and colleague, Niv Rizaki, who couldn't attend. You can see both our slides and the card on this link. I apologize if it's not very clear, but it's GitHub, Niv M, learning chess. So what are you going to talk about? Today is learning chess from data. All everyone wants to make computer play chess smarter. We're a bit modest and we just want to make the computer play chess. So what's on our mind? We want to know if a computer can learn chess only by looking at data of chess games. So there are many questions that can be asked in this domain. We're going to focus today on two of those questions. One is giving a board state. Can we make, can we do a specific move? Is it a legal move? And that one is game over, giving a board state. Is it a checkmate? Has the game ended? Of course, if those are possible, then the sky is the limit. And what else can we empirically learn about other systems? We have some physics and other things. I want to mention that this is a work on progress. We're still working on it. We have additional and further ideas, but I came here today to show you what we have done so far. So let's start. And what we know about chess, and there is, first, tell that there is constant tension between features that we allow ourselves to know when doing this learning process and features or other things that we want to know. But, first, we know that there are two sides to parties who play the game. We know that a game could end with either one winner or a tie, no two winners or other situation. We know that the board is eight by eight and doesn't change through the game. We know that there are different pieces that have different unknown properties such as how can those pieces move? Can they eat other pieces? What happened to them when they get eaten? Maybe promotion for pounds and so on. Okay, so the data set we worked on is given in a logistic chess notation. If we'll have some time in the end, I'll show you how it looks like. But the idea is that every square on the board is represented by a letter A to H and the number 1 to 8 and the move is basically done from one square to another. Usually only the two square is written and while there's only one piece that can do that move or if it's not clear, then both the two and the from square is written. We ignored the metadata on this set such as player ranking location and so on. We had just a bit more than 100,000 games with full or partial discussion. There were many games that didn't end either with a checkmate or a tie just ended in the middle and we've had a bit more than 8 million moves with distribution between the different pieces. We used a Python library package which is called chess that allows to parse chess library notation and provided the board the status, provided methods like is this chess, is this checkmate and so on and some pilot, mainly sci-pi, some matplot for plotting and non-pi. Basically we thought we would have enough or big enough data for doing smart produce and all we build it as we thought we were going to do more produce but for this time it was enough to do it on a single machine, maybe some in the future. The first question we addressed before, the game on, can we do a simple move? The most naive thing would be, okay, have we seen that move before by saying this move? The board status and the move I want to do, if so, yes, good, do it, no, try again. Maybe there's not enough data so I haven't seen this move or maybe it's not legal and therefore I haven't seen it. And it's not efficient on neither running time or memory. So well, and of course there is no learning done here. So let's move to our second try here. And so for each move we made, we checked the difference from the two square to the first square and two square and we drove the diff histogram. For example, if the pound move two steps on the first time a pound can move then the x difference is zero and the y-diff is two and we did some adjustments of the black and white so it would be relatively. And now you can see those histograms. So this is one for pound, a pound can move either one step forward, two step forward or one step forward into the side and to each side. This is how the bishop move. This is how the rook move on the stretch. This is how a knight move. It's kind of nice. The king and you can see that the king can move one step to each side and castling to one of the sides. Okay, so the pros of this approach is it's very good for common moves and it's getting better as the data size grows of course and it's fairly time and memory efficient. We can code all this really, really simply. However, it doesn't take into account the board status. So if there are pieces in the way, I cannot answer this question or I can answer it wrongly. So it's a necessary condition if we have enough data but it's not sufficient. So the next take we did on this idea was that for each move we not only looked at the move of diff but also the surrounding of each piece so you can see here this is roughly the idea and we have three possible states. One is occupied, one is free and one is out of the board. If we're standing off the edge of the board then some of the squares can be out of the board. And this is some of the results we got aggregating those histograms and doing some grouping on it. So for example, for the queen, if the queen wanted to move on this direction, moving two, at least two steps then the square above it and on the right must be free and that makes sense knowing the chess rules. Another thing about the queen, if she wanted to move seven steps downward and right then this means that she's moving across all the board. Therefore she must stand in the corner and this square must be free. Okay? Cool. So if the king, if there is castling and the king move then the one near it must be free. Also for the pawn, if the pawn goes forward and oh, surprisingly nothing for the knight and knowing chess rules we know that the knight can jump over pieces. However, not having this rule doesn't tell us anything because maybe there is not enough data, maybe there is nothing relevant but that's nice for us knowing the rules of the system that the knight can skip over pieces. Okay? So the purpose of this approach is also we keep it efficient, not too much data that we store and of course run time. We take the surrounding into account so we can argue whether the surrounding is one radius, two radius more. But also doing this says tell us that we have the trade-off. I talked before that we have some external knowledge about the game and about the environment we're in. So again this trade-off and the main comment about this or this plan is that we assume that moves are independent of one another. And while we can usually say that it's not true for all the moves, for example, castling a king cannot do castling if there was chess before or if the king moved before. And there are several more moves that are limited by this limitation. So okay, this is all we're going to discuss about moves today and we still have an idea to improve it but we know that this gives roughly good results. And it's of course let's generalize which I mentioned before. Okay, so now for learning checkmate and here we ask giving state of the board is it a checkmate or not? We're not asking whether if it is a checkmate who won the black or the white, we might be asking that in the future. Okay, we used several datasets, 10K, 30K, 800K, the training set we used, 40% of it we used for training, 64 testing and we had 50-50 of true and false samples. Of course the real distribution of the probability is much less because you only have one checkmate at most, at each game maybe less. And we use SVM classifier with linear kernel. We probably won't use it in the future although we had some nice results just with this name classifier. Now, crash course about classification for people who don't come from this domain, really crash course. I know I speak too fast, I apologize. We have a lot to talk about today. Okay, so we start with data and then we extract features and we'll talk about the features we used in a minute but features can be count features, Boolean features, categories, many others, maybe a combination, maybe the features depend on one another, there are models for each problem and then there is a classification. Some of the data is used for training, some for testing, some you predict. We use SciPy for this mission and actually we were able, SciPy is very, very general and we were able to use very, a code that we used before for a totally different task, just applying our feature extraction and pushing it to the classifier we had and actually another good feature of SciPy, it is very easy to taggle between different classifiers, they all have terrain and estimate and fit functions so just play with it. Okay, so here again we have a few versions. So the first version we had was simple count features. But that means first we counted the total number of pieces that were on the board, then we counted how many white pieces, how many black pieces, for each type of piece we counted how many pieces, for example there were five white pounds and three black pounds so we had total of eight pounds, we also counted the number of different white pounds, five and three black pounds. So this brought us to something that is a bit better than a monkey, with a accuracy of 70, we had for the cases for checkmate we were able to say 80% that this is a checkmate and we were, for not checkmate we were able to say in 59% of the time that it's not a checkmate but then we had some misclassifications. So well we want to be much better than a monkey, so we moved to the next thing and the next thing was using the previous features and using data about the first degree neighbors. In this case we didn't look out of the board and we excluded it but we'll do that on the next versions and so we looked on the data of empty, is it of the same side of the piece we are looking on or from other piece around it and we aggregated the data for all the different pieces on the board from each party and we also built some Boolean features based on this data for example, is there more pieces around me from my side or from the other side, is it mostly empty and so on, such features and we did head improvement, we can say that the checkmate raises to 86% and no checkmate were now able to classify well on 87%, remember we had 59% previously so we are doing much better now and the third version was doing taking the same as before but extending the radius to 2 and 3. This makes much more features however 300 features is not that much and maybe in the next versions we can add more features without, it's not that much and but as I said before it makes it less generalized as we assume something bigger about the game and the board and indeed we had improvement and now our QRC is 89.5 and both has improved, we can ask further questions if increasing the radius to 4568 would improve it, I personally don't like this approach and don't want to do it because we assume more about the board and about the game and about the system as a whole and I would rather think about or suggest different features. So having this benchmark, what we think about or what we suggest want to do in the future. So test different classifiers, here we used SVM, maybe changing the kernel, maybe think of nearest neighbor maybe using some deep learning as a buzzword, I don't know. This is a small change but I think that it would have some interesting effect and result integrate out of the board, the edges of the board into the different counts we are doing. Okay asking who is the winner, which I mentioned earlier, is it the black, the white, we can either approach it as multi classification problem, whether the white one, whether the black one or it's not a checkmate or we can use it just as black or white. Like if we have a checkmate, is it black or white? Who won? Okay asking whether a specific situation is chest, not necessarily chestmate, complex move detection, okay history, something writing the history or maybe we can think of other features that represent us what we have done. Maybe for example counting how many times this specific piece moved or something else, of course as I said, as I mentioned we want to reduce the data we have, the external data we have on the game. Okay more efficient parsing, we use the chest package which is nice but on some cases we did something like bootstrapping, we took the data, we put it into the chest, we produce what we want, maybe we can not do this lap and just do it ourselves. Scaling so for classifying the 800,000 samples, it was really hard for our computers and the sci-fi, it eventually happened but it was hard so maybe we need to think about distributing it, about using something like Shogun that was mentioned here earlier, I think there are many tools that we can think of. And surprising we have time for questions and thank you for listening so far. So actually we are sprinting in the last two weeks about this as well as working at the same time. I forgot to mention earlier that even we both work in a tech company which is a poly-Israeli technology company but so as well as working for doing this overnight. Have you analyzed if there is a current pattern in the wrong estimation? No, we think of it, we want to say okay why, we haven't looked for it, we haven't looked into the black box that say okay this is why we are wrong, we do plan to do it because we want to go further, we want to improve it and we want to make it general. And of course the big question is whether we can apply it to other systems, can we learn physics just by looking at it? For example? You know, Nargon chess and giants generally are stateless, they just get a position and make estimations, not based on data but based on their own algorithms and they don't take account neither except for the first part of the game which is pretty easy to do, they don't have any state, just a position and not that. But knowing what they can do. So we are not focusing on doing better strategy, we are focusing on lesson number one, this is how the pieces work. Thank you very much for listening, have a nice day.
|
Niv/tomr - Learning Chess from data Is watching a chess game enough to figure out the rules? What is common denominator between different plays and game ending? In this presentation, we will show how Machine Learning and Hadoop can help us re-discover chess rules and gain new understanding of the game. ----- Can empirical samples unveil the big picture? Is chess games descriptions expose good enough data to gain understanding of chess rules - legal piece moves, castling, check versus checkmate, etc. Which features are important in describing a chess game and which features are not. What is a good representation of a chess game for this uses. What is the minimal sample size which is required in order to learn this in a good enough manner and where this learning can go wrong. **Ne3 => E=mc2** Looking at the bigger picture - Can we understand big systems based on empirical samples. Can we reverse engineer physics and discover how physical system work based on no external knowledge beside empirical samples.
|
10.5446/20012 (DOI)
|
Please welcome Nicola for her talk, Eve, REST APIs for Humans. Good morning, thank you. So first I'll tell you a story. Two years ago in 2012 it was at Europe, I was at the European and I gave a talk about building REST APIs with Flask. That was kind of a training, a very long talk and there was a lot of interest about that project and the code that I showed back then. People were asking if we were thinking about releasing that kind of application as an open source project. So Eve, which is the project that I'm going to show you today, is basically the offspring of that talk. It's very cool for me to be here in Europe and presenting the result of that event. So REST API for Humans, I guess 100% of you know that I stole this tagline from Kenneth Rice, which is basically the client side of any Python REST API. The reason why I'm doing this is because basically the idea behind this REST API framework is the same as RECAST, which is make things as simple as possible. So here it is. Well, I will just keep on this one since you already heard from my chair, speaker, what I do for work, for job. And what is the philosophy of Eve? Basically you have some data stored somewhere, some data, and you need the REST API to expose your data to some, I don't know, mobile client, maybe, or website or what have you. And what you do is just install Eve and hopefully in a few minutes you get working API for you. It is powered by Flask, as I told you already, MongoDB ready for a few features and a few other things, but these are the big three guys in town, let's say so, and a very quick start. So you get an idea of what working with this framework means. How many of you are working with Flask already or have an idea? Oh, great. So if you work with Flask, you can recognize this code is basically the quick start from the Flask website. The only difference is that you are using Eve instead of Flask. This is because Eve basically is just a subclass of Flask. So everything you can do with Flask, you can do with Eve. And this is probably a good idea because I see people using Eve as a, yes, as a REST API, but also as Flask. So they are using Blueprints, for example, for adding new features to the API and stuff like that. Then the other thing you need to do is you need the launch script that we just saw and you need a settings file where you basically design your API. And the idea here is like Django and other framework, you just have a text file. And in this case, we are giving two endpoints to our APIs, people and books. As you see, we aren't defining anything for these endpoints. So we are basically just saying, hey, I want two endpoints on my REST PIs and these endpoints are named people and books. And then you just launch the API and your API is up and running and ready to work for you. For example, you can access the people endpoint. You see that even if we didn't define anything for the endpoint and we, did you notice we didn't define any kind of database connection actually. But the API is working anyway. What you get here is a few metadata, metafields, so items, which is supposed to be the list of items from the people collection, which is empty, of course. And the links is a different metafield and we will cover it in a few minutes, but you can already guess what's going on there. Yeah, it's this age 80, 80, yes, I don't know how to speak in English, but basically they are just the links to the API endpoints. And you can, if you want to, you can write your client in a way that it can navigate these links and build the client API and the client UI based on these links. You can turn this feature off if you want to. Okay, I'm just keeping it. Let's connect the database now. Very simple, of course. And then while we are here, let's also define some schema for our endpoint. So here what we are doing is defining a few fields. And we are using, we are defining some data types and some validation rules. So the name fields is a string, it has a max length, it is unique, and email field, it is a string, of course, but we can also set a reg X for validation of this field. Don't use these reg X in production because it sucks. But just to give you an idea of what you can do. And you can, even if you look at the schema keyword down there, you can even nest the dictionary within dictionaries, list within dictionaries, and list whatever you have here. And then we can, by default, an API is read only, but of course you can change that. In this case we are enabling writing to the API endpoint. We are also allowing an edit of the items, replacing the items, and deleting the items. You have to do this explicitly. Otherwise the API endpoint will be read only. This is of course for safety reasons. And yeah, a few more toys just to show you what you can do. You can set cache control on that point, additional lookups, a lot of stuff. And so we defined our API endpoint. We brought that launch script with a few lines. And what do we get from this? Well, first of all, you have filters, for example. Your clients can query the endpoints. They can do that using a Mongo syntax of sorts. So here you have an example where we are querying the endpoint for the last name, though. But you can also use a Python syntax if you prefer. This is because maybe if your client is being write in the right road by you, you can use Mongo and there is no big deal. But for example, if you expose your API to a website or to people actually using it, they don't know anything about Mongo, maybe you prefer to use a different syntax. You can do that. You can have sorting on your endpoint. In this case, we are sorting around the sending order. You can use pagination. It is enabled by default. So for example, here we are asking, give me page two and only 20 results at maximum projection. This is very nice. You can say you have the document with, I don't know, 50 fields. You can say, don't give me these fields in this request back to the client because I want to save on bandwidth or on performance, for example. In this case, we are telling the API, don't send me the pictures, for example. Don't send me the avatar because I don't need it. And here we are doing the contrary, only return me last name, for example. Which is very handy if you are writing a mobile application, for example. You want to optimize the data being sent on the wire. Another very cool feature is embedded resources. So basically, here we will see an example. Here we are asking to embed the author field. Let's see what does it mean. By default, when you get the document, you will get for the author field, it's for NKEY, for maybe another endpoint. This is what you would get by default. But if you send a request with the embedded keyword, what you get is an embedded document with the full author. This is again to avoid sending two requests for the data that you need on your client. By default, your API will support both Gison and XML. And here you have an example of the resources with Gison. You know that very well. By the way, all the field names, for the meter field names are configurable by you, so you can change whatever you see here to suit your needs. And this is the same resource in XML. We already saw HyperMedia, the engine of application state at work. Let's quick look here just to have an idea. You get the link to the same item, to the parent item, next page, if the pagination is enabled and you have more pages, you get the link to the next page and even a link to the last page, of course. Again, all these features are enabled by default, but you can switch them off. For example, you don't want us to support XML, switch it off. You don't want HOS, you turn it off, et cetera. You can customize the API however you want. Document version is something that we just released with the last release. It is basically a git for the documents, if you allow me. And what you do when you switch this on, basically, you get versioning for your documents. This feature has been interpreted by a SpaceX engineer, actually, so I am very proud of that. And you see here, we are asking for version three of one document or give me all the versions of the documents or you can even ask for the diffs or between the documents. Maybe this is not something that everybody needs, but it's very cool to have it at hand. File storage. You can store files within the API since Baydefault is supported by Mongo, is using Mongo. We are storing in GridFS, which is basically optimized storage for files in MongoDB. How many of you are using Mongo or think about using Mongo in the future? Okay, not quite a good number. We will see that there is also a SQL Alchemy branch for you later, so keep your hopes high. And here, an example of how you do storage of a file. When you define your endpoint, we saw earlier that you can define a string type, but you can also define a media type. And then when you send your data, what you do is just use a multi-part data format, sorry, a post, and you send your pick along with the other fields of your document, your picture. And when you get that document back, you get the picture as a base 64 string. And you can also enable this extended media info setting, which basically is going to give you not only the file itself, but also the metadata about this field. So, for example, content type, the name of the file, size, et cetera. Again, you can disable and enable for storage however you wish. Rate limiting, this is powered by Radis. What you can do here is set the number of requests that a single client is allowed to perform on your single endpoint per minute, or I shouldn't say per minute, but per time window. Here we have an example where we are setting the get method limit at one request per minute window. So you can have different limits per method and different time windows as well for every single endpoint. So the first get back is answered by the API and in the header section, so you get information about rate limiting. So you only have, there is one request allowed on this endpoint per minute, you have zero remaining, and the next reset of the time window is at that point in time. The second request within the same minute will get two, two, nine, too many requests. This is just to give you an example of how this works. This is supposed to work. This is, of course, useful if you have performance issues or if you want to avoid your API getting hammered by some client, maybe a buggy client or somebody trying to do some kind of weird thing or attack on your API. Conditional request, so the client can send a request using the if-modify-sins header and, for example, say, please return me the data from this endpoint only if it has changed since. So we don't get back always the same, the same, I don't know, in the people endpoint example we saw earlier, when I get back the first reset, the next request I can get back new data only and on all the data load. If none match is similar, but we are using e-tags here, so this is mostly used on the item endpoints, so not on the people endpoint, but on the single person endpoint, and what it does is the same things, basically, give me the personality if something has changed on it. There is also support for bug inserts, so you can send multiple documents on the API endpoint with a single request. Here we are sending two documents, for example, and when you get back as a response, there's an array of responses, actually, because validation is performed on every single document, and there are two options here. You can switch coirins mode off, which is the default. In this case, you are only going to back the metadata back, which they will be useful for sending a subsequent request later and stuff like that, or you can basically say every time send me back the whole document included metafills. This is something new that was just added to the API, to the feature set. Quickly, on that integrity, concurrency control, we are basically using e-tags for that integrity checks, so when you try to modify the document, if you don't provide that if match header, you aren't going to get that patch in. In fact, you are going to get a four or three back. If you send an e-tag, but this e-tag is not matching the e-tag on the server, you are going to get a precondition failed error. If you give me the e-tag which matches the document on the server, then the edit will go in. Why is this? Because we want to avoid a client with an old version of the document overwriting a newer version of the document on the server. Only the client who already knows about the latest release of the document can update it. Data validation, of course. Here we have a response example of a bulk insert where the first document got an error, and Clinton is not unique, but the second document was accepted. There is support for authentication, so basic token authentication, HMAC, which is basically what Amazon S3 is using on their platform. It runs on all kinds of Python, 3.4 and PyPy include. There is a lot of more stuff. We don't have time to go over that. Here, versioning means API versioning and non-document versioning, so you can have basically your endpoint version one, version two, version three of your API. It's BSD liaiscensored, open source, you can do whatever you want with it. You don't owe me any money. Okay. What we saw so far is what you get for free without any line of coding. You just have to pull the switches on and off. You have this feature as we are turning on and off, but what about developers? How can I customize my API? Here I have a few examples. For example, you can have custom data layers. In this case, what we see is the code from the SQLHME branch, which is a working project in progress. What you do is basically subclass the base data layer, and then you go off and write your own data layer. For example, we have an extension, which is called Eve Elastic, and it's using Elastic Search. Here we have SQLHME and whatever you want to use. This is an example of the SQLHME. By the way, here you see that in this approach, what you do is using SQLHME classes, and you just register the schema so you don't have to write the resource schema in the settings file because SQLHME is already providing you the classes with basically the implementation of the URL endpoint. Here is an example of the Elastic Search data layer, and the MongoDB is doing exactly the same. It's just subclassing the basic data layer. Authentication. This is where you actually have to do some work because you need to subclass the base class here and provide the authentication logic by yourself. This is, of course, because this is something you want to be in total control of. You can do a lot of stuff with authentication. You can lock the whole API. You can lock only certain endpoints and leave other endpoints open to the public or read-only, write-only, read-and-write, whatever. You can do all this with access control. There is a lot of stuff here, but we have to look at it. Just three steps of the tutorial to give you an idea of how you do this. You basically just import your class, basic auth class. You override this check-auth method. Here we are basically saying, hey, whatever request comes to this endpoint username and mean, password secret, let it go. It is good to go. Otherwise, we will send back a not allow response. Then what you do is when you create your instance, you just pass your custom class to the toEve, and that's it. Your API is now protected. Of course, in this case, we are just setting this protection for the whole API, all the API endpoints, but you can actually change your class for every single endpoint, or as I said before, you can even leave some endpoints without protection and other width. All that you need, custom validation. You can add the custom data types, custom validation logic, if you need to, and that is very nice because, for example, in the next release, we will add support for GeoG zone, and so you will have point, multi-point, polygon, polygons, and all this kind of stuff. Then we have event hooks. This is, sorry if I'm going very quickly on this, but the time is short on us. This is very nice because when something is going to happen on your API, you can hook callback function on basically every event. So here you see that you can set, add a callback function every time an item is being inserted, for example, or after it is being inserted. The same happens with get, patch, put, delete, and whatever have you. There is a simple example here. What we are doing in this example is update the documents, the client is sending us with a new field. So you just define your callback function. Here we are, and you see that the functions is getting the resources, basically resources, the endpoint, and the documents is the collection of documents that are going to be inserted in the MongoDB. And what I'm doing here is just adding a new field or overwriting this field if it exists. I gave this presentation and forced them in Bruce Selling for everybody. This is why there is forced them and I didn't have the time to update it. And then when you are about to launch your API, you just hook your function to the callback as you can see down below. And then you have custom file storage. As I said before, we store on GridFS by default, but you can change it to whatever you want. There is a guy who did an S3 class, for example, so he's an instance, he is storing the data on S3, on Amazon S3, or you can store on file systems whatever you want. And then there is the community. Just a few words on it. There are a few extensions available already released by the community. For example, if docs is a very cool project, it generates documentation for your APIs. And what it does is this, you get a docs endpoint and when people access that endpoint, instead of getting a JSON or XML, they are getting an HTML page with documentation of the API. And it is actively maintained and there are a lot of contributors on this project and what basically gives you an automated documentation for the API. If Mongo Engine, this is basically a connector between Eve and Mongo Engine. If anybody you're using Mongo Engine, you can do what we have seen with SQL Alchemy with this ORM for Mongo. Eve Elastic we told about, we saw it already. EveMocker is a mocking tool for Eve. And the thing that matters to me is that the community about Eve is starting to grow quite a bit. At the moment we have about 50 contributors to the project, but what I'm really looking for is more contributors joining the project. So if you are interested in this kind of stuff, you do know that you can actually contribute to the project as it is on GitHub, of course, and we have a few tickets open. And what I'm specifically looking for is people willing to work on the SQL Alchemy branch because it is now feature-complete, but before merging it, what I need is people actually wanting to work on the SQL Alchemy branch even after it's been released. I don't want to merge the branch and then I have people complaining because something doesn't work, I don't want to look at it personally because it's not my kind of job, I'm doing something else. So if you want to join a nice project and you're interested, please do so. We are working on GeoGison for the next release, so you will be able to define a Geo data point or Geo data point and validation will work. You can do queries on this kind of stuff, etc. So if you are a customer in the crisis, a lot of stuff is coming up, it's on the pipeline basically. This is the URL for the project, so you can go there and read the documentation, get to the GitHub repository and see the changelog, get in touch with me or my Twitter account, of course, you can get in touch with me even at this account. If you go to GitHub at nicolayar.com, you find the source code for the project, and if you go on Twitter, you can follow me, and I usually use Twitter to update on the mergers, the new coms, it's coming in and commenting, and sometimes even, I don't know, like complaining about stuff and stuff like that, but if you want you can follow me on Twitter. And basically that's it. I wanted to give you a little demo, but I don't think we have time for... Yeah, I think we'll show. So if anybody wants to see something working, just one thing, there is actually a... I'll cheat here. There is basically an online demo of an API, you can consume with your clients or even with just Chrome, for example. This is Postman, but if you go to ifdemo.roquap.com and the slash people, what you get back is a NixML because Chrome is requesting XML data, but basically you can consume an API, play with it, send the get, put request, and stuff like that and play and see whatever... For example, here I am asking for the people endpoint where people have a white last name, and what I'm doing here is using the... actually the Heroku application in remote, so you can play with it basically. And get the first-hand experience of the API. Thank you very much. APPLAUSE Okay, thank you, Nikola. We have a little bit of time left for questions, so please raise your hand and I come with the microphone. Thanks for the talk. It looks really interesting. Thank you. Do you have any support for testing your REST API? Testing? Yeah. What do you mean by testing? I mean, there is a huge test suite on the repo, so every feature is being tested on every commit you submit to the source code, actually. So what I mean is it's easy to create your REST API, but what if I want to have some tests for it? For my API that I built with your framework? Yeah, it is easy to do because you can... basically, again, it is a Flesk application. It's Flesk, so whatever you can do with Flesk, you can do with it, which means that you can do tests very easily on your own API, actually. You can also use our tests to see how it's done, it's done, and then implement your own tests on your own API. It's very easy to do. The test suite is something which I'm kind of... I miss the feelings about it. I'm kind of proud because there are, I think, 500 tests, and I'm testing everything, but it is in need of some refactoring, actually. So that's another area when you could think about joining the project. Thank you. Go ahead. Thanks for your project. Do you also support URLs instead of IDs? For example, you showed us the book and the author of the ID. Could we have a URL on 790? Yes, what you can do, what I didn't show here, is you can actually have nested the URLs. So, for example, you can have cities slash... city ID slash people slash people ID, for example. So the person in the city, this is not a very good example, actually, but you get the idea, and you can design your own URLs since, again, it is a fresh application. So you can even define an additional URL for the same endpoint. For example, if you don't like the ID, which is a row number, of course, you can define a new URL based on the last name, for example. So api.com slash Smith will get your person with the last name of Smith. So, yes, you can play with the URL quite a bit. Okay, thanks a very... Thank you again, Nicola. Thank you, guys.
|
Nicola Larocci - Eve - REST APIs for Humans™ Powered by Flask, Redis, MongoDB and good intentions the Eve REST API framework allows to effortlessly build and deploy highly customizable, fully featured RESTful Web Services. The talk will introduce the project and its community, recount why and how it's being developed, and show the road ahead. ----- Nowadays everyone has data stored somewhere and needs to expose it through a Web API, possibly a RESTful one. [Eve] is the BSD-licensed, Flask-powered RESTful application and framework that allows to effortlessly build and deploy highly customizable, fully freatured RESTful Web Services. Eve features a robust, feature rich, REST-centered API implementation. MongoDB support comes out of the box and community-driven efforts to deliver ElasticSearch and SQLAlchemy data layers are ongoing. Eve approach is such that you only need to configure your API settings and behaviour, plug in your datasource, and you’re good to go. Features such as Pagination, Sorting, Conditional Requests, Concurrency Control, Validation, HATEOAS, JSON and XML rendering, Projections, Customisable Endpoints, Rate Limiting are all included. Advanced features such as custom Authentication and Authorisation, Custom Validation, Embedded Resource Serialisation are also easily available. In my talk I will introduce the project and its community, recount why and how it's being developed, show the source code, illustrate key concepts and show the road ahead.
|
10.5446/20011 (DOI)
|
Nick Douglas yn fwy roi medision per Michel Giamwy. embargo, y producedor yn medzirif my Family i blкиad a gondi ei ddeakaeth ymgyrch o'r cyfnod. Yn ymgyrch o'r cyfnod, yma, sy'n credu cyfnod o'r cyfnod sy'n credu cyfnod o'r cyfnod. Felly mae'n fyddi'r prydau o'r idea ac mae'n gwybod yn cael ei gwybod o'r cyfnod cyfnod ar y gyrdd yma. Mae'n gweithio'r cyfnod o'r cyfnod fel mae'n gwybod yn cael eu cyfnod o'r cyfnod o'r cyfnod, ac mae'n gwybod ar y gyrdd. Mae'n gweithio'n gwybod o'r cyfnod o'r cyfnod yn cael eu cyfnod yma o'r cyfnod o'r cyfnod. Felly mae'n gwybod yma. Mae'n gwybod ar y cyfnod o'r cyfnod o'r cyfnod. Mae'n gwybod ar enw o'r cyfnod yma o fwymo diwethaf i'r ffyrdd ymgyrch i'r ffordd iawn i'r ffordd iawn, fel mae yna'r gwybod i gydag o'r cyfnod o'r cyfnod o'r cyfnod byddwn yn zu'n llawer o'r Un해wen oedd na i gynoes. Onid, yn g dumhyn wy nad yw, bythreddiad Budwellaf 你wrdd y fwrdd ar Explorer python, a hwnny was gendag y lle bl Mat dim ymlaen, roeddwn yn eu gennymau Yn postedion sydd wediALL cyn zwarn o gyflwyngine Usuallyón. But I don't want to talk about Edward Stonan, really. Because he's been dealt with elsewhere. What I want to talk about is the result. The result was a moral panic. Apart from the United Kingdom. I'm not sure why. Suddenly, privacy became a hot topic. Up until Snowden, privacy was a topic that was dealt with more within the domain of corporations fears,.. cobbwy modau neu trans bumper. часol, a i ail yn y pwdwys i...茱wch arachten yn mynd ar gwein oportun mewn Bai Canada. Rwy'n hollwch nesaf yn rhaed i hyn o fe unseren gan......kawn ymddwyd hyn? Na vertically wnaeth ie y diwrnod wrth gwrdd yma......saib ar published marching ond organise o darlo�... Sığ 있adol ar y ddigon nhw. Felly, yn gallu gwneud o satgenas ar fanệtring gyda gwn myfio'n ysgrifennu ph brace nad ydych chi dcrioedd yr amser lle i hynny y myth mrych himfio. Rwy Standards Arけどwch am Erich Schmit. Fy entitled Elect pulses, w righteous a licen o arwgar angen yr oedd yn hoffa, Deciw ty mae ei wnaeth ymlaesi hen o'u enw. Bydd o facebook yn deby hyd rhyw f yer Pelain o'i wneud y traed hynny am wexODd mewn gwel belladau fel Ùs mewn byddoo iawn slad iawn ti wda i ôl grannu strings w teachers ac uchaintwch teulu i'w meddwl am tanto, a mae Prefusi fel Reynțiwch i'n cyhoedd kindergartenau a gwneudio ei cael iNog. Gyda ni,rictingu yn natys imperfectό Syria ac yn dweud i ni fel Reyn waxengol y tair turtles yn ei parfa wasp mwysbildung sy'n gallu FESS-Nog yn y grip theorem. A fe brifように churches ar fy f Waarigau did chopped cadw yousio yng nghym yn cynal dictionary fizioaeth y ilgili ac Tirgellian youtube ac yma am gyfudd i y ta corporations dadiff ac yn seguidebu iawn yn gwidadleach ar gyflogi i g weld diarrhea, i chi surprising nhw'n literatureagain owners axeion hereu o gyflogi Yong yn y gallu gw teachers. Yma yw Ymddag William Hagueh yn ystllud a chyma ffraithol ymyrwyr o'i gael. Codion negod shrinksix proffesiad a hefyd, hefyd yn cerdd li bydd mwynt agaringo. Felly bydd sicrhau sy'n mew'r anghy**a gyda tbspunag ymblogaeth, todo ar 我們 Argh quella Gweithod yha. Da ni weithio argynwism Unrhyw hwn Mae Hahegwyn yesen nid curiosityau i blwyddyn Ond rhaid i obiwch chi ilivellu ti. A llwelu nhw'n go iddyn nhw'r gweithio y'r han yn mi. A i chireadraeth gynnwys? Ni'r boblod gete yn ei braf gan gweithio chyfnod yr un ar rhai'r gw center wedi'i mynd i gydag yma gan um comedianhau cyllidol honno'i gw mRNA-ros i chi'n gweithio set antiquadu yn blwyddyn bob because this is a terrible case because there is a terrible animal that will capture and identify. This sort of argument, nothing to hide or fear argument is often trotted out with other classic defences like think of the children, terrorists, extremists and heaven forbid for us hackers. So this is blatantly wrong argument. For a start it's a false dichotomy. What do I mean by that? I mean that it turns a very nuanced and complicated subject into a simplistic black and white subject. Okay, you've got nothing to hide, you've got nothing to fit. That's it, black and white actually. It's a lot more complicated than that as I'm sure we all know. It's also lazy thinking too, and it's manipulative as well because you're framing the argument in a binary way when in fact it's a very nuanced argument. And putting that aside, it's also an argument that hides several uncomfortable truths which I'd like to explore now. So the first uncomfortable truth is that it's not you who determines if you have anything to hide or not. For example, these gentlemen who are some prominent American Muslims who are law-abiding citizens. They are political candidates, civil rights activists, academics, lawyers, people like that. Yet the NSA and the FBI have covertly been monitoring their emails and other communications. And this was done under a law intended to target terrorists and foreign spies. How do you think that makes American Muslims feel? As an aside, I read at the Guardian website this morning that the Metropolitan Police in London have been monitoring the communications of the family of the man they mistakenly shot on the tube train soon after the July 5 bombings. This was a grieving family, yet they were monitored. They had nothing to fear, yet they still had their communications monitored. You have nothing to fear because you've got nothing to hide. It assumes that surveillance results in correct data and sound judgment. If you live in the UK, you'll be very familiar with this particular tweet. But some poor unfortunate gentleman who lives in Yorkshire was trying to catch an airplane one winter, and he tweeted, crap, Robin Hood airport is closed. You've got a week to get your shit together, otherwise I'm blowing the airport sky high. This was obviously a joke, you would think. And then the police turned up and he got carted away under terrorism. I can't remember precisely what it was, but he got carted away anyway. He was imprisoned and the result was that they ended up going to the UK's highest court at much expense and getting thrown out. So you know surveillance, you might have a bit of a problem if the police get the wrong end of the stick, for example. Or they're just collecting the wrong sort of data. If you're doing nothing wrong, you have nothing to hide. Well, rules and governments change. For example, in the UK, obviously I'm British, so many of these examples are British. In the UK there's a law called RIPA Rupa, and it's a UK law to monitor the communications of people for national security reasons. You could understand why people might want such a law. And the way the law works is that it allows certain stated public bodies to be able to use such a law for such a reason. And since the law was introduced at the beginning of the 2000s, that list of public organisations who are allowed to use that law has increased four times and now includes local councils. So local councils have been found to be monitoring their citizens' communications to track incidents of dogfowling. If your dog craps in the street in the wrong place, you might be tracked. If you have nothing to hide, you have nothing to fear. Well, breaking the law isn't necessarily bad. If we look at this rogues gallery of people, some of them might not be familiar, especially the second one in. So that's Socrates right at the very end who was executed for corrupting the youth of Athens with philosophy, no less. The second one in is Emmeline Panchurst, who was a suffragette, who changed herself to Buckingham Palace in the cause of women's rights and getting votes for women. Obviously, I'm guessing you all know Mahandas Gandhi, who was imprisoned for basically trying to make India independent from the British Empire, and obviously Nelson Mandela, a recent example, imprisoned for his protest against apartheid. Now, these are widely regarded as people who acted as beacons of hope. And I guess, you know, hindsight is a good thing. But what I'd like to ask is how would their causes have survived in a digital penopticon if the authorities that imprisoned and in some cases executed these people were able to view their communications. And finally, if you've got nothing to hide, you've got nothing to fear, we should be able to watch you. Well, actually, you know what, privacy is a fundamental human right. There are many examples that enshrine this right, but the one I've chosen is sort of the big daddy, as it were. And this is from the United Nations Universal Declaration of Human Rights. And I'm believing, and I guess that you do too, that things like intimate decorations of love and doctors discussing the patients and engineers, working on a new top secret project or journalists planning an expose of the government. These are just a few scenarios where privacy is both a reasonable and legitimate requirement. Yet, of course, people want to surveil you. So am I saying that privacy trumps all? Absolutely not. Openness of public institutions, governments and corporations, I believe, is a fundamental requirement for our society to be able to function. Otherwise, how else are we going to be able to hold such entities to account if we don't know what they're up to? I also believe that surveillance is legitimate given probable cause for concern. And I'm not the only person who believes that. Can anyone identify where this comes from? It's the fourth amendment to the Constitution of the United States of America. The right of the people to be secure in their person's houses, papers and effects against unreasonable searches and seizes shall not be violated. And no one shall issue but upon probable cause supported by oath or affirmation and particularly describing the place to be searched and the persons or things to be seized. I guess I'm not the only one in the room who sees the great irony of the fourth amendment. So you're sitting there thinking, hang on, we're at EuroPython here. This is a technical conference and who is this British guy ranting on for the last ten minutes about politics? After all, what has politics got to do with programming? We're engineers. A straw man engineer might ask questions like, we're engineers, we like to solve engineering problems and I don't really worry about the politics of stuff and things like that. I'm far more interested in the hard problems of technology and servers and code and things like that. For example, we ask questions like what is the best way to organize computational resources and we answer them by thinking about architecture and design. We also think about how should such arrangements be created. What tools are we going to use? We use Python, we use databases, we use methodologies like test-driven development and agile methodology. We organize ourselves. And also we ask who is responsible for making such things work? In a team we have people who have particular responsibilities. There's the QA guy, there's the DBA, there's the developer, there's the business analyst. There's all these different roles and each is responsible for doing something. Each of them also has authority to do certain things. Perhaps only the QA person is allowed to deploy the thing to the server, the website to the server because they're the one who signs off that the QA is done. We also have people who create standards that we use so that we in some ways delegate responsibility for making things work by following standards that are made in public. So if you contrast these with problems in political philosophy after all, what has engineering got to do with politics? So political philosophers, I'm not saying politicians here, I'm talking about political philosophers, people who think about politics, not the politicians who are the ones involved in the political system itself. These guys ask questions like what is the best way to organize humanity? That's a pretty big question to ask. What's the best way? What forms of governments should we try and promote? They think about the problems of democracy, they think about things like corporate structures within the public sphere, things like that. How should such arrangements be created? They try and define concepts such as duty and rights and they think very carefully about how the law should come to pass and how it should be enforced. Talking of enforcement, who is responsible for making such things work? Who has the power in a society? Who has authority? How does governance work? This is political philosophy 101. So I would say, I'm asserting, that programming is politics quite simply because we are asking and answering questions about organization, process, power and control. We're writing, implementing in some respect, we control the laws of the digital world, if you look at it that way. So part two, questions. Assuming that these things are important, that politics are programming important, how do we explore this program? What questions as developers should we be asking ourselves? So we turn to Holger, who I notice is sat at the back of the room. Last year Holger focused on these political aspects of programming by asking several pertinent questions. What digital world do I want to live in? What sort of software do I want to create as a developer? And if you're a parent, what legacy do I leave my children? How would you answer these questions? And remember my aim at the beginning, which is to give you a context in which you can think. And part of having that context is being able to answer such questions. So one of the conclusions that Holger and I, many others believe is important is the answer to this question. Is peer-to-peer and ubiquitous cryptography a way to address the concerns over power and control in a digital world? So I'm going to, because I don't have that much time, I'm going to brush over cryptography, assuming that you can go and read a book about it somewhere. This is talk about peer-to-peer. So let's examine what peer-to-peer means and how this affects the political aspects of the talk that I was just talking about. So what do I mean by peer-to-peer? Well, this is my back of a fag packet definition. Peers of equal status, devices running appropriate software, co-operate in a loose, decentralized network for mutual benefit. And also peer-to-peer is the antithesis of hierarchy where some have elevated status and power over others. And one way to visualise this is the taxonomy diagrams, the very simple taxonomy diagrams over there. On the left is peer-to-peer and on the right is the client server topology that we use on the web. Notice that the red spot is the point of power and control in the web. And wherever there's power and control, well, that's where politics is. So let's just think very carefully about how this affects, for example, the worldwide web, which is probably the most ubiquitous technology platform of the day. So the client server architecture of the web is fundamentally unbalanced because the server always has power over the client. You authorise yourself, authenticate yourself against Facebook's servers, for example, and then Facebook decides whether you are allowed to see this content or that content or the other content. And of course, the server can decide that it's just not right for you to see certain content at all because it's illegal. Also, a server is a single point of failure that is obviously, that it is also an obvious target for attacks. We all know about the Twitter fail well, but where did the NSA go when they wanted to try and hoover up lots of people's emails? They tapped into Google because lots and lots of people used Gmail. So am I saying that hierarchy is bad? No, I'm not. Sometimes hierarchy is very good, especially when it's efficient and it saves lives. If I was having brain surgery, I would like to know that the person in charge of that team had trained for several years and was acknowledged as an expert in their field. I wouldn't want to have surgery from a democratic group of, I don't know, hippie doctors who would vote at every point in the operation as to what to do next. I'm more likely to be dead as a result of that. So it's important in certain situations that there is definite power and control. But the important thing to notice is that in an ideal world, such a hierarchy is best when the obvious skill, knowledge and capabilities of the person or the entity are acknowledged to bring about greater benefit for all. In an ideal world, those with elevated status and authority would have earned it by reliable and consistent public displays of such skill, knowledge and capabilities. So everyone knows this is a good surgeon because not a lot of people die when they're on the slab with him. For example, him or her, it's a him in this photo. In an ideal world, the responsibility and trust associated with such status and authority would be a serious yet welcome obligation. But we don't live in an ideal world. We live in a digital world where architecture in some sense defines power and control as I just tried to illustrate with the client-servant model of the web. If Facebook changed their terms and conditions, we have no way to challenge them. Not only because they're the ones in control of the servers, but also because they, in some sense, they've trapped us in their walled garden of data. All our photos, all our lives, all our social life is within this walled garden controlled by Facebook, for example. So I'm about halfway through the talk and I want to summarise. So programming, I believe, is politics because we're thinking about process and power and control of digital assets. We agree, I hope, that strong cryptography protects against surveillance. And we agree, I hope, that surveillance is in some forms not a good thing to have. Peer-to-peer, decentralized, distributed, federated systems mitigate points of control. And authority derived from architecture is bad. However, authority derived from evidence is good. So what can we do to address these issues? So part three, actions. So this time last year, I didn't know Holger and I was moving house and as Holger was giving his keynote, I kept getting tweeted by friends in the audience saying, you should contact Holger, he's doing this peer-to-peer stuff, Nicholas, you're interested in peer-to-peer stuff, you should get together, which is what I did. And Holger got together with lots of other people at EuroPython and the outcome of that is that we decided that we would get together a group of us and organise some sprints where we would be able to explore the ideas surrounding the peer-to-peer and cryptography and so on and so forth. And we'd try and do something as well. Obviously Jonas needs an avatar because he doesn't really look like an egg. So at this sprint, what were our aims? At this sprint, we grew a community interested in re-decentralisation of the internet. We also have people from re-decentralised.org in the audience, which is a fantastic project. Here, I'll put your hand up. Talk to her. Promoting non-safael communication is another thing we're interested in. Exploring existing solutions because we're not the first people to be worried about this sort of thing and doing something practical as well. We're programmers. We can do stuff with digital assets. So at this sprint, at the first sprint, we asked ourselves two important questions. What are the fundamental elements of a secure peer-to-peer system and what can we build that is useful to this end? So at the sprints, we looked at existing technologies, Bitcoin, peer-to-peer messaging, et cetera, et cetera, et cetera. And at the sprint, we also plugged Holger into the matrix as well. And the point I'm trying to make here is that seriously, you don't need to do silly things to enjoy yourself at these sorts of sprints because these are fun and interesting and challenging engineering problems. And you don't need to be plugged into anything to enjoy them. We also decided we would try and organise ourselves at conferences and gatherings like this one. I'll come on to that in a minute. But most importantly, I guess, is that we wanted to prototype at conferences and gatherings like this one. So we had something tangible, some code that we could point people at. So at least people could say, that's wrong, you're doing it wrong, or that's good, I might join you. This stuff isn't going to happen by itself. So talking about outcomes, what were the outcomes? So, like I said, prototypes and hacks. And there are two that I would like to talk about. We explored the problem of a peer-to-peer cryptographic message passing system, completely decentralised. And we also looked at a universal distributed hash table as a platform, which is based on some work that I've been doing on a project called The Drugless. I'll give you a very high level view of both of these projects now. So, the peer-to-peer decentralised crypto messaging. Holger calls this the test card, because if we can make this work, we've solved many of the fundamental problems of a cryptographically secure peer-to-peer system. We also had expertise within the group. We had Jorgis, who is one of the developers of the Cryfo project, which is in-browser cryptographically safe chat system. He was there, it was very good to have him and his expertise there. And we also looked at existing solutions. We found many, but the most interesting was one called RetroShare, that met many of our requirements for such a system, but not all of our needs. So we tried to work out what are the gaps that we can fill in. So to give you a sense of some of the thoughts that we've been having at these sprints, I just want to pause a moment and describe one of the problems that we have. The problem is, in a secure decentralised message delivery system, how do you communicate with offline peers? Now, with email, it's very simple. I just send my email to your email server, and the next time you come online, you go and collect it. It's almost like a sort of a postbox for you. But that's a centralised point of control. It's somewhere, as we know with Gmail and so on and so forth, that people can intercept your communications. So we wanted to make this completely decentralised if we could, with no single point of failure, so that the message could get through in a secure way. So what we've been looking at is building a system that allows trusted online friends to sort of pass the message, like a baton in a relay race, until the message is delivered. What we're trying to work out is, can this be done in a completely decentralised way? It's early days, and we'll have to see how it goes. The other important thing that we realise is that signalling and discovery are the key. How do you know when this person is online or offline? This leads me to the second project that I talked about. You could use a distributed hash table to do that. So let's have a look. What is a distributed hash table? Everybody knows what a dictionary is in Python. Yes? OK. It's a distributed one of them. It is literally a distributed and decentralised key value store. There's no single point of failure or control. It scales to a huge number of nodes as well. Lookup is relatively efficient, although obviously it's done over the network. And it also, depending on which algorithm you use, the one I'm using is one called Cadmilia. It has good handling of fluid network membership because of course there are nodes joining and leaving the network all the time. And it's also tested in the real world with distributed hash tables because BitTorrent and FreeNet and other similar projects use distributed hash table for lookup. But they use a distributed hash table for just their application. And what we were thinking about doing is a universal distributed hash table. And that application could store key value pairs in this DHT. So the universal DHT, it's my current obsession, programming project obsession. I work on it on the train when I go into London and those late nights when my kids are finally gone to bed and so on and so forth. So development is a little bit slow. But it solves the problem of discoverability and signaling because people can leave their status within the key value pair, within the dictionary. And friends can look up. We also had a quick look at... Sorry, we didn't look. We had a think about how we could make this work. I'm not going to talk about this very much because I'm not even sure we understand it. But we were discussing a platform called P4P2P which is distributed hash tables within distributed hash tables. So these are namespaced in some ways. So that particular applications can use particular parts of the network that best meet their needs. So you're probably sitting there thinking, well, he's had about half an hour now. This sounds far too utopian, hippy. And it'd be quite valid for you to ask why. You're obviously crazy, you guys. And that's usually quickly followed up by... What about the economics of this sort of stuff? How is development funded for peer-to-peer systems? How do you put food on the table? So, well, let's think. Serendipity. There's a good example of that that happened last year at EuroPython. I kept getting tweets about Holger from friends who just happened to be in the room. I met people that I'd never met before at these sprints, but they're here and they're my friends now. So serendipitysly. We met and we're collaborating together. Why is that? Well, perhaps it's because we share the same values. We actually care very passionately about our privacy and working in a world where a peer-to-peer system is some way of enabling us to build the digital world that we want to live in. It's also fun. That's a good reason why you might want to work on this sort of thing. These are fun engineering problems, as I hope I've demonstrated. And it's also that itch for me. Everybody has a different sort of an itch, but that's my particular issue. That's my particular itch. It might be yours. It's also important to remember that there was no economic argument made when the web was born, as Tim Berners-Lease said, that the web is more of a social creation than a technical one. And he designed it for a social effect to help people to work together, and therein is the value of the web. He didn't sit down going, hmm, what world-dominating hypertext system could I invent? It sort of grew from the bottom up, which chimed in with the keynote from yesterday morning. It's also important to remember that even as far back as 1996, William Gibson, the science fiction author, I'm sure you're all familiar with him, said in an article that the worldwide web is the test pattern for whatever will become the dominant global medium. The reason I'm saying this is because it's very easy for us at this time, after 20 years of the existence of the worldwide web, to have worldwide web goggles on. Everyone seems to see things, we must have a website for that, we must use a RESTful API, we must use HTTP, because that's what everybody uses. Perhaps it's time that we might be able to think outside the box and think, well, what should come after the web? What post-web solutions and digital architecture should we be using? Which leads me onto Alan Kay. Alan Kay is very famous for saying the best way to predict the future is to invent it. We're in a very privileged position as developers because we could actually build that future with Python. I actually like this quote more. I believe that the only kind of science computing can be is like the science of bridge building. Somebody has to build the bridges and other people have to tear them down and make better theories and you have to keep on building bridges. What's the next bridge after the worldwide web? The penultimate slide, so I've nearly finished, don't worry. This is some cuniform in the British Museum. It's 5100 years old. It's one of my favourite places to be the British Museum because you can't help but get an enhanced sense of perspective. In two years, it's a huge amount of time. This is 5100 years old. It's one of the earliest examples of writing that we have and it records the allocation of beer. You'll be pleased to know. By administrators in the city of Uruk. The symbol representing beer is actually an upright jar. I'm not sure if I can find one yet. Okay. Like that one. With a pointed base. The amounts of beer that these workmen or workladies have been having is denoted by the circles and the crescents. That's their counting system. If you look in the bottom left, there's actually a person drinking from the bowl and that's kind of like the receipt to say that the goods have been received. There's a bowl. It's a bit of a small aspect. So I would like to end by asking you, is the worldwide web our cuniform play tablet and what should we be building afterwards? If you would like to discuss this more with not just me, but my friends that went on the sprints as well, because what I presented here is very much a group effort. Meet us in the foyer at 5.30 this afternoon. We'll have a chat and we'll probably go afterwards for beer and food. The end. Is there time for questions? The key signing is arrived parallel to the time. I can't hear. The key signing. Is almost in parallel to that time in the basement. So it's a bit unfortunate somehow. Okay. I didn't realize that. It actually we looked into key signing when we were our first sprint. Now you've got to remember there were about nine highly technical people in the room and we managed to do it wrong. Which says a lot for the key sign that there are lots of ways that you can improve security. But we got it totally wrong. No more questions. Okay. Oh. How did I guess? David who's going to ask the question is my colleague. So what is the best way to organize humans to organize humans? That was your your big question. Okay. So. I should say that David has a philosophy degree. As do I. And my answer would be if you come along at 5.30 this afternoon. You can work out the details then. Actually we could put a note in our sprint plan. We can create a ticket to discover what the best way to organize humanity is. You said there is a small community of people working on that stuff. But the only thing I find on your website is contact is your Twitter account. So what medium do you use to communicate except meeting in meet space for some sprint. There should be some more effective way without driving through Germany or whatever. Yeah. You're probably quite right. I'd say in our defence this is early days. We're a group of people who are just exploring the ideas. And it's not as if we've announced a political party or a new free software project or something like that yet. We're getting together to think about these ideas. We communicate on IRC in a channel that has nothing to do with peer to peer. Because it's run by one of the guys. It's his company IRC channel on FreeNode. So I'm not sure he might not like me sharing that channel. But it's informed. Say something now. Okay. So the first thing you can do is come along at 5.30 and we can talk and share email address. You can't. Okay. The second thing you can do is you can prod me on Twitter and I will get back to you. The third thing you could do is probably annoy Holger and send him emails because he's quite a high profile person as well. That will be able to disseminate information because people follow him an awful lot. But you are quite right. This is something that we need. We've been enjoying ourselves with distributed hash tables rather than IRC channels and Twitter accounts. I'm afraid. Arena, do you want to have the microphone over there in the middle? I haven't recorded that microphone. It needs to be recorded for posterity. Hi everybody. I just want to say because it sounds like you guys have a bit of a clique out there hanging out. For those who want to join a broader movement, there is a redescentralised mailing list. Readescentralised at Libralist.com with public archives where there's a huge community of people who are interested in redescentralisation technology, adoption, how do we change stuff. That is a really good discussion list and there's a whole website as well which you can join and follow and get involved with other people and have discussions. I would encourage people to do that. Definitely. Join Arena's list. I'm a member of it. Libralist.com. If you search for redescentralised.org, don't go to Google, type it into your URL. Redescentralised.org. Through the magic of the internet, this web page will appear. They have interviews with various people who are doing very similar projects in the PISFIR. I think because we need to turn the room around for the next talk. Now's a good time to finish. One more question then. Okay. Hello. The five questions that you had there on the blackboard. I think the most interesting is the fifth one. Some political philosophies in Poland recently said that it's not about who we give the power to. It's about how we can remove them when they fail to the river, what they were supposed to do. It seems that this distributed peer-to-peer communication is about taking away the power from everyone so that nobody holds the power. The power is not centralised. There are many instances of stuff where, for example, when you make a standard, you need to have a centralised power to make the standard happen. You need some kind of standardising committee. For example, like English, for example. Do you have any ideas of how to organise the removal of people you don't trust anymore? Just for your interest anyway, this blackboard, that's half the blackboard. The bottom half is even better. But it was only the top half that was pertinent for this part of the talk. This was created by a UK politician called Tony Ben, who recently died. If you look up Tony Ben's five rules, you'll see the whole picture. He's very, very good. That's it, I guess. Thanks a lot, Nikolas, for your great talk.
|
Nicholas Tollervey/Holger Krekel - The Return of "The Return of Peer to Peer Computing". At last year's Europython Holger Krekel gave a keynote called "The Return of Peer to Peer Computing". He described how developers, in light of the Snowden surveillance revelations, ought to learn about and build decentralized peer-to-peer systems with strong cryptography. This talk introduces, describes and demonstrates ideas, concepts and code that a group of Pythonistas have been working on since Holger's keynote. We asked ourselves two questions: what are the fundamental elements / abstractions of a peer-to-peer application and, given a reasonable answer to the first question, what can we build? We will present work done so far, discuss the sorts of application that might be written and explore how peer-to-peer technology could be both attractive and viable from an economic point of view. ----- This talk introduces, describes and demonstrates concepts and code created during sprints and via online collaboration by a distributed group of Pythonistas under the working title p4p2p. We asked ourselves, as frameworks such as Zope/Plone, Django, Pyramid or Flask are to web development what would the equivalent sort of framework look like for peer-to-peer application development? We've tackled several different technical issues: remote execution of code among peers, distributed hash tables as a mechanism for peer discovery and data storage, various cryptographic requirements and the nuts and bolts of punching holes in firewalls. Work is ongoing (we have another sprint at the end of March) and the final content of the talk will depend on progress made. However, we expect to touch upon the following (subject to the caveat above): * What is the problem we're trying to solve? * Why P2P? * The story of how we ended up asking the questions outlined in the abstract. * What we've done to address these questions. * An exploration of the sorts of application that could be built using P2P. * A call for helpers and collaboration.
|
10.5446/20004 (DOI)
|
Next speaker is Matt Williams. He has worked on the computing infrastructure of CERN and he's going to talk about how the LHC's computing grid works. So please give a warm round of applause for Matt Williams. Thank you. So I'm Matt Williams. I recently finished my PhD in particle physics. I was working on the LHCB experiment on the LHC for four years, recently graduated. And I'm now working at the University of Birmingham, working on computing resources for the scientists who themselves are doing the analysis now. And it's part of that work that I'm helping to help develop this tool, GANGA, which is an interface used by scientists to interface with a huge amount of computing power and storage available to them as part of the LHC computing grid. So a brief little update in case anyone here doesn't know anything about CERN or the LHC. It's the world's largest particle physics experiment, or at least the world's largest man-made one. It's arguably the world's largest man-made structure as well, being a 27-kilometer-long ring on the meets underground in a tunnel dug specifically for the purpose. It's a proton collider, so it's accelerating protons to near the speed of light and climbing them together at four locations around the ring. And each of those, there's a detector which studies the outputs of those collisions and analyzes the data that's given to them. Given the huge amount of collisions that are happening every second, billions and billions are happening, it's outputting a huge amount of data. I mean, the amount of data that it is producing is way beyond what we would actually be able to collect. But the stuff we do collect to date equals something like 200 petabytes, though already it's probably a bit higher than that, and it's only going to grow as the accelerator gets more and more powerful in the future. So in order to be able to process that huge amount of data, alongside the design of the LHC was a corresponding project called the grid. The idea of this was to produce a computing environment which would be able to handle the large amounts of data and processing power that would be required. So it works in a tiered system. So at CERN, there's a central hub, a tier zero grid site, which has a large amount of computing power. And they then defer down to a single site in each country that's involved in the LHC. There's about 12 or 13 of those tier one sites spread around the world, one in each country that's involved. And the level below that are the tier two sites, there's around 160 of those. Each of those is generally something like a university or a search institute. There will be a dozen or so in each country, for example. Some countries have more, some countries have less. And it's at the tier twos and the tier one where the largest amount of data processing is done. And the sort of data that we study at the LHC in the sort of analyses that we do really does lend itself to this sort of distributed nature. You tend to end up, if you're doing an analysis, with a list of collision events, maybe you've got 10 million, 100 million events you want to look at. You can very easily take a small chunk of those and process them independently of any other chunk of data. There's no real interaction between the events. So you can very easily chunk it up, send it out to wherever it needs to go, and then collate the results at the end. So as I say, the project was, the project was evolved alongside the LHC. So even in the early days, well before the LHC actually started, people were looking into building these computing systems to provide the services to the scientists that need them. So in 2001, the LHC project started work on GANGA. This was their in-house specific interface to this grid infrastructure. Each of the other experiments were also working on their own personal projects in order to interface with the grid, since everyone was convinced that they had their own special problem that only they could solve in the way that they needed it done. However, the LHC project GANGA was designed using a Python system with the explicit goal of being pluggable and extensible and so on. And so it was very easy in the intermediate years to take the parts of it that were LHCB specific and remove them and allow other experiments on the LHC to plug in their small part of experiment-specific logic that's needed. So the Atlas experiment, there's a number of scientists on the Atlas experiment who are using GANGA for doing their data analysis. And in fact, outside of that whole ecosystem as well, there's the T2K experiment, which is the neutrino experiment in Japan. Some of their scientists are using GANGA for interfacing with the grid resources which are provided to them as well. Of course, all the software that we create at CERN, or as far as I know, all the software is completely open source. GANGA itself is GPL, and the vast majority of software that comes out of CERN is GPL or other more liberal licenses. So how does that actually work? So if a scientist has a bit of code they want to run, they can use this tool, GANGA, to interface with the grid system. Or in fact, not just the grid system, they can interface with any other system that GANGA has an interface to. So in this case here, you see on that second to last line, we're setting the back end to be equal to local. That's telling the GANGA system, don't run this on the grid, just run it here on my machine. That's something that's often done by scientists when you're testing a bit of code. If you've just written a new piece of analysis software, you don't want to immediately throw it up onto the grid infrastructure, run it 10,000 times and have it crash within three seconds because of some bug you've put in. So it's a good idea to test it locally on a small set and then later on be able to submit it up with the grid. So all sensors around this job object at the top, you can set some parameters on it. Here we're setting the name parameter to give us a string which we can use for bookkeeping, keeping track of what jobs we use for what, since all the job information gets stored into a persistent database where you can see it all later. The real workhorse of the job system behind the scenes is the application. So the application is what is actually going to be run where this thing is going to be run. In most cases, you just want to run an executable. It can be an executable binary or it can be a Python script or in this case, it's just a small shell script. So you just say to GANGA, this is the thing I want to run, this is the actual code that's going to happen and this is where you can find it in this file here. In this case, this script is just going to create a file called Out.txt. And so we're telling GANGA the output files from this job. These are the ones that are going to be made by it. These are the ones we want to make sure end up back where we are now. We want to make sure we've got a copy of those in our local output directory, wherever the job was actually run and whether that file was originally created. And so we specify that it's a local file. In output files, local file means copy it back to where I am locally. Once we've set up our job object, we just call submit and at that point the GANGA subsystem is coming to play. The monitoring loop comes in, it starts submitting the job to the system. In this case, it's just going to start up a local shell instance somewhere else on your computer. But if you were accessing the grid, it would be uploading it to the grid somewhere. It would then keep track of its status and make sure it's downloaded any output files at the end of the job. So once it has finished, you can just access the output files directly inside the IPython based GANGA user interface. So you can just call the peak method on the job you just had and it basically does any else of the output directory. You see it's created a file for the standard out on the standard out and most importantly the out.txt we asked it to give. And if you want to peak further inside, you can pass a name of one of those files to the peak and it will open up a pager directly inside IPython and you can have a scan through and look at the output files to make sure that everything worked the way you wanted it to. Obviously, that was just a toy example. That's nothing more that you can do there than simply running a local script on your local computer and looking at the output file. So it would be good to be able to leverage the power of the grid. And it's as simple as changing the back end on that last step from local to LCG, where LCG stands for the LHC computing grid. It's the acronym that we use for that. So with one small change of one line to the other, you could run exactly the same script and that code would be uploaded to the grid system. The grid system would take over, distribute it, run the code wherever it ends up running. You don't even worry. It could be in China, it could be in America, it could be in Amsterdam, it could be here in Berlin, it could be anywhere. And it's completely seamless to the user. At the end, the data will be copied back and everything is the same. You don't have to worry about it. But Ganges is more than just that. It's more than just locally running stuff and the grid. It can interface with anything that you can access via an API, basically. So there's a series of back ends for, you see here, PBS, LSF and SGE. Those are batch systems. Often universities have got local batch systems or a batch farm of some kind, which they use for running jobs which are somewhere between running on your local computer and you want to upload to the grid. And again, you could just change it to PBS and it would be submitted to your local farm and you wouldn't even have to worry about any of the details. These last series of one here are a set of experiment-specific back ends. So various experiments have got their own middleware interfaces sitting between Ganges and the grid to make an onion layer type situation to provide extra features that maybe that experiment particularly needs. But again, it's all a black box as far as user concern. You don't have to worry about what's going on. It's just going to work. So now that we're using the grid, it would be good to really make use of the huge amount of power it provides. So let's say, for example, you have sitting on your local hard disk, directory containing a whole load of files. Maybe you've got 3,000 files or something in there. Each of them some number of megabytes. So adding up to a gigabyte, let's say, a data or something like that. So there's a lot of data you're going to want to analyze. You can tell Ganger that these are the input files you want to run your job over. From that point on, Ganger will keep track of those input files. It will make sure they get copied to wherever the job runs, whether that's locally on your back system or if it needs to be copied out to the grid, it will make sure those files end up where they need to be. Of course, to be left at that, it would be pretty useless because you'd be taking one huge chunk of files and copying them to one place on the grid and they would just be run in one single compute node somewhere. It would be good to be able to distribute it around and make sure we're running things in parallel. Ganger provides a tool for this called spitters. So again, you define on the job object a splitter parameter. In this case, we can use the spitbyfiles object. This is an object which knows how to spit the files up into a smaller set of data. We simply take one parameter, files per job. It's going to take this list of however many thousand files you have, chunk it up into chunks of 10 or maybe less files if there's not enough to fill a chunk, and take each of those chunks, add in your analyzed data script that you want to run with it, submit it, and the grid will put it somewhere. It will take the next 10, that will go up, and that will be sent off somewhere. It will keep doing that all the way through the list and you'll end up with some number of hundred subjobs that Ganger will keep track of you, keep track of for you. So you won't have to worry about how many subjobs are made or doing it manually. It's completely automated. At the end, each of those subjobs is going to create a histogram.root. Root is a file format we use at CERN that's basically a table of data as far as this stuff is concerned. It can also contain histograms and so on, basically a table of data. By specifying the local file here, this isn't saying that file is going to be made locally. It's being made wherever the job is run and you don't care where that is. But you're asking Ganger to copy it back to your local computer so you can have a look at it, open up in your text editor or analysis software or whatever you are going to use to analyze the data. But that's not ideal even then because you're going to end up with however many hundred copies or variants of this histogram.root. They'll get put in a subdirectory structure but still they're going to be separate files that you're going to have to go through manually and look at. So to solve that problem, Ganger provides something called a merger. It provides a whole suite of mergers but the one in particular here is the root merger. So this is a little bit of Python code which understands how to concatenate together root files. You can stick them together and combine them into one single file. And this again is completely automated. Once the job's been uploaded, it's been split up, sent out all over the world. Ganger's downloaded all of the results from each of the single sub-jobs. Once they're all downloaded, Ganger will automatically kick in, combine them together and start turning it into one single file which you can then look at. So from that point of view, you don't even have to worry about the fact it was split. You started off with one single analysis script and one single set of data. You spit it and merged it, you ended up with one result. You don't even have to worry about the fact that it was distributed around. It's completely seamless. There is much more than just the root merger. You can write any sort of merger you might wish. Anything which post-processes the data, basically. As a class in Ganger which lets you pass in a single function which simply takes the output file directory and you just list through there. You could, for example, look at the log file for each output job, grep through it for a single string and find the average of the numbers or something like that. You can do anything you can think of to post-process your data. So once you've been working at CERN for some number of years, you're probably going to submit several thousand of these jobs over your lifetime. Many of which you're going to have deleted because maybe they broke, but many of which you're going to want to keep around for log files to check that stuff's working how it used to make sure your data's being reproducible. So Ganger provides a persistent database of all the jobs you've ever run through that system. So you can see here the three jobs that we've submitted so far. The first one we just found locally. That's all finished. It's showing up there as competed. The last two, because they were sent off the grid, they've been distributed around and they're still running. You see here each of them, Ganger has created 324 subjobs. That's how many it decided to spit it into. You don't worry about the number too much. You just have to know that they're there. We don't have any more details about which of them are running, which of the subjobs are finished or anything like that here, this is just a very high level overview. But it's very possible to get that information because Ganger provides a full API access to everything that's inside the Gangers API. So inside the Python interface, you can access any of the information. You can access job information. You can resubmit things. You can do anything you want. So the most simple, we call that jobs object again, like we had in the last slide. We give it a parameter. We ask for job number two, which is the bottom one here, the merger job, which is as far as we're concerned, overall still running. We ask for status. And again, it tells us it's running. It's the same information. We can delve in a little bit deeper, though. We can ask that job for a list of all its subjobs. So we just give it the dot subjobs parameter. That's going to give us a list of jobs. We can loop through each of those subjobs and ask each of them what their status is. We get a list of all the ones that are completed. We find the length of it and we find that 24 of those 324 subjobs so far have finished. If we waited half an hour and ran it again, it would be a high number because gang was constantly keeping track of how many subjobs have finished. But jobs don't always just be running or have finished. Quite often you'll get random failures. On the grid, your data will be sent to run at some particular site. It could fail without any real reason. Maybe there's an outer memory error on that particular location or things like that. So as long as some of your jobs are passed, there's a good chance that those that failed, it's simply a transient failure. So you can loop through all the subjobs once more, check if the status has failed on that particular subjob and resubmit it. It will go back into the monitoring loop and keep going around and eventually it will be redownloaded once it's finished. This is the sort of thing you might want to do quite regularly. You might want to have a function defined which loops over a job object, checks all the subjobs, resubmits the failed ones. So you can take any bit of Ganger code, stick it inside a function inside a dot Ganger dot py file in your home directory and all those functions will automatically be available inside the user interface which is based on a Python. It's a slight fork of IPypy Python to provide this sort of functionality. So the last thing I want to talk about is dealing with very, very large files. So the example I gave at the beginning, I was saying you might have a directory on your computer which has got something like a thousand files in it or something like that. Even if it's only some number of megabytes running up with gigabytes of data. In fact, quite often when you're doing data analysis with the LHG, you're going to be dealing with at least gigabytes if not terabytes of data. You're going to be wanting to run your analysis over. So it's nice not to have to keep those files locally on your local computer and upload them every single time you want to have to do an analysis over them. And then at the end, if the output's big, you don't want to always have to download the output. Maybe you just want a summary file. Maybe you just want to find the number of events that pass some sort of criteria. So as well as being a distributed compute network, the grid is also a distributed file system or at least it provides a number of distributed file systems. The one in particular here is using this DRAC file system which is again originally an LHCB specific grid interface. But the important point here is that it deals with a remote distributed file system. You don't have to worry about where these files are. They're out there in a way in the cloud. So for the input files here, we tell Ganger that we want as our input file a file called input.root and we're saying DRAC knows where it is. So I don't know the exact physical location, but the file catalog knows where to find it. For the output file, my program is going to create a file called histogram.root. That's going to be made locally on the worker node wherever my job is run. And I don't want that copied back to my computer here. I want you to send it off to the remote storage. That will keep track of where it is. That will keep a record. I can access it later if I want to. So for now, I don't want to be dealing with all that network traffic coming up and down. And in fact, it can even be a little bit cleverer than that. Using the DRAC backend, which is basically a layer on top of the OCG backend, it's got a bit of extra logic in there to deal with this sort of file system access and so on. One of the clever things that can do is all you have to say to this with this exact script here, you upload that script, you submit it, DRAC would automatically take your analyzed data, like program you want to run. It will look around, find the physical location where input.root is stored, and it will send the job to that site. And it will run it there locally rather than submitting that analysis script somewhere and copying the files over. It will try and automatically reduce the amount of copying that's going on in order to make things as efficient as possible and avoid clogging up network bandwidth. In the same way, the output is going to be stuck somewhere. And so you could then run a second job. You could change together jobs. You can say, this is the output of job one. I want that same output to be the input of job two. You just have to pass in input files equals DRAC file histogram.root. That job will be submitted to the grid. It will go up, it will look around, find out where histogram.root was saved to, and again, it will be sent and run there. You never have to deal with those files on your local computer at all. You let the experiment or CERN have to deal with all that storage and file management. So, yes, so using the grid like this, you can deal with hugely large files. I mean, I've never had to deal with them. Of course, for each sub job, you'll get back a standard out file. You'll get back a standard error file so you can make sure your jobs are running correctly. You can always have some files being sent off the DRAC, some downloaded locally, some sent off to some mass storage place. You can have as many input and output files as you want coming from whichever source you want as long as GANGA has an interface for it. And GANGA being extensible, you could very easily write a new plugin which dealt with any other file system type that you might want to use. We do have a file system type which uploads things to Google Drive, for example. Quite often people just want to be able to share files to Google Drive and so you can access uploading, downloading files from there. So you can write basically an interface to any infrastructure you might want to be using yourself. So you can find out more information at the website, cern.ch.ganga. Like I said, all the code is completely open source so you can go to the download link and have a poke around with the source code. It was started, the project was started in 2001 which for reference is about the time that Python 2.0 came out. So some of the code is quite, has been around quite a while but on the whole it's quite readable and you can see what's going on as far as the job flow goes. So take a look at that if you want to have a little poke around and thank you. Questions? Thanks for the nice talk and the nice tool. I have two questions. So the first question is can you target some schedulers such as Slurm using the library? Yeah, I don't know if there's a Slurm backend yet but there's ones for condor and talk and so on so there could easily be one for Slurm. I mean it's a simple case of writing a bit of code to call the right commands at the backend so yeah, Slurm could absolutely be interface with if necessary. And my second question was, I forgot the second question. Okay, I'll ask you in the break. We got another question over here. Thank you. There was on your merge slide. Can you go back there? I don't understand the line with the J.input files where there's actually a list comprehension on... So in this case, input.cst is in index. So you open, overload it somehow or you get a file handle. In this case, input.cst contains a list of file names. So that's an index file containing a list of all the file names that you want to include as your input for your job. So you're going to loop over each of the lines in that file, each of which is a string which is the name of the file. But that's not what Open usually does, isn't it? So you just have a file handle. If you have Open.readlines, you get, of course, the lines. When it's looped over in this list comprehension, it does produce a list of the files. It has a slash n at the end of each one. But it does work. I did check this line. Yes. Hello. Thank you. I wanted to know how you handle code that runs parallel and needs to communicate to other processes or in different computers? So inter-process communication between analysis jobs and things. Or internet work communication. On the whole, there's very little scope for communication. I mean, Ganga is blind to that. If you submit to a supercomputer which has got some inter-process communication that you need to do or some sort of communication of any kind, it will handle that because Ganga doesn't care about it. Really jobs on the grid, you don't have any sort of communication between them. Each job's siloed, very much so. So I suppose you don't submit jobs that need to be run across multiple processes? Mostly not, no. No, not in the sort of work that we do. Okay. How does Ganga find files? I mean, you can't just use the name, right? So each of the local file or DRAC file, for example, each has got a little bit of logic in there. So local file, by default, will look in the working directory that the user's in. DRAC file will, so obviously you can't just say input.root, it's going to match it no way it is. By default, it's authenticating with the DRAC system. So each person's got a local user area. So it will look in their local user area for the file. And likewise, it will be saved to their user area on the file system. So you have to be careful not to use the overnight name? Yes, yes. You can overwrite files in a previous job and things like that. You can give multiple output directories and stuff, yes. Okay. Thank you again, Matt.
|
Matt Williams - Ganga: an interface to the LHC computing grid Ganga is a tool, designed and used by the large particle physics experiments at CERN. Written in pure Python, it delivers a clean, usable interface to allow thousands of physicists to interact with the huge computing resources available to them. ----- [Ganga] is a tool, designed and used by the large particle physics experiments at CERN. Written in pure Python, it delivers a clean, usable interface to allow thousands of physicists to interact with the huge computing resources available to them. It provides a single platform with which data analysis tasks can be run on anything from a local machine to being distributed seamlessly to computing centres around the world. The talk will cover the problems faced by physicists when dealing with the computer infrastructure and how Ganga helps to solve this problem. It will focus on how Python has helped create such a tool through its advanced features such as metaclasses and integration into IPython.
|
10.5446/20002 (DOI)
|
Hello. You'll have to bear with me. The presentation will definitely fit on this screen because I adjusted it for 700 and it's 1900. So I have a lot of extra space. So who am I? Well, I was ready to juice just now. Been doing a lot of stuff and most recently more path. So why would somebody create a new web framework in 2014? I started in 2013. I'm going to try to show you why you can still do new things with more or less traditional routing web frameworks. And in order to do so, I have to contrast more path with other web frameworks. So I'm going to do a laundry detergent commercial. You know, when you put the stuff you put in your washer, you put this white powder in. And in the traditional commercial for that, you have the shiny box there and you have the evil brand X, which really sucks. So I'm going to be this annoying sales guy who is going to tell you all that the shining box is awesome and brand X is horrible. So yeah, brand X, what do you mean with brand X? Brand X is one of the popular routing web frameworks in Python that you probably use. I will not name them. And by this talk, you might learn something about more path, but also more about brand X because the contrast works both ways. That might benefit you as well. And I really won't name brand X. Okay. It could be bottle or flask or Django or Pyramid in its most common configuration. Pyramid is kind of special. Well, they're all special and they all have their benefits. I'm not trying to put them down. I'm just contrasting. And of course, I'm going to say that brand X sucks and I'm better at more path and all that. But Pyramid is especially special and I learn a lot from it. Who here uses a brand X web framework? Like routing and fuse and stuff. Yeah. Lots of people. So a little bit about the more path origins. I won't go into this very deeply. So this is exploding planet. Soap and just at the last moment, they shoot out this sort of hero with superpowers. As long as you're not exposed to old pieces of the planet, soap, right? That's really bad. So they shoot it out. That's actually not the first thing they shut out. They already shut out a whole pyramid before. It was only crumbling then, but you know, the pyramid people were a bit smarter. They sort of crumbles earlier. But anyway, we just got out in time. It's okay. And the soap pieces are really easy to recognize because they sort of have this weird alien green glow. So it's simple. Anyway, so what are the goals of more path? More path is focused on the modern web. And the modern web means rest and it means rich JavaScript based client side applications run in the browser. They use some kind of rest backend. And more path tries to be easy to develop with to be powerful so that when you are sort of trying to do something more sophisticated, you can still do it. And it tries to be small as well because it needs to be embedded in an existing system that didn't want to take too much on board. So yeah, I claim that more path has superpowers. It looks like to this innocent clock and guy who definitely doesn't have superpowers on the outside. So it looks like your average flask or something. And then you know, it pulls open his shirt and then suddenly the superpowers come out. And it's important to realize that the superpowers of more path are not a different mode of working. Clock and actually has those superpowers too. And doing the normal clock and things, you know, routing and views and all that stuff in more path automatically sort of doing the super things is sort of doing more of the same. So you only need to learn the primitives that you don't need to learn something new in order to use those superpowers. So I'm going to discuss three topics. I'm going to compare routing with brand X. I'm going to compare linking more path linking with brand X. And I'm going to talk about reuse in more path, like how you reuse code. So let's discuss routing. So we have a route URL path and it goes to a cat. If you're hoping for more beautifully drawn pictures, I didn't have enough time to make more. So we're going to get boring slides now. So it takes time to draw that well. So anyway, so we have a route to a cat. And that sort of means get the representation of an animal with ID cat in a typical routing framework. And we have one extra requirement. If the cat ID does not exist in the database or whatever, then we want to get a four or not found error from our web server. Pretty simple case. So this is what you do in a sort of typical, I mean, this is hypothetical code, but it's sort of what you typically do in a brand X framework. You declare some kind of route with a variable in there, it's ID there. And then ID is used in a function that does two things. It queries the database for an animal with that ID. And then it represents that animal, in this case a JSON, but you could be using a template or whatever. So yeah, that looks simple. But we're actually not done. We haven't fulfilled our requirements. If we query an animal that the ID, an ID that does not exist, if we're querying for a T-Rex and it's not in our database, then we will get a 500 error because it tries to find a T-Rex and maybe this thing returns none. All of many ORM mappers will return none. And then you try to get a title of none and that doesn't work. And there's an exception and the web server will make this into a 500 error, but that's not what you want from your API. You want a four or four error. So you have to add code. You have to say, well, if there's nothing there, you have to say, okay, we're going to return not found. Maybe you can do an exception. Maybe you just return a response. So this is spent X routing with error handling. So this is more path routing. We split the code up into two functions. If I knew I had 1900s horizontal, I would have put the first line on one line, but I thought I had much less available. So but there's two functions. It's the same amount of lines of code if you write normal lines. You don't have normal line lengths. It's the same amount of code, at least same amount of lines. So first you say there's a path to a cat or to an animal. You have to say what kind of class you're going to return from that function first. That will be used later. And then you do the query there and then this returns either none or an animal object, an instance of animal. And then you have a way to represent that animal. You have a view. You say, okay, this is the default way to represent an animal. You make its title. You split it up like that. And yeah, so it could first go to the model and then a viewer's looked up for the model. It's a two-step approach instead of a one-step approach. Now a nice benefit of this, that not found actually happens automatically because this will return none the first function, getAnimal, and then the system knows, oh wait, we don't have an animal to represent, then four, four. So you don't have to do anything special. You just work in a normal way. You get a four, four. That's nice. And you can also have multiple named views for the same model. If you also want to have an added view or whatever, you can give it a name and then Animal slash cat slash edit goes to the added view. So this biases, yeah, it's harder to do it wrong because it's easy to forget. It looks simple, but then you forget the special case. And you get better linking. So now we go on to the next topic. So let's do a little rant about linking. So links make the web work, right? Animal, HTML, websites work with links, web applications work with links, REST work with links. I mean, there was a talk about hypermedia APIs on Tuesday talking about how useful it is for loose coupling and scaling over multiple servers to use links and let your client follow links just like somebody clicking on a link in a website. So why then do brand X web frameworks barely care about link generation? They do hardly anything. What they do is this. So you have to give your route a name so you can refer to it later. Then you have to use that name and you generate a link somehow, you know, some API. And you have to know that an ID needs to go into the template of the URL to make the link work. So you have to know the name introducing tight coupling between your routes and your code that uses the routes. So if you change the routes, you might have a problem. You don't want that tight coupling. I thought routes were for loose coupling. So you could change things. And you have to know what parameters go in and you have to extract that information from the object in order to put it in there. So yeah, I just discussed this. This is more path linking. So there's no change to our previous code. It's just exactly the same. And then you just have an animal object and you say, give me a link to it. And it works for any object that we know how to make links for. Any object that has a route declared with a path decorator will be linkable. And this is loose coupling. You can just make a link to an object without knowing what it is, making it possible to write generic code that doesn't need to know about what kind of links you want to generate. And it's just easier. You don't have to remember all this stuff anymore. You just do it. So let's look about a linking with query parameters. So imagine instead of doing what we did before, we do slash animals and then we have a query parameter in a URL, it's called ID, and then we give it the cat. It's very similar to the last case. And of course, this is a bad example. It's a very simple example. That's why I put it there. A better example would be like some kind of filtering search API. The idea of URLs is that in a good, restful web design or a traditional HTTP website, the client does not construct any part of the URL except for the query parameters. So that's why I'm giving you this example here. So yeah, you want to get a representation still the same. You want to get a four or four if it's not there, if you ask for the TREX. And if you don't supply the ID, well, you want to assume some default, like, okay, the default animal is a cat, why not, right? Or you want to say, no, there is no default. If you don't supply an ID, that's a bad request. And you want to get a bad request error from the system. So in brand, actually, do that like this. You have to add another special case there. You have to say, okay, well, first you have to know that you have to extract this from the request. You have to do that. If it's not there, then we have to do something special. You have to do all that extra work. And your function is getting less simple. So and if you were to have no default, then you want to raise a bad request. Now Flask does automate that, actually. If you ask for something and automatically raise it, and that doesn't exist in your request parameters, we're sorry, automatically raises bad requests. So that's kind of nice. But most of them, you have to implement it yourself. In more path, this is the more path way to do that. So we've actually not changed the last three lines of the code at all. We've barely changed the top. The only thing is we've changed the URL path there to just slash animals. We've added the parameter to our getAnimal function, and we have given a default there, like Python. And that will do this for us. So now you can just... So, yeah, this is the same so you don't need to do something special here. And now we can look at linking with query parameters. So in brand X, in some frameworks, they have different things, but they don't do very much. So either you refer to the route name and then you give it some keyword parameters, and that will be added to the query parameters. Or perhaps they recommend that in your template, you start adding things there that's really ugly, or you do it by hand, basically. They just sort of drop the ball on that, typically. They don't really think about query parameters very much, because after all, they're a fundamental mechanism of how the web works. So I would think about that, right? So in Moorpath, it just works like that, like before. There's no change. This link will still generate a link. It just will generate one with the cat ID in there now, if you... My animal happens to be a cat. And Moorpath also knows about the type of the parameter involved. So if you give it... If you say this parameter needs to be an int, and often it's enough to just give the default parameter and make that be an int, the system will assume, okay, well, this should be an int. So if you then give something that can all be converted to an int, a string, as a parameter, it will try to convert it, get a value error, and say, okay, wait, we cannot convert this 400-bit request. So we'll do that for you for free. And it will do that for all kinds of parameters. You can actually plug in your own. So by default, it does dates and things like that, timestamps. So that moves on from linking. I hope I've shown you that linking can be done better. So let's look at reuse. So Moorpath offers a lot of facilities for reusing code. Because when web applications grow, or you have different pieces of web applications, you want to combine, or you have an application, and it's perfect, and it's developed by somebody else who just want to make few changes, you want to do reuse. And you want to make that easy. And Moorpath reuse is not like a special case. You don't have to go to a special kind of subsystem and learn all these new things. Reuse is just there as part of the system. So let's talk about view reuse first. So here we have a collection of animals. That's maybe on slash animals, you know, without this cat bit there. And we want to return some kind of JSON that has an array of animals, a list of animals. And instead of creating a list of links to animals, which we could have done with request link, we want to actually embed the information about the animal. In the JSON. So we can just say, okay, give me the JSON representation, or at least the Python representation that translates to JSON for the animal. And it doesn't care what kind of animal it is if you are doing, if you're getting a list of whatever, and you don't know what they are in this code, this code can be generic and still embed them or link to them. So that's view reuse. You can just reuse views. And there's again, you have loose coupling, and you write generic code by default. It's actually easier to write generic code by default. So here we also have a generic view. So if you have a live form base class, you have an animal subclass that subclasses from live form, and you make a view for live form, that view also exists for animals. It's just inheritance, basically, right? So you can make a generic view for all live forms. We say, well, for elephants, we really want to add some extra information. You can make an exception. You say, okay, well, for elephants, we have this extra thing going on here. So you can do stuff like that. And that allows you, again, to write generic frameworks. You can write a generic collection-based class that then has a set of generic views, and you just have to fill in a few details. And that sort of flows from the primitives of the system. You don't have to do something else and learn new things. You can have more than one application. So a mobile application recently, actually, in the most recent release, became a class, and you just have two classes here, and then you can do all these paths and routes and stuff to the, you can add them to the classes. And they're independent from each other. So they don't interfere with each other. They don't share anything. So you can just have two of them. And that's actually very useful. But we'll see a little bit more of that later. First we're going to talk about inheritance. You can also just inherit derived application from a base application, meaning you will share everything with the base application. So you just get that. You don't have to, it's just Python, basically. But you're sharing everything, not just methods, but all these path registrations, all that stuff is also inherited. And then you can do override. So I can do extensions. So let's look at the extensions first. You can say, well, the base, and just inheritance is just copying. It's boring. You want to add something. So you can say, OK, in the derived application, everything is as the base application. But you have one extra view. It's called slash animal slash cat slash extra. And that does some extra thing for you. So you have the same application, but one extra little thing added. Which is kind of nice. You can think about framework applications that offer a sort of set of framework views that you can then reuse in multiple applications. You can do override. So it's still kind of like Python. So we have a base application here. And the base application has a default view for animal. But then you say, OK, well, our base application is great. All its routes, views are all great. We want to change one thing. OK, you can do that. You can say, OK, I re-register the default view for animal. Then in the derived application, that's the view you'll get. But if you use the original application that's maybe maintained by somebody else who doesn't want that override, it will still work. It will still work as before. You can do composition of applications. So you have an application, a user application. And you have a Wiki application. And they're independently developed from each other. Or maybe you are developing both of them. And you don't want to think about users when you develop Wiki. So you don't want to think about Wikis when you develop users. So you don't want to have a URL space with users. You don't want to have to think about worrying about users when you are developing the Wiki URLs. So you just have two applications here. The one special thing we've done there is to say that the Wiki application has a variable, expects to be, it's parameterized. And it's parameterized with the Wiki ID. You need to, in order to create a Wiki application to instantiate it, you actually have to give it the Wiki ID, otherwise it won't do anything. And they're just completely independent from each other. They just, one does a Wiki, and Wiki pages, and the other one does users. And now you want to combine them in each other. So you say, OK, we have the user application. And we want there to be a slash Wiki on every user that has the Wiki. So you say, OK, we mount the Wiki application onto the user application. And then we give it a, this is the path where we want to mount it on. So we want to mount it on Wiki. And then we have to say how to get the Wiki ID from the context of the user application. So we know the username, and we need to have some way to find the Wiki ID for a username. You look it up in some database. But that's only in the mounting code. The Wiki doesn't need to care about user names anymore. And you can merge them together. You actually still have access to the username if you want to in the Wiki application. So Moopath has a bunch of other features I won't go into details about here. There's a built-in identity and permission system. So you can protect views with permissions. And you can define rules for your specific existing models of this kind, like animals. People only have the edit permission if they, you know, this table in the database says so, whatever. You can just do whatever rules you want. It's a very flexible system. It doesn't assume anything. So the basic core of Moopath doesn't make many assumptions. It does let you to come up with whatever rules are appropriate to your application. Moopath is extensible, the new view predicates. We've really only seen the name predicate where you have multiple names views. Those are get requests, but you can also make a request for the post view. So you can have a request method that's built in. You can extend it to say, okay, this view only matches when the HTTP accept header in the request says this and this. So you can extend to that in sort of the normal Moopath way with a few decorators. It's also extensible with new converters. If your application has specific, you know, data type, like a car or whatever, and you have a way to represent it in URL, you can define a converter for it to just parse that and also convert it back again from a car object to a representation that you can use in your URL, either in your path or in your request parameters. And you can extend Moopath itself. So you can say, I mean, I haven't documented this. So it's sort of a special thing. And a Moopath application is a bunch of generic functions. You can actually override those generic functions in your application. So if you don't like the way Moopath does routing or whatever, you can actually override little bits. And it's sort of using the same mechanism to implement Moopath. That Moopath sort of allows you to override and extend it. Moopath does have a few extensions. I have a more dot transaction extension that I sort of copied from the pyramid version of that. That basically integrates with the transaction module. And the transaction module integrates with SQL community and whatever else sort of integrates with that. And that automatically commits when a request is handled, unless there was an error or you're returning explicit status code that indicates error, then it will not commit the transaction. It will abort the transaction. So that's kind of a nice feature to integrate these systems in a general way. And recently I released the more dot static, which is an extension that adds the ability to publish static resources like JavaScript files and CSS files in a cache-friendly way, but also in a developer-friendly way, busting the cache when you need it to the web. And it sort of plugs in as a whiskey middleware. It's similar to fan static, but it's oriented around the Bower tool. So you can use the Bower tool to install these JavaScript libraries. And then you can just start using them without having to do any extra special work to wrap them. It's similar to fan static, but there you need to do that extra work. And all this stuff is documented. Well, some of the extending more path itself is not documented, but the rest of that is all documented on morepathredetox.org. I just checked the PDF version of the documentation, and it's about 90 pages. So I ended up writing quite a bit. Let's look a little bit at performance. Maybe there's a huge performance cost that makes morepath very slow. But I did a benchmark. I did a very simple Hello World benchmark, and it's raw whiskey calls. There's no real whiskey server there. And morepaths, well, it's sort of in the middle. Some systems are a lot faster. But of course, in reality, when you look at the real web application, the overhead of actually implementing your stuff, like doing the database interaction or generating views, it's going to be so much higher than the web framework. It's negligible. But I just wanted to check that morepath is not ridiculously slow, and it's not. It's faster than Flask, but a lot slower than Bottle. Code size, like maybe morepath is enormous. It's like, whoops, Zope. And no, it's not. I checked, and I was kind of surprised that morepath, depending on how you measure it, is smaller than Bottle. Though if you realize that Bottle has no dependencies whatsoever, I don't know why they don't have a, yeah, I don't see the reason not to use dependencies. But morepath is a few dependencies. If you add them all up, it's about, well, it's going to be smaller than Flask, or definitely, and they'll call the code base. Reg is sort of the main library that morepath depends on, which is sort of a rewrite of the old Zope component architecture in terms of generic functions. The tests, I excluded the test in the doc strings. In measuring this, the tests are a lot bigger than the actual code that's being tested, which is how it should be. Yeah, we, WO is Webop, and Zi is Zopelot interface. I didn't list all the dependencies, and WZ is, I would have spelled it all out if I had, if I knew I had 1900 horizontal instead of 756 or something. Conclusion. So, yeah, I hope to have shown that routing, link generation, and reuse in morepath is better than your brand X. Morepath tries to be, and I hope I've shown that morepath is both easy to use, right? It's not much harder than Flask. It's not so intimidating. At the same time, very powerful in its reuse and overwrite and potential like that. And it's also still small, so it's not intimidating in sort of what's there. And I hope it didn't make you frustrated with brand X that you may be using. So are there any questions? Yeah, that's what happens when you... That's what happens when you watch your tiger with brand X, right? Yeah. Martin, thanks for a very... If you have the questions, would you go to one of the mics? Yes. Thanks for a very interesting talk. What is the kind of status and also the outlook or so? Is it kind of like stabilizing and you think that the next things are going to happen with them? What are your next plans basically with that, right? Because there's also the question of, you start to use it now. What changes do you need to adapt? So maybe a few weeks ago, I thought it was pretty stable. It was not going to change any massive way. And then I decided to make applications classes instead of instances, which was a rather big way, though not a high burden for a developer to actually adjust any code in. Now it's not. So I don't think there's going to be any changes that may give you a huge problem if you start developing with it now. My plans are mostly involving writing extensions for it, like looking at some of these REST fall standards for making links to things, building on top of more path, not drawing more path itself. There's also a lot of potential for writing an extension that sort of reimplements some of the pyramid authenticators, the ticket-based authentication, but that's all extension. That's not core stuff. So I don't think I'm going to change it very much anymore, but you never know. Within three weeks' time, I have some brilliant idea. But even then, I don't think it's going to be a giant burden on whoever has a code-based stand. So it's getting pretty stable, I think. We are starting to use it ourselves in our own project now. So people are starting building real-world in my customer project, real-world code with it. So it's ready, I think. Yeah. If there are any more questions, can you take them over there? We've got the next speaker. Okay, sure. If you can come and set up, please, and Martin will take any more questions. Thank you. Thank you. Thank you.
|
Martijn Faassen - Morepath: a Python Web Framework with Super Powers Morepath is a server web framework written with modern, rich client web development in mind. Why another new Python web framework in 2014? Because it can be done better: Morepath understands how to construct hyperlinks from models. Writing a generic view in Morepath is like writing any other view. With Morepath, you can reuse, extend and override apps as easily as you can construct them. Even if you don't end up using Morepath, you will learn something about how the nature of web frameworks. ----- Morepath is a new server web framework written with modern, rich client web development in mind. In the talk I will be discussing some core features of Morepath that make it different: * Its different take on routing and linking. Morepath has support to help you construct hyperlinks to models. * Its view system: plain views, generic views, view composition. * Morepath's approach to application construction allows application extension and overriding, and composition. This talk will attempt to convince people to try Morepath. For those unable or unwilling to try, I will communicate some design principles behind Morepath which can be of help to any web developer.
|
10.5446/20001 (DOI)
|
Okay, again, welcome to this 10 o'clock session on this Wednesday. With us is Markus Zapke-Gründemann, active member of the jungle community in Germany, and he will tell us something about multi-language documentation in Swings. Morning everyone. I hope everyone can hear me. Yeah, okay, great. So nice you all came to hear the talk. I hope I can tell you something new. Yeah, as Fabian already introduced, I'm a member of the German jungle community. And I'm doing software development since now, nearly 15 years. Started with Pearl, then doing some PHP now, mostly I'm doing Python, jungle stuff, and also involved with the open data community in Germany. And I also do a lot of training with people, basically, jungle training. And I have a company called Transcode, and I'm a board member of the German jungle association. So if you have anything about jungle in Germany, you can also talk to me. And you can find me on the internet on a kind link day or at kind link on Twitter. Yeah, so some basics. First, short introduction of Swings. So who is already using Swings here in the room? Okay, so nearly half of the people. Okay, so introduction is good. So Swings is a Python documentation generator, which means it is written in Python. It's not only for Python. You can use Swings for any project you want to. So yeah, any language you want to, you can even document stuff which has nothing to do with programming at all if you want to. And the Maka language is called restructured text, which is something which was, if I remember correctly, was invented by the Python community. It is something comparable to Markdown, which many people know already. And the interesting thing is that you have a lot of output formats. So you have your restructured text documents, and you can create a lot of output formats as you can see here like HTML, Latech, and of course PDF, EPUB, text, manual pages, plain text. So you also have the opportunity to display your documentation properly on different devices. And yeah, Swings can be found on Swings. Also the Swings documentation. Yeah, and of course, as the title of the talk says, I'm talking about internationalization. So it's often referred to as I18N, so because it has 18 letters and first it is I, last it is N, so this is often the abbreviation for internationalization, especially if you look at module names in programming. And the idea about I18N is that you can translate these strings inside your software without having to change your software all the time, because if you had to have your software and inside all the different texts in the different languages and you always have to switch there in between while writing your regular code, this would really be a mess. And of course, you also need a transparent system to exchange the messages, or the language of the messages, the people see that use your software. And the most or the best, maybe best known tool for that is a new Get Text, which is open source software. And this is also used in Swings to create the translations. So in simple Get Text example looks like this. There was a bit missing on the left side, but yeah, I hope that's not that critical. So it's simply that you create a translation catalog object. This example, the catalog is called example, and it's in a directory called locale. And 4 by 2 is simply for changing the way of exception you get if it's not existing. So it's not important at the moment. And so what is missing here on the left side of the screen is that there is an underscore there. So underscore equals to Get Text. This is usually done everywhere. So underscore is used as the alias for the Get Text function so that you have to write less code for the internationalization. Then you can see this print statement, always look on the bright side of life. And you can see that the string that is inside the print statement or the function caller is again inside this function call which uses this underscore, which is an alias of the Get Text. And this way you collect this string for translation so your source language would be English in this case and you would have English as the source of your translations. And then you can create inside this locale directory different example catalogs for each language where it translates this English thing to something else. So this is how Get Text basically works. You collect the source strings and then in the end you translate them to your target language. Of course this is what's happening behind the scenes, what the swing is doing for you so you don't have to do anything like that if you want to translate your documentation. Just for short explanation of technical background. But why use Get Text for translated documentation? Why not simply copying all my files to a new directory like from EN to I don't know DERGP and simply translating all the text documents again, bits of documentation again and then I have the documentation in a different language. So first problem is if the documentation is updated in the regular language so the English documentation is updated you have a hard time on finding which stuff has been changed because you can't easily differ between them because all your content is different because your content is a different language. Especially if you have a language like Japanese which has totally different characters like Latin languages. Second problem is if you leave something out because maybe you don't have the manpower to translate it at the time or you don't know what is the proper translation for this thing then it's simply gone. Or you have to copy and paste this paragraph to your documentation in English and remember that you still have to translate it. If you use Get Text, Get Text will replace all strings that haven't been translated automatically with the original string. So if you don't translate a single string or paragraph and you build your translated documentation you will still see the original string. Another thing or advantage is that your markup is not duplicated because if you think of a document with any markup like with such a text or markup on whatever, HTML and you copy this whole document and you translate it, you copy all the markup with it and if the markup is also changed because the documentation evolves and the way it is displayed is also changed with the next release of the documentation. You have to do this in your translation but you only want to do the translation in the markup so you can have another problem. And another problem of not using tools like Get Text is that you exclude people because for this usually you have to use tools like Get or Mercurial which they have to understand how all of this works, maybe do a pull request then somewhere to get your translation merged in and so forth. Sometimes people do translations are not familiar with this enough to get involved in this process. So if you can use professional translation tools, it's much more easier for the translators to translate something for your project. And these professional translation tools also offer you opportunities you normally don't have with regular text editors like these tools can for example show you, oh, it seems like that this paragraph has been translated before and here I have a text which is 90% the same the translation, 90% the same like or similar to the text that has been before so you can simply then get these 90% similar translation from your translation cache, put it in and simply change a few words. So also your translation of already existing and evolving documentation is faster. Okay, yeah, so how does it all work? The Sphinx number I explained. This is an illustration from the Sphinx documentation and this shows how internationalized documentation is built in Sphinx. So you see on the far left the RST files which are then built into POT files which are templates. So POT is the extension for the source files of get text and POT is a template for that so it's the source language. And then you translate these POT files here to like PUTL is mentioned but there are also other tools to a POT file. The POT file is then again transferred into, yeah, compiled into a MO file and the MO file is a binary version of your POT file which is faster on XS so you can get all these translation strings faster. And in the end then the RST files are taken again together with the MO files and then all the original strings are replaced by the translated strings and you have your translated documentation. If you want to do this you have to do a few things. First thing is if you haven't done this in your, so usually you have a docs directory, inside the docs directory you have a conf file and swings where you configure how a swing works. And first thing you have to do you have to say which is my source language, like the main language of my documentation, in this case it's English. And you have to say okay where do all my translations go like the other catalogs with the different other languages and in this case it's the directory locale. And then there's also a fairly new configuration value getXcompact and if you set this to true and your documentation has several subfolders all documents in a subfolder are collected into a single catalog, if you set this to false you get a catalog for each file. So because if you have a documentation which consists of 50 different files and five different folders you may only want to have five catalogs instead of 50 because it makes translation easier. So you have different opportunities here. So to do the translation in an easier way there is an extension or plug in for swings called Swings INTL and you can install this via pip so that it is available in addition to Swings itself. And this will then help you to deal with all the translation files because it is possible to use Swings itself and some getX tools to create the PO files and MO files and all the stuff but this is a very nasty way and also complicated way of doing this and Swings INTL makes this all easier. So yeah if Swings INTL is installed first thing you do is make getX which yeah, calls the built in Swings command to create all these PO files so it extracts all the strings from the RST files and put them into these catalogs and then you say Swings INTL update minus L which is the language so the new language I want like German and where should I, where am I located, where are my PO files like this build locale directory so they make getX stores them in this directory and so this Swings INTL update command will fetch all these PO files, make getX created and create new German catalogs for that. Of course it will be empty because this command is only like copying them and preparing them so you can insert the German translation strings. So the next thing you have to do is translate the documentation. I will talk in a moment about the details for that and then you, after you translate everything you can say Swings INTL build and then it will create all these MO files that had been shown the image before and then in the end you can again say make HTML which creates for example the HTML documentation, you can also say make PDF. Lartech or something else make EPUB and with the Swings opts, yeah, options you can say that the language should be something different so you override the language in the conf pyify to create a German HTML instead of an English one because English is your default language and now you want something different. So this is the four commands you need to create the catalogs and then translate them and build a different documentation language. So what you can use for translation, one, two of course and also others but what is very helpful at least to me is trans effects. So trans effects is a web service so it's not a software you download, it's more like a website you use and it's free for open source projects but if you use commercial projects you have to pay them of course because they also need to have some income. And trans effects really works like a professional translation tool which features I said like show me similar strings I had before and other things and it's also very nice tool because you can collaborate with many other people on the same translation and people can work together to translate all the stuff online. PIP has trans effects as a command line tool, trans effects line, you install it, you set TX in it and you can use the name and password so you can talk to the platform and then you can use Sphinx INTL together with trans effects because Sphinx INTL has trans effects support and as you can see you first run this update TX so TX is the abbreviation for trans effects update TX config resources which prepares trans effects so that it can be used together with your Sphinx documentation. Finally you have to do a lot of steps manually and Sphinx INTL is doing everything for you here and then you can say TX push minus S and this is a little bit similar to Git or MacUriel so you push your stuff to trans effects and minus S means sources so I push my English catalogs to trans effects so that other people then can create translations and translate it to German, Spanish, Japanese, whatever. All of the translations have been translated on trans effects, I can pull them again with TX pull and I can explicitly say which languages I want, for example here again I want to pull the German translation some people created on trans effects and then again say okay like before Sphinx INTL built, Sphinx make HTML for German and now I have my German for translation and can see everything in German. So this was like the workflow you usually use and during the years I'm already working with Sphinx IITNN I discovered a few things that are helpful in the everyday work with international ideas documentation. So if you use code inside your documentation, here's a snippet of template code for example, you should be careful and always use English inside these code examples because otherwise if for example you use German here and something would be translated to Japanese, the Japanese people would have a hard time understanding what the content of the template really is about. There have been requests on the Sphinx mailing list to also translate code but this hasn't been decided on how to deal with that. And if this is possible because at the moment only like the written text is extracted from the Sphinx document and not the code of course. Another thing is there are different ways to handle URLs and RST. You can use the syntax above which puts URL directly into the text and you can use the syntax below where an alias is used and if you use the alias version, URL gets lost because it is not longer part of the catalog and especially if this sentence is translated, the name Python sources would be translated to Python code or something similar in another language and so the name of the reference is lost and the URL can't be inserted. So always put the URL inside your documentation so that it can translate this properly. Another thing is you often want to refer to external URLs and often these URLs are referring to documentation which is versioned like you see it's referring to Django documentation for 1.5 and if you then switch your documentation to refer to Django 1.6, you don't want to go to all of your documentation and change it and especially not all the translators want to go over all of their translators stuff to simply update the version number. They can use a feature of Sphinx so the thing about the dictionary X links is something from the Sphinx configuration and they define these aliases and then as you can see below in your text you can simply use these aliases as prefixes for the rest of the URL and so you like magically can construct these URLs and only have a single point where you define the prefix on the left side. And how to handle special cases. There is a if config construction in Sphinx. We always can check against the current language and so you can have every documentation like special cases for example here is a link to the German community and of course you want to have this part of the documentation only in the German language and not on the other languages so you can always have some kind of special part of your documentation only for a specific language. Last tip is how to run link check stuff. For example there is a link check target and also the link check target and many other targets except the language switch and link check simply goes over all your documentation and checks the URLs and of course you want to check your for every language so this is also possible with all the different languages. So what is still missing? For example a translation setting is missing so that you can build all translation with one command. At the moment you have to execute all the steps for every single language and if you have a documentation of 500 languages there is a lot of work you have to do so of course you want a translation setting like language with your original language and translation is for other languages and you can simply build all of them together on the row. Another thing that is missing is a landing page for the HTML version because you need something like a page where you refer to all the different languages if someone comes to your page where your HTML documentation is. Another optimization could be you do use getXcompact to create a single catalog instead of many catalogs and the rest is not that important. I think we use the rest of the time for at least a few questions before time is up so thanks for listening and please ask a few questions. Okay Q&A there is one microphone over there. Yeah. Hi thanks for the talk. So if you have several source files and documentation and you choose to have several catalog files but then you have small words that occur in every file do you have to translate them in every catalog file or are they shared within one Sphinx? Oh yeah. Yeah. I was sort of on that maybe excluding an example slide where I show how this catalog looks like so we should do that. So with Sphinx the always a full paragraph is extracted so every paragraph is a single message so you never have single words that can be translated. This is only true for the index so in Sphinx you can have an index of keywords where you then which is in the end of the book which can link them to different parts of documentation and these index words are translated separately but all the other stuff is simply translated per paragraph. I hope this was your question. Yeah. Any other questions? I mean. Yeah. So I'm available all the day here and I'm here to Saturday and on the sprint on Saturday we will also have a sprint on Sphinx so if you have any ideas on how to optimize stuff or I don't know I have any other questions you can always come to us on Saturday and talk to us about improving Sphinx. We'll be happy to hear from you. Yeah. Okay. I hope you've enjoyed the discussion.
|
Markus Zapke-Gründemann - Writing multi-language documentation using Sphinx [EuroPython 2014] [23 July 2014] How to write multi-language documentation? What tools can you use? What mistakes should you avoid? This talk is based on the experiences I gathered while working on several multi-language documentation projects using Sphinx. ----- Internationalized documentation is a fairly new topic. And there are different approaches to do this. I will talk about how Sphinx internationalization support works, which tools and services I use and how to organize the translation workflow in an Open Source project. Finally I will have a look at what the future of internationalization in Sphinx might bring.
|
10.5446/19999 (DOI)
|
Okay, we will proceed. Mark Andre Lemberg will talk about advanced database programming. He is a season Python developer, being around since 1993. Also founder and CEO of Egenix.com, one of the founding members of the Python software community, Python Software Foundation, and a board member of the real Python Society which brought this lovely conference to you. Give him a warm welcome, please. Thank you very much for coming. I'm going to give a little talk about advanced database programming because we've in the past heard a lot about the easy stuff. So I thought it might be a good idea to talk a bit about the more advanced things. A bit about myself, I'm Mark Lemberg. I've been using Python for a very long time. I've studied mathematics. I have a company doing Python projects. I'm a core developer, Python Software Foundation member of the European Society and I'm based in Düsseldorf. So this is the agenda for the talk. I don't know whether I can actually do everything that I have on this agenda because of the time constraints, but I'll try. First I'm going to start with a short introduction of the Python database API 2.0. How many of you know the Python database API? Interesting, not that many. So the Python database API was designed in the, well, the design started in the mid-90s. That was the 1.0 version, which is now deprecated and we're now at 2.0. So it's a very old kind of standard that was developed. The development is ongoing on the Python DB SIG, so if you want to join the discussion there, you just have to subscribe to that mailing list and you can add your thoughts to the standard. It's supposed to be a very simple kind of standard. It's supposed to be easy to implement so that we get as many database modules as possible and I think that has worked out really well. The two main concepts in the Python database API, one is the connection object and the other is the cursor object. So you use connection objects to actually connect to the database and also to manage your transactions and then if you want to run queries, you open a cursor object and you run your queries on the cursor object and the cursor actually works like a cursor in a text processing system. You actually scroll down or like in the spreadsheet, you scroll down in your result set and then get your data into your application. So this is how a typical application looks like that uses the DB API. So first you import your module, you get the connect API from that module, you open a connection, you pass in the database name, the username and the password, then you create a cursor object on the connection object and you run your queries on the cursor object and finally you free the resources by closing everything again. So that was a short, very, very short introduction to the DB API. The next part is going to be about transactions. Transactions are a very, very useful thing in databases. You can do stuff on your database and if you find that you've made a mistake, you can just roll back your changes, which is very nice to have. You need it in production systems to work around bugs or input errors from users so that your database doesn't become corrupt. So it's very useful to use these transactions. However, there are a few mistakes that people often make and this sometimes causes people to not like transactions. One common mistake is they forget to commit their changes. So they apply a lot of changes on their curses and connections and then close the application and see that the database hasn't really changed because the database API defaults to transactional. It doesn't actually store the data if you don't do an explicit commit. Now a workaround for this is to just disable transactions, which of course is possible in databases as well, but it's not a really good workaround because instead of losing your changes, you get data corruption for free. Another common mistake that people make is they keep the transactions running for too long time and I'm coming to that later in the talk. The transactions are basically your units of locking things in the database. So you want to keep transactions short to not lock other processes from accessing the database. So what you have to do is you have to try to make transactions short. Now the best practice is, of course, like I said, always use transactions. Even if they are sometimes annoying, don't use auto commit. Always try to make use of them. Keep your transactions short. If they get too long, you can run them in batches. For example, if you're loading data into your database, it's much more convenient to do that in batches, say a thousand rows of time and then you do commit. That also keeps the transaction lock of the database short and the performance will stay just fine. So you won't really see the overhead that is caused by the transaction mechanism. And if you know that you're not actually writing to the database, it's a very good practice to set the read-only flag on the database. You can do that in the connection. You can usually do that in the connection options. And then the database will know that it has a read-only connection. So it will basically not work on this not-use-the-transaction lock and make the whole query mechanism run much faster. So that, again, was the simple kind of level of transactions. Then we have a more advanced level of transactions. If you want to connect multiple databases and you want to have transactions span the different databases, then you have to think about what to do. When you read data from one database and then put it into some other database, of course, you only want that to succeed if all the databases have actually received the data. And that's what's called distributed transactions. Typical use cases are in accounting, for example, you debit from one account on one database and you credit the amount to some other database. You only want that to succeed if both databases have actually made the change. And you have similar things in Q-processing or if you want to integrate different applications. Now the typical buzzword that you'll hear when you're talking about distributed transactions is two-phase commit, which is the kind of the standard method of approaching this problem. So it works like this. You have a first phase and the first phase, the commit is prepared. So all the different databases are asked whether the commit would succeed with a high probability. And if all databases say yes, then you go to the second phase and you actually do the commit. Now there's a tiny probability there that the database, some database in that process may fail in that second phase. And then I can say I'm sorry then your data is corrupt. You have to work around that in some way because there's no easy way of undoing the second phase. But most databases just handle this fine. To make it easier to deal with these transactions across multiple databases, there's something called a transaction manager. This is not in Python. This is usually done in some other system. For example, there's MQ series from IBM or J2EE or Microsoft has this DTC mechanism, some database systems, DB2 and Oracle offer these transaction managers. You can sometimes hook into them from Python. There are Python APIs for some of these and you can then use them or you can use a database specific one like for example the one that's integrated into Postgres. In the DB API we have addresses with a new set of APIs, the TPC API. Those are modeled after the XOpen standard for these transaction managers. But unfortunately not many databases support this and not many database modules actually provide these APIs. So you have to check your database whether it supports this or not. Okay, next point, concurrent database access. Am I going too fast? Too slow? Concurrent database access is very important if you have multiple processes accessing your database. For example, if you have a web application because of the GIL you would normally want to have multiple processes set up to talk to your database. And so it's important to think about how the databases deal with the problem of concurrent access. So you have typical setups, what I've written down here. The most typical one is of course again the many writers, many readers so there's no real special case that fits in the common case. You definitely need to make compromises in these setups. So when writing an application and thinking about this you have to ask yourself some questions. For example, should readers immediately see changes made by other transactions or made by other processes? And should they even see things that went into the database even though the transactions in those processes have not yet been committed? Or should the reader just see everything as it was when it started the transaction so it doesn't see anything that came into the database after it started its transaction? And the databases can handle all these different situations. They provide different what's called transaction isolation levels. But they have to implement this using LOX. And LOX is something that you usually try to avoid in your application because they keep your whole application from running in parallel and using most of the resources that you have. So I'm going to walk through the typical sets of transactions, isolation levels that you have. The first one is the easiest one to implement. It's read uncommitted, which basically says you don't have any LOX. With read uncommitted isolation level, all the processes that you have that talk to your database will immediately see everything that changes in the database, even uncommitted changes. So strange things can happen. You can have dirty reads, which means that you read data from another process that hasn't been committed yet and you're not sure whether it's actually going to be committed later on. You can have phantom reads, which is something that basically says you add something to the database and some other process and then you remove it again. Your process might have read this row that was added by another process and later on it's removed again, so it's basically a phantom that you're working with. And there are some other things that you have to watch out for. If you want to read up on these things, there's this URL down there. It's going to be in the talk slides. You can click on it. It's a very good explanation of these things. Then the next level is the read committed. This is the default level in most database applications. So when you open your connection, you will usually get this isolation level. This isolation level basically says you're only going to see changes that were committed to the database. Now you can still see changes that were made while you're running your current transaction. So if some other process commits while you're running your transaction, you will still see those. But you will not see any uncommitted things from other processes. So the way that it works is you have this cursor. I drew this table up there with the yellow bar in it. It'll lock the current row. It'll put a read lock on it. The database will, if there is a write lock on a row, it'll wait for that write lock to be removed by the database. So if some other transaction has written to that row, that other transaction will have put a write lock on the row, and the write lock is only removed if the transaction is committed. So if the other transaction has committed to change, then you can actually go ahead and actually read this row. So this is basically you just looking at one row. Then the next level is repeatable read. This basically says that your data won't change in the transaction. So everything that was returned to your application by the database is guaranteed to stay the same throughout the whole transaction. And this, of course, requires more locks on the database. So you put locks on everything that you actually pass back to the application. And then the highest level is serializable, which basically means whatever you do on your database, the database will stay like that, will stay exactly like it was when you started the transaction, and nothing will change. And this requires lots and lots of locks. The locks will not only address the things that you've read from the database or you've written to the database, but everything that you've ever touched in the database. So even whole tables, if necessary. All of these levels are necessary for some applications. For example, if you want to run a report, you may want to avoid inconsistencies in the report. So you may, for example, want to use the serializable isolation level. And these other levels can be used if you have situations that are not as strict about data processing. So how do you do this in Python? There are two ways to do this. Well, actually, there are three. You can usually have an option in the connection settings that you can set to set a default isolation level. But you can also do it dynamically in your application, and you can actually do it on a per connection basis. So you can have multiple connections to your database and different isolation levels. You can run a statement, set transaction isolation level, for example, or you sometimes some database modules have special ways of directly setting the option on the connection. What's important to know is if you want to change the setting while having a connection open, you need to make sure that no transaction is currently running on that connection. So the easiest way to do that is just commit your connection or roll back. Right. Organizations. So you have a database application. Of course, you want to run it as fast as possible. The first thing that you should do is you should ask yourself what kind of system you're running, whether you're running an OLTP system, which means online transactions. So you're interested in putting lots and lots of data in. You're not so much interested in making complex queries on the database. Or you want the other thing. You want to have data analysis. So you already have all the data and usually huge amounts of data in your database. And you're interested in doing complex queries, multi-dimensional, faceted search, drill down all these things. And then you use an OLAP system. Now just like with transactions, the situation is often that you actually want a mix of both. So you want to have both run fast. So there's one way you can do this, depending on the size of the database that you're talking about. One way is to put an OLTP system in front of the system that's actually taking in the data. And then every now and then you copy over that data into your OLAP system to analyze it. On a more lower level, you can, in Python, you have a certain number of problems that you can address directly. One is, for example, your queries run too slow. And the queries are simple. So you're just doing, for example, the select on a few tables, a few columns. The usual way to address that is just you add more indexes. Because adding indexes is very easy in database. Some people, they add indexes to everything they have in the database. This slows down things. Because every time right into the database, the database has to update all these indexes. And so you should really only put indexes on columns or on combinations of columns that you actually need in the database. And the best way to find out which tables and which columns you have to index is best to use a query analyzer. The database usually offers a way to get the information about how a query is running the database. And you have a look at that. You analyze the data. And then you check which indexes you should put on the database. And it will increase the performance enormously. If you're using Python, you can, in some situations, also add caching at the Python level. So you basically read your data from the database. And you store it in memory for subsequent use. You can even use SQL Lite for that if you have smaller datasets. And they do in-memory processing. Now the next point is complex queries run too slow. For example, you have a report that's running on millions and millions of rows. Those will usually take a few minutes to run, depending on how complex they are. Of course, users don't want to wait a few minutes for this. So a common strategy for doing this is to pre-process some parts of those queries. So every, say, 50 minutes, you run the queries, you put them into a separate table, and then you run your reports on those query tables. And again, if your queries themselves are too complex, you can address that in Python as well. What you do is you simply split up your queries, make them easier to handle for the database, and then you combine the results from those queries in Python. A typical example of that is if you want to run a report that has aggregates right in the result set. So doing that in SQL is really hard. And you can do it, but it's really complex. And it's much easier to do in Python. So for this example, you just run two queries. You run query for the details, and you run one query for the aggregates, and then you combine everything into a single table, and you set in your application. Tips and tricks. This is just a collection of random stuff that I just thought might be interesting for you. Typical problem you have is record ID creation. So you want to put in a new row into your database. You need an ID for that, for that record. And you have this kind of chicken and egg problem, because the typical way of doing that is to use an auto increment column in your database and just have the database deal with that, where you have a sequence and get your IDs from that sequence in the database. The problem there with the auto commit, for example, is race conditions, because with the auto increment, the database will take care of adding the increment value. And then you have to, of course, fetch that value again, because you want to continue working on that row. And the usual way is to ask the database for the last used ID, and depending on how you do it, you run into race conditions or you run into context problems, because it's not really clear what the last ID is. There could have been some other transaction, for example, that also just inserted something. And so it's not clear where to get that last ID from. Another way of doing that is you just let the auto increment field insert your ID for you, and then you just query back the row simply by knowing what's in that row, and you query the database for that row, but that introduces the performance overhead. So it's not really ideal. Something that we always use in our applications is a very simple approach. It's kind of a randomized approach. We simply have a big integer field for the ID, and then we use a good random number generator to generate the ID for us. And we just bet on not having collisions. So we use that ID in the row, we put it into the database, and it usually succeeds. In the very, very rare cases where it doesn't succeed, you just create a new random number and then you try again. How does this work in Python? Well, first you have to set some constants, so you have to have a row range ID. What we often do is we set the highest bit, so the IDs look nice. Then you have an API get random row ID. This needs to be thread local because every thread should generate its own ID, so you don't get any overlaps. And then you have to deal with setting up the random number generator. And the best way to do that is to use system random to get a good seed for the random number generator. You put the seed, you put it into hex, and then you feed the random number generator with that, and then you use that in your thread local. Right. Next point is referential constraints. People are usually very happy about using them. Does everyone know what a referential constraint is? Oh, very few. So it basically means this. Instead of, for example, you have a table up there with a, you want to reference a product. Now instead of referencing the product name directly in your table, what you do is you reference the ID into the table that has all the product names. And then you just put the ID into your table instead of the name. And then in your reports, you combine all those things into a nice looking output. And this process of referencing from one table to another is called a referential constraint. Usually you use foreign keys for that. You can implement one-to-n mappings, you can implement n-to-m mappings. The constraints are enforced by the database, and that can sometimes lead to problems. Because if you, for example, you have lots of references in your database schema, and you want to load data into your database, then it will often fail because the way that the data is loaded does not actually match those referential constraints. And with some database, you can switch off the checking for those constraints during the load phase. But it's not really ideal. Another thing that can happen is if you delete something, then you can get cascaded deletes, which is not necessarily what you want. So what we do for these things, we just completely leave out the referential constraints and put everything into the Python database abstraction layer, which has much more knowledge about these things. It's also a good idea to put those things into the business logic. And then you avoid all these things. So if you wanted to immerse into an emergency situation where you have to quickly load data from a backup again, you don't have to think about how to turn off the referential constraints. You just load everything and it works. My availability is one of those things that you have to think about. So you have multiple databases. One database server breaks. You want to switch over to the other one. The various systems for doing that tend to not always work perfectly. So something that can happen is you can have automatic failback, which means that you have a failover situation where the system switches to a different database and then it automatically comes back, but you're not necessarily sure that all the database servers have actually synchronized by then. So you can have the split, what's called a split brain situation. So you have the data is spread across different servers, but you're not necessarily sure whether all the servers have the same data. And of course, some clients may miss the failover event. So some clients may continue to write into a different database server than the one that is actually currently being used. So again, the best thing that you can do is you move back to Python and manage everything in Python. And then you can also handle the communication between the clients and make sure that all the clients know about this failover event. You can do the failback in an optimal way. And you can also use this for doing read optimizations. You can have, for example, the database, the application part writing to the database, write to one database, and then have the synchronization between the databases, take care of moving the data to the other server. And at the same time, while writing here, you can read from the other database, and you also avoid some of these locking issues. So I actually made it, 42 slides in half an hour. That's it. Thank you for listening. Any questions? Yes? Thanks. So thank you for your talk. You said transactions shouldn't be too long. What's a good measure for the size of a transaction? If you're writing to the database, I'd say just maybe like 10 to 100 rows that you write in a single transaction. Okay. Thanks. Hi. You may not want to use random numbers for your IDs, since you may break the internal awareness of location and the internal balance tree. So it's maybe a good idea to use auto increment for the awareness of location. Otherwise, you may have some performance issues during writing to the database. Okay. Yeah. Well, we've measured it, and it worked out fine. But in some databases, we're using sequences, for example, from the database to generators. It depends on the implementation of the database. I think it depends on the implementation in the database. So my SQ and NUDB would may break them. Okay. We're mostly using Postgres, so maybe. Hi. I have a question. Maybe you can help me with that. Maybe I'm missing something, because when you are talking about problem with slow, complicated queries, you suggested that you may split it into multiple simple queries and move the weight of the computation a little bit to the Python. And my question is, why would it even speed up the process? Because I imagine that the database is constructing such a way that such queries should be done in the maximum performance. So why would Python do the same logical thing faster? It's not so much about speeding up the operation. It's about making it possible in the first place, because there are some things that you want to do in the reports that are not possible in SQL, because SQL is still limited for that. And even though sometimes you can do things in SQL, but you get really huge SQL statements to do your processing, or you have to resort to procedures and everything, which makes things a lot more complicated. So we found it's usually better instead of just wasting time thinking about how to structure your SQL properly and making it very complex, it's easier to just have a few simple SQL queries and then just do the combination of those in Python. Okay. Thanks. The slow world may be confused a little bit. Thank you. Sorry. Hi. Thanks for the talk. It was very interesting. Just a suggestion based on experience, because I've been a DBA in a former life. If you people do plan to do applications, talk with your database people beforehand. And just they're usually nice. They don't bite you. Just talk with the database people. They can make your life much easier, because there is so much implementation detail to be considered. So maybe that's a good question. Just to consider, buy them lunch, get out, talk to hours of them, it saves you about eight hours of programming time. Yes, thank you. Thanks. Very good suggestion. By the way, many of these things are database specific. So you really have to know your database if you want to make proper decisions. So they don't necessarily apply to all databases. Some databases are better, some are worse. We found that Postgres is a great database, so use it. That's our recommendation. Thank you. Thank you. Thank you.
|
Marc-Andre Lemburg - Advanced Database Programming with Python The Python DB-API 2.0 provides a direct interface to many popular database backends. It makes interaction with relational database very straight forward and allows tapping into the full set of features these databases provide. The talk will cover advanced database topics which are relevant in production environments such as locks, distributed transactions and transaction isolation. ----- The Python DB-API 2.0 provides a direct interface to many popular database backends. It makes interaction with relational database very straight forward and allows tapping into the full set of features these databases provide. The talk will cover advanced database topics which are relevant in production environments such as locks, distributed transactions and transaction isolation. ---- The talk will give an in-depth discussion of advanced database programming topics based on the Python DB-API 2.0: locks and dead-locks, two-phase commits, transaction isolation, result set scrolling, schema introspection and handling multiple result sets.
|
10.5446/19997 (DOI)
|
Okay. So here with us are Fabrizio and Marche who are going to tell us about their Python-driven company. So please give a warm round of applause. Hello. Hello everyone. I'm Marche. That's Fabrizio and here's the rest of the team from TBG, the company that we're representing today. So it's not our companies, we just work for them. But so what's a Python-driven company is about. So basically we were challenged with quite a big problem at some point in our career. We're working at TBG and the solution that we found was actually branded by us as a Python-driven company. So before we actually start taking you through all of the challenges that we have to meet, let me actually tell you the story that we're fighting with. So what was the challenge? So we are quite a modest size team. It's only five of us, three developers. It's here and those three guys. And basically three developers, one QA and Scrum Master. And then our management came to us and said, hey guys, by the way, so would you be able to provide some tools for the management of data, producing report, helping to do the statistical inference and help all the bunch of our analysts that know actually nothing about programming. They actually would like to have shiny UIs. And there are quite a lot of them. There are some people from finance, actually they would need your help as well. There are people also even in HR. That would be nice if you could help them as well. So then we just looked at each other just doing the simple counting from one, two, three, four, five and then just realizing that it's actually a little bit too much for us. By the way, we started quite a lot of new features, new projects and we started to move to the area that was kind of unknown to the company. So when it comes to specs, so we had this incredible requirement, this point number two and the second point was actually like we really don't know what the specs are. So be prepared for everything, which is quite easy if you have five people. Please deliver frequently naturally and of course quality first. So when you deliver an application, that would be nice from management point of view to deliver the highest quality possible. So we started to scratch our heads and saying like should we actually look for a job and actually should we actually take that challenge? So now the big challenge. So we actually started to think about how we could approach that. And the first thing is actually we started to think about the basic foundation of working in a company because we realized that actually, so there is no way that we'll be capable of delivering reach UI application that will help automate some of the task of the people that will actually make their life much easier because it's simply not enough of us and simply we don't have enough time and we won't be able to fulfill all the points that I just outlined a few seconds ago. So we asked ourselves, all right, so what's the current? So we asked ourselves what's the current structure of a company? And then we realized that actually the structure of companies is even though we use all the agile or the buzzwords of the scrum inside detectives, we are pretty isolated units. So basically if you think about it, you have a big guy in the middle and surrounded by small managers and then of course you have a team themselves. So basically if you want to automate or automate some of the tasks of the guys that are doing the main work that are doing actually the heavy lifting, we should go to them. So we shouldn't actually rely on the usual communication channels. We should actually think about the structure of a company like that. So the first impression that you can have is actually, so this is like one big mess. So please imagine that each color represents people coming from the different departments and of course color in the size indicates importance of a given person in a car key of the company. So detecting actually should be blended into the landscape of a company. So we thought how about we actually go to people and talk with them and just ask them like, all right, what do you actually need right now? What's your problem? We don't want to know what is a high level specs of actually a very advanced application that maybe in a few months having the team of this size we'll be able to deliver. So then we started to ask even more questions to ourselves, the others, in order to define our approach. So the first question is like, could we find a platform that would combine flexibility of Python and some of the web tools? So why are we actually decided to ask Python oriented questions? Because the first thing we realized no UI. We went to our management to say, you want to have those tools? Let's forget about the UIs. No user interfaces, no graphical user interfaces. So they say, all right, so how can you interact with the platforms with the application? We say, you know, we can code. And he's like, whoa. But wait a second, like those people know nothing about coding. So he's like, yeah, all right, so that's another challenge. We'll teach them. So basically we just realized how about we use iPod on notebook where they can actually, you know, inside those really shiny cells, put some of the snippets, share those snippets, share the solution, automate their work and we can by being blended into the company actually help them at every stage of the development. So actually we thought we'll provide some high level tools and they will provide more of them. So in other words, we'll go with the mashup kind of approach. And suddenly our development team actually exploded to tens of people because it's not five of us anymore. But if they will gain even like a slight knowledge of Python, they can actually build some apps for themselves. Some of the questions that they were asking and some of the struggling that they were having was mainly data oriented. So this is the world of Excel. So Microsoft Excel is like ruling the world of any analysis. And the biggest problem that they had with the big data, they were already calling it big data is that they're not capable of opening the Excel files because they're just too big. So they already arrived to the world of big data. And we said like, all right, so the first solution that we deliver was called the three analysis interface. So here is just very primitive and simple snippet. But actually this is all you need in order to fetch quite a lot like around half a gigabytes of data. And of course, we don't, the user do not have to know basically what we are using at the back end, whether it's Mongo, whether it's Postgres, how we actually transform those data, how we actually deliver them. They don't need to know all the primates. They just say like, all right, so I have this data store, which is like name of my client, let's say company A, and they say fetch me data out that. And of course, they will fetch the data in the form of data frame. So it's already more of Excel world. So basically we said, and when you have the data that you just fetched using three lines of code, you can either export them to Excel or perform your analysis in the notebook itself. So, and they actually interact with some of the services. And they ask us like, is it possible actually to automate those? So we deliver them a few solutions that enable them to just interact with the services, not only to fetch data and do the usual transformations, but actually to interact with the services to post some of the service in the server-related platforms to actually fetch the results, to combine them and so on. And at the end, and of course, as always, we provided like a back end for storing the data. Okay. The biggest challenge that Fabricio will tell you about was actually to teach them that all. So we provided the high-level interfaces, the programming interfaces, so that they could just log into the software that we exposed to them for hosting the notebooks. They could actually just go there, but what to type. So in order actually to achieve that, the first try was like, maybe we will try this guy and being crazy enough to teach them Python programming, or maybe we need someone with superpowers and that's Fabricio. So, but that was totally not easy. So maybe the story that I'm telling you right now is like, it sounds trivial. And actually when I was presenting those slides, I thought like, oh, is it actually a good topic to present during your Python? Because we are saying like, yeah, we just decided to teach programming to a bunch of non-programmers. But presumably they have no association whatsoever with anything programming related like anything. So, all right. So the thing that we had to actually keep in mind all the time was wise and house. So in other words, it's not only to show them the snippets and show like, hey, this would be super useful. You need to reinforce this idea that this would be actually useful and helpful all the time, every single small step. Like, hey, this is how you can automatically take or get your all of your data, do the analysis, and that's faster than any other solution that you know. Then for instance, you can use this and this library to draw stuff to actually plot some things. Then you can use this and that to actually produce your reports. So by simply just introducing those elements of the usual pipeline, we were able to enforce those house and wise at the same time. The other problem was and realized that we realized that we got it for free is that the usual question that we are getting from them, can you say that in plain English? So in other words, like, they were coming with some problem, they were like outlining the problem, like using their hand waving and just high level terms. We were giving them response, answer to that, and they were saying, oh, could you say that in plain English, actually, because I didn't get that. So then we thought, if we would teach them Python, Python is a language, we know it's programming language, but let's stick only to the language as a buzzword. Let's actually treat Python as a foreign language. So if you have a problem, draft your problem in the notebook and use the language that we really understand and then we'll deliver your solution much faster. And of course dealing with the big data, as I already mentioned, let's use Python again. And if you combine their knowledge, if you actually introduce the multitasking of the Python that we are all aware of, they are just starting to just discover totally different world. Okay. So naturally there were like a lot of people being like incredible non-believers, quite lazy, didn't want to actually engage themselves into something new, and they were keep challenging us with some problems. So for instance, like, yeah, you know, there's this notebook and they were using this special tone of voice while asking those questions. So hey, can I pull my data using notebook? Yes, use request or whatever. Like, is it possible to scrape websites? So of course they were always waiting for the snow saying, like, okay, our system have limitations. So, but they never ask actually the right question. So naturally, is it possible to scrape websites, naturally request, scrapey, whatever. Can I pivot my data? Pandas, can I create plots? Yeah, there is a whole bunch of solutions for that. Can I work on millions of rows, which is till now like totally not possible? So or you could actually use like, you know, the very nice solution of splitting your data into 10 Excel files and then trying to do big data like MapReduce. And yes, Pandas again, okay, but I really need to work with Excel because like, you need to have Excel otherwise you won't be able to survive in the jungle of data. So read docs of Pandas, you know. All right. What was the result? This is a work in progress. We are still struggling and fighting and trying to enforce our way of thinking. The result is far from utopian. But I think that lots of people actually learn something new. Lots of people started to automate their work. Lots of people started to actually gain and do something in the free time, this time that we free up by just introducing the solutions. And the main question was like, why did we do that? So we have like very huge open space and I'm actually sitting there and observing people and we are like making jokes, like the tech team, we are quite happy guys. And then I am observing those analysts and the photo that I showed you of this crazy Brazil fan. So basically this is kind of the faces that are doing when they are struggling with the big data or whatever. And you can clearly see that those people are not enjoying their day. So we thought like we need to give them something that will enable them to express themselves, some expressiveness. And the programming languages are actually the tools of the sort. If I will give you any kind of problem, you will start to actually putting the keywords and the grammar of Python all together and give me the solution. They are the slaves of the UIs, which are actually just enslaving them and giving them no sort of expressiveness. It's about actually the scenarios, the clickability scenarios. That's what I'm actually achieving. With the expressiveness comes creativity. So I was thinking like, I'm a software engineer, I don't know what kind of application you need. But if I will give you the expressiveness and free up some time of yours so that you can actually use your creativity with those new tools, you will come up with really awesome applications because you know what you need in your industry. And of course at the end of the day, the most important is freedom. So I think that those people, the ones that actually leverage this knowledge started to be much, much more productive and hopefully happier. So let's welcome Fabricio. That will actually take you through the biggest challenge ever taken in the UK, which is actually true training non-programmers to program in Python. So hello everyone. So basically when we submitted the abstract, we were asked to have a separate section about the educational aspects of this journey that we had. So these are my slides about it. You can find them here. It's open and public. And see a little bit about me. So a little bit. So don't have much time. So why this part of the presentation? The short answer is that they asked us. And the short answer is we hope it's interesting. Actually, some people teach Python, some people train other people in Python. Some people have colleagues that they go, how do you do this? How do you do that? And so I'm going to talk about a set of guidelines that hopefully will inspire you in your work, whatever you do. What we are going to talk about. What a trainee should do. What a trainer should do. And basically about my experience in Manila, being in Manila two weeks in March to train 20 people. It was a very nice experience. I asked to have four groups of five people each. As much differentiated as possible. Because I didn't want to have like one group of super smart people and one group of less smart people. I wanted them to be able to help each other. So wait, wait, wait. These are the guys. These are my Manila guys. Amazing people. So what a trainee should do. Two things. Listen with the capital L. Which means you have a trainer at your disposal. So squeeze him as much as you can. If you are ever going to take training in anything, get everything you can. Get the knowledge, the experience, what they think, why they think, the way they think. How do they get the solution to the problems? Try to get everything from them. Use that opportunity. And work hard, of course. What a trainer should do. I think it should achieve excellence in the following. First of all, smile and be patient. Because if you don't do that, you will put a wall between yourself and the people you are trying to teach something to. And they won't be open anymore. And you need them to be open. If you want anything to pass to them. Never take anything for granted. So you can go and teach non-programmers and start. So we are going to write a function and then a method and then we are going to use the class. And then we are going to iterate over this stuff. And they go like what? Because they absolutely have no idea what the function is, what the method is. So you have to give them the concept, the lingo, all the words that we as coders use every day. And we are not maybe even aware that this is a special language that we know and we use, but they don't know and they don't use. So the idea is why do we need to do what we need to do and the reasons. The reasons behind all the techniques that we use, like code reuse for example, why is important. Use the GPS technique. This is something that I always say. So you are trying to bring someone from point A to point B through a bunch of roads. And at some point you can see by their faces that you have lost them. Because there is some concept that they are missing, so they can't go from here to here. In this case, it's you. You have to find another way to explain to them. So ask what is not clear. Choose different words. Provide other examples. Do not repeat the same stuff in the same way because they are not going to go through that road. Set some goals. They should be reachable. Start gently so they are not scared off at the beginning and then gradually increase difficulty and skip the boring trivial stuff is possible. So they don't start wandering off cell phones and stuff. And deliver outstanding quality material. You have to put your best effort into it when you prepare the materials. The first reason you have to enjoy it as well needs to be a pleasant experience for you as well. Give them only what they need to know because we are all pressed in time, especially for me it was the London people. Manila people, they are very open, they are very hard working people. London people are a bit more difficult because they have, at least in our company, they have a lot of things to do. And so if they get the feeling that you are telling them something they don't need, they get, I'm not saying upset, but we don't need this. We don't have time for this. So just give me what I need. So just give them what they need. Focus on the needs, which means if at some point you get to a point where you have to explain something, again, do that instead of trying to stick to your original plan and refer to the real world, something easy to relate to. So this is just one example. When I was explaining how to handle files, it's all very, you can't just go and say, yes, you have a handler to a file, a pointer here and there. You have to open it, do something with the file and close it. They won't remember it. So if you give them the example and you say like a file is like a fridge, they all know how to use a fridge. So you open the fridge, you put something in and you close the fridge. And they already know how to handle a file because they have the association with something they've been using since they were little kids. This is very important because they will not forget that. And adapt. So adapt speed, difficulty, and so have extramarticles ready if you have a group of people that are actually fast and know what you can cut if you have a group of people that are less fast. And entertain them, communicate your passion, make them laugh and have a good time. Don't be too strict if they make a mistake. Use those mistakes to explain things again instead of saying, oh, you made a mistake. And the next one is only for the brave. So if you have to take a slide with you home today, just take this one, flush your ego down the toilet. This is because you're doing it for them, not for you. And if you do it for them, it's just going to get so much better. The tools that I used, well, Bunto, then a bash shell, iPython, console and notebook, libraries, pandas, numpy and the analytical interface that we brought. My paint with the cheap Wacom tablet that is really, really useful because if I have to explain what the function is, I just go crazy to someone who is not a coder. But if you can draw a box with an arrow that goes in, with an arrow that goes out, and you see a bunch of lines of code that takes some input and produce some output, they all go, oh, okay. And Skype or Googling out when you have to go remote. And this is an overview of what is it possible to deliver in about 12 hours. So we had an introduction session, four Python, iPython sessions, three data-oriented sessions and one QA session at the end. So the introduction session is giving them all the points, all the concepts. So they have to understand at least high-level what's looping, branching and so on and so forth. And it's really important because it provides context so they don't get lost when you then start to explain this stuff, the basics of Python. So code reuse and functions, the looping and branching, handling files, Python data structures, the main data structures, and the main built-in functions. Extra materials that I delivered. So advanced stuff with dictionaries, function arguments, list comprehension, slicing, and the broader introduction to built-ins. And the data-oriented sessions, basically when you work with data, you have to do three things. You fetch data, you work with it, you clean it, you mangle it, and then you provide some sort of output that can be text, statistics, or, you know, those pretty graphs that everybody likes. And you have to use, of course, JSON because stuff that comes, we work with Twitter and Facebook, so we've got a lot of JSON to work with, some regular expressions, some string manipulation, and of course the lovely daytime objects, especially with time zones. They are nice. So that's it. I hope this was interesting for you. Feel free to contact us. And before we close, I have a question, I have a favor to ask to all of you. There's a lady that I love to beat, and she's coming out of surgery in this very moment. So if we could have a round of applause to wish her good luck, I would really appreciate that. We have a few minutes for questions now. Please remember to use the microphones. Hi, thanks for the talk. So I have a similar problem in my current company, but what would they ask for us normally are just queries, basically, just how many of these, these months, how many of that. So maybe for us it would be even simpler. We could just somehow make it simple to make queries. Do you have any suggestions how we can do that in a way that they don't allow, are not allowed to make some crazy queries that kill the database or delete all the records? Yes, this is exactly what we've done with the D3 analytical interface. Basically we provide an interface to them, and this interface limits what they can do with the database. Basically shields the database, and it has two basic advantages. They can query the data regardless of where the data is stored or how the data is stored, and we don't have the problem of someone saying, oh, you know what, I deleted everything. So because they can't, there's no delete everything through that interface. How do you catch up with the questions which comes afterwards after you leave Manila, for example? Yeah, basically, after I left for Manila, some people were still asking, how do you do this, how do you do that? And of course they always make time. If my boss is happy, I do a remote session. If for any reason I'm very pressed, I do a bit of overtime, maybe an hour or so, and because I'm always happy to do this, I really, really love teaching Python or whatever. So it's always a pleasure to help someone, especially when they are eager to learn. I just can't say no. So yeah. And there is also one more thing, because please remember that as soon as you introduce notebooks, this is the communication actually, item that is used. So whenever there is a bug or a problem, they just share a notebook with you. It's not high level email sent to you like, hey, by the way, this is the problem. They just give you a notebook. You just run it and say, all right, this is a bug. And it's very simple to actually, you know, that's why we said that Python would be used as a foreign language for us, even though we are all foreign as using English. English is not enough to actually discuss their problems. We need to go into more formal one. Hi. Thanks for the talk. You nearly answered my question. Could you have done it without iPython notebooks? I think that we actually were at the, you know, it was a perfect time to introduce kind of solution, because I think that if we would introduce them like a console based, you know, solution, which was like the bare bones Python, even without iPython, you know, like it's showing, giving all the nice colors in the console, we will never actually be able to introduce or even push this idea in a third. iPython notebook gives this enough UI that you need. You can plot things there. And this is mainly what they do. They want to get data. They want to actually perform some stuff on the data. The other thing was that they all mostly work on the laptops. And the thing is that the amount of data that I'm dealing with are just too big for the laptops to handle. So what we did, we just bought this super powerful server and just hosted the notebooks there. So since they are working in different time zones, basically they could have like 32 gigabytes of RAM or 64 gigabytes of RAM just like that and just perform. And since they were not, you know, building the graphical representation of the Excel file, they could actually handle even more and more data with that. And then at the end, it was synced with their laptops. So when they were done, they could actually just save their results and even carry on their laptops with much smaller, already cleaned up data or plots. You said that you were using Ubuntu for the trainings. What did you do afterwards if some people were using Windows or some other operating systems? You said that you have a server with notebooks. So do all the people use just that server or do they have local environments? So they were using their own computers with Windows, unfortunately, but everybody uses Windows. And the place where we host this technology is on Debian, right? So the notebooks are on Debian and I use Ubuntu. So if you want to be a trainer, Ubuntu gives you all the tools that you need. You can do anything you want. And the beauty of the Ipyto notebook technology is that it's not really important what kind of system you're using because basically you do everything in the browser. So as long as you have Firefox or Chrome and you don't use the other one, can't even pronounce it. So yeah, if you are on Windows, though, there is a small problem when uploading stuff with Chrome. So if that happens, don't go crazy debugging. Just use Firefox and it will work. Yes, but how do you deal with people that want to install packages on their laptops or something like that? Especially if you do data analysis, NumPy and this stuff can get kind of hard and explaining them like package managers and everything could be hard. So do they just not do that and use the server or do they actually want to use Python locally? Oh, basically they didn't have to do anything on their laptops. Everything was provided on the server side. So where we host the notebooks, we have installed all the libraries. So the three analytical interface, NumPy, SciPy, MacLocklib, Pandas, everything they need is there installed for ready to use. So they just open their notebooks. We have an authentication system for each analyst and they go in and they do from blah import blah and it just works for them. So they have a slow window box or super nice Ubuntu machine. That doesn't really matter because the heavy lifting is all done by us in the server that Machu was talking about before. Are there any more questions? So I understand that the data analysts are quite happy. You were also talking about the human resources people and the accounting department which have quite different needs like payroll, which was it handled also? If you think about it, one project that James developed for Finance Team was basically they had this super annoying thing that they were doing manually which was just combining Excel files that were delivered by the sales guys and basically they had to build some reports on top of that. So it sounds trivial but for them it's just a manual work. There is no other alternative. It's just wasted time during which that person will go irritated and then just one of us in this case James just provided a solution for them. So basically we gave them the tools to talk with the data to actually get the data but at the same time we are listening for some new stuff that we can actually deliver on top of that. So we're just basically delivering them one magical black box function that they could actually call a create report and of course under the hood lots of stuff was happening. But since we were free up from delivering QI because nobody required that anymore we could actually just focus on the business value on the business algorithm that was actually delivering the solution. Please use the microphones. The use of Excel several such programs decrease after you are training. So basically it didn't decrease because the thing is that they were not able in some cases to use it at all. Just imagine they had this very nice adjustment mechanism. So they were opening an Excel file and if the person was smoking that person knew that it needs around 30 minutes to open that file. So that was a break time basically. So they just stopped opening those. So maybe less breaks maybe less bad actually. I don't know. But in Manila actually things changed a bit. They are starting to introduce I Python they introduced the I Python notebook thing step by step in the process. So they were doing a bit and then switching to Excel and do the rest and then another bit and then switching to Excel until it was all. I think the most important is that even if you train them even if you show them all the solutions you still need to wait for this discovery moment when they actually send you to see me like wow they actually realize that this is powerful stuff. Now they will I think alright some key words something is flying around but then they actually automated a pipeline you know and they actually realize that they can do something creative and actually contribute to something interesting rather than you know just just clicking through their life. Okay thank you very much. Much hey and Fabrizio.
|
Maciej/Fabrizio Romano - Python Driven Company Adopting Python across a company brings extra agility and productivity not provided by traditional mainstream tools like Excel. This is the story of programmers teaching non-programmers, from different departments, to embrace Python in their daily work. ----- By introducing ipython notebook, pandas and the other data analysis packages that make python even more accessible and attractive, we attempted to adapt python as a core technology across our whole company. We’ve challenged the dominant position of Microsoft Excel and similar tools, and dared to replace it by pandas-powered ipython notebooks. During this transitional phase, we have been inspired and sometimes forced to develop multiple packages that extend pandas, numpy etc., in order to enable our colleagues, in other departments, to access all the data they need. Moreover, we are developing several high level functionalities for the notebook environment. The notebook environment is allowing us to be extremely responsive to the changes our users are asking for, since, for part of the work, we don’t have to go through the whole traditional development process. The talk focuses on challenges and problems we’ve solved and managed in order to achieve our long term goal of creating highly agile, data-driven non-tech teams, free from the constraints imposed by mainstream technologies, and all of this thanks to python.
|
10.5446/19993 (DOI)
|
I work for Red Hat and I work in QE department. I'm not a software developer. I'm a QE person. And I will talk how we test our system. And it's not about every Red Hat program. It's only about my team. So it's only about Red Hat Enterprise Virtualization Manager. So this is downstream version of Oviad. So this is tool which helps you to manage your virtualization. So on which level we are. You mentioned that you are a developer. You develop your application. It tests all your unit tests. It tests all your functional tests. It was already built. So you have RPM or, in our cases, RPM. Maybe you have another packages. You want to distribute to user. But before you do that, you want to test if everything is OK. So you want to test how it will work on real system, in real cases, and so on. So you give it to us. And what we do? We develop your application on real environment. We try to do it on as production system as possible. So for example, this is system for virtualization. We will run what we can on bare metal machines, not on virtual machines. We will use official APIs. We will not take a look on your code at all or almost at all. We will simulate user. More or less, we will do what your user will do. And we will test both negative and positive cases. But we will do much more positive. It's like, OK, there's a problem if something is broken. But it's much bigger problem if you cannot do some case. So what are problems in such situation? First of all, we touch environment. Actually, we break this environment. And we do it on purpose. We want to test how the system will behave if we break network connection. If we break disk, can it recover? So the problem is you have to maintain this environment. And it's hard. What else? It takes time. You have to set up this environment. It takes time. It takes, it's not a fraction of seconds. It's minutes. Sometimes it's half an hour. Then you have to run your test. It is quite short, usually. And then you have to clean this environment. And because this test may fail at any moment, you have to have this cleanup quite well. You don't want to reinstall everything starting from the operating system every time. OK. So as of course, we want to automate it. We may do it manually. But imagine that you will have to do it every week, the same, install everything, click on it, click on everything. So why we want to use Python? First of all, it's easy. Quite a lot of people in QE departments are not software engineering people. They have some background in IT. Quite often they are from admin departments. Only language they know is BASH. And you have to somehow force them to write code. You don't want to teach them Java or C++ or anything like that. Python is easy. They can understand it. On the other hand, it's normal, mature language. You have libraries for almost everything. And it's popular. So if you need people who really know this language, for example, you want people who will write tests, and you want people who will write libraries and so on. And for these people, you want someone with coding experience. And finding someone who knows Python is not that hard. So one such test does, it must get resources. You don't want to run your test on the same machine on which your system will be developed. Because quite often you will want to, for example, reboot this machine and test what will happen. So you have to get a machine on which you will run the test code. You have to get test definition. You usually keep it in a repository like Git. And you need resources. The resources like BEMeter hosts or virtual machines or some kind of storages. Then you have to install product packages. You have to take them from somewhere. So you have to specify from which repository you want to take. Then you have to configure this product. And your test may want to change this configuration. Then there's what's specific for the test. This is a setup because, for example, when you simulate a user and test if you can add virtual machine on your, to your installation, you have to do everything which is required for this step. Then you have to run this test. Then you need to perform a theorem and this part quite often is more complicated than the test itself. Because these two parts, setup and training test, they may fail at any moment. And it may leave your environment in a really crazy state. And you want to clean it up. Then of course you have to collect logs. You cannot, okay, you don't run this test manually. You run them from Jenkins overnight and when you come to work in the morning, you have results of this execution. But you cannot assume that you will have access to this machine where you run this test. You have to copy every log which may be important and you have to copy it. You have to run clean versus sources because you will run next test. And you have to report results. You keep your tests in some kind of test management system. So you have to take execution of every test and see that it passed on this release, failed on this release, failed because of something. Okay. So how we do it? We use tools which we found whenever we can. So for example, we use Jenkins, we use Forman, we use Git. The problem is that we didn't find anything which will help us for managing the resources and for leasing them to this test, then checking in which state they are and leasing to another one. And we have our library, we call it art and this is library for running this test. So okay. So what this library can do? First of all, we use official APIs to use all of them. So for example, in our case, you have Python SDK, Java SDK, CLI, GUI and the API. And you don't want to run the same test for every API. You want to run it once and do your library to run it for every API. What else? We don't run a test collecting functionality by ourselves. We use knowside. So everything which you have in knowside, tagging, you can use here. And when you take a look on the scope, one test is just a unit test. And usually it's not that long because everything what is in this test is only when we go back here, it's only this set up, run test teardown. Sometimes even without set up and teardown when it's common for a lot of tests in one suit. Okay. We have a lot of helpers. We have helpers for pure open Pyramid car. Sometimes it's too hard for people when you want to run something quickly. It's a little bit complicated. We have tools for managing remote files, directories and things like that. You don't want your user to run I am minus I f or anything. Some slight mistake, they will break something really, really bad. And of course helpers which we use in our tests, you don't want to run IP tables every time you want to simulate network connection failure. So we have just the helpers for it. Same for disk simulation, same for load simulation. And multi-fading. Multi-fading is hard. Multi-fading in tests usually means that you want to do something. And in other things you just want to wait for some event to arise. And that's it. That full executor is almost easy enough. Sometimes it's not easy enough. Okay. So yeah. This was the library which we use for running the test. So when we go back, yeah. So we have this part from setup environment to T-down. But still you have to install product packages, install configure the product, collect logs, release resources and so on. So getting machine, getting test definition, we use Jenkins, Jenkins can do it for you. We use also Jenkins for collecting logs. We use our BIOCAs for getting resources and then for cleaning them. I will go to this point later. But still we have this point, install package, install configure. So we have something which we call job planner. And this is actually what is run by Jenkins. And we have something called jobs for every stage of the test. So this is installation, this is running the test. This is for upgrade because sometimes you want to test what will happen with your system during upgrade. And each of such jobs defines its cleanup. So for example, if you have job for installation, it defines that cleanup is removed of the packages. Okay. And it has plugins for getting resources because sometimes you need resources on this stage just for installation. Sometimes you need it within the test. Sometimes you need it also here. So what we do for getting resources, we call it resource brokers and we actually have two types of resources. Some of them you can just create on demand to some level like VMs or NFS shares or ice casi shares. Some of them you just have a pool, nothing else you can do. For example, bare metal hosts. And of course, when you have bare metal hosts, all of them are not the same. So your test has to specify that it needs to host one of them with such setup, another one with another. Okay. And of course, when you have resources which you can create on demand, when they return to the pool, you just disturb them. You don't try to clean it. But when you have something you cannot create on demand like bare metal hosts, you want to check in which state they returned. You try to clean it. If you cannot do it, you just reinstall everything. This is why we use Formant. You just mark this host built in Formant and reboot it. So if we cannot fix it, we will not spend time on it. Okay. Okay. So we have integration with external tools like bugchackers. If you know that there is a bug in which is found by this test, sometimes you don't want to run this test. Because of this bug, environment may be in such state that you cannot do anything. It may be so broken that it doesn't make sense to run another test. Okay. Version control system and so on. So my main thing is do you know such libraries, such tools? As you have seen, we use a lot of tools written by ourselves. And at my previous job, we were in the same state. We also use what was publicly available. But then we still didn't have anything for getting resources. And we also use the same tools. And I don't believe there are no tools for it. Because if it happens twice in two totally independent companies, it must be a common problem. So if you can, if you know anything which can help, I will be very thankful for any suggestions. So thank you. Thank you. Thank you, Katrina. So we've got plenty of time for questions. Is anyone got a question they'd like to ask? Hands up. Okay. I'll bring the microphone. Hey. Do you know of any solutions if you have a bare metal server and you maybe need to push a button or, well, anything that wouldn't require a robot like taking out a disk or something like that? No. Okay. We don't do things like that. Okay. Thanks. Thank you. Any more questions? One at the back. Maybe guys can follow the... I think we can hear you. Yeah. Yeah. So you asked the question and Katrina, you repeat the question for the microphone. Is there a product for those time, for managing those kind of tests? Just like in the situation between Jenkins, Format, and you have customized to reduce the workflow that you have to require. Okay. So the question is if there is package which enables to reproduce this flow and gives you integration with Jenkins and Format, so I don't know about anything like that. If I knew, I would use it and not write again something. If you know something, please share it. Okay. I hope one day we will open source it because it's not true that everything which is in Red Hat is open source. Some of our tools is not open source. I hope it will be open source but one thing is politics. Another thing is code state. Now I would have not show this code publicly, at least parts of it. Any more questions? Are you doing all this stuff remotely? So you are using Paramiko to execute stuff on the remote VMs or servers or Jenkins runs the unit test and uses your framework and the SSH to the server or how does it work? So in Jenkins you get a host on which you run this code and then when you want to run something on remote machines because one test usually uses at least two machines, then you do it via SSH. And this is painful. This is for example, Paramiko is not very good. It's not also very easy to use. Is there something better? No. Unfortunately, we use Paramiko. We just have our wrappers for it. So it's like you specify what you want to run on this host with such credentials, this command and you just get output. And we can go to the cost. Anything else? Any other questions from the audience? One thing I was going to say was that I know that where I'm working we've set up, we've got code to set up to run IP tables and then tear them down again after the test is finished. I was going to ask you if you'd open sourced any of those helpers that you listed on your slide, but the answer is no. The other thing is that we've recently open sourced something to allow you to set up network namespaces in Linux. We're running IP tables in namespaces, which might be useful to you. That's called nomenclature. Okay, thank you. Okay, any other questions? Okay, well thank you again, Katrina. I'm sorry for mispronouncing my name. I'm wondering whatever you did for this25 day where a comment like that, but I didn't get that right forравствуйтеS. Kitties.
|
Katarzyna Jachim - Python in system testing When you think about Python+testing, you usually think about testing your code - unittests, mostly. But it is not the only case! When you have a big system, you need to test it on much higher level - if only to check if all the components are wired in the right way. You may do it manually, but it is tedious and time-consuming - so you want to automate it. And here comes Python - the language of choice in many QA departments. ----- When you think about Python+testing, you usually think about testing your code - unittests, mostly. But it is not the only case! When you have a big system, you need to test it on much higher level - if only to check if all the components are wired in the right way. You may do it manually, but it is tedious and time-consuming - so you want to automate it. And here comes Python - the language of choice in many QA departments. I will tell about differences between unittesting and system testing which result in totally different requirements on test management/running systems. I will tell how we use Python (and a little why) to automate our work. Finally, I will tell a little about my "idee fixe" - a framework for system testing written in Python.
|
10.5446/19992 (DOI)
|
Okay, no, Euro Python or any Python conference is complete without at least one talk on packaging and this must be at least a second. So let's see what your key has to tell us. Thank you. If you don't mind, I'm going to wait for a tiny bit for clock to hit three, four and then just depressed. No one comes. Before you can run away, I'm going to take a photo of you. This is a great in a zero chance that it happens when I start speaking. Okay. Good to talk, Euro Python or Helsinki will come in. That's my outcome of two years of German studies while in elementary school. Unfortunately I can't do better. But welcome everyone to a talk about a tool called DH virtual end or as I label it packaging in packaging. But before we look into that deeper, let me introduce myself. So my name is Jurke Poulienen. I'm from Finland but I'm living in Stockholm, Sweden nowadays. I work for a music streaming company called Spotify. In there I do kind of do two stuff. I am a content engineer. So I build like a pipeline of new music to the service and on the other hand I also fiddle around a lot with our internal Python stack and answer people's questions about Python. If you want to reach out to me, there's my email address and there's my Twitter handle. So please do if you have any questions. Now about this talk. This talk will be in three different sections. First we're going to look into some of the existing deployment strategies you can use on the PyNU-based machines. Then we're going to look into what actually is DH virtual environment and how does it differ from the existing deployment strategies. And in the end we're going to go through an example of how you package software and we're going to package Sentry mostly because it's not the simple piece of software. So I'm going to show you by example how you can use DH virtual environment to package something like Sentry in production. Now let's start with that tiny quiz. Who here runs a Debian or Ubuntu-based system? Good. You're from the correct token. Who here deploy stuff using the so-called native packages on Debian system, like relying on the native libraries on those systems? Who here is kind of frustrated with that? Yeah. Okay. Who here uses virtual environment? Like you set up a virtual environment and then you install everything in there on your production host. Cool. Now these both have their good sides too. If we take for example the native Debian packages. So this table, the stuff that gets in the Debian, especially stuff that's in main, is stable, it's well tested. People usually don't change them so that the backwards compatibility breaks. So if your Debian system gets an update for let's say Python requests, you know that that request is something that's backwards and compatible with the previous version. Now Debian has enough or Debian packaging has another nice bone side. You can declare non-Python dependencies. So say your software requires SQLite to be installed or your software requires MySQL to be installed on the same machine. When you create the Debian packages, you can say that your package depends on MySQL and you get that one installed on the machine. Everything nicely contained in one package. It also has pretty neat existing infrastructure. So not only you have like dedicated build tools or separated build environments like sbuild and other CH root solutions, but you also have the possibility of running your own app through repository which means that you will have your own way, or like own network of deploying your stuff and production. Plus all kinds of CI tools like Jenkins or Team City have at least some rudimentary support for dealing with Debian packages. And the last good thing I think in the Debian packaging is that you have a quite nice scripting support. What that means is that if you need to remove a cache when you upgrade your package version, you can write the script that before you upgrade your package, it will clear out the cache. Or if you want to restart your service after you've installed it, you can write the script that after installation, it will restart your service. You can do crazy borderline stupid stuff like database migrations and postings. I don't necessarily recommend them, but you can do all kinds of stuff there. It gives you quite a lot of power. Now that was the good parts. Then if we start talking about the bad parts, you probably have run into this case where you see that, oh, let's say back in the days, can network hard and push out request 1.3, and that has just the feature you need, but unfortunately your ancient web-filled box is running request 1.2. And what do you think? Are you going to wait for Debian to package that? What happens, you're just going to rot in front of your computer waiting for the newer request, especially if it's a backward incompatible to come on the current system you're running. So a lot of stuff in Debian, even at the release of that particular Debian or Ubuntu release, is already outdated. The packaging itself is kind of complex. So what we're talking here is that someone created a packaging system that is built for building a whole operating system. So it covers all the possible corner cases, all the possible scenarios, if you want to deploy Perl, Haskell, Python, you name it, it has everything covered. That means that the whole system is really, really complex and the documentation is far from simple. In addition to that, all the documentation that you usually can find is also geared towards Debian package maintenance. So okay, you want to ship this thing within the operating system. So this is how you should do. And it does not necessarily resonate to how you yourself would deploy your tiny service in the host. And what I think is the worst part in Debian packages is that you get a global state. So if you're only deploying all your libraries as Debian packages and then using them, eventually you'll end up having a case where you would like to upgrade one library, but it's also used by some other software on the same box and you don't know if you can do it without breaking anything. We've had this at Spotify a few times when we have rolled out like new common Python utils libraries on the host and we're just too afraid of deploying that because we might or kind of afraid that something down the line might break even we have been testing it for months. So it kind of slows you down and it's really, really annoying. Now if you think about virtual environments then, so what you get in the virtual environment is somewhat the opposite. So you get the whole new stuff. You just do pip install and you get whatever is available in PyPI, like the latest release. And you can go to the extent of like you can open your Git or Mercurial and pull the stuff in your virtual environment and run it in there. So you can always get the new, hottest stuff. It has become kind of a de facto method in the Python world. So every guide usually contains a word or two how you run stuff inside the virtual environment. You can do the same virtual environment stuff on your laptop as you can do on your servers. It kind of works. It's also battle tested. So nowadays so many people are running it in production so it's fairly safe to use that one. But the best part I think is that it's contained. So if you take a package and install it inside the virtual environment, a Python package, you can be sure that it won't affect anything outside the virtual environment. So updating a simple packages doesn't mean that your whole system crumbles because something was relying on all this older version or that you would interfere with the underlying operating system. No, you're only poking the actual virtual environment and you get this nice contained fuzzy feeling when dealing with that. Now on the downside, I bet some of you might have seen this line. That means that you can't have any native dependencies if you're using virtual environments and people install. You need to know what MySQL libraries you have to have available. Like in this case, you need to know that you have to install MySQL client to find this MySQL config on your host. So it requires you to do some manual digging through things. Even more, you end up doing source installs. So you can have wheels or even eggs to avoid source installs. But if you don't have wheels for your platform or you haven't set up your own wheel riposter, you end up doing source installations. When you have your virtual environment in your production server, you end up installing all the dependencies for that source installation on that production environment, which is not necessarily bad in the sense that you could break something, but you'll just clutter your production environment with developer headers and other unused files. But the worst part, what I think is with the PIP installs is that you're basically executing a bunch of random scripts. So sure, wheel gets around this, but if you run set up the PIP install, you probably haven't looked into what all those files do or what all those files that those packages depend on do. So you're just relying on a good faith of people and you're running random stuff in your production environment. It doesn't even need to be malicious to hurt you. Someone might just accidentally release a package that wipes your whole ETC or your whole home directory or something and you by accident end up crippling your system. Now this brings us to question then what is DH VirtualN. So DH VirtualN was about two years ago my attempt to combine the best of the two worlds. So it is a virtual environment that is placed inside the PN package. It supports both Python 2 and 3. It is kind of version agnostic. I won't say that you can't install Python 1 stuff with that, but it doesn't execute or import any Python code. So it doesn't really care if your tool is written with Python 2 or Python 3. You can use it anyway. It even supports using the new virtual environment package with 3.3 so you don't even need to install virtual N to run it. It's also open source. So it's GPL like all the different build tools. It has a good documentation. Now I'm the guy who wrote it. So I might be a bit biased here. But I think the documentation is good. It's at least better than the average open source documentation. But the best part which seems to be actually very functional is that it has a simple tutorial. So if you go to the DH VirtualN.readthedocs.org you find there's a four step tutorial that you can run through and boom, your package is inside the virtual environment inside the Debian package. Under the surface it is a depth helper extension. So depth helper is this... How would you make it pretty? Like a pile of purl scripts that Debian executes when you're building packages. It's a certain fixed sequence of purl scripts that you install and then there are different extensions for Debian to know how to build packages for Python packages or how you build purl packages or how you build bash completion stuff. So what the virtual environment does, it just injects itself into that flow. So there's 12 lines per included there that does the magic to inject the virtual environment there and then it just runs as a part of the sequence. Now this is kind of like the implementation details but for you who already have existing Debian build environments, like if you're using sbuild or something else or just playing the build or the package build package on the command line, this means that using the virtual N is just going to work in your existing workflow. So what I basically did back in the days was that I found a great blog post by Hinek Slavak. I adapted the idea a bit to fit our built environment and that's why I ended up writing depth helper extensions. Hinek system works well too but it uses a thing called FPM which back in the days didn't fit our build system. Now in practice, the DH virtual environment is a packaging builder that it creates a virtual environment, you can define what Python you want to use with it. So if you have multiple pythons installed in your machine, you just pick the one you want. It installs everything you have listed in requirements of TXC and this is the exact same format you get with pip freeze. It installs those inside the virtual environment, then it takes your project and runs set up the py install on that. So it just doesn't dump your sources in there, actually installs your project inside the virtual environment and then it does a bunch of magic which is like set scripts where big thanks actually goes to Hinek about those and other stuff like rewriting, activate, so that you can actually run like instead of having all your build system paths, it will contain your production system paths and you can use things like activate or activate this in the production and end up with the same virtual environment. So that's nice and cool. But let's see. Let's take a project. Let's package something with DHR term. Let's package sentry. So who here knows what sentry is? Cool. It's a really good exception tracking tool. We use it in our production systems. It works like a charm. The best part for the example part is that when you install sentry, if you have ever done pip install sentry, it pulls down half the Python package index. It pulls like what ever you can find. It depends on a lot of stuff. It's not because it's bad software. It's because it's a complex software in a good way. That sounds a bit bad. Well, anyway, let's see how we do the DHR term. So first step is that we need to install DHR term. If you're running a modern operating system like Ubuntu trustee or the Debian testing, the DHR term is actually available inside those repos. You can just say apt-get install DHR term. And all of a sudden you have the DHR term available in your system. It's as previously discussed, a bit old on Debian and Ubuntu, but it still works. Then you need to create a Debian directory inside your sentry installation. So Debian directory is a custom directory Debian uses to figure out what stuff should it, like from that it figures out what package should it build, what stuff should it run on the build time. In that directory you have to create a few files. There's a minimum set of four files that you need to create there. And don't be afraid. All of these are covered in a tutorial also. So when you start packaging your sentry, you add a control file. This is the place which Debian uses to figure out what does it need for building. So you can see that sentry requires the Python development headers for building. But for running it, it doesn't require anything special. So in this place you have to just build dependent DHR to run and then write the required dependencies like Python and stuff like that in there. But this is basically how I did this today when I built the sentry with the DHR challenge that I just copied over the tutorial stuff and changed the fields that I felt need to be changed, like package names. Then you need a change log. The change log is a, well, it's a change log, but it's the file that is required for the Debian package to figure out some version. And it just tells Debian that we are packaging sentries 6.4.4. Cool. And the third one. This is why the packaging is the complex black magic thingy. The third one you need to define what is called the compatibility level so that Debian knows how we should build the package. What's relevant to you guys is probably just the echo 9 in that combat file you're done with that. If you don't do that, it will pick up some engine compatibility level and won't build your package. So that's pretty much it. And the last part is the glorified make file, aka rules file, which just tells how you build stuff. Now, if you build Debian packages before for Python, you probably recognize this file, and you can just see that we changed the Python 2 to be Python virtual end. This basically tells Debian that built this package using DH virtual end instead of the default way of building Python stuff. And that's it. If I are D package, build package, and it rolls through, you get the nice matrix-like output of stuff building, and all of a sudden you see, hey, look at that. It's actually great to virtual environment, puts pip, whatnot in there, and starts pulling half the internet down into your package. And once that's done, all that's left for you is just take the Debian package, copy it on your production host, and install it there. Now if you have defined some build dependencies in the control file, they get installed at the same time. And the best part is that you haven't executed any random scripts on your production system because all of them were done in your build system. You end up with deploying the whole thing without cluttering any production system with any development here, stuff like that. And you have a nice contained virtual environment in your production host. Okay. So once you've done that, let's look at the kind of nice parts of the DH virtual end then. So what it gives you is that it gives you the non-Python dependencies or the possibility to define non-Python dependencies just the way you could do with normal Debian packages. It also leverages on the existing infrastructure of Debian building. So you can use your existing build agents, you can use your existing CI systems if you're already using them for Debian, or you can have your own app triple still in use. It has the new hotness. So it's not limited by what you can find on Debian. You can actually just install what's the newest stuff available. And as I told before, it's contained. So we end up with the virtual environment in certain place in your production system and that's it. Of course, like any solution, there's also some negative sides to this. The build times can be slow. So especially if you're not running your own PyPI mirror and you're not using wheels, you're basically downloading all the requirements from the internet and then building them, which means that, yes, your build time will become longer. But it can be also substantially mitigated by running your own mirror which cuts down the network latency and using wheels which cuts down the build times. It still requires you to dig some requirements. So you need to know what requirements you need, what native system requirements you need to have on those systems. It doesn't let you out of that loophole. So if you're running, let's say, you're parsing XML using LXML, you need to make sure that your control file depends on installing LIP XML on your production system as well as having the development headers for the build part. And the build system needs to have the exact same Python. So like, because it's virtual environment, what virtual end does, it actually links outside the virtual environment for, well, link or link stuff on the production system. So you need to have the same Python available on your build system as you have on your production system. But it rarely is a problem if you are already having an existing build that's pitching stuff. For the future of the virtual end, I'm at least trying to add some of this stuff. Like I'm looking into cookie cutter templates. Unfortunately, I didn't have time to prepare them before this talk. But it would be sweet if you could just use cookie cutter and boom, you would have your DH virtual end packaging done without you, you don't need to go and echo nine on random files. I'm planning to add a trigger support, like if you get a minor update of your Python on your production host, it refreshes the virtual environment to make sure it still runs. And I'm also, this was actually a tip from, damn, I'm so bad with names, from Adam, from this, because when he was rehearsing his talk and we talked about this and he gave me a tip, like what if I could actually break out the dependency on system Python, like use PyEnd or something to incorporate the whole Python into the virtual environment. But that's pretty much it. If you want to find out more, the source is available, open source, under the Spotify umbrella in GitHub, and there is the good, emphasis on the word good, documentation, on the read the docs, and then there's a blog post, we posted when we released this. So with that, I thank you for your time. Thank you, Joachim, you've just saved me a lot of time. Any questions to the microphones, please? Thanks very much for that. The other time I was looking at your presentation, not sure, but your colleague from Spotify about using Docker for the deployment and managing the infrastructure, I myself am right now struggling between Docker for dev and production, and right now in my company we are building the Debian packages, actually, so this would be really, really cool to use. Do you actually do that you install some software based on this Debian packages and some other deployments based on Docker, or do you mix it? The current plan with Python stuff is to mix it. While Docker is great, it still doesn't solve the problem of defining dependencies, so it basically boils down to two options. Either you write a Docker recipe that says, like, pip install this, pip install that, apt-get install this, which kind of does the trick, but then you have the same problem that you would probably need two different Docker images, one for building your package somehow then extracting that out and putting it into the other Docker image. So the benefit of the Docker is that you don't necessarily need the virtual environment because the Docker already provides the isolation. But we are leaning more and more towards to build Python software with this and then use the Debian package to dump that one into the Docker images. And my other question is, what do you use for your local APT repositories? I have no idea, but it's something fairly off the shelf. I looked at it, but it's like, it's ancient installation. Hi. Oh. Yeah, I've got a question. Does that support any great mechanism as well? Can you repeat the question? Yeah. Does your system have a kind of great mechanism? If you want to, I know, like, let's take the example of Sentry and let's say you database needs some migration and you want your package to be able to provide some, I don't know, whatever type of a great, your packaging need to have some changes from one version to another. So, because it's a Debian package, so it will have all the post installation scripts and stuff like that that you can run. So if you want to run scripts before removal, after removal, before installation and so forth, you can use the existing Debian infrastructure for that one. So it's not too complicated to do those, but of course, we require some knowledge of the Debian packaging at that point. Okay. One question related to that. Do you use these post installation, pre-installation files for database migrations? So you said earlier you use something different, but what do you use there? We do have, we actually do have some project that used the post-instit files for database migration, but it's kind of, I feel it's kind of scary. Like if something accidentally triggers a package update and we get like an unwanted migration, even if it would be tested safe. So usually we do that in database migrations with some sort of manual steps, depending on the project. But yeah, you can do post-instit data. That project has been working. Fingers crossed with the future. Can you use system Python packages or everything is installed in the virtual world? For example, Alex and Pilo. Currently by design, everything is installed in the virtual environment. But it shouldn't be too hard to add a feature to the virtual world where you can shoot in your both legs at the same time to allow the system packages to be. Because yeah, I see your point because Alex or Pilo are kind of annoying. Alex is fairly easy, but Pilo is really annoying to install inside virtual environments. So yeah, it's probably something for future to add a flag that if you want to use the site packages or the existing ones, then why not? But currently no. Maybe you can remove from requirements TXT and just add to the pens. Yeah. But then if you add it to the pens, it gets installed on the system level. But the virtual environment is filled with no site packages flag. So it won't see them. One intermediate step you could take is that you could depend on Pilo, let that install on the system level, and then you could use Python path when you start your software to point that in addition to the virtual environment's Python paths. So that could work, but I wouldn't go for, I wouldn't say sure if it's viable solution or not. But nevertheless, like if you want that feature, please open a ticket in a GitHub. It should be fairly simple to implement, so I can just build that. Thank you. Or if you want to make a pull request, that's even better. Sorry. Did you hear about FPM, so-called FV package management tool? Yeah. So this is a Hinex blog post, which is actually doing this exact same stuff, but using FPM instead of injecting the depth helper sequence. I wanted to use that originally. It works. It gets the job done, but it didn't fit our build systems. So then I ended up building a depth helper sequence instead. Okay. But if someone starts from scratch and doesn't have an legacy that he wanted to use, then how, why he could use your product instead of FPM? What's the advantages for him? I don't know if there's any specific reason for it. In that case, I would say go read Hinex blog post, check out the excellent documentation of my project, and decide on which one seems to be simpler. So I've aimed to cater people who don't really know the PMPackaging at all, so that it should be easy. But it's the same with Hinex blog post. You just follow steps and do stuff. So it's rather a matter of personal preference rather than features. Yeah, pretty much. Okay. Thanks. Just a quick one. Is there any plan on your side to port it to Weezy, for example, using back ports that even the log? Yeah, that could be done. I'm already planning to, like, because trust is having 0.6, so I'm planning at least to set up a PPA for trusty so that you can, the newer releases on trusty. This stuff builds fine on Weezy, so I've built it on Weezy, but Weezy was already stable at that point when I released this. So it shouldn't be too hard to add it to the back ports for Weezy. Okay, so there are no technical reasons not to do that? No. Should work? No, this is really simple. And, like, we build stuff with DH version on top of squeeze. So if it doesn't work on Weezy, then we've done something really bad wrong in that case. Okay, thanks. Thank you. I think that's all the questions. Thank you for a very clear, very useful talk, Joachim. Thank you. Thank you.
|
Jyrki Pulliainen - Packaging in packaging: dh-virtualenv Deploying your software can become a tricky task, regardless of the language. In the spirit of the Python conferences, every conference needs at least one packaging talk. This talk is about dh-virtualenv. It's a Python packaging tool aimed for Debian-based systems and for deployment flows that already take advantage of Debian packaging with Python virtualenvs ----- [Dh-virtualenv] is an open source tool developed at Spotify. We use it to ease deploying our Python software to production. We built dh-virtualenv as a tool that fits our existing continuous integration flow with a dedicated sbuild server. As we were already packaging software in Debian packages, the aim of dh-virtualenv was to make transition to virtualenv based installations as smooth as possible. This talk covers how you can use dh-virtualenv to help you deploy your software to production, where you are already running a Debian-based system, such as Ubuntu, and what are the advantages and disadvantages of the approach over other existing and popular techniques. We will discuss the deploying as a problem in general, look into building a dh-vritualenv-backed package, and in the end, look into how dh-virtualenv was actually made. Goal is that after this presentation you know how to make your Debian/Ubuntu deployments easier! [dh-virtualenv] if fully open sourced, production tested software, licensed under GPLv2+ and available in Debian testing and unstable. More information of it is also available in our [blogpost]. Talk outline: 1. Introduction & overview (3min) * Who am I? * Why am I fiddling with Python packaging? * What do you get out of this talk? 2. Different shortcomings of Python deployments (5min) * Native system packages * Virtualenv based installations * Containers, virtual machine images 3. dh-virtualenv (10 min) * What is dh-virtualenv? * Thought behind dh-virtualenv * Advantages over others * Requirements for your deployment flow * Short intro to packaging Sentry with dh-virtualenv 4. How is it built? (10 min) * Debian package building flow primer * How dh-virtualenv fits that flow * What does it do build time and why?
|
10.5446/19989 (DOI)
|
Hi, everybody. Thanks for having me here. I hope everybody can hear okay. So testing design. This is me on the Internet. You can find me on GitHub. My email's on the last slide. You can come talk to me afterwards. I work for Magnetic. We do online bidding on real-time bidding on online advertising. We use Python, PyPy, in production, lots of fun stuff. You can talk to me about that also. Okay. So luckily, this talk is fairly simple. So simple that the core ideas fit on a slide and a half, which is basically that we know a lot of things or we think we know a lot of things about what makes designing software good. And the good thing about testing is that most of those things translate quite easily over into what makes writing test suites good. And so tests are just code like anything else. So we have all these principles that we think help us out when we write software. And they're obviously something that you think about as you're writing software and they translate pretty well over the tests. So we have these principles like make sure your objects only do one thing or make sure that you try separating things so that things are fairly simple and composing objects together and those sort of things. And most of those principles translate pretty directly into testing. So try to keep unit tests, integration tests of any sort down to testing one specific thing. And to make sure that your tests are both simple and also transparent because you're not testing them. So all of these principles that we have for regular software design apply pretty well to testing. Getting down slightly to specifics, we have this three-step process that gets drilled typically into our heads when writing tests, which is that there's three steps to a test. First, you set up some stuff, then you do some exercise, you do whatever it is that you think you want to test in that test. Then you make some verification that what you expected to happen actually is what ended up happening. Our supply is fairly uniformly across all the different types of tests that you can end up writing. As a three-step process that if you actually think through it as you're writing tests, your tests end up clearer, they end up being more self-documenting, all the sorts of nice things that we like out of our test suites. One particular thing that people sometimes say when we have this three-step process is make sure that your verification is only one assertion. So you write a test, make sure it has only one assertion. It's kind of a peculiar thing to hear the first time that you hear it. First of all, most of the time people's first thoughts are how do I actually make that happen? Because they remember back to times when they've written tests and they had this long list of assertions. So you make it happen. But even more than that is like when you hear this statement, when you hear someone tell you this, your first thought is like what's the actual benefit? What am I going for by keeping my tests down to a single assertion? You stare at this example here. It was a fairly simple function. It just takes a bunch of dictionaries and adds them together, adds all the keys and associated values together. Then you have this alternative. It's green. It's better. So what's better about like what is the actual difference between the two examples here? Why is this any better? And of course, this is a simplified example, but it's the first representation of this idea that like make sure to keep your tests down to one assertion. And the most obvious benefit just to answer that straight off is that the most important things about tests are their failures because tests are destined to fail but meant to pass. So first you got to see the failure. And the difference between those two slides is basically how much context you get when that test fails. So the main benefit that we're aiming for with this sort of idea is that we want more context when our tests fail. So rather than seeing stuff like well, this isn't this, which is what you get when you have assertions that look like that, if you make these sort of larger assertions, then you get extra context. And that's useful for a lot of reasons. One of which is that while the first one tells you that what you got is not what you expected, the second one tells you not only what you got isn't what you expected, but possibly the extra information that you got is telling you in what way the actual implementation in your code differs from what you expected it to be. So for example, you're swapping values for some reason for keys. Obviously, in this example, that's pretty unlikely from an implementation point of view. But in real world examples, it's quite common for things like that to happen. And if all you see is just well, this isn't this, it gives you sort of less information, less ability to just be able to look at it and say, oh, well, that doesn't look right in multiple places. And the combination of the places where it doesn't look right means that what I did wrong is something. And in particular, this applies to unit test two, which has this type of quality protocol where you can get all sorts of nice context like that. So moving on a bit, so now we sort of like the idea of having one assertion in a test for the reason of getting this extra context. But sometimes it in fact turns out to not be possible. Oftentimes now we're shifting over from unit test to integration test. And sometimes what happens is that there's sort of these two sort of worlds of assertions that you want to make. Sometimes you want to make assert equal or some assertion of that sort, which are basically data comparison assertions. They're like, I have some stuff, some values, and some expected values, and I want to make sure that those two values match up with each other. But a lot of times when you're writing applications, what you actually want to assert at some point when you're writing your test suite is more like, I want to assert that some collection of things are true, that the state of my object application something is true. And unit test won't have assertions for those because unit test doesn't know about those, they're part of your application. It's about basically the difference between making assertions about some data and making assertions about the meaning of some application-specific thing. So to take another specific example, you want to compare some strings. Cool. Unit test can help you. You just make an equality assertion. But if in your application those strings are actually HTML, unit test probably isn't going to help you because while the standard library has a bunch of parsers for things, it doesn't have assertions for that. So if you want to make assertions about some two pieces of HTML being equal as HTML, not necessarily as string literals, you're sort of out of luck. And that's kind of unfortunate because that's kind of useful to do. I'll talk about it a bit more in a minute. But in the test suites that we write, this is what we're doing a lot of the time. And sorts of things like changes in white space and things like that are just annoying. So if you can just compare some HTML, that's way more useful. It makes the test way less brittle. But it's not something that you're going to find out of the box. So here's a pretty specific example that I have this sort of fake but very much similar to real example of a test for an ad server, some web application that basically you give it an ID. It shows you the associated piece of media with that ID. And we have our three steps here. We do a little setup. There's some hand waving here that basically there's some in-memory database that we're adding this advertisement to. And then we hit the URL in our application that's actually supposed to be showing that. And then we make these assertions about, OK, I expect these three things to happen. I want to make sure that I got the right status code. I want to make sure that I'm actually properly setting headers for content type or whatever. And then I want to actually make some assertions about the body of my response. And you stare at this for a moment and you try to apply the rule that you had before, which is one assertion per test. And it's sort of not necessarily obvious what the way to make that happen is in this case. Because all three of these assertions are useful, they're all things that are basically does my response actually work correctly. And so the same mindset that we had before basically leads some people to basically split this up into three tests with the sort of same set up and exercise, which is not ideal for reasons that I'll skip over at the moment, and instead just skip straight to something like this, where I have nothing highlighted because this is way more code on a slide than I expect anyone to read. But so what is this? This is basically the conversion of these three assertions into one assertion that actually encapsulates some meaning. And the meaning that this assertion is actually trying to encapsulate is that responses have content. I sometimes want to assert against that content. And there are a couple of things that need to be true when I'm making that assertion, and I want this assertion to just handle taking care of all of those things. So for example, we're checking all of the same things here along with the addition of content length. So we're checking all these things, making all these assertions in the assertion method. And so you end up with something that looks more like this. And obviously, we've cut down the number of lines in the test. I think also we've sort of gotten some extra clarity. I think the assertion with this name has more of a direct, it's telling you directly what you're trying to assert against. And we also have the same benefit that we get anytime that we take a bunch of code and refactor it into one place, which is anytime that you want to come back and make some improvement to this slide, that's going to immediately go and affect all of the tests that you have in some positive way. And you actually obviously have to be careful with that because you can silently break tests in that way. But if you are careful with that, what it means is that things like if someone comes and makes a change to a test and it starts failing in this assertion, and they notice that it fails in a particular order. So for example, it'll fail first for the content length order. And they decide that, you know what, that doesn't really as helpful as if I reordered these assertions. So if an assertion fails potentially, they want to see the body first. And have that be the failure message if you got a different response than you expected because that'll tell you more information. They can go in here and reverse that assertion. And now that particular thing, which is probably true across your test suite, you now get better failure messages everywhere. Just because someone noticed and went there and made that improvement. So it's the same benefit that we get anytime we take some software or we take some particular operation that we're trying to perform and refactor it out into one place that we can basically concentrate on. And so what happens? What happens is if you do something like this, the proposal of starting to build assertions on top of the data comparison assertions is that you end up with a sort of hierarchy of assertions. So at the bottom, you have your data comparison assertions because at the end of the day, you are just comparing two objects. So that has to happen somewhere. But on top of that, you can add these layers of meaning. So rather than just being about comparing strings, now you have an assertion built on top of that. That's really about comparing HTTP responses. Even though at the end of the day, it's just comparing a bunch of values, but it becomes a much more powerful assertion that's able to add all of the nice messages and things like that that you might possibly want to do once you know that you're actually dealing with comparing HTTP responses. And then on top of that, you can even build more interesting things. Some of these things, in our case, are only in theory. They don't actually exist yet, much as I would like that they would. But if you come back and decide that, you know what I want to do, even more fancy things or even more useful things, depending on your perspective, you can come back and layer on top of these assertions that are making assertions about HTTP responses and start making assertions about, well, is this valid HTML or layer on top of that something like, well, now I want to start doing assertions about how these things render in the browser, given whatever conditions I actually feel like placing around them. And so we're actually doing well on time, so I'll actually spend a bit more on this slide, which is, so where do we go from here? Assuming that we've all agreed that this sort of using assertions to build more meaning on top of the data comparison assertions, so what comes out of that? And so it's kind of interesting what comes out of that. I think after a while of doing this, what comes out of that is lovingly mix-in hell, I think, by which I mean we build up all of these assertions in a bunch of mix-ins, in our case, and that's great. It gets you a ton of benefit, and I'll specifically tell you what some of these things have before I actually tell you why there's a nicer way of doing all this. So for example, we deal with GDBM. It's a GNU database in memory key value store thing. It's in the standard library. But for example, GDBM doesn't have ways of, there's no object layer on top of that. There's no way of basically comparing GDBM databases, there are files. So we have a mix-in that will take a GDBM database and compare it to a dictionary and tell you if those two things are equal. We have a mix-in for logging. This is something that I think everyone has written at least once. There are some packages that actually try to make this helpful for you. It basically attaches a handler to the logger for the standard library logging module and then lets you make assertions about things that have been logged. I think everybody has written that. Well, I think that assertion is quite often written and rewritten by people. We have a mix-in that has assertions about our own proprietary log format, which is quite crazy. So we parse that into something more sane and are able to make assertions against that. We have this response content mix-in, which I mentioned, which has this content assertions, something has content, doesn't have content, has content that looks a certain way. We use statsD with Datadog. StatsD is like a, you send it metrics and it basically puts a nice UI on top of that and shows you graphs and things like that. So you want to make assertions about things having been, a metric having been incremented, all those sorts of things. So what happens is basically you end up growing this companion to your test suite, which is I have all the things I want to test and then I have all these useful assertions that have either don't exist yet or have some meaning specific to my application and then I'm able to basically use those in all the places in my application. I called it mix-in hell because it's sort of annoying. Just the coupling of inheritance to actually adding these assertions to your test is kind of annoying. So I will mention that there is something else. It's up on the slide, which encapsulates this idea of I have this collection of things that I call an assertion and I want to just be able to use that all over the place. Those are test tools, matchers. They sort of claim to solve this problem. So they're worth checking out if you're convinced. And so the last thing that I want to say is sort of a call to action, which is that in part to myself, which is that I think a lot of people are writing these sorts of assertions. There are things that are useful. If you want to compare some HTML, where are you going to go and look for the assertion that actually does that in a way similar to how I described? So I think we need to start sharing these things that we're writing. When was the last time that you downloaded and installed a package that whose job it was to add a bunch of assertions? I think it's a bit more common to do that for test tools, matchers. And there possibly are some packages that do that, but I'm not sure how widespread they are, so I'm sure someone will tell me. But regardless, I think there are a whole bunch, there's this whole layer of things that is possible to build on top of the simpler assertions that we have for doing these sorts of comparisons. And it would be nice if we sort of built out a bit in sharing these assertions so that they only have to be written once. We can sort of distribute that benefit across everyone. That's what I got. Thanks, everybody. Thank you, Julian. So we have at least five minutes for questions. So there's one microphone here. Is there another one in the room? No. Okay, so has anyone got any questions? And I'll bring the microphone to you as close to you as I can. One question here. I would like to know how strict are you with this one or third equal statement, because if you go back to your first example, what if your data structure, say you have a dictionary, is more complex than that? But in this test specifically, you only want to test for cats and dogs. If you do a direct comparison, you will get a big diff, which is not telling you anything, right? So you would need two assert statements that look into that dict. So what would you? Yeah, so sorry, didn't mean to cut you off. So I gave this talk last week in London, and it took 40 minutes, not 20, which is why I thought I'd be close. And someone asked that same question. I think he's actually here. So the answer is, or there's no general answer. We're all going to just cave at some point sometimes. The ideal answer I think is usually that I try to actually use those cases as examples for when to try and look at the code again and see if there's a possible way to split whatever it is that's doing that apart. Because if you have this thing that's being outputted and there's a bunch of assertions that you're making on pieces of it, then possibly that means that what you actually really have in code land is this thing that really should be spitting out a whole bunch of different results, and then something later that's combining them. So I try to be fairly strict because I think that's turned out well in that particular case. It sometimes tells me useful things about the actual code that I'm writing and tells me how to break it up differently. But sometimes it happens. Yeah. Yeah, so if I catch you correctly, it makes less assertions, but makes them more meaningful by using custom assertions. And that works well, I guess, if you have generic things that you test for, like HTML, for instance, output, and you can even share them and use them. But I guess in a lot of cases those will be specific to your application. And I've had something recently with a pie test fixture, and then I realized, geez, I'm putting so much logic into this assertion, into this fixture, I need a test for this. So then I ended up writing a test for my fixture, and then I suddenly noticed the afternoon was over. And then I went home and thought, is that really the way forward? I'm making my tests more, they're smarter and they're certainly better readable, but now I have to test the test assertions. Do you see any way out of that? What's your experience? It's the same thing, and it's the same thing with anything in testing, I think, which is always a balance. So that's obviously not good. I don't want to be writing tests for my tests. I don't think anybody does. The nice thing, I think, is that often the assertions that we have that we've written come out of, like, I have a bunch of places where I want to make the same set of assertions, and I notice that, and I say, OK, that means that that's a good place to factor out. We try not to start from the other side, which is, I have this thing that I want to assert about, start writing the assertion method and then using it in places, because then it often does turn out like that. You try shoving a lot of things into that assertion at the end of the day. There'll be flags in the assertion to do different things. Somebody tried to, it actually got merged, I think. There's a flag in our test suite now where, sorry, this is embarrassing, it's going to be on video. There's a flag somewhere in our test suite for doing comparisons on HTML, and it sort of switches on whether or not to do an assert in or an assert equal, and I hate it so much. So it happens, unfortunately. I don't have any other, like, I don't have any thing that can help out with that other than just use your best judgment as it's happening. Any other questions? Hold up your hands if you have a question. No. Well, since we've got some time, I've got a question. Oh, no. So we both work on the Twisted project, and I know that I've been told off, umpteen times for having multiple assertions in my tests. But I wonder, when you've just moved all your assertions into a custom assertion wrapper, then you're going to fail, what JP will always preach to us is that the problem is that then if the test does fail, you have to run the test multiple times to get to find out, and if it fails on multiple assertions, you have to run the test multiple times to get all the failures. So would it be worth instead of having those multiple assertions in the wrapper, just collecting all of the information and putting in one single container? Yeah. Yeah. You couldn't have gone with the softball question? Yeah, so I agree with him and with that. That sometimes it's nicer to just shove it into a container. I think it's very easy to be lazy in that case, because it's sort of a non-obvious thing. Until someone tells you to do that, I think people probably think of that when they try to apply this rule, like, okay, I'll just shove it in a tuple and make tuple comparisons. And they say, no, that seems ridiculous. And I don't think that's a solution that people are likely to come up with unless someone tells them. So to be perfectly honest, I think that the actual solution for that problem, which I understand, is that TestTools has this nice other thing, which is failures that don't actually stop execution of the test. Right. That's what he always talks about. I haven't actually seen it in action. Yeah, they're pretty great. So if you actually have a bunch of assertions and you want to also get this thing of, I want to execute all three of these, and then I want all of the context, then I think that's the actual way to go, is to get something like that. Okay, great. Any final questions? Oh, there is one more. Have we got time? Yeah, I think so. Did you have a look into PyTest? Because I think it solves some of the problems you described. Yeah, I'll be perfectly honest. I don't know of the surrounding things that PyTest adds. So I believe you, I'm sure that it has like similar, like other than just the eliminating test case thing, I know it has like fixtures. I don't know of the other components that PyTest has. I mean, you don't need a Metro library because PyTest shows much more of the context. Yeah. Yeah, I know that it shows nice things like, it'll show locals and frames, right, when tests fail and things like that. I know it has some nice things. I don't know which particular, like the, I can't imagine it can provide the same sorts of, like it's not going to give you layer assertions because those are things that you're probably defining in your application. I know that PyTest will give you some nice things, but I don't, I haven't used them. Okay. Yeah. Okay, thanks Julian. That's a great talk, very informative. We'll be on to the floor. Outro
|
Julian Berman - Design Your Tests While getting started testing often provides noticeable immediate improvement for any developer, it's often not until the realization that tests are things that need design to provide maximal benefit that developers begin to appreciate or even enjoy them. We'll investigate how building shallow, transparent layers for your tests makes for better failures, clearer tests, and quicker diagnoses. ----- * Life span of a test * 5 minute - why does this fail? * 5 day - what is this missing? * 5 week - do I have coverage for this? * 5 month - what's *not* causing this bug? * Transparent simplicity * one or two "iceberg" layers for meaning * Higher-order assertions - build collections of state that have meaning for the domain in the tests * bulk of the details are in the code itself * show an example * grouping for organization * Mixins * show an example * unittest issues * assertion/mixin clutter * setUp/tearDown tie grouping to the class layer or to inheritance via super * addCleanup * weak association / lookup-ability between code and its tests * package layout * other conventions * Alternative approaches * testtools' matchers * py.test `assert` magic
|
10.5446/19987 (DOI)
|
So yeah, good morning everybody. I'll be talking about Amanda, our distributed services platform. I won't be showing any other pretty pictures. I just hope that the talk can live up to the amazing work everybody else has been doing. A couple of things about myself. First, I've been software developer at MPC since about 2010. I've been working with Python since 2009. Love services and everything that's plug-in based. Slightly obsessed by monitoring after various phone calls at 3 in the morning. And I had the great opportunity to actually hold on to OSCO for Life of Pire. It was a great opportunity down there. So I'm part of the infrastructure team at MPC. I've been working there since, like I said, 2010. And we create visual effects for advertising and feature films. These are a couple of the movies that we have been working on recently. And we actually do this across eight sites with what we call fully integrated cross-site pipeline, which makes sure that our data flows from one site to the other, depending on what the departments are and where they work. So I guess not everybody here might be specifically familiar with what visual effects are. So this is a quick quote from Wikipedia. It pretty much comes down to everything that is either expensive, dangerous, or would hurt an actor during filming. So we're trying to avoid it. But I guess a couple of actual images of the work that we do is probably going to be a bit better. So this is a shot from what was the years we got it in from the clients. And this is the actual work that we did to it. So everything that you see there in the background is absolutely fake. Same thing here. This is a shot that we got in from Godzilla, one of our latest movies. Same thing here. This is what we got in, and this is the actual work that was done to it. If you look well, you can even see that the guys in the tank got replaced with CG characters. That's how far we push things these days. So to do this, we work, of course, with a lot of assets, where an asset is something like a creature or a texture or whatever else that we need that is fake, and we actually make sure that it flows through the whole system. To do that, of course, the artist first does a bit of his magic. Once that is done, he creates what we call a daily, which is a short movie to actually show the work that he has been doing, and that can then be reviewed by the supervisors. Once that is done, he can approve the asset, and he can, of course, add some comments and things from there. Once it's approved, we actually go through a releasing stage where a lot of things happen. We actually create directories where we store our data. We add some actual metadata about the assets as well and make sure that everything flows into the next department. So here, for example, we've got our modeling team, which, for example, creates an actual character, and then some textures. And while we release, we actually make sure that we update all of the dependencies, we make sure that we notify all of the different artists that new things have been released, and we actually make sure as well that we sync any data that we have to all of the different sites. So, of course, we have to keep a couple of things in mind. Doing this, there's not one artist working. There's about 1600 working. And we release thousands of versions of assets a day. So we have to keep that in mind, but also an ever-changing schedule. So one day it might actually be quiet, and the next day we might have a completely different schedule with a trader that needs to be delivered in a couple of weeks. Which means that we have a whole lot of different sources that we use, coming from a database, third-party APIs, storage, a whole lot of different locations. And they're used by in-house tools that we have been writing, third-party applications that artists tend to work with, and, of course, we have a whole lot of actual multiple environments. So we don't work with one single environment. We've got a whole range of different environments, which means that for every single show we can have a soft set of tools that they use with a specific version, where a different show might be using completely different ones. Other things to keep in mind are users themselves. The artists themselves want something that's quick and easy to use, something that's consistent. They don't want to have to worry, oh, I'm using this API, so I'm going to have to use this way of doing it, or I'm using that API, I'm going to have to use that way of doing it. We do also have to keep in mind that these artists are not necessarily trained developers, but they do write code. They hack around quite a bit. We need to make sure that we can present them data in a safe way for us and a safe way for them, so we want to expose only certain parts of our data to them in a nice and consistent way. Similar for developers, we have developers of any level coming in. Some of them are trained in more visual effects side of things. Others are trained in asset management, but they're not necessarily trained in anything that's distributed or, you know, scoping across eight different sites around the world. So to do this, we developed a service-based architecture called Amanda. We provide that as a platform as a service to all of our different artists and developers, and it's a multi-protocol setup with multiple transporters and multiple concurrences. I'll be going into every single bit throughout the different slides, but it's just a small introduction to what it is. And we try and provide an ecosystem to write a service for developers of any level. So anybody that comes in on the first day should be able to write a service during that day and get into the production by the end of the day. So we're currently running our second generation that was written in 2012 and has gone live in 2013, and it replaced our first generation, which was a push model, which caused a lot of problems. So as soon as the request would come in, it would actually start scaling with extra threads and start running and running and running, and there was no way for us to actually limit that in any sort of way. So we have now moved to a Q-based model, which just allows us to limit things a bit nicer and actually make sure that we have a specific flow and can control that flow in a way, way nicer way. So just some stats. Godzilla, which is one of the latest moves, like I said before. Like I said, we have a render file, which has thousands of CPUs, but if we would have rendered, which means creating that final image on one single machine, it would have taken 444 years to actually render, which I guess a fair bit amount of time. And we've got 65, 650 terabytes of worth of data that went through the system as well. And that generated during our peak times about 250,000 demand requests a minute, which is 120 million requests in eight hours. And for those of you, I guess, since we're in Germany, most of you have seen the Brazil-Germany game. It's about four times the amount of tweets that were about that specific game. And Congress-Germany are winning, by the way. So I'm not going to just step into how we actually have been setting up the whole system from the ground up. I'm going to be starting with the actual service. And the way we have done that is that the service is nothing but a class. So we're going to make here a make movie service. We've got 20 minutes to make a movie, which is probably going to be a bit short, but let's try anyway. So we're going to start with green, the director, because we need to get some work in, which is your typical Hello World scenario. And the important bit here is that it's a class. It's absolutely standalone. It's completely testable. And you don't depend on any of the tools or any of the scaling features of Amanda, which is very important for us, because we don't want people to have to worry about any of these things. We have these little decorators here called AdPublic. We also have an AdProtector and AdPrivate, which allows us to actually expose what methods are available throughout the system for other people to use. So an AdPublic would mean that an artist and the developer can use it from outside. An AdProtector would mean that you can only call it from a different service. So, cool, we have a service now, but it's not actually doing anything useful. And it's definitely not going to help us getting the kind of ratings that we've been having on running Tomatoes. So let's actually make it do something. To do that, we provide into service calls, as we call them, and it actually allows us to call different services. And the way we do this is by actually declaring a dependency inside that class. So I can say I have a dependency on the storage service, and here I'm using the storage service to actually check if the data is on disk. And I can do the self-destroys and check-exist and pass in the parameters that it needs. At that point, of course, we also need some information about our database itself, or from our show itself, sorry. And we can do that with what we call infrastructures. And infrastructure is really a form-alizing, a way for us to formalize our access to the back-end, such as databases, login, configuration, sessions, and you have those things. And in here, you see the underscore DB that is an actual infrastructure. And it just provides the users with a nice, clean, and consistent way to actually access databases. They are in themselves services, but they're stateful services, so that we can do things like pooling and caching and those kind of things. And they are local to the service. So these services are actually not spread across the system. They are inside that same Python module, which allows us to do the pooling, of course. And the really, really nice bit about this, and I'll be hamming on this quite a bit throughout the whole talk, is that we can swap any of those services with other services. So it means that, for example, in this case, here at the bottom, you've got our config, where we're getting something out of the configuration. In the development environment, this could be a dictionary. In production, this could be an XML file. This could be a YAML file, it could be whatever file. And we can swap that in and out with different services. Without, once again, the actual developer of the service having to change anything to his code. So now we've got something that does something, but it's not very useful in any sort of way. It's not scaling. It's still local in one person's machine. And we also don't have that bit where we can actually provide a consistent interface. But we did create all of the abstractions that we need so that we can change any of the parts that we already have with other parts that we might want to use in the future. So let me introduce you to the service provider. And this is how you actually create one of these service providers. And this actually allows us to get that consistent interface. It hosts the services for us. And at the bottom, you can see, you know, we created Make Movie Service here and our story service. We passed them in and we can then call it with services.makeService, you know, make my movie magic happen. Or logging and actually change the logging level. So this is the kind of things that we actually allow them to do. But we still don't, like, we're still not able to scale in any sort of way. And we came up with the idea of proxies. And proxies are stand-in services for the requested service. So they pretend they are the service that you want, but they're not really the service that you want. And underneath the hood, they just stick your data into a queue and the queue can pass it on to whatever. We also tend to call these queues transports. And once again, they're completely transparent to the user. The user doesn't have to care. The service developer doesn't have to care where his data is coming from. So queues or transports, they allow abstract away technologies like RabbitMQ, 0MQ, UDP, any of those things. We can all abstract them away. And it allows us to transparently swap out things like adapters. So if at some day I want to use LibRabbitMQ and the next day I want to use PyAMQP, I can. Without once again having to change my service, my service provider, anything else, I just have to pass in the transport bit, which is configuration. So at this point, we can scale a bit, but it's still going to be expensive to run 250,000 requests simultaneously because we need a whole lot of these services running. So of course, you want to do some parallel processing and some concurrency kind of things. Service developers for us shouldn't have to worry about how they're doing concurrency and how that works. And they need to know if something is going to be CPU intensive or IO intensive. There's something that we do want them to think about, but we don't want them to think about, oh, I'm going to have to pull a thread there and do this, that way, that way, that way. We think that we should accommodate for both because some tests can be CPU intensive, other tests can be IO intensive, and we don't want them to worry about that. We want to be able to use threading in one way and dreamless in another way or multi-processing even if we wanted to. So far we have been building this little block here, which we have been seeing, and what we did is actually stick a worker pool in front of it. The worker pool provides a simple interface across various concurrences, and the pool is fed with requests from internal queues that is filled by consuming from our queues in there. Once again, workers can be changed, can be extended, and they can actually be changed, just like you would do with middleware. So you can just build a whole nice set up here. At this point, we've got a nice little building block that we can reuse everywhere. So at this point, we really have something, like we have all of our building blocks that we need to start building a slightly larger system at this point. And the nice thing about this is that we can actually start chaining these blocks together, and that's what we did in production. So in production, we have a cross-language pipeline. We don't have Python. We've got 95% of Python at MPC for most of our tools. But of course, we need some C++ for anything that is really, really heavy. This is just really heavy to do. At that point, you might just want to use C++ for any graphics. We have some JavaScript laying out for some of the web tools. We have Lua. We have a whole bunch of other ones. And we actually want to be able to present all of the data that Amanda has and all of the services have to all of these different languages in a nice and consistent way. So what we did is actually our first work of pool, we replaced it with MicroRisk in Flask. Nice and lightweight and simple. Just a little zoom in here so you can actually see where it changed. And that allows us to actually use HTTP quite effectively. It allows for simple clients on every single language. I mean, any language these days should be able to make an HTTP call. And it's a nice, simple client that people can use. And people don't have to worry, ooh, I'm going to have to do threading to use this transport or that transport. We take care of that and we take that away. It does limit us to native types because our HTTP transports either transport JSON or XML, because JSON and XML are pretty much available across all of those languages as well. But it does limit us to native types. Oh, we need to start extending the encoders and decoders to actually start dealing with those issues. So our front-end here is a MicroRisk and Whiskey Flask worker. And actually, we don't really do any work in Flask except for session handling, which itself is an actual service. The rest is just being proxied across to RabbitMQ where RabbitMQ takes care of the distribution across all of the different services that we might have running. So at this point, we've got a system that can be distributed that is available to all of these different languages. Of course, we want to make sure that it's full tolerant as well. So what we did is actually we run two instances of those MicroRisk and Whiskey Flask workers and we stick NGNX in front of it to do load balancing and failover. Nice and easy. And we actually run a non-clustered RabbitMQ setup. So rather than actually clustering RabbitMQ for those who are familiar with RabbitMQ, we run multiple instances of RabbitMQ. And what that gives us is that we can actually use our proxies to consume from multiple queues and transport at the same time. So like I said before, we can swap in any of these transport with a different transport and we can go as far as running RabbitMQ and another RabbitMQ, but we can also run RabbitMQ, 0MQ and Redis at the same time and we can start consuming with one single proxy from all of these different transports at the same time. So if in the future something nicer comes along or something better comes along or our whole setup changes, we don't have to rewrite the services. We don't have to rewrite anything else. We can just swap all of these bits in and out. So at that point with that going, the last bit that is left is monitoring, which I'm quite keen on. And there's something that needs to be done. So what we did is we assigned an actual ID to every single request. As soon as it comes in, we'll make sure that it has an ID and the ID is being followed throughout the system. So if I go from service A to service B to service C in Vancouver, if it blows up in Vancouver, I'll have a trace that it blew up in Vancouver because every single request is logged and I can actually start searching on those request IDs throughout the system and find the whole trace of all the different requests. So since we really love our services and service-based architectures, we actually made sure that we have a statistic service and a logging service and the data that we had in here, for example, at the bottom, where we have a calculation of how long it takes to get from the front end, so from Michael Whiskey to the end of RabbitMQ, or the amount of time it actually took to execute the request, we can map these onto the system itself and we actually send those to a statistic service or a logging service. And what that allows us to do once again is that if at some point we're now using Carbon, if at some point we want to use StatsD, we can change the statistic service without once again having to change everything else. The nice thing that we did with our workers, since they can be wrapped in, can be done a whole lot of things with, is that we have one single worker that executes the requests and that worker is wrapped in a statistics worker, so as soon as the request has been done handling it and since we have TransposingQs, we already have the result going back to the client, that is the point where we actually start doing our stats calculation. So we don't have the overhead of actually doing our stats while we're still executing the request. It's done afterwards. Of course, there's a bit of calculation up front, but all in the debt, it all happens afterwards. Same thing for logging. All of our logs are going through a logging service which allows us to actually dynamically change our logging levels by an Amanda request, I can say, change my logging level to debug for this specific service and we can make those changes on the fly as we need them. So maintenance-wise, we use Salt. For those who don't know Salt, check it out. It's a really cool tool. Similar to Poppet and Chef for those who know those. It's in Python and actually we extended it within Amanda module, so Salt can now run up Amanda services and can run up a whole framework for us. And we actually wrapped the Salt client itself in a service, so we actually use Salt to investigate the system on the fly via actual Amanda requests to know what's going on in the system without really having to log into the masternode. And what it really, really gives us is that predefined, repeatable configuration that we need because we've got eight sites to look after. We want to be able to make sure that what's running inside A is going to be the same as inside B as inside C. We want to make sure that it's all the same. So we've got an adaptable, extendable, configurable system at this point. We can change services, swap them in and out, like we want to. We can swap our transports with whatever tools that we need. By the way, a big thank you to everybody who has been writing a lot of these modules like LipRabbitmQ or, you know, SimpleChase, we use them a lot and thank you for that. It's very extendable and configurable and it's all configuration based. We can abstract the whole system from, you know, system level all the way down to service level. And we really have a best of breed system at that point where, you know, we might build a system for a particular show or pipeline or for any of our specific use cases. So there's a couple of things that we're still looking at. Containerization is one. We don't want service A if the CPU is going crazy on it to actually take out service B. So we're looking at containerization. We're looking at auto scaling as well. If you have done investigations just like us and you want to have a chat about it, it would be great. And we're also looking at the possibility of actually open sourcing the whole system. So that's pretty much it for the whole technical thing. Sorry, I didn't go a lot into actual Python code itself. It's 20 minutes, so actually digging into it would be quite tricky. Just a couple of slides. We are actually looking for people and we've got a lot of things in production at the moment. The Jungle Book one, keep an eye out for that one. It should be a really, really cool movie. And of course, we are hiring as well across all our studios, across everything. So yeah, either have a look on the website or have a look at recruitment and or come and talk to me after the talk, of course, as well. Thank you. And yeah, any questions really? The microphone for questions is over there. Just stand up and go there if you have any questions. Do you deal with versioning of these services in any way? Do we do it, sorry? Versioning of these services? Yeah. So every single service as it's deployed, get assigned a version. And we have an actual, that's why we use SORT as well. So we have a configuration set of those services. And if a service change, we actually run up a different version of it. And we can actually have a staging and a development mesh where we can push those changes out first and run a bunch of tests against and spread it out to a couple of users to start using before actually pushing it into production. So every single service is versioned. I have a question. How Amanda differs from, let's say, a standard enterprise service bus? Because I do get it. I don't understand why have you rolled the code from scratch and not used, for example, let's say a service bus where you can plug in different services and so on. You mentioned salary, right? I'm saying ESB, enterprise service bus, because when you do want to do, let's say, service oriented architecture, you just use an ESB and I don't know why you haven't done that. I don't know to be honest. I'm not too familiar with ESB, to be honest. This is a technology, not a tool, like enterprise service buses when you want to integrate a lot of different various environments and so on. You just use an ESB with multi-protocols and so on. And this looks quite same to me. Maybe we can chat about it. Yeah, let's do it. Of course, yeah. Interess it a little bit about that one. Hello, it was a great talk. I have a question about load balancing. What do you use to do it? Have you got any algorithms and metrics? Sorry, I couldn't hear you. What about load balancing? What technology do you use to do it? To do load balancing? Yes, approximately, or something like that. So the front end we've got engine X which we use for load balancing. So we've got multiple micro riskier flask instances set up there and engine X load balances between that. And in production we use, on the other side, we use RabbitMQ to actually do load balancing. So we have our proxy set up and they have a certain amount of requests that they can handle simultaneously. And if we see that the queues are getting too long, we just start spinning up more services. That's why we're looking at auto scaling as well to actually deal with those issues. What do you do about the large data amount? You have a service which operates on data like source images and they're not available on your other location around the world. How do you make sure that the data is available and how can it be pushed around the world? So we've got various things that we do. We've got one of our infrastructures that we have is what we call a cross mesh infrastructure. And of course we cannot check if something is on, say, if we look at storage, we might not have an on storage in Vancouver. And we can actually make what we call a cross mesh goal. And you can do with self cross mesh with this site and you can then use the same service interface to actually go and call that specific method, say in Vancouver, to go and check with the storage service in Vancouver if it's available down there. And then we've got of course our syncing queue which takes care of actually syncing all of the data across all of the different sites which happens at release time. We have specific rules set up as part of a service that says, okay, this method or this asset has been released, doesn't need to be synced to any of the other sites. So there's all service based. And the same for generating data. How do you prevent generating like terabytes of data by some artists who you just do monitoring and you look at the operations. So the requests going through it, man, are very small and very lightweight. So we wouldn't be sending terabytes of data through there. We just use our sync services, we call it, to actually detect and make sure that the data that needs to be synced is going to be synced. We have large dependency trees in these assets where we can say, oh, this asset has this texture and this texture and this texture and you know, these kind of rigs. Go and check if we need them in the other sites as well. Or are they just doing something like lighting where they just need to render frames or example. Any more questions? Not. Thank you very much. Thank you.
|
Jozef - Amanda: A New Generation of Distributed Services Framework To help create award winning visual effects, MPC developed a distributed service-oriented platform, Amanda. Amanda allows developers of any level to write a service that is presented to users across 8 facilities globally without them requiring any knowledge of building large concurrent systems. It allows artists and developers across different domains to work with clearly defined API's and gives the service developer control over what and how data can and should be accessed. The talk will cover how to set up such a platform from the ground up. Starting at the service level building it out with additional modules and technologies until the fully distributed system, covering topics such as concurrency, componetisation and monitoring that allow the fine tuning of setups depending on the type of work being undertaken and changing business needs.
|
10.5446/19986 (DOI)
|
Joseph Heinrich is head of the scientific IT systems group at the Forschungszentrum Jürich in Germany and he will be giving a talk on the visualization system. Okay, thanks. Have a welcome. So okay, first of all, thank you for coming here this morning. My name is Joseph Heinrich and together with my colleagues I'm working on different projects at Forschungszentrum Jürich. Most of them are visualization systems and I'm proud to have the opportunity to give this talk here at EuroPython about GR, a framework for visualization systems. So let me start with a question. Who is already using some scientific software with Python such like Medplotlib, Mayavi, VTK or... Okay, that's a big number. As mentioned, I'm working at a research company and in the past years it turned out that there is a growing need for better and faster visualization software, especially scientists need easy-to-use methods for visualizing two and three-dimensional data sets, possibly with a dynamic component. And they want to create publication quality graphics and videos for their publications probably in the Internet. And they want to make glossy figures for high-impact journals or press releases. At first glance, those methods don't seem to be very challenging but we are talking about it later. There are a lot of scientific plotting methods we need such as line bar graphs, curve plots, scatter plots, all these things you see on the slide here. In principle, this is nothing challenging and there are dozens of solutions for all these kinds of plotting methods. There are also powerful software libraries for scientific applications in Python. Those listed here on this slide are the most popular ones, I think. Maybe I've forgotten one. But we all know Medplotlib, which is the workhorse and the defect of standard concerning graphics in Python. And there's even MyAvi for three-dimensional applications, MyAvi, which is very powerful and based on VTK. And it offers an application interface called MLab, which can be used in your own scripts. There's VTK, it's very versatile, but it's difficult to learn because it's a very low-level system. And we have tools like Wispy and OpenGL, which are both very fast and which are limited to 3D and which are really the lowest-level APIs for graphics with Python. And there are also some GUI tools, just like QWT, with its corresponding 3D equivalents. And the problem with this is that they are currently unmaintained, at least for what my information is. So there are some problems so far. And the main problem, I think, is that the 2D world and the accelerated 3D world are separated. You won't find a tool which provides services for both 3D and 3D graphics. And another problem is that some graphics backends only produce kind of figures. So it's not possible to present continuous data streams from life sources. And also, I've made the experience that there's only a bare minimum level of interoperability. So user interaction is somehow limited with these tools. Also, if we are talking about analyzing large data sets, we often see that there's only a poor performance. And also, these APIs are partly device and platform-independent. So your own scripts will suffer from some system dependencies after the time. Okay, so let's Python get up and running and push for Python. There is a very nice distribution which has been introduced to you in the keynote this morning. It's called Anaconda. And I would really like to recommend this distribution as it's very easy to install a complete scientific Python stack. But I think we need something more. For example, we need some more performance. And this can also be achieved by Anaconda at once. For example, number, which is also be mentioned this morning, which is capable of accelerating number of your Python applications based, which contain NumPy code, even on GPU hardware or multi-core processors. And I will give some examples later. There's something more. We also want to achieve more graphics performance and interoperability. And for this purpose, I would like to introduce our GR framework, which is a universal framework for cross-platform visualization. And the main key points are that it has a procedural graphics back end. So you can really present continuous data streams. And it has built-in device drivers so you can visualize both 2D and 3D scenes in one canvas. And there's a very good interoperability with GUI toolkits so you can establish a very good user interaction. And as you can see in the bottom part of the slide, it's also very easy to install. So this would be our complete scientific Python distribution. I think we have everything we need, especially we have more performance and interoperability. So let me give some examples how this looks live. You can see here a numeric simulation of a damp pendulum. The calculation is done in the RKV, RK4 function. You can see which is simply a numerical integration of this differential equation. And you can see that you can mix graphics with test formulas. And you can do all these things live while your script is running. You don't have to produce figures or something like this. The same works for 3D. You can see it here. In this case, visualization is done with an API which has been written by a colleague of mine, Florian Riem. And he has written an openGL layer for GR, which is called GR3. And you see it's very performant and it does its job. You can even visualize live signals from ray files or from the microphone. And with lock x-axis, and this all runs in real time. These are all things which are very hard to realize with other tools. You can do this also in 3D. Adjust, push the audio away so you can focus on the graphics. So the frequency spectrum is, in this case, visualized by a surface plot, which is realized with openGL and it's that fast as you could see. You can also produce graphics with user interaction. You can see here an MRI application which renders some MRI data through a margin cubes algorithm, which is part of our software and which can be rendered very, very fast and moved with a mouse. So let's talk again about performance. We not only have some needs for more graphics performance, but also for numerical performance. And as mentioned before, there's something called NUMBER which is part of Anaconda, but you also can install it for your own. And there's NUMBER PRO which has some additional features. It's part of Anaconda Accelerate, which costs a few bucks. I don't know the actual price. It's capable of calculating NUMPAI expressions on the GPU, so you can write your own GPU kernels in Python. And it's a very nice tool and it's worth to look at this software. And there are even other tools like KU, BLAST KU, FFT, KURAND, those tools are just dedicated to CUDA hardware. So in this case, you can see how you can profit from such software. You see particle simulation which is very slow, currently running at three frames per second. And just by adding some decorators and an import statement for sure, you can increase the performance by times 15, I think. So you don't have to change your code and you can speed up your application enormously. This is calculated in real time. This would not be possible with poor Python. If you run this simulation in Python, I think each frame is about three seconds. In this case, it was paralyzed or vectorized. This is several examples in our demo suite. Just take a look at the website and you will see how the different optimizations work. So let me introduce some of our success stories. We have integrated our software in several of our applications. We are working both for experimental physicists and for theoretical physicists. And this is something for our instruments. It's a live display for a small-angle neutron diffractometer. And as you could see, you can set the region of interest and the surface is generated in real time. You can rotate it, you can flip the axis, and there's even more. And all this can be done in real time. So this is another example. Here we are processing a huge data set. And it's also done in real time. And this was formally done by a proprietary solution. And with a GR framework, we could embed this into a QT4 application, which was a replacement for the existing solution, and which is much faster, which can produce movies and all these funny things. There's another example here. Nicos is a very complex network-based control system, which is used at the Proxima's reactor mention in Munich for all the instruments which do neutron scattering. And in this case, we replaced by QVT, QWT application with a QTGR application. And it was much faster, it was more responsive, it had some additional features which we didn't have before. So this is a case study to see how fast you can simulate data. Born is a software for simulating neutron and X-ray scattering. To compare it, it was a replacement for medplotlib at this point, and it uses a single call. It's the line just in the bottom of the left side. And if you look at the old code, if you compare the old source code with a new one, well, that's only one line and an export statement to generate a movie, for example. So it's not such complicated to produce movies with a GR framework. So what are the conclusions? The use of Python with our GR framework and Numba, and perhaps Numba Pro extensions, allows the realization of high-performance visualization applications, both in scientific and technical environments. And the GR framework can seamlessly be integrated into any Python environment. So I would suggest to use Anaconda. The integration is simply done by a C-type mechanism, so you can also use it in your own Python distribution. And the combination, Condor and Anaconda, provide a very easy to manage and ready to use Python distributions that can be enhanced by the use of our GR framework, especially with its functions for real-time or 3D visualization functions. So what's next? We are not far from implementing a molecular dynamics package, which will produce such results. We have already all this stuff written in C, and we simply have to write some simple wrappers which will then be integrated into our GR framework. And with this framework, you will be able to do things like this here. This is a simulation which has been calculated on a very big machine, and the data is read with a simple Python script, and then rendered with a GR3 library. You can then export this scene to, for example, Povray, and produce high-quality graphics like shown in the right side of the slide. And you can even do this in highest resolution if you give the correct parameters to those routines. And you can see here it's a very realistic presentation of a DNA. So what are our future plans? Well, we have thought to combine the power of MADplotlib and GR, and we think it should be possible, and the basic idea is to use GR as a MADplotlib back-end. So this would speed up MADplotlib, and all your MADplotlib scripts would profit from this speedup. I think it's possible we didn't start this development, but I think there's a good chance that we get these things running. And there are even more challenges. You learned about Bokeh this morning, and I think this should also be possible. Once we have the MADplotlib integration, it should also be possible to connect those scripts to the Bokeh back-end, which Travis mentioned this morning in the keynote. At this point, I think we should talk to Travis to cooperate. On this slide, you find some resources. There's a website for our framework. There's a Git repu. It's posted on the Py package index. We even have first Beanstar binary distributions for the GR framework, and the talk should be online on this link, which you can find out later. So some closing words. Maybe you hate me after this, but I think that's important. I think that visualization software could be even better if the prerequisites of an application would be described in terms of usability, responsiveness, and interoperability, instead of a list of software with module dependencies. We should use native APIs on the different systems instead of GUI toolkits, and release updates should not break version compatibility. This is something that I have observed very often. So let me end up, and thank you for your attention. Thanks for this great talk. One of the features of Macplotlib that I find very convenient is its integration with iPython notebook, and I play with the visualization before I integrate it into some application or save a high resolution copy for publication or something like that. So I wonder if is the GR framework compatible with iPython notebook, and I use it from there? Right now it's not, but because there's some discrepancy. On the one side we are talking about immediate mode graphics, and with iPython notebook there's just a sequence of commands, and maybe if we get our Macplotlib back-end running as we consider that it could work, then it might be possible to use it in iPython, but I'm not sure about this. Thank you. We'll do our best. Thanks. I'm doing a lot of training of neural networks inside CYthon, almost completely outside of Glober and Teptoloc. Would it be possible for me to bind to CAPIs or CYthon APIs that are excluding Glober and Teptoloc, or do I have to bind back through Python then? As I want to visualize the training of the network during the training process, I think this would be really cool for that. I'm not sure about this. I think we have to talk about this after the session. Sorry. Are there any more questions? Okay, I have a question. When you use the Specterize decorator, and you use the same code in the limited or the basic number version, will this just do nothing, complain with the name error or something, or import error? Is the Specterize decorator available? I mean, even if it does nothing in the basic number version. So, do you mean if it's not present on your machine? Yeah, okay. I mean, for example, you have the Specterize, which I think I understood only works for the pro version, and if you have only the basic version and sort the number. The problem is that the pro version is capable of pushing those LLVM code on your GPU, and the public version is only capable of parallelizing on your own CPU. So, the pro version is only needed if you want to use your GPU for the computing of NumPy operations. What I meant is if I get code from you, which has the Specterize decorator, and I only have the basic version installed, will it just not vectorize but otherwise ignore the vectorize decorator? No, this would not work. So, if you want to try those demos, you really have to purchase the pro version. There are a lot of other demos which don't depend on the pro version. Okay, thanks a lot. Okay, if there are no questions, have been added? Okay, so thanks again. Yeah.
|
Josef Heinen - Scientific Visualization with GR Python developers often get frustrated when managing visualization packages that cover the specific needs in scientific or engineering environments. The GR framework could help. GR is a library for visualization applications ranging from publication-quality 2D graphs to the creation of complex 3D scenes and can easily be integrated into existing Python environments or distributions like Anaconda. ----- Python has long been established in software development departments of research and industry, not least because of the proliferation of libraries such as *SciPy* and *Matplotlib*. However, when processing large amounts of data, in particular in combination with GUI toolkits (*Qt*) or three-dimensional visualizations (*OpenGL*), it seems that Python as an interpretative programming language may be reaching its limits. --- *Outline* - Introduction (1 min) - motivation - GR framework (2 mins) - layer structure - output devices and capabilities - GR3 framework (1 min) - layer structure - output capabilities (3 mins) - high-resolution images - POV-Ray scenes - OpenGL drawables - HTML5 / WebGL - Simple 2D / 3D examples (2 min) - Interoperability (PyQt/PySide, 3 min) - How to speed up Python scripts (4 mins) - Numpy - Numba (Pro) - Animated visualization examples (live demos, 6 mins) - physics simulations - surfaces / meshes - molecule viewer - MRI voxel data - Outlook (1 min)
|
10.5446/19985 (DOI)
|
Okay, so I'm going to talk about, as we, I guess we're almost have everyone in and this should still just be introductory material about scalable real-time architectures in Python, what do I mean by that? I want you to walk away from this session with a couple key ideas, okay, specifically around partitioning and fault tolerance and how we can achieve that in building such scalable real-time architectures. My focus will be on storm, but the ideas are applicable to other tools as well, such as Spark, streaming, or tools that you might roll yourself. And I think the reason that we want to be doing this is of course we're dealing with more data, but we also want to be more responsive to that data. I will show Python, but more at the end. So I'm core developer Jython, you may have seen my state of Jython lightning talk yesterday. Let me plug this book if you're ever interested in Jython. This is a great book. I work at Rackspace on these types of issues. I have had the chance to work in distributed computing for a while and especially failures at scale. I teach the principles of programming language course occasionally at the University of Colorado, which is fun, but actually is done in Scala, not Python. And I work with this user meetup on storm. We're probably going to be changing this to real-time streaming in the near future. What sort of real-time architectures would you like to build? Well, I can think of a few and I'll just mention a few since we are time constrained. Maybe real-time aggregation. This is your classic approach that you'd be doing in Hadoop, but again streaming. So you're not looking at this on a batch by day basis instead as it comes in you're updating your counts or other aggregate aggregations. Perhaps you're building the dashboard. So some extension of this of the real-time aggregation. I'm particularly interested in the idea of decision-making where you will be responding to information in your environment and taking some action. So what are some of the common real-time characteristics of such systems? Well, you're consuming streams of events. You are being event-oriented. As an event occurs, you may take some action. You may want to compute. You may be doing something downstream. You want to go and minimize the latency from the arrival of that event to that potential computation. Certainly not hours. Ideally, going down to seconds or below in terms of that threshold of latency. Such systems are often called complex event processing or you could call it stream processing. It doesn't matter. These are the concepts. The one thing that we are not doing here is doing hard real-time systems. Oh, one last thing. You might have written one of your own type systems. In fact, I'm going to show you what you might have written in the past. So you might have written something like this. Does this look familiar? Well, it should because it is about as generic as you get with a Unix pipeline. But what I wanted to show here is that you have compositions. You're being able to go and build this pipeline out of reusable filters. Another nice property. You can rerun this pipeline if any intermediate step failed. But some problems. I obviously didn't show some of the details around tailing, whatnot. But there's also these aspects of how do you go and implement these other things such as joins, windows. But most importantly, how would you scale that thing up? I mean, I think we can all think about how that could be done. You might go and in some way describe one set of files as being processed by this machine. One set of files by some other machine. But you have to go and actually manage that partitioning yourself. So at Rackspace, we're using this framework for complex event processing called ESPR, which has very much the same properties or again, this homegrown code. And here's the problem. You need to ensure all relevant events about that given customer. It's in one place. If I'm going to be able to know something about this given customer, I need to bring in all the relevant information. I have to go and put that in one place for that to happen. We have to have some locality. Again, in terms of this Rackspace example where we're implementing global alarms, we want to allow customers some degree of customization. So we don't want to make it too hard coded. Then run some computation. So again, a very classic complex event processing system. And simple sharding, the one that we might just know, be able to do quite readily by say, well, maybe we'll do it all by all by customers and whatnot. It doesn't necessarily get us as far as we'd like because you might have again sharded on some of the key. You might need to shard on multiple keys. So what observations can we make? Obviously small problems are easy. How do you make a small, a large problem easier? You divide and conquer it. To divide and conquer, you need to have some horizontal scaling. We're no longer building systems such that they always will require just that larger machine and said we build so that they can scale to a cluster of such machines. But what do we know? The more machines we add to the mix, the more likely we're going to have failure as well, especially since we like to use commodity systems. And once we have failure, then we have to go and coordinate. So I have seen, and maybe you've seen in your own environments, that sometimes people will propose add Zookeeper or add some coordination system. And that just doesn't go and make the problem go away. Even if Zookeeper is awesome, you still need to go and consider how do you manage failure. So yes, Zookeeper can go and give you all this in its toolbox. And it's fantastic as a consequence. But I'll tell you that that doesn't, you know, just assuming that you have distributed locking in your environment doesn't tell you how to recover from a failure and then going and releasing those distributed locks. The solution is not reboot the cluster. How many have done that? So that's not a solution. We want to be, it's okay for a given node in the cluster to fail, but rebooting the cluster defeats the purpose of running the cluster. So Storm has some terminology which I will introduce by, as we go through, but there is this idea of an events source which we call spouts. We have event processing nodes called bolts and some topology to link it together in a directed acyclic graph. There is strong support for partitioning fault tolerance. These key elements that we started with in terms of thinking about how would I go and be able to build up a problem that I could divide and conquer on. Storm is written in closure, but it exposes a Java API. Hence my use of Jython here, although of course you could use some IPC mechanism to talk with, you know, a CPython for example, and that was actually done in the talk on Tuesday where they were looking at the parsley system for that support. It uses ZooKeeper to manage things, but you don't necessarily see it, although you can use it as a resource for your own coordination and as part of the Apache incubator. Actually there's some other Apache incubator projects that are competing in this space. Probably the most notable would be Spark Stream. It actually looks great in terms of the support you can do with that, especially with its Python integration. I think it's the top contender and has nice properties that one can look at that in terms of being functional that you do not see in Storm. These others, SAMHSA and S4, I don't know, I think that these are not nearly as competitive and there's some interesting stuff around there like 0VM as well that you could potentially be doing. So Storm lets you partition streams so you can break down the size of your problem, that returning to that idea of being able to divide and conquer your problem. If the node fails, Storm will restart it. But here's an even more important aspect of what that means. Oftentimes when we are thinking about building systems, especially distributed systems, we want to think about what are the invariants provided by the underlying framework. What does it give us? And this actually helps explain also why you have this distinction between an event source and Storm and event bolts, these processing nodes. Because sources are the things that you are producing events that have to eventually be acknowledged in Storm. So that you can ensure that all events are in fact eventually processed. And when I said eventually, by the way, it actually means in an eventually consistent scheme. And likewise, this idea of bolts, distinguishing that, because you can always replay for them. Hence this topology and the topology in terms of how many nodes you might have. So this is describing how your cluster is being split up in terms of resources. Here's the perhaps one more important invariance that's being provided. The number of nodes for a given spout, maybe a Kafka event source. The number of nodes for a given bolt, maybe that something that is actually doing some interesting processing of that event is held constant. So during the lifetime of this topology, you know that you always have seven nodes for your Kafka spout. What does that mean in terms of your ability to know exactly what's going on in your environment in order to maintain appropriate counters in other state? It's very strong. So Storm is taking into account one extremely important aspect of how you divided that problem up. Because during the lifetime of this topology, if you always know that someone is handling this problem, and you can always go think about it if we had people here, you can always go to that person. Now you don't have to think about, oh, what happens if we add some additional people? It's nice to be able to scale up problems, but that is a separate issue in terms of scaling up the size of a topology. You would do that outside of a given run at your real-time computation. And if that isn't sufficient motivation, I'll give you an example in just a moment. Okay, I think I'm running a little bit slow, so I'm gonna try to speed up now. But these are most important concepts, your takeaways. So you have computational locality. You know that the events for a given node are the ones that are supposed to be there, because that's how you define your routing. You can route on some sort of what they call field groupings. Again, an example is your customer, a tenant in a cloud, a region of some kind, some way of breaking up the problem space in terms of your events. And again, what possibilities do you have if you knew all of the information you needed to know about a given customer, and it was in one place? It changes how you think about things. So normally when you go and write your queries and say, you know, using a relational database, the data is over there and you bring it to you for a given computation. Instead, in systems like Storm, you move the computation to the data. That is the fundamental idea in a map-reduced system, like Storm, like Hadoop. But unlike something like Hadoop, this is in real time. So you're able to keep all of that data in memory, compute on it in real time, and do something interesting. So I should mention, of course, since it's mapping, you might have multiple customers on the given node, so you have to consider that. But that's an easy distinction to make. So again, you will know that it will be on this node and the only on this node. There are other ways of routing, such as random shuffling, global grouping, which means that there's just one node. Everything goes there. Obviously not scalable, but useful for getting totals. Storm will track that success. All you have to do is consider what your retries are. And there are other ways of doing this in terms of doing exactly once event processing, but knowing that at least once all events were successfully observed and computed on is pretty fantastic. Again, think back to that pipeline, that UNIX pipeline I was showing you. That's that same idea. If the pipeline goes down, I can retry it. Of course, you have to handle retry. But what does that look like? Here's the first Python code. It's easiest. If you've already seen that record, then you can ignore it. Otherwise, process it. If you re-try this computation, then you, if you haven't seen it in the context of actually being successful with it, then you can do that retry. That's a merge function. It can be that simple. But it will depend on the nature of your problem. For instance, if I am doing something that's transactional in its nature, you are wiring me $1,000. You can retry that many times, and I'll be very happy about that. So don't think that retries are always going to be successful. That's the nature of eventually consistent systems. Sometimes you actually have to have strong consistency. There are other ways. You could go and have some sort of balancing compensation or something like that and say, well, give me back that $1,000. Actually, in real wiring of funds, that in fact does happen. Another thing, and you see this in any type of streaming systems, your streaming should not capture everything that comes through. This is that old thing. In order to do your computations, you have to download the Internet. If you're doing web crawling, sure you do that. But again, that's done in an appropriate system. So you have to window. Another caveat would be around there are no query languages. You have to build your own. That's actually kind of interesting. Again, with Parsley has some approaches. I expect that we will see these things emerge over time. Think of these as building blocks. That give you these capabilities. But again, these capabilities are a lot like a UNIX pipeline. Imagine, and we know what we could build with those. So you may have to make your own, maybe use some of these other libraries. Okay, I'm going to go and just skip through this a little bit. Except that I will say that Zookeeper is an important aspect of most systems that you would build. And so you probably will be interacting with in some way. So your spouse are ultimately responsible. They're event sources. They're responsible for ensuring that all events were eventually played through successfully, eventually acknowledged. And you need to go and ensure that appropriate handshaking. So as you are consuming from Kafka, you are updating your offsets in it accordingly. I should tell you that again, as some experienced Zookeeper, if you actually try to do one thing I didn't mention at the very beginning, this is that people run a million events a second through a storm node. And of course, you can have many of those nodes running. If you try to do a million events or more per second against Zookeeper, that's not going to work. You have to do some sort of clever batching just like a storm does. Okay. So you get into things like this in terms of the Kafka handshaking, what that would look like. And I will go back through this in just a moment, but I need to move on. So you can have spouse like this. And some of these are already written mostly in Java or in Scala, but the advantage of storm is it's multilingual in the sense that it all runs on a JVM. You could run feeds like Twitter through. You can pull in data from Cassandra. The important thing to know is it's not that difficult to implement. I'll show you in just a moment. Another thing is that you can push or pull events into the system. There isn't something where storm is just saying it will only accept events at a certain time. You can always push events in. But it will ask you when it wants events as well. So you can use that to help balance things. Again, you're going to be using that topology sizing invariant to know how your work is being split up. So you might go and have something like a real time dashboard. You might be using Rackspace Cloud files or some other cloud provider to send things out. You might have some sort of real-time decision-making. I'm interested in auto scaling. In an environment you might often have contradictory information about how things are working. Converge on some action. That's the point of why you're bringing all that data in one place. You may have something along this line where you are doing some sort of real-time aggregation. A couple minutes I will describe Python on Storm using Clam, which I briefly described in the lightning talk yesterday as well. Let's face it, Python is a great fit for writing your Storm code. We have this system called Clam that allows you to readily wrap your Python classes so they can be readily used from Java. In some cases in just one line of code. I discussed this yesterday where you could go and have some bar-clamp class. This is sort of a hello world example. You can readily in a couple lines that are unique to Clam. Just go and add it. You can construct an Uber jar. I'm in Germany. I can say this. An Uber jar, your one jar or a single jar, build it, has everything and distribute it to Storm. You can use it in this fashion. But here's what it really looks like. Here's some code that you could readily use. Say you have a monitoring spout. You're going to be opening up your connection to Kafka. Again, I actually gave this talk at Rackspace. It looks like there's an internal system that, of course, we talk about publicly called Adam Hopper. But again, some others have a feed like Twitter. You can go and read parse events by, as it's asking for it, from next to next to but again if you have something where you're pushing events you can emit at any time and then you're responsible for managing these callbacks fail and act. What do you do under those circumstances? Perhaps you want to go and do some sort of computation on this. What would it look like? This is the pseudocode associated with it. This is all that you need to, this is the basic thing that you need to implement. You need to implement these three methods. You're pretty much done. How complicated is that? It could be like your UNIX pipeline. It could be something that's really simple or could be more complicated. Again, if I'm trying to weigh events that are telling me, oh, it's going this way and it's going that way and I have to figure out what really is happening. But that's the advantage of having it all in one place. Conclusions about this. Storm, less you horizontally skate out your real-time architecture. You have to consider partitioning and fault tolerance. In fact, to me, these are the key questions you answer. This is how you actually think about what it means to divide and conquer your large problem so that it works on this cluster of machines. Answer these questions and you get to go from what you were previously doing in terms of a simple UNIX pipeline, which was analyzing log files, something very similar, to something that can be scalable and, again, real-time. You can choose your favorite language, but I know what your favorite language here is. It's Python. You could again use some of the mechanisms that are out there in terms of communicating via C Python or whatnot, but you could also use, again, Jython and type directly into this big data system. So I definitely would advocate for it. I should. Your strategies will work with Storm, such as Test Urban Development. Have fun. So this talk is available on my GitHub. Repo Jim Baker talks and any questions? Yes. Yes. Yeah, so that's a great question. Yeah, sure. So the question is the long lines of they currently have multiple data centers. They're running a storm cluster in each of these and they need to go and consolidate that information in one place. Now, the first thing to know is that you do not want to, there are very few systems that will span multiple data centers. And Storm is not one of those systems. Okay. You are going to be running a storm cluster in one data center, another storm cluster in some other data center. The way that you would do the spanning problem is use some queue to ensure that data is pushed to some central data center. You may need to go, of course, consider what failure you might have. It gets more complicated. So you basically move that problem to your queues. Use something like Kafka, for example. Yes. That's the best. Yeah, I mean, there's certainly not something that you can do in Storm because, again, it really is depending on the fact that, you know, it's running on ZooKeeper. And ZooKeeper doesn't span multiple data centers. I mean, yes, you can in a theoretical sense. That's not going to be great. So don't do that. Use systems that actually are proven to work. And you're not just doing some interesting innovation along those lines, which you will find is not fantastic innovation. So we are, we are, so yeah, I don't know if there are any other questions. This is something that I can take out afterwards because I obviously would love to spend lots of time on it. Anything else? So I guess we're done. So please ask me those questions that you might have.
|
Jim Baker - Scalable Realtime Architectures in Python This talk will focus on you can readily implement highly scalable and fault tolerant realtime architectures, such as dashboards, using Python and tools like Storm, Kafka, and ZooKeeper. We will focus on two related aspects: composing reliable systems using at-least-once and idempotence semantics and how to partition for locality. ----- Increasingly we are interested in implementing highly scalable and fault tolerant realtime architectures such as the following: * Realtime aggregation. This is the realtime analogue of working with batched map-reduce in systems like Hadoop. * Realtime dashboards. Continuously updated views on all your customers, systems, and the like, without breaking a sweat. * Realtime decision making. Given a set of input streams, policy on what you like to do, and models learned by machine learning, optimize a business process. One example includes autoscaling a set of servers. Obvious tooling for such implementations include Storm (for event processing), Kafka (for queueing), and ZooKeeper (for tracking and configuration). Such components, written respectively in Clojure (Storm), Scala (Kafka), and Java (ZooKeeper), provide the desired scalability and reliability. But what may not be so obvious at first glance is that we can work with other languages, including Python, for the application level of such architectures. (If so inclined, you can also try reimplementing such components in Python, but why not use something that's been proven to be robust?) In fact Python is likely a better language for the app level, given that it is concise, high level, dynamically typed, and has great libraries. Not to mention fun to write code in! This is especially true when we consider the types of tasks we need to write: they are very much like the data transformations and analyses we would have written of say a standard Unix pipeline. And no one is going to argue that writing such a filter in say Java is fun, concise, or even considerably faster in running time. So let's look at how you might solve such larger problems. Given that it was straightforward to solve a small problem, we might approach as follows. Simply divide up larger problems in small one. For example, perhaps work with one customer at a time. And if failure is an ever present reality, then simply ensure your code retries, just like you might have re-run your pipeline against some input files. Unfortunately both require distributed coordination at scale. And distributed coordination is challenging, especially for real systems, that will break at scale. Just putting a box in your architecture labeled **"ZooKeeper"** doesn't magically solve things, even if ZooKeeper can be a very helpful part of an actual solution. Enter the Storm framework. While Storm certainly doesn't solve all problems in this space, it can support many different types of realtime architectures and works well with Python. In particular, Storm solves two key problems for you. **Partitioning**. Storm lets you partition streams, so you can break down the size of your problem. But if the a node running your code fails, Storm will restart it. Storm also ensures such topology invariants as the number of nodes (spouts and bolts in Storm's lingo) that are running, making it very easy to recover from such failures. This is where the cleverness really begins. What can you do if you can ensure that **all the data** you need for a given continuously updated computation - what is the state of this customer's account? - can be put in **exactly one place**, then flow the supporting data through it over time? We will look at how you can readily use such locality in your own Python code. **Retries**. Storm tracks success and failure of events being processed efficiently through a batching scheme and other cleverness. Your code can then choose to retry as necessary. Although Storm also supports exactly-once event processing semantics, we will focus on the simpler model of at-least-once semantics. This means your code must tolerate retry, or in a word, is idempotent. But this is straightforward. We have often written code like the following: seen = set() for record in stream: k = uniquifier(record) if k not in seen: seen.add(k) process(record)
|
10.5446/19984 (DOI)
|
Please welcome Jayir who is going to tell us all about random module. Thank you. So nonsecutor, an exploration of Python's random module. Nonsecutor means it does not follow in Latin. And I chose it as the name of the talk because it sort of describes the behavior of random sequences but also because my own interest in the topic was completely random. So it does not have much practical relevance for me but still I think it's an interesting and beautiful topic worth talking about. So my name is Jayir Trejo. I work for Pink Orbeez, a small development shop in Mexico City. And I want to talk to you today about randomness in computers and in the Python standard library. So I like the English word random because besides its basic meanings of unpredictability and impartiality, it also has a connotation of spontaneity or suddenness. In fact, it lightly comes from all French words that mean things like speed or violence or impulsiveness. The Spanish word is azar. It comes from Arabic and it refers to an old dice game. So even now we call chans games, juegos de azar. So the mathematical term is very much related to the gambling meaning. So it is no coincidence that the first explorations of probability, the mathematical theory that measures analysis into point predicts random outcomes, has its roots on trying to understand gambling and what goes into predicting the outcome of gambling. One of the first examples of probabilistic analysis comes from a series of letters between Isaac Newton and Samuel Peppes, the president of the Royal Society, concerning some dice bits that Samuel Peppes was going to make. We think of the rolling of dice as a process with a random outcome. For a fair die, we hope that each of the six phases has an equal probability to come up when we're rolling it. So before throwing it, we don't really know what is it going to come up. And even when we do a series of rolls, the information about past outcomes of the dice does not give us any insight into what is going to come next. So if I roll the dice and the number four comes up, is four random number? Well, it certainly is a number chosen at random, but just by looking at it, we cannot know whether the process that produced it was actually random. So we can't really talk about the randomness of individual numbers, but about sequences of numbers. And sequences of random numbers have many applications in real-world situations. They are often used for reducing the size of a problem by sampling it at random points. This can be seen, for instance, in statistics, where you take a representative sample from a population, like when you pick people to call for an election poll, or in simulations where you want to approximate the probability for some event or property, we can randomly generate events, and then statistically measure the probability that we're looking for. These applications require the sequence of random numbers to be uniformly distributed. So this means that every number in a certain range needs to come up with roughly equal probability. Or otherwise, our result is going to exhibit those same biases. For instance, this sequence, it looks pretty random, reasonably uniform. So it can considerably be used in simulations as a source of numbers between zero and nine. In fact, if we try and take the average, we will see that it's reasonably around 4.5, that's what we will expect for such a short sequence. And every number comes up roughly with the same frequency. But random numbers also have important applications in cryptography. Many secure communication algorithms use random numbers for secret generation, so that only trusted parties will notice the secret random numbers. Like signage schemes also use random numbers for generating signatures in a way that doesn't reveal information about the key, even if you have a lot of signage messages. And for instance, in Django, there is a long random secret key that you need to use for every website that is used to sync sessions and encrypting them so that users or attackers cannot tamper with them. And there's a scandal about the NSA making a trapdoor in the random number generator used by RSA in many of their products. So apparently they can predict which random numbers the RSA products are picking, which has disaster security implications. So these cryptographic applications require more than just an unbiased sequence of numbers. They require the sequence of numbers to be actually unpredictable. An attacker that knows which random numbers I'm picking, or even has some insight on the word to look for the random numbers that I'm picking, has to open to all of my secrets. So the sequence we were talking about just before, it looks unpredictable, maybe, but it really isn't at all. So they are the first deeds of Pi, which of course is a completely fixed predictable sequence. If an attacker knew that I was using digits from Pi for generating random numbers, he will only have to compute Pi and he will know all of my future pickings. So keeping in mind these two requirements of impartiality and unpredictability, what can we use for getting suitable sequences of random numbers? One option is to use a natural phenomenon that we know to be unpredictable when measured with sufficient accuracy. For instance, the website random.org uses atmospheric noise. It measures it and extracts random numbers from it. In the UK, there is a machine called Ernie that uses transistor noise measurements to pick winners in a national lottery. And we can also use radioactive isotope decay, which we know is pain-intensive and predictable, and independently of the precision of our instruments or even the quality of our models. But what this case is having in common is that it is often slow, expensive, and require of specialized equipment to measure these sort of natural quantities to generate random numbers. So it might be useful to generate this large quantity of random numbers once and then compiling this into a table of random numbers that we can draw from in the future. As a matter of fact, in 1955, the Rand Corporation published a table of a million random digits obtained from specialized hardware. This enormous book came to be widely used in simulations for engineering and science. And of course, large numeric tables also have some disadvantages of their own, especially with the computers from back then. It is very hard to store and efficiently access such a large table, which led researchers to look into techniques for random number generation on the fly. So of course, computers are deterministic artifacts. So the future state of the machine is completely determined by the present state. So how can an algorithm actually generate random numbers? Well, it turns out that only as we incorporate input from outside devices, we can only generate pseudo random numbers. That is, random number generators, output numbers that look random when statistically measuring them. But they are not actually hard to predict if you have enough information about the state of the generator. In the 1940s, Jambon humans was doing simulation work that required a stream of random numbers. He came up with the idea of generating it by taking a random number, squaring it, and then taking the middle digits to produce next. The output of the generator looks reasonably random, but it is crucial to pick the right seed because, for instance, if we get a zero somewhere in there, that means that from then on the sequence is going to generate a zero. And it also has a tendency to fall into short loops, which of course, there is no way to get it out of. In fact, we can use different seeds and measure just how long does the generator run before starting to repeat numbers. And we can see that even for 40, the sequences are not very long. But if we take one of those long sequences and check the average value and other statistical properties, we can see that they look reasonably random. So is it possible to evaluate randomness more precisely? If we want to mathematically evaluate randomness, we need ways to formalize what is impartiality and its unpredictability aspects. What the formally measure unpredictability is to look at the entropy of our output, which is a measure of the space of possibilities that it can take. It is important to know that this cannot be immediately tell from looking at the numbers. You have to actually look into the process that generated them. For instance, if we see those numbers and I told you that I picked them at random, you might think that I picked them from zero to 100. But if I told you that they are actually prime numbers, then you will see that the actual space from which I draw them was much smaller than we thought. It is similar to how my bank asked me to pick a password, but it can only be like A characters long and I cannot use repeating patterns or consecutive numbers. So in general, they are reducing and reducing the space of possible passwords that I can pick. Although it might be worth it if it is subscriptable from using passwords as a ranking password. As for impartiality, we can look at the statistical properties of a random sequence and see if they are consistent with probabilistic predictions. When checking the randomness of the digits of pi, we used the very informal test, sort of a roll of thumb. We checked that the average of the values was what was to be expected in random sequence. We strengthened this test by looking at individual frequencies for different digits and see that they are roughly the same. But we have much elements to assess whether this is sufficiently right or disasterally wrong. We need something a little bit more quantitative. A much better evaluation is the G-squared test, which is used in statistics to see if a set of data confirms the certain distribution. The general idea is to measure the squares of the difference between real and expected values weighted by the expected value. And summing them, this gives us a measure of how much our observed frequencies deviates from what we will expect probabilistically in the real one. With this measure, we go to a table like this that gives us the likelihood of observing different values of this quantity. If it is too big, then we conclude that the sequence is too different or is too much deviated from what we will expect in a random sequence. But also, and different from the application in statistics, if it is too low, then the sequence is also suspect to be too uniform for being random. Other tests check the sequence for more complex patterns. For instance, in a random sequence, pairs of numbers need to be as uniformly distributed as numbers themselves. Or we can also check if there are gaps between successive appearances of the same number and whether the length of these gaps is consistent with what we will expect probabilistically. And there's also a number of other patterns that we can use. As a matter of fact, there are standard tests that can be used to check random sequences. There's one by the American NIST, which group a series of conventional mathematical tests that can evaluate a list of random numbers. On the other hand, there's some more exotic tests like the Marsaglia dihar battery tests that prove that tests random numbers in some quicker situations like the spacing of birdies in a random population or it tries to place circles and see which circles overlap in the plane. And many other random experiments that we know what values to expect. And we can check that against the performance of our random or supposedly random sequence. Taking into account these tests of randomness, better generators have been devised. One very popular is the linear congressional generator, which is our recurrence, where we take an initial value and use this equation to produce subsequent values. Of course, this is going to eventually repeat itself with a period no greater than m. But if we pick the right values for a, c, and m, we can get reasonably large sequences that exhibit very good statistical properties. The problem with this algorithm is that it's very easy to fall into this situation. Even if numbers look random when seen linearly, when you plot pairs of them, they sometimes exhibit this kind of behavior. They are all falling in the same straight lines. We can choose better a, c, and ms to get rid of this behavior, but it always ends up happening in higher dimensions. So how can we get rid of even these deviations from truly random behavior? Well the Merchant Twister is an algorithm proposed by Macondo Matsumoto and Takuyo Nishimura, which consists of a large linear feedback register. And it operates in a way that permits the sequence to have a very, very large period of 2 to the 19 or 20,000 potency, more or less. It is also interesting that this sequence has internal state, and it uses that internal state to produce the actual random numbers. So even if you know the random numbers themselves, you cannot predict immediately the next number in the sequence. You need a large sample of them. And if you actually measure the statistical properties of the sequence that it produces, they are very, very close to randomness, and they don't exhibit or this weird correlations in many dimensions up to 623 dimensions. So it is a very good random number generator. These desirable characteristics have made it a very popular generator. It is baked in in many languages. It is one of them. The Python random noodle uses Marcin Twister as its underlying default random number generator. There is also the question of how to get random numbers that are cryptographic as a queue. This obviously cannot be obtained from an algorithm because algorithms are deterministic. So they have to be gathered from system activity. And some other unique systems provide a source of random numbers in depth random that is feed by an entropy pool that derives randomness from various sources like keyboard inputs or the timing of mouse movements, noise in sound or network interfaces, et cetera. So when users need random numbers, they can get true random numbers from this pool. Of course, getting random numbers out of the pool sort of drink a bit of our entropy milkshake. So we need to replenish the pool with more entropy. Besides the regular sources for a consumer system like the keyboard or the mouse, modern computer systems often incorporate some form of hardware random number generation. Intel chips, the EVP Bridge family, dedicated random number generator in the hardware. So now we have finally arrived to the actual random module. The Python's random module starts from this generator of numbers between 0 and 1, uniformly distributed, and provides a lot of other interesting distributions based on that. So the way to use it is there is a class in the module, random, which can be seeded, and that provides a method random that is going to produce a sequence of numbers from 0 to 1. If we can use the seed if we are going to repeat or we need the same sequence, or we use the same sequence several times, if we don't, then we can just let Python seed it with a number get from the view random or from the number of seconds at the time of the call. From real random numbers between 0 and 1, it is very easy to get real or integer random numbers up to a certain number. We just multiply the random real by the maximum value. If you have a specific range, you have to generate a random number up to the width of the range and then offset it by the start. So it is still very easily derivated from the random real. And if you also need a certain step in this sequence of possible random numbers, you just generate an integer up to the number of steps and then offset it by the start and you have your random integer. So we can generate random reals, we can generate random integers. And if we need to include the whole interval of numbers, so we need a number between A and B, including A and B, we can use this special function, randint, that just calls randrange with the appropriate arguments. We can also pick or we can also perform some operations or random operations in a sequence. For instance, we might want to get a random element in the sequence, which is very easy to generate an integer in the range of indexes for the sequence and pick the element corresponding to that index. If we need a sample, we just repeat this process several times. If we want a sample without replacement, we need some form of tracking of which numbers we have already picked. There are two ways to do it. You can track which numbers you can still pick in a list and then remove from them every time you pick one. Or you can use a set to remember which numbers you have already picked, which is more efficient, depends on the size of the population compared to the size of the sample that you want to get. And the Python random module actually computes this on the fly and uses the more efficient method. You also might want to shuffle the list. The algorithm used by the random module is the feature jade shuffle, which just goes to the list and exchange every item with another one that is randomly picked. If we need a simpler way to do it, that gives us a new list, we can just sort by a random key. This is not as efficient, but it's much more simple. Now we may be interested in random real numbers that have another distribution other than the uniform one. How may we go about it? Let's consider the normal distribution. It is determined by two parameters, me and sigma. In this plot we can see that for each real number we can know the probability of picking it with this normal distribution. But we can also plot the probability of a random variable with this distribution falling before every real number. This is called the cumulative distribution function for the variable. We can see that it's always increasing. This means that to get a sequence of normally distributed random numbers, we can generate a uniform random number that we will use as a probability in this plot. And then we can check to what x does that probability corresponds. And the result of that selection is going to be normally distributed. This does not only apply to the normal distribution, but to any distribution for which we know the distribution function. But it is not always obvious or easy to generate the inverse cumulative distribution function just from looking at the distribution function. So there are many mathematical tricks that have been devised to ease these computations. For instance, for the normal variant distribution, we use sort of a mathematical trick where we pick two random numbers, use them to generate a point in a circle, and the x and y coordinates of that point end up being normally distributed. And from there, we can get a number of interesting distributions that we might know from science and engineering, like the triangular distribution, gamma and beta distributions, the Pareto distribution or the Weeble distribution that is very popular in engineering because it can be used to approximate the other ones. Another one of note is the von Meissner's distribution that is sort of like the normal distribution, but for angles in a circle. Because when we have angles, we can see that several angles may actually correspond to the same point in the circle. So the von Meissner's distribution is wrapped around the circle to consider the effect of these like double angles for every point. And finally, the random middle creates the following instance of the random class and provides the bounded methods as module methods. So you can just import random and if you don't care about the state of the generator, you can just use the module functions. If you need separate generators like for multi-threading applications or because you need two independent generators for different experiments, you can actually instance the class and cheat them individually. You can also subclass the random class to provide your own random number generator. The Python random middle comes with the Wigman hill for backwards compatibility reasons and as an example of how to provide your own random number generator. There's also system random that will get numbers from the system provided random number generator in Unix systems and there's even a library that will connect to the random.org server and use that as a source for random numbers. So if you need true random numbers, you can use this. And since all of the other methods rely only on the generator of real numbers from 0 to 1, they will still work even if changing the source of the actual numbers. So, concluding, the definition of randomness is more philosophical than a mathematical problem. But we can use mathematical definitions that are useful for our purposes. If you need sequences that are deterministic but behave as if random, we can use pseudo random number generation. But if we need numbers that are completely unpredictable, we need sources of entropy like input devices, noise measurements or other external natural phenomena. And for most of our random number needs, Python provides more than adequate capabilities. Finally, I would like to talk about a book that inspired this talk. This is a very good book that takes this very short basic program and uses literary criticism techniques to analyze it word by word. It sounds far-fetched, but it's actually a very interesting book. And the chapter of randomness is what got me interested into this very, very beautiful topic. The art of computer programming volume two, half of this book is about random numbers. It is very theoretical, but it's also very fun. Lots of really nice mathematics in there. And finally, you want to read a little bit more. There's a series of really good articles about randomness in cryptography by Clothfair that might help you understand why is randomness important in cryptography. There is a very good description. In the second link, there is a very good description of how statistical testing of random numbers work. And if you want to read more about the possible backdoor in the RSA's random number generator, RSA's technical article is also very good. So that will be it. Thank you very much. Thank you. Thank you. Thank you. Thank you.
|
Jair Trejo - Non Sequitur: An exploration of Python's random module An exploration of Python's random module for the curious programmer, this talk will give a little background in statistics and pseudorandom number generation, explain the properties of python's choice of pseudorandom generator and explore through visualizations the different distributions provided by the module. ----- # Audience Non mathematical people who wants a better understanding of Python's random module. # Objectives The audience will understand pseudorandom number generators, the properties of Python's Mersenne Twister and the differences and possible use cases between the distributions provided by the `random` module. # The talk I will start by talking about what randomness means and then about how we try to achieve it in computing through pseudorandom number generators (5 min.) I will give a brief overview of pseudorandom number generation techniques, show how their quality can be assessed and finally talk about Python's Mersenne Twister and why it is a fairly good choice. (10 min.) Finally I will talk about how from randomness we can build generators with interesting probability distributions. I'll compare through visualizations thos provided in Python's `random` module and show examples of when they can be useful in real-life. (10 min.)
|
10.5446/19982 (DOI)
|
Okay. So welcome back. I hope everybody had a nice lunch break. And now I'm very happy to introduce Jungle Core Developer, but more importantly for this talk, Employee of Elasticsearch, the company, Hansa Kraal. Hello. Hello. So I'd like to tell you a story. It's a story about how we developed five clients for Elasticsearch in five different languages without losing our minds in the process much. And so as any good story, it starts a long time ago in a galaxy, no, no, no. It starts actually when we looked at the current landscape of the clients for Elasticsearch. And there were some things that we liked and what we've seen are good and some things not so much. For example, in the Python landscape, there are many clients, but none of them actually implemented the entire set of APIs. None of them did everything that we would like to see in a client, and none of them did it on a scale that we would be comfortable with. As a result, users had inconsistent experience with Elasticsearch itself, and naturally they blamed Elasticsearch because of the way they interfaced with it was not ideal. So we decided to create our own clients, sort of to control the last mile how people talk to Elasticsearch so we can make sure that their experience is good and consistent. So we started with the design, obviously. And we sat down and said all the things that we want our clients to be. And for that, we need to start with Elasticsearch itself. Elasticsearch is distributed. That brings a lot of problems with it and a lot of opportunities. It talks via the rest API over HTTP, which is both good and bad because it's good that it can be deceptively easy to create your own clients. The number of clients just in Python or just in Ruby was staggering just because everyone thought that, oh, it's just HTTP, right? I can just do an HTTP request and everything will be good. But there are a lot of corner cases there that they didn't count with. There's also a lot of diverse deployments of Elasticsearch. Some people just deploy the cluster in their own networks and talk to it directly. Others would use load balancers. Some people would use alternate transport like thrift to gain some speed or would use a set of client nodes and a distributed multi-rack setup or something like that. And also just the set of endpoints that Elasticsearch has is quite staggering. It's almost 100 API endpoints with almost 700 parameters that the clients only need to support and document. But more than that, we wanted the clients to be true to their language. That's why we only developed the four or five in the beginning because those were the people that we had. We had a Python engineer, yours truly here. We had an excellent Ruby developer, a Perl developer. We even managed to find a really great PHP guy, believe it or not. So those were the clients that we started with because we felt confident that we can actually make it feel like a Python library, not like a library written by a Java guy in his spare time in Python. And we wanted the client to be for everyone. We wanted people not to have any excuse to use it. So that's my first lesson that we learned. No opinions, no decisions. In order to make sure that everyone would use the client, we had to abstain from any making any observations, any decisions because whenever you have an opinion, there is someone out there who would disagree with you. So the only way how to make sure that that won't happen is not to have any opinions. So we decided the client should be low level, just essentially a one-to-one mapping to the rest layer. And they should be extensible. We should design them in such a way that where you don't like some aspect of it, you should be able to replace it or just hook into it and change it. So we came up with this. This is a foreign for an HTTP client. This is a kind of complicated diagram that specifies how the client works. You have the client itself that has a transport class, which has a serializer to serialize and deserialize data as they go over the wire. Then you have a connection pool that actually stores a list of connections. Connection pool in this case is a misnomer because it actually doesn't pool collections. It just holds a collection of connections to individual nodes in the cluster. You can see why we would have the naming problem there. And then we have connection. By default, we use yourLab 3 because that ended up to be the best one for Python. And we also have a connection selector. So when you connect to multiple nodes, they control the strategy on how do you do load balancing. Do you use random or do you use round robin? By default, we do round robin over randomized lists of nodes. And the goal why we did it this way is so that we are able to give you the option of override any simple component in here just by subclassing the default implementation and filling in all the blanks. So some examples. If you want to create your own selector, you just create the class and you pass it in. Everything is essentially using dependency injection. So you can just pass it in as a constructor parameter. And we will use that instead. So you see three examples here. The first one is not really injecting your own code. It's just setting up the options. So the first one instructs the client to talk to the cluster and get the current list of nodes on startup. Then whenever a node fails, and then also every 60 seconds. This is excellent for a long running process, let's say a web server. So that even when you keep changing your elastic search cluster, you keep adding nodes and nodes keep dropping out, you still talk to all the nodes that are available. The second one is where you want to control the load balancing. For example, imagine a scenario where you have two racks and you want to by default only talk to the elastic search nodes in the same rack as the application server and only fall back to the nodes in the other rack if none of those are available. You can do that. You can just write a simple class that will do this. So that's what we mean when we say that we are modular and extensible. The last example is just using a thrift connection which we actually provide as an optional plug-in and using a different serializer, in this case YAML. Because why not? Some people like YAML better than JSON for some reason. So that was sort of the first lesson. No opinions, so people have no excuses not to use it. The second lesson was to prototype everything. Because you just don't come up with something like this without some preparation, without some prototyping. And more importantly, you don't come up with it for a single language and have it be applicable for all the others. It's very difficult to find a pattern that would work for both Python people and Ruby people, for example. That's why we created a prototype implementation in both Python and Ruby to make sure that the design will work, that the design will hold for both of these languages. And also so that we have a reference that we can talk about, that we can have the same terminology and then when we talk about connection pool, we know what we mean even though it means different things in different languages and even how we use it is not exactly correct. But we had code that actually showed what it does, so we could have at that point a clear conversation even with PHP and PURL people, even with the JavaScript people that came on and even later with.NET people who are developing the new client now. So prototype everything, not just to see if your design works, but also that you have something to talk about that you are absolutely certain that you're on the same page. Because you should never trust humans and just their understanding if you can do more. So that's the next lesson. The next lesson is don't send a man to do a machine's job. Humans are amazing. They are amazing in a lot of things. Consistency not one of them. Consistency and repetitive tasks, you really want to have something that doesn't get tired, that doesn't get frustrated, that doesn't mind doing the same thing over and over again. To me that sounds like a computer. So this lesson states that you should automate as much as possible. And this why I'm talking about this is I already mentioned we have almost 100 API endpoints with 700 parameters. That's very difficult to track. That's a lot of work, a lot of boring tedious work that you don't really want to be doing. You don't hire a decent Python being and force them to maintain a list of 100 APIs and 700 parameters if there is any other way. So what other way is there? So first we thought that we could do a reference implementation. We could just arbitrarily choose one of the clients and decide this is how it should. This is the reference implementation for our APIs. This is the authoritative collection of all the APIs, all the parameters, their possible values and descriptions. But that doesn't really scale that well. First of all we only have one person per language. We only have one Python developer. We only have one Perl guy. We only have one PHP guy. So what if he leaves, what if he's on vacation and there is a change that needs to be made? And also how does that person make sure that everything is synchronous? We found out even with the spike implementation of the transport layer that maintaining it when we add more features and we need to add it to both Python and Ruby even though it was just two of us and co-incidentally the two of us that lived in the same city, which is not true for any other two people on the project, it was very difficult. It was difficult to keep in sync. So we discarded reference implementation as an option. Next we looked at documentation because obviously elastic search has documentation and all the APIs are documented. All the parameters are documented as well. But again, they're documented for humans. It's a documentation that's intended for the developers to read it and to understand it, to make sense of it. So again, it would require a tremendous manual labor just to make sure that everything that we need is there. Someone would have to read actually all the documentation, collect all the stuff and not just one person, but each and every author of the client would have to do that. That's a very tedious job, a job that I haven't signed up for and I doubt we would ever find a person who would sign up for a job like that. So what other options were there? We found the progress from reference implementation to documentation had some promise, but it wasn't there yet. So we decided to take it one step further, to actually extract all the information that's already in the documentation, that's already in the code and present it in a structured format. So we chose a format that's human readable and machine parsable. And we at elastic search, we really love JSON. So we just decided that we should document everything in JSON and create a spec of former specification for our APIs. This is the one case where I was super happy that elastic search is written in Java as a statically typed language because it provides you with a bunch of tools so we were actually able to write a tool that would just parse the source code for all the APIs and extract 90, 80% of all the APIs and its parameters in an automated fashion. And we then just had to go once over it and we could actually share in this effort all of the client people and just fill in the caps, fill in the documentation for each of the parameter, fill in the option and the type whether this option is required or not, whether it's a list or a single value, a Boolean or an integer. So that made the effort so much easier and going forward also so much easier to define. So what did we choose to capture in this document? First of all, the URL path, all the different variants of URL path if it was dynamic, which most of the URLs in elastic search are. We can optionally include an index name or a list of indices on which to perform the action. It can include other dynamic parts. So we have to document those including all the different options how the URL can look. As part of it also, we had to do the HTTP methods. So is it a get or post? We decided to do very little than that just to list all the parameters, list all the ways how to combine them into URL. We didn't actually capture all the dependencies that this parameter is only valid if this parameter is set to blah. We just found out that we don't really want to have this information, this validation in the client, and we'll instead decide on our users to use it directly. So that way we would have less overhead with maintaining it, and the code would be much simpler. So how does it look? This is an example for the Suggest API. This is just a fraction of it. And you can see that we have a link to the documentation. We have all the possible HTTP methods. So in this case, post or get. We have all the different forms of the URL paths with the description for each part. So here there is only one dynamic part, which is the optional index. And we have description of all the parameters. We also have a description of the body, what it contains, and information, whether it's required or not. And this is all the information I need to write, or in my case to pre-generate, and a Python method. I have the name, I have the list of parameters, I know which ones are required and which ones are optional, so I can actually choose that these ones will be positional, these ones will be keyword. And I have the way how to actually put all those information together to create a URL and send it over to the server. The last nice thing about this is it minimizes the effort to maintain it, because we stuck it into the same repository as Elasticsearch code bases, the same repository where the documentation is. And that meant that updating this just meant that whenever I make a change in Elasticsearch itself, whether I add a new API or just add a parameter, inside the same commit or pull request. It also provides changes towards the specification. And then all the client people, all they need to do is just monitor this one directory on GitHub to see all the changes that they need to implement in their client. But again, we approach the difficult thing, like people need to watch something and do something, and whenever you rely on people, you will get into trouble sooner or later. But you will probably get into trouble. So that brings us to our last lesson. Test everything. Don't trust. Just verify everything. In our case, we needed to verify that all the clients are consistent and that they work well with the server. So again, we created our own solution and we created a Unify test suite. We again took a machine-parsable language, in this case YAML, and created a simple test suite with a setup and a bunch of actions and a bunch of assertions that enabled the code to run not only against Elasticsearch itself, but also against all the clients. So this is how it looks. So this is, again, a test for the Suggest API. And you can see there is a setup that will actually call an action called index with the parameter of index type ID and body. And it will then do a refresh. So it will make the document available for search. And then there is one test, the basic test for Suggest API, and it will actually perform the Suggest operation and then run to assertions. So this test validates that the Suggest API is actually capable of correcting our typos. And this is a test that is run as part of the Java test suite, as part of the integration tests. And so that means that it's version-specific. It's in the same code base. So whenever you have a branch of Elasticsearch, it can have its own test suite, just like any other tests. And also all the clients have an interpreter for these tests that makes us sure that we have the same naming, we have the same API coverage, we have the same exception handling because we can know just one assertions, but we can also run assertions that this should fail with this error code. And just by specifying it once, we can make sure that all the clients are consistent. This set of tools together led to that when we decided to develop the fifth client after the original four, it only took a few weeks for the JavaScript to write an interpreter for the test suite and the parser for the API specification and make sure that everything is working as it's supposed to. So these were my lessons that we learned during this process. It was good times, it was bad times, but we made it through. And I believe that the clients are working well for people, and now we're sort of approaching the next stage, and that is to actually create a more high-level opinionated clients that will be more helpful to our end users. But also for those clients, we are OK with people not using them. So thank you very much, and if you have any questions, I'll be happy to hear them. Thank you. So questions? Hi, thanks for the talk. Did you try to think about if it was possible to generate the clients completely, since you already have everything in theory? You have the parameters, you have the return value. So some clients actually do that. For example, the JavaScript client, it doesn't actually contain much source code, it just actually internalizes the JSON specification and it generates the method on the fly. For Python, I actually wrote a script to generate the entire client, and I used it as the first draft, and then I edited it manually because it's great to automate everything, but usually it's OK to automate just 90% and don't try to catch it all and just do the 10% manually. This is the classic 80-20 problem. So I started with the generated code and then I filled in all the exceptions to the rule. So now when there is a change, I run the generation process again and I manually look at the diff and see what parts represent an actual change and what was just a manual edit that I had to do in order for the API to actually fear more like Python. Have you been trying or considering to use the protocol buffers from Google? So yes, we have actually considered that as an alternative transport. Currently we are fine with just HTTP and JSON, though we provide some alternatives. There is an experimental transport with a memcache protocol and Redis protocol and Thrift. We haven't looked that much longer on protobuf because the trade-off didn't seem to be that huge to warrant that investment. However, we are still looking for more effective transports and encoding schemas, so it might still happen. This definitely is something that you can implement yourself in a plugin for both the clients and the server. Questions? No more questions? Okay, then that's it. Thank you again. Thank you very much. Thank you.
|
Honza Král - Lessons learned from building Elasticsearch client Lessons learned when building a client for a fully distributed system and trying to minimize context-switching pains when using multiple languages. ----- Last year we decided to create official clients for the most popular languages, Python included. Some of the goals were: * support the complete API of elasticsearch including all parameters * provide a 1-to-1 mapping to the rest API to avoid having opinions and provide a familiar interface to our users consistent across languages and evironments * degrade gracefully when the es cluster is changing (nodes dropping out or being added) * flexibility - allow users to customize and extend the clients easily to suit their, potentially unique, environment In this talk I would like to take you through the process of designing said client, the challenges we faced and the solutions we picked. Amongst other things I will touch on the difference between languages (and their respective communities), the architecture of the client itself, mapping out the API and making sure it stays up to date and integrating with existing tools.
|
10.5446/19981 (DOI)
|
Okay, so like Fabio mentioned already, I've been doing a number of tools in testing, and then after some time I decided, okay, there's this whole thing about unit tests, and py tests, and nose tests, and whatnot. And actually it would be nice to have a really unifying experience when running tests against the Python application. So that's why I went for also writing talks, which is kind of like a meta test runner, and that can actually invoke nose tests, unit tests, or py tests. And after a while I thought, yeah, that's all very nice, but the real problems when you want to have something like quality assurance in your projects is really also about release management. So you actually have several packages, dependencies, and I have that with my own open source projects, but also with people and companies I consult for. And that's why I also went for basically the next level to have something that manages the packages, then also get tested. But all the time coming very much from this kind of like QA and testing perspective. So that's when Defpy actually was born. The Defpy system is basically there to help you with PyPy-related release workflows and quality assurance. It currently consists in version 2.0 of three main components, which is the core Defpy server. And I'm going to talk about all of these components in detail. The server that actually provides the PyPy caching index and your private indexes, where you might not actually want to publish from, but you actually want to use that within your organization. The recently released is the Defpy web plugin, which provides web interfaces also for your documentation, a few other things and search across metadata and documentation. And then there's the third thing that you don't have to use, actually, but it's helpful if you have to deal with development and production indexes and so on. And that's a command line tool that basically drives the well-known other tools like pip and easy install and setup.py upload and things like this. So Defpy served indexes. One of the main purposes at the beginning that was before PyPy from org grew content delivery network was that you can have a local self-updating PyPy cache. So you basically work against your local index. If a package is not there, it goes off to PyPy from org, grabs it, and the next time you don't even need to be online. You don't even need to have online connectivity. It will just satisfy everything completely offline from your local cache. So everything that you install basically gets cached, including the index information, and it uses the change leg protocol with PyPy so that from time to time it asks PyPy, is there anything new for the projects I care for? If so, it basically invalidates the cache, so the next time you ask, it's going to update the cache. As with every cache, cache invalidation is a very important topic, and this is actually using the official PEP 381 API. It also manages multiple private indexes for you if you want to implement staging, and each of these indexes supports running against it with PEP or easy install or build out, and it supports the typical setup Py upload, upload docs, and so on commands, how you can then get packages into DevPy. There's one feature that distinguishes DevPy from other indexes that you may know in that it provides an aggregation or inheritance feature. So here, this is one possible layout that some people use. You have the so-called root PyPy, that's the cache I talked about. You can directly use that if you don't care for private indexes and forget about the rest, but here we actually have a production index which contains the private indexes, the private packages that you don't want to publish on PyPy from org, which might depend on PyPI release files that you don't have in your private index. So you may have a web application that depends on Pyramid, and Pyramid depends on lots of other things, and those all come from root PyPy, but if you work against the company production index, you're going to see one unified view of your private packages and all of the PyPI, Py from org packages. And then if you want to do some kind of QA workflow, you also can do a development index, for example, team-based, that's what some companies are doing, and there you just put your in development releases that are not ready to be deployed on your web servers maybe, but they can be used for further testing. And one important thing here is that your production index is actually somewhat protected from malicious PyPI packages, and I'm going to tell this, which is also interesting if you don't use DevPI, something which I call the higher version attack, there's also variants of this attack. Let's say you have a credit card release file that contains your credit card processing in your web application. You put this on a private index, and somebody, that's the attacker actually, uploads credit cards with a slightly higher version number to PyPI. Now if I install against the production index that inherits from root PyPI with this install command, I'm actually going to get the PyPI version, because I didn't know that somebody actually went and occupied my private name on PyPI. PyPI is a package wiki. Anybody can basically publish any kind of package. So if you have private package names that are not yet registered at PyPI, somebody can go there and do that. It's very easy. And the... I don't know, I didn't try myself. But I'm pretty sure I could get, I guess, something like 100 bots per day or so with something like this. That's not the only problem that is there, but I'm just saying that if you have something that somehow merges the world of the PyPI Py from orc wiki with your private indexes, then you get into this kind of problem. And that's also the case, actually, if you forget about the FPServer, also the problem if you use pip install extraindex URL. Because then the merging is actually done on the client side, but it does exactly that. It actually takes the higher version. So you end up, you thought you install something from your private index, but you're actually installing something from PyPI. So that's a bit of a problem. And PyPI in version 2.0 prevents that because it says by default, if you upload anything to a DefPI private index, any kind of further look up, even if you inherit from the PyPI cache, will be prohibited. And you have to whitelist it. If you actually have a package that comes from PyPI Py from orc because it's an open source release of your company, then you have to whitelist it. Otherwise, all PyPI is ignored, basically. If you basically install from the production index credit cards, and it's not whitelisted, so by default, PyPI is not considered because there is a package in your private index. So it's basically trying to prevent this kind of error. That's not the only way if you want to be a bit more careful because there's other attacks. For example, if you have typos, somebody in your company on the laptops installing Pyramid without a D at the end, or what I do sometimes, pip install PyTest. So if you want to get hold of my machine, it's very easy because you just need to register the package PyTest without the T. For some reason, I sometimes forget this last letter. It's not currently registered, so it's a good chance. You get my machine. So if you actually want to, this is really a problem because, I mean, you can imagine there's some very popular packages, if you register variants of this kind of package names, you will eventually, from the millions of users, literally across the world, you will get some people, actually. I checked with the PyPy admins. There are actually, you can see that in the server logs, there are actually a lot of instances of mistyped things, so it's clear you can actually exploit that. Okay, but this is not about attack vectors against PyPI. Would be a fun talk by itself. This is about, if you want to be more careful, then you probably should not inherit directly, but you rather have root PyPI as the self-updating cache, and you work with that in development, but then when you want to have a package in your, including dependencies in your company, you actually push it explicitly into your production index. And sorry, in your development index, right? And then basically you just push packages around the indices, and that's something that DevPy makes easier, or somewhat easier, and you upload your own packages to company Dev, and you won't have any kind of these attack problems, like typos and so on. Suddenly, if people, your production machines cannot be easily compromised. Okay, this is just some background on how you can organize and what you might want to be careful about regarding indexes. The way how you can organize indexes for your teams, and also maybe platform-specific indexes that contain wheels for your deployment platforms and so on, there's several variants about this and kind of best practices in merging, which are not yet documented, but this is kind of a start on this. So one feature that came out last week actually is replication, because that's what one funding company who actually gave some money for development, for the open source development wanted to have, is that you can now run DevPy server in replication mode. That means the first command actually starts the server, it's the full command that you run on port 3000, and then you start replica somewhere else. In this case, I just run it also on localhost, I specify that the server state goes into a separate directory, the replica one, and then I say, okay, my master actually is this. So the second invocation actually starts a replication instance, and this works by HTTP between the replica and master, and it maintains a full failover copy, so that when you actually upload something to the master, you can also upload something to the replica, it has the full interface, and that will only complete if the package is also at the master. So at any point in time where you upload something, you will have it at least on two hosts. And all rights, it's kind of like a simplified replication model, always go through the master, and that kind of seems to work quite well already, although there might be some buckets out loud last week, I've been running it myself in instances, and now some companies are starting to use the replication also in their settings. The defpy web is the second big feature that came out last week, mostly implemented from Florian, where is he? There. We have a refactor defpy to use pyramid everywhere, and defpy web actually is a very nice web interface now that shows you metadata and summary information, description and documentation, so it's your basically read the docs in the company, basically server as well. And maybe I show that quickly, so this is my semi public instance, this is like for example my development index, and one of the things you see that for example the defpy server 201 release we did, that's the release file, and here you see tests that were performed on the various, on the truth platforms here, RIN32 and Linux on the different interpreters, and I can basically look into that and see that this was executed, and the same way of course I would see if there's a failure somewhere. Also if I have documentation, I can go in here, or I can just say show me, okay what do you know about defpy and Jenkins, and that's a full index, a full defpy server search, and then I see okay, there's some links to that, and I get to the integration part with Jenkins on the defpy documentation, and that is just there because I uploaded the documentation to the index, it gets unpacked, you get URLs for that, and it's indexed in the search. So that's also quite powerful facility. So the last component is defpy client, it's a relatively thin wrapper around PIP and some setuppy invocations, it also performs the actual upload, so it always uses SSL and some other bits, and it maintains on your local machine any kind of login information, so you basically say okay I'll log in, and then I use a certain index, I upload something, and then I don't need to re-log in all the time because that token I get from the server is going to be valid for 10 hours, and defpy client basically stores this temporary authentication information. It also has experimental support now for SSL client certifications, if you want to step up your scenario to have encryption and authentication through SSL. The commands that defpy client offers are used to actually set the index you want to work on development or just root PIPI or your production server, upload is for helping you with the uploading files and docs and so on from a checkout, test is the one that produces the test, it invokes docs actually, and push is the operation that actually pushes a release including all of its documentation and release files from one index to the other, and PIP or other installers you can just directly use. Then there's some configuration administration commands that you can use for index configuration, user configuration, and also accessing the JSON interface, so defpy server has a full JSON interface on all of the resources that you can use for scripting. Typical release workflow looks like this, you basically go to your development index, you upload a release file, either you implicitly build because you are in the setup.py directory, you just implicitly build with defpy upload, or you already have built your release file then just say defpy upload this release file, and you send it off to the index, and then from the same machine or from all kinds of other machines that you might manage with Jenkins or something, you issue this single line defpy test package name, and that actually gets the latest release and performs the tests and attaches the test results back to the release file. That's why I could see in this web view, okay, this release file, what kind of tests has it seen. That was produced by this client-side defpy test command, and when it's ready actually, when you're happy, then you push it to another index. Of course, you can also automate this kind of like Jenkins job and just invoke these commands to on success of something posted to an index that says these are all the tests passing packages and things like this. So this is a release file working that gets slide shortly into TOX. TOX is a tool that allows to define how you want to, what kind of tests you want to do against your release file. It's basically in the release file it expects to find a TOX.ion. And then it invokes TOX. I have the next slide discusses what that means. It produces something called TOXresult.json. And then I can actually, from the command line, I can say defpy list the package name and see what the status is. If it was tests passed or what kind of test failures there were and show me the traceback from the command line. And then I take the release file once I'm happy with it. This is then pushed bit by bit verbatim to the next index. So I know that this thing I actually tested against on the different platforms actually works and I put this thing, I don't basically re-upload something to production. I really take the same thing that works and push it through to the next stage. TOX for automating test runs. It's kind of a standardized testing. I'm not going to talk much about this because my slot was exchanged for a 30 minute talk. It was originally a 45 minute talk. Was scheduled wrongly so I can't talk too much about it here. But you can go to the web page to actually get some more information about how you configure your test runs with different test runners. The server you already saw that you basically just installed the server, you have the typical host port and some other settings that you can and the data idea where you want to have your server state. And then from different clients that don't need to install the server, of course, you can just install the client and then say, Defpy use my company server and just work against that. What you usually want to do is that you want to have an engine X base deployment. There's an example file that gets generated from your settings, host port and so on, which is basically more or less complete engine X or basic engine X site config file that you can just include in your engine X configuration or use as a template to work further from. And this actually happens in such a way that engine X directly serves the static files. So, some things actually, the service doesn't see anymore. Once you upload something, the whole URL structure is such that the engine X directly serves that file. So, for that, the service doesn't even need to be running. So I'm going to conclude the DevPy systems developed since about a bit more than a year, I think, a year and a couple of months. It's MIT licensed, it's test driven development a lot, surprisingly. And also it's a bit funding driven. So there's some users cases that are interesting to me, myself personally, but it also depends, I mean, one of the upcoming things maybe is a company who funds some LDAP integration, authentication integration, but kind of like feature development and some things and consulting is provided by Florent and me and of course, pull requests are a good way to contribute. Okay, that's my brief overview of our DevPy on talks. Thank you. Okay, we have a good five minutes of questions. Thank you. You just briefly talked about LDAP authentication. Does that mean that you can integrate DevPy into an active directory domain and use this information to authenticate users? Well, if the funding realizes, I guess so. Okay. Then I ask my employee if he can give you some money. I'm sorry? Yes, I mean a sprint or something like this is also possible, but even the sprint, I mean, you know, takes some time and organization and in order to get something release ready and documented and everything, I mean, you probably know that it's kind of some work involved, right? But just to give you a brief idea on how the feature discussion around LDAP is currently such that we say we want to have, basically, we want to have EngineX deal with LDAP server integration and just pass a certain username header and group header into DevPy server and basically have an optional DevPy server that just says, okay, my upstream EngineX is going to pass me the right thing and EngineX does the integration because there's nice plug-ins for EngineX that actually do this. And then we need some client-side support to handle the login part, but that's kind of like the current implementation plan. The alternative, obviously, is to actually have direct LDAP support in DevPy server itself, but, well, we don't have to reinvent every reel, I guess. Yes? Hi. Thanks for all this hard work you've done. And the question is about testing run by DevPy server. In particular, is it possible to configure some workers which are remote to the server itself because it's a bit kind of overload for the server? Yes. I mean, maybe I wasn't clear enough. The server and the running of the tests, for example, they are completely separated. So where you issue DevPy test is completely separate from where the server runs. The DevPy test command goes to the server and gets the files, performs the testing on whatever host, and then attaches back the test result. So on the DevPy server instance itself where the server runs, there's nothing. There's no setup.py or anything ever executing. Otherwise, it would be bogged by, I mean, if you have to execute something like setup.py, you basically run risk of compromise. Yes? No, no, the pushing is really after you test that. You test, like what you saw in this. The upload. Well, the upload you also do on the client machine. I mean, the client machine does the building and like you do a Vio, for example, for Linux, Ubuntu, 14.4, 64-bit, blah, blah, UCS2, whatever your platform is. And then you actually upload the resulting file to maybe a platform-specific real index. No, it doesn't. No, it doesn't. Although there is an upload trigger. So if you upload, you can define on a per index basis. I mean, I didn't talk about all the features. You can per index, actually, if you upload something, it can, for example, trigger a Jenkins job. That's kind of like one path that is documented. I showed it. You just go to the documentation and then the MISC section about the Jenkins integration. Okay. You already answered my question. I already have about a plugin system or signaling stuff like this Jenkins plugin. Is it already as generic as I can maybe generate Debian files from this upload trigger? No. I mean, DevPy tries to solve a few problems, but only those. It's not yet something like, it doesn't have like all kinds of events. It has this upload trigger for Jenkins, but not a generic web hook, whatever. So I mean, that's not very hard to do, but it's basically very much, DevPy is very much driven by actual real world use cases, not by all the features I can possibly think of or so. So when somebody actually comes along and wants to have a certain feature and discusses the use case, it's much more likely that it gets implemented. That's kind of like my general development approach these days. Okay. One more. Okay. Then that's it. Thank you very much. Thank you. Thank you.
|
holger krekel - packaging and testing with devpi and tox This talk discusses good ways to organise packaging and testing for Python projects. It walks through a per-company and an open source scenario and explains how to best use the "devpi-server" and "tox" for making sure you are delivering good and well tested and documented packages. As time permits, we also discuss in-development features such as real-time mirroring and search. ----- The talk discusses the following tools: - devpi-server for running an in-house or per-laptop python package server - inheritance between package indexes and from pypi.python.org public packages - the "devpi" client tool for uploading docs and running tests - running of tests through tox - summary view with two work flows: open source releases and in-house per-company developments - roadmap and in-development features of devpi and tox
|
10.5446/19976 (DOI)
|
Our next talk is about graph databases. Welcome Francesco Fernandez Castanio. Hi, my name is Francesco Fernandez. I'm from Madrid in Spain. I work as a software engineer in Beacote. I also run the CC++ user group there in Madrid and also Neo4j user group. Today I'm going to talk about graph databases, a little connected tool. Let's start by the beginning. There's a lot of people talking about NoSQL, big data, why relational databases don't scale, but these kind of databases are based on graph theory. Graph theory is a bit old topic. Let me introduce you this guy. Probably you will know him. He's a scholar. He was a mathematician from the 18th century. He's the guilty of the graph theory. He developed a lot of mathematical stuff, also the graph theory. He has a lot of time to think and question things to himself. He used to live in Brassia in Coningsburg. I think that I've pronounced it well. He asked himself, okay, the old town of Coningsburg has seven bridges. Can you take a walk through the town, visiting each part of the town, and crossing each bridge only once? Does somebody know the answer? Well, the answer is no. But this is not the interesting part of this question. With this problem, he started to develop the graph theory. Thank you to his work. We have these kind of algorithm, these graph databases, and everything. He ended up defining a graph in this form. It's a very concise form. A graph is just another set of vertices and edges that connect that vertices. I have to read it. It sounds scary, but we are used to dealing with graphs every day. Even my mom is used to dealing with graphs. Here we have an example. Here we have a map from the Manhattan Underground. In one place, we have the stations that are our nodes. The connection between the station and the relationships or the edge of our graph. Most of you have come here to Berlin, and you probably have run some graph algorithm to find how to come here to Alexanderplatz. I am in this place. How can I go to Alexanderplatz? Probably it's not the best, the shortest path, but you have found a solution. What is a graph database? Does somebody know what is a graph database? Any idea? It's a very simple concept. It's just a database that uses a graph as a main data structure. Today I want to talk about Neo4j, and Neo4j implements a property graph. What is a property graph? Here we have the definition of a property graph in a form of a graph. A property graph stores nodes and also relationships. These relationships connect our nodes, and both of them could have properties. What are properties? Just a pair of key values. As I told you, today I'm going to talk about Neo4j. Neo4j is a graph database. It's written in Java. Sorry, it's not Python. It provides some exit transactions, a REST interface, a Cypher language that is a declarative language to query the database. It's open source and it's a no SQL database. Probably you are questioning yourself, why should I care about graph databases? I usually work with MongoDB or probably Postgres, MySQL, and everything is okay. Why should I learn a new technology? I think that probably there are a main reason to take care about these technologies. I think that the traditional way, when I'm working with relational databases, if we're dealing with highly connected data, this approach is a bit artificial, because relational databases weren't designed to deal with connected data. Probably we have some problems, because we have to deal with some meta information. We have to deal with foreign keys. If we are working with many to many relationships, we even have to create a new table to hold this meta information. We have to take care that this information is consistent. I think that we are mixing our data with our metadata in the relational case. If we are working with a documental database, we have the same problem. If we want to work with connected data, the scenario is even worse, I think. We have to run some Hadoop process or whatever to get some information. We cannot get insight in real time. We probably face some scalability problems in highly connected domains. Probably we will have some problems of performance. Some guys, the Neo4j in action authors run an experiment. They wanted to compare the performance between my SQL and Neo4j in a highly connected environment. They ran this experiment. They modeled a domain, a social network with users that follows between them. I think that they store a million of users and a lot of relationships between them. They wanted to know, give me the friend of my friends until at the depth of five. Here is the table. We can see that there is, at the first level, the times are similar. But when we go deeper, the times are far away from my SQL. It takes a long time to finish. Why? Why this happens? Probably we will design our relational database in that shape. We will have our user table and then many to many relationships, that this is the relationship between users in another table. Each time that we are looking for the friends of one user, we have to look in this table. It's an index lookup. It has a complexity of log of n because we are looking for an index. When we are working with our database, they are designed to get the neighborhoods for free. They are stored in a shape that we get in a constant order of complexity. What happens when we go deeper? In our relational environment, we get this complexity because per each depth that we have, we have to look into our table. We have to have an index lookup. It is multiplied by the depth of our lookup. When we are working with graph databases, we end up with this complexity because we only have to traverse our graph. The other reason to think about using graph databases could be that we can transfer our domain model in a natural way. When I face a problem, I usually grab a paper and a pen, and I finally end up with this kind of drawings. I have some entities that are related to each other. The relationship has some semantics. This is some kind of UML diagram. If we are using a graph database, we can translate this to our storage directly. We don't have to take care about normalizing my model and blah, blah, blah. This kind of thing that we have to do when we are working with relational databases. Probably using a graph database for a, I don't know, storing documents is not the best solution, but for other scenarios, it could be rational. What are the use cases for graph databases? For example, we have social networks, the well-known use case. Someone follows. This is the model of Twitter, for example. Then we have other use cases. For example, just partial problems. We go from point A to B. This is a classic algorithm that is used in graphs. For detecting fraud, authorization, network management, to build recommendation systems in real time, and there are a lot of other use cases. Okay. Now I will start talking about Neo4j. I will introduce you to Cypher. Cypher is a declarative language. It's ask oriented. In some way, we translate our what we are representing to ASCII code as drawings. You will see better in later slides. We look for patterns. Neo4j gives us these layers to access to the APIs. On the top of it, this is Cypher. Then we can access to other APIs, traversal API. We have to write using some JVM language to access to these APIs. We can use jython if we want to. Okay. What is the simplest thing that we can represent using Cypher? This thing. This is related to another one. A is related to B. On the top we see a drawing. Below we have the Cypher representation. The translation is very straightforward as far as you can see. Okay. Then we can represent other things. For example, here I'm telling that Eric Clapton playing cream. We have one node that is Eric Clapton, and we have cream that is a band. We have a relationship with some semantics. We have two entities using a graph. Then we have our example of a social network. We have some users. In Neo4j, we can label our nodes because probably we want to categorize our nodes. Here I'm saying, okay, I have some users and they are related. They follow each other. Then I can also add properties to my nodes, to my relationships. Here I'm representing that Eric Clapton has some properties. In that case, a name that is Eric Clapton, and also the relationship has a property. That is a date when he started to play in that band. Here I'm trying to represent what bands, musicians that play in bands and styles that these bands are labeled. What is the simplest thing that I can query to Cypher? This thing, I am asking to Neo4j, give me all the nodes that are related with this relationship, with a relationship that is labeled with playing. It will give me all the nodes that are related with this relationship, and it returns all the nodes. I can look for other things. Here I'm asking to Neo4j, okay, give me all the nodes that are related with playing, and also in the other side are related with labeled. Basically, it will return me all the nodes that a musician plays in a band and the style of this band. It returns some properties. Okay. But we can look for some particular nodes. Here I'm asking to Neo4j, look me in your index a node that has a property name with a value Clapton. We will have an starting point. We have the node with this value that represents Clapton, and I want to know all the bands in that very Clapton played and the style of this band. This is the goal of this query, and I return some properties of this node. In that case, I get the name of very Clapton, the name of the band, and the style of this band. Okay. Then I can look for more patterns here. Here I'm saying to Neo4j, okay, find a node with an name Eric Clapton again, and give me all the bands that have the style Blues, and looking for two nodes in that case. I'm asking to Neo4j, look me for the node with this property name Clapton, and also this node with this property Blues. I look for the bands that have these properties, have this relationship, and I return order by some field, and by some, okay. We also can have optionality in our relationship here. We have evolved for the model, and we also have the relationship between a musician and a band, and musician can produce also bands. Here we are looking for all the bands that Clapton play or produce, and we are filtering by some date. As you can see, at some point it's similar to SQL. Also we can have an optional depth. Here I'm saying to Neo4j, okay, look me for all the nodes that are related with this property at a maximum depth of five. So he will look for me, and he will give me A1, A2, A3, A4, A5. All the paths, if they are paths until depth of five, he will give me all the nodes, okay. Here we have a more developed example. It's a just partial problem, and my goal is going from a metro station in Madrid to another. So I look for a station, I am in Sol, and I want to go to Retiro. Okay, so I look for these two nodes. I ask Neo4j that find for me these two nodes, and then I find all the connections, all the paths that exist between these two stations. Okay, so probably I have one, two, three, or four, I don't know. And paths to that connect Sol with Retiro. And then I reduce, I add all the weights between all the stations that is composed of paths, and I get the shortest path. Just notice that Neo4j has implemented all these kind of graph algorithms. It provides a shorted path, distra, A star, all of these kind of graph algorithms are implemented in Neo4j. This was just an example. As I told you, Neo4j give us a REST API to query, to create nodes and everything. There are some occasions where we need to extend this REST API, so we can extend Neo4j using extension, manage or unmanage, so we can write some algorithm using the API, the transfer set API for example, and we can expose this as an endpoint in our API. This is some example, right in Java, sorry. There are drivers for almost every language. As I told you, we access via REST API. If you want to use using Python, I recommend you Python now. It has a module for Django I think. And I also, my conclusion, I want to quote Martin Fowler, and instead of just picking a relational database or probably MongoDB, because in Hacker News is the trending thing, we have to think about our data and what we have to do with this data. Probably we have to tend to polyglot persistence, have two, three or five databases in our systems to explore this data. If you want to know more about this topic, I recommend you these three books, NoSQL Distilled by Martin Fowler, No4j in Action and Graph Databases. Also, if you want to try it without installing it, I recommend you GraphMDB, that is a No4j as a service. There are some free plans to try it. And, okay, questions? So this is all very new to me, and I have only very big idea about that, but from what I've seen, my impression is that we basically store records in notes, right, and we label the edges with the relations. Okay, so in SQL, when I want to create a new record, I have to put it into a table for which I define the type, right, so I define all the attributes in advance, and I define how they should look like. Am I required to do it here as well? So I actually have to define the type of data which I can store in the notes, or can I just do anything with the cipher statements? You can do anything that you want. There are no predefined schema. Okay, so this reminds me then of difference between dynamic and static type languages. So what happens if I write a statement in cipher that actually doesn't make sense? I would ask for relation between, or I would create two notes and connect them with the relation, and then I would create other two notes that would carry different type of data, and I would connect it with the same relation. I could create many statements that probably wouldn't make any sense. What happens then? Nothing, no. It allows you to store whatever you want. Okay, so basically the issues or the problems are solved during the runtime when I run the statement? Probably it will return nothing if you are querying something that doesn't make sense, or something that you didn't store before, but there are no type checks. Are there any advantages that this brings to us? Like dynamically typed languages definitely have some advantages out of this. Do we see something in the data? There are advantages. You can evolve your model as well as you evolve your program. You are not tied to a schema. For example, if tomorrow I want to, in my example of musicians, I want to ask the engineers that engineer the albums of these bands. My old voice will still work and it can evolve without touching anything. It's more agile. It's like a no SQL philosophy. There are some real scenario where disadvantage can play a role. It would be interesting to me. Thank you for your answer. Hi. In the example that you had where you were searching for two kinds of relationships, it was an artist and producer or musician and producer or something like that. Yes. That one. In that query, can the result contain the type of connection? Yes. So here you are. You are storing in the R variable. You have information of this relationship. So you can get that. Yes. Thank you. Sorry, this is a silly question. You're adding all your objects in their relations and then you have a database full of stuff. Are there tools that can introspect that to then just sort of not UML but dump out the relationships that you actually have within your database? I can hear you. Can you repeat the question, please? So once you have your database full of data, is there something that can output sort of a summary of the relationships that are stored within the database? Yes. You have a web interface that represent graphically what have you stored in your... And that's part of Cypher or part of... It's part of Neo4j. And there are other tools like Link torus, I think, that explore in this way visualization of your data. Thank you for your talk. You said that the relationships you get for free, there are no indexes. And there is just on the slide, I wanted to ask how it is implemented that our date is greater than 1968. So there are actually some internal indexes for comparison or it is linear search. Yes. When you're looking for properties in the background, Neo4j use Lucene. So when you are looking in that case for name Clapton, you are using Lucene. So probably this could be a handicap of this kind of databases because you have to go to the index. Yeah, okay. Thank you. You're welcome. Are there any more questions? Okay, thanks a lot for your talk.
|
Francisco Fernández Castaño - Graph Databases, a little connected tour There are many kinds of NoSQL databases like, document databases, key-value, column databases and graph databases. In some scenarios is more convenient to store our data as a graph, because we want to extract and study information relative to these connections. In this scenario, graph databases are the ideal, they are designed and implemented to deal with connected information in a efficient way. ----- There are many kinds of NoSQL databases like, document databases, key-value, column databases and graph databases. In some scenarios is more convenient to store our data as a graph, because we want to extract and study information relative to these connections. In this scenario, graph databases are the ideal, they are designed and implemented to deal with connected information in a efficient way. In this talk I'll explain why NoSQL is necessary in some contexts as an alternative to traditional relational databases. How graph databases allow developers model their domains in a natural way without translating these domain models to an relational model with some artificial data like foreign keys and why is more efficient a graph database than a relational one or even a document database in a high connected environment. Then I'll explain specific characteristics of Neo4J as well as how to use Cypher the neo4j query language through python.
|
10.5446/19974 (DOI)
|
We will talk about out of core columnar databases. He is a creator of PyTables, a developer of Blaze, and a performance enthusiast. Welcome, please. So thank you very much, Oliver, for the introduction. So in my talk today, I am going to introduce you to out of core columnar datasets. And in particular, I will introduce big calls, which is a new data container that supports in memory on disk columnar chunked compressed data. Big calls seems like a strange name, but you can think of it like a big columnar. And the final LZ stands for Lempel Cif Codex, which big calls use a lot internally. Okay, so just a plug about me. I am the creator of tools like PyTables, Bloss, now big calls. And I am a long-term maintainer of NumExp, which is a package for evaluating NumPy expressions very quickly. I am an experienced developer and trainer in Python, because I have almost 15 years of experience coding full time in Python. And then I love high performance computing and storage as well. So I am also available for consulting. So what? We have another data container, right? So yeah. In my opinion, we are bound to live in a world of widely different instances of data containers. The Nosecule movement is an example of that. We have a wide range of different databases and data containers, even in Python. And why? This is mainly because of the increasing gap between CPU and memory speeds. That if you understand this fact, you will understand why this is so important. So the evolution of the CPUs, it's clear that the CPUs are getting much more faster than memory speed. And this is creating a gap between memory access, and the CPU is mostly doing nothing most of the time. And that has a huge effect in how you access your data containers. If you want more details, you can see my article, why modern CPUs are starving, and what you can do about it. So why columnar? Well, when you are querying tabular data, only the resting data is accessed. So that basically means less input output required. And this is very important when you are trying to get maximum speed. So let me show you an example of that. Let's suppose that we have an in-memory row-wise table. This is the typical structured array in NumPy. It is stored like this. So for example, if you are doing a query, the interesting column is the second one, the integer 32.1. So due to how computers work with memory, you are not accessing only the interesting column, but you are accessing also the bytes next to this column. This is for architectural reasons. So typically, if this is in memory, you are not accessing, you are not bringing to the CPU just n rows multiplied by four bytes. But you are bringing to the caches n multiplied by 64 bytes. And 64 is because it's the size of the cache line, typically, in modern CPUs. So we are bringing 10 times more data than is strictly necessary. In the column-wise approach, if you store the data in the same column sequentially, you will be only bringing to the cache the exact amount of information that you need. So this is the rationale behind why column-wise tables are interesting. Now why chunking? So chunking means that you store your data in different chunks, not in a monolithic container, but that means more difficulty handling that data, right? So why bother? Well, the fact is that chunking allows efficient enlarging and shrinking of your data sets, and also makes on-flight compression possible. So let me put you an example. When we want to add data in a NumPy container, for example, we need to reserve to do a malloc in a new location, then to copy the original data in the original array, and then finally copy the data to Append at the end of the new area. So this is extremely inefficient because of this gap between the CPU and memory. Now the way to append data in because is chunked. So if we want to append the data, we only have to compress the data because because is compressed by default, because containers, and then you don't need the additional copy because basically what you are doing is adding the new chunk of chunks to the initial list. So it's very efficient. And finally, why compression? Well, the first reason for compression is that more data can be stored in the same amount of data, of media, sorry. So if you have your original data set and your data set is compressible, and let's say that you have a compression ratio, you can reach a compression ratio of 3x, you can store three times more data using the same resources, which is great. But this is not the only reason. Another reason is that if you deal with compressed data sets in memory, for example, or on disk, whatever, and you have to do your computations, typically they execute in the CPU cache, you will need to transfer less information if your data is compressed in memory. And that could be a huge advantage. Now if the transmission time of transmitting the compressed data from the memory or the disk to the cache plus the compression time, we can do that time, the sum, less than the time that it takes the original data set to be transferred to the cache, then we can accelerate as well computations. This is a second goal. And for that you need an extremely fast compressor. So Bloss is one of these compressors. So the goal of Bloss is bringing data much faster than a memcpy memory copy can work. Here is an example where the memcpy is reaching a speed of 7 gigabytes per second, and then Bloss can reach a performance of 35 gigabytes per second. So Bloss would be interesting to be used in Bicol. In fact, it is part of Bicol. So goals and implementation. One important thing, an essential thing I would say in Bloss, in Bicol, sorry, is that it is driven by the keep it simple stop it principle in the sense that we don't want to put a lot of functionality on top of it. We just want to create a very simple container, a very simple iterators on top of it. So what Bicol is exactly. So as I said before, it's a columnar chunk, compressed data containers for Python. It offers two flavors for containers. The first one is CRA and the other is CTABLE. And it uses the Bloss compression library for the compression. And it's 100% right in Python and also for accelerating the interesting parts. So for example, the CRA container, which is one of the flavors of Bicol, is just a multidimensional data container for homogeneous data. So it's basically the same concept that numpy, but all the data is split in chunks just to allow this easy to append and also to allow compression as well. So the CTABLE object is basically a dictionary of CRAs. It's very simple. But as you can see, the chunks follow the column order. So queries followed on several columns will fetch only the necessary information. And also, adding a removing columns is very cheap because it's just a matter of inserting and deleting entries in a dictionary, Python dictionary. So persistency. CRA and CTABLE can live not only in memory, but also on disk. And for doing that, the format that has been chosen by default, it's heavily based in Bloss Pack, which is a National Library for Compression Large Data Set that Vality Hennel has been working on for the past years. And tomorrow and Sunday, he will be giving a talk on the PyData conference. So because and the goal of because is to allow every operation to be executed entirely on disk. So this persisting thing allows because operations to be executed entirely on disk. And that means that all the operations that you can do with objects in memory can also be done on disk. So you can add very large data sets that cannot fit on disk, on memory. You can do these operations on disk or even queries. So the way to do analytics with because is, as I said before, because men's strives to be simple. So because basically it's a data container with some iterators on top of it. And there are two flavors of iterators, then iter and where, which is the way to filter data, for example. And there is the blocked version of the iterators where instead of receiving one single element, you will receive a block of elements because in general, it's much more efficient to receive blocks and to work with blocks. And on top of that, the idea is that you use the iter tools in, for example, in the standard library, in the standard Python library to use these building blocks. Or if you need more machinery, you can use the P tools, the excellent PY tools on PsyTools packages in order to apply maps, filters, groupby, sortby, reduceby, joins, whatever, on top of that. This is the philosophy of V-calls. Also I recently implemented V-calls. If you cannot create V-calls from existing data containers, then you are lost. So I created interfaces with the most important packages when you are talking about big data. So for example, by default, V-calls has been always based on non-Py, but there is also support for PyTables. So for example, you can do indexed queries. For example, using PyTables, just start V-calls and produce HDF5 files with that. But also you can create, you can import and export data frames very easily from Pandas that we give you access to all these backends as well. Okay, so let me finish my talk with some benchmarks with real data. And in particular, I will be using the MovieLens dataset. And you can find all the materials for the plots that are going to show in this repository. So let me show you the notebook. Basically what I did is a notebook. So this is the notebook that you can find in the repo. And here it's all the parsing processing and everything. And here are the results. So you can access to this, go to this repository and reproduce the results by yourself if you like to. Reproducibility is very important, as you know. So the MovieLens dataset, it's basically people that rate movies. And there's a group of people that collected these ratings and created different datasets. There are three interesting datasets, one with 100,000 ratings, one million and 10 millions. The numbers that I'm going to show are the biggest one, the 10 million ratings. So this is the way to query the MovieLens dataset. So typically what I am doing here is using pandas basically for reading the CSV files and then produce a huge data frame containing all the information from the data files. Then the way to query in pandas is like the recent versions of pandas. I can use the dot query, which allows you to use this simple way to query the data frame. And for example, in the big calls Ctable from data, I import the data frame and create a new container, which is a big calls container. It's a Ctable container. And then this Ctable container is queried through the word iterator, as I said before. So you can pass exactly the same query than pandas. In fact, these queries are using numx behind the scenes. So they are very fast. And then you are selecting, you are saying to the iterator that we are interested just in the user ID field for the query. So here we have a view of the sizes of the datasets. It turns out that this dataset is highly compressible. So we can see that pandas takes around a bit more than one gigabyte and a half. And the big calls container for the same data frame, it's a bit larger. In fact, we have compression. But if you apply compression, your size or the size of the dataset will be reduced to less than 100 megabytes. So that's a factor of almost 20 times. So that's very interesting. But perhaps the most interesting thing about that is the query times. So pandas, you know pandas because it's extremely fine tuned for getting high performance queries, right? It's, in fact, pandas, the data frame, it's column oriented. It's column wise container in memory as well. So it's a perfect match for doing a comparison. So the time that it takes pandas for doing this operation, this query, is a little bit more than half a second. And for the big calls without compression, we can see that the time it's like maybe 60% less or something like that. And the most compelling thing in my opinion is that when you are doing the same query, but with using the compressed container, the time that it takes is less than using the compressed container. And this is essentially because the time that it takes to bring the data compressed into the CPUs is much less than the time that it takes to bring the data uncompressed. So the last, the upper row, the upper bar, means that big calls is on disk, but using compression. It is a little bit slower than in memory case, but it's still faster than pandas. And this is probably due to the fact that the big calls container, although it is stored on disk, the operating system probably has already cached that in memory. So it has a little bit more overhead because of the file system overhead, but the speed is very nice as well. So this has not been always the case. So for example, when I run the benchmark in a laptop which is three years old, for example, which is the one that I am using for the presentation, MacBook Air, we can see that pandas is the fastest. Then when big calls is a little bit slower, but when you're using the compressed container, it has an overhead. This is because BLOCK is not as efficient running in all architectures. I mean, new CPUs are very fast compared with older ones. In that gap, that increase that we are seeing here in my other laptop, my Linux box, we are going to see these kind of speedups more and more in the future. So compression will be very important in my opinion in the future. So let me finish with some status and overview of big calls. I released version 0.7.0 this week, so you need to check it out. So we are focused on refining on the API and tweaking knobs for making things even faster. We are not interested in developing new features probably, but just in making the containers much faster and also the iterators. So we need to address better integration with BLOCK. I am in contact with Valentin in order to implement what we call super chunks. So every chunk right now, it's a file on the file system when you are using persistency. And when you have a lot of chunks, that means that you are wasting a lot of fine nodes. So the idea is to tie together different chunks and to create these super chunks in order to avoid this overhead. And the main goal of big calls is to demonstrate that compression can help performance even using in-memory data containers. And that's very important because I mean I produced BLOCK like five years ago. And although my perception was that compression would help in this area, just five years later is when I am starting to see actual results with real data. But this is, this promise is fulfilled. So we would like you to tell us about your experience. So if you are using big calls, tell us about your scenario. If you are not getting the expected speed up or compression ratio, please tell us. You can write to the mailing list there. Or you can always send bugs, patches. Please file them in the GitHub repository. You can have a look at the manual, which is online, big calls dot BLOCK dot org. Then you can have a look at the format that is using big calls by default, BLOCK SPAC. And the whole BLOCK ecosystem lives in BLOCK dot org. So thank you. And if you have any question, I will be glad.
|
Francesc Alted - Out-of-Core Columnar Datasets Tables are a very handy data structure to store datasets to perform data analysis (filters, groupings, sortings, alignments...). But it turns out that how the tables are actually implemented makes a large impact on how they perform. Learn what you can expect from the current tabular offerings in the Python ecosystem. ----- It is a fact: we just entered in the Big Data era. More sensors, more computers, and being more evenly distributed throughout space and time than ever, are forcing data analyists to navigate through oceans of data before getting insights on what this data means. Tables are a very handy and spreadly used data structure to store datasets so as to perform data analysis (filters, groupings, sortings, alignments...). However, the actual table implementation, and especially, whether data in tables is stored row-wise or column-wise, whether the data is chunked or sequential, whether data is compressed or not, among other factors, can make a lot of difference depending on the analytic operations to be done. My talk will provide an overview of different libraries/systems in the Python ecosystem that are designed to cope with tabular data, and how the different implementations perform for different operations. The libraries or systems discussed are designed to operate either with on-disk data ([PyTables], [relational databases], [BLZ], [Blaze]...) as well as in-memory data containers ([NumPy], [DyND], [Pandas], [BLZ], [Blaze]...). A special emphasis will be put in the on-disk (also called out-of-core) databases, which are the most commonly used ones for handling extremely large tables. The hope is that, after this lecture, the audience will get a better insight and a more informed opinion on the different solutions for handling tabular data in the Python world, and most especially, which ones adapts better to their needs.
|
10.5446/19973 (DOI)
|
Roedd y cerdd 댁 i fy orbitu ychydig. Felly, mae'r ffictiwyr yn cyfnodd. Cyfnodd ffictiwyr yn cyfnodd yn ymddiadau yn ymddiadau. Mae'r ffictiwyr yn cael ei ddweud i'r ffictiwyr yn ymddiadau. Mae'n cael ei ddweud i'r ffictiwyr yn ymddiadau. Mae'r ddweud i tam test that we find useful building a big test tree. I'm going to assume some knowledge about Pyto test itself and you got the basic idea of how fixtures work, so how did pens injection sort of works. But a very quick reminder of that. So fixtures, it's basically so any test function can request a fixture by just taking a parameter, a named parameter, the name parameter will then be looked up to a function that's decorated with a fixture marker, which is just a function with terms of value, and that value is then injected by Pyto testing to your test function. You can create fixtures on different levels in your files, so if you do it in a class it will only be visible to members of the class. So that's sort of very quickly the basics. So the first thing to start sort of extending that is caching a few fixtures. This sort of, so basically by simply giving another keyword argument to this Pyto fixture decorator, you can change the scope, the lifetime of how long that fixture lives. Normally it's being, if you don't provide an explicit scope it will just be torn down straight away after the test function has run. But you can change the scopes into like session scope, you also got like module scope and class scope, as well as function scope. And that basically means that you'll only call that function once for the scope. So in this example we got two differently scoped ones, one function scoped, one session scope, and we got two tests using both of these, this is fairly contrived I guess. But if you run Pyto test in the output I'm using minus s here, which kind of stops the, make sure that the print statements I put in my test code actually can get shown up because normally Pyto tests will capture that and only show it in case of failures. But with minus s you can clearly see the order that things happen here. So the session setup happens first and then the function setup happens then the dot, which is basically the test that's been run. And then the function finalises it there down, kind of happens and then that happens again. So you can see how the caching sort of works there. So if you've got expensive fixtures that you don't want to recreate a whole time like a populated database with a schema or something or create web browsers or those sort of things, this is a very common thing to start kind of using. Yeah, and you have sort of available scopes there. Next on you can sort of, fixtures can just, can be used in other fixtures as well, not just in test functions. And that makes it really composable because it means you can have a, so in this example, creating a database connection fixture and that might be used in some tests directly. But then, for example, if this was a functional test and one functional test needed a table to be there or something, I can just build on those existing fixtures and you can puzzle them together in that manner. One of the things here, the request fixture that's being asked there is, it's the way to out finalise it to places. So you've probably already seen that, but it's essentially no more than just another fixture in pilot tests. It's just a build-in one that's provided. So that kind of happens quite seamlessly. The test function here simply uses the latest one, the DB table fixture, but it could have been using both as well. So if it needed both of them for some reason, there's no reason not to combine it that way either. Another thing you can do in fixtures is, so normally when you're on test, in pilot tests you can mark test as, oh, I want to skip this test or something for whatever reason. But you can also do it based on a trigger this from inside a fixture basically. If you do it out, it means that any test that will be requested, it will be using that fixture, will automatically be skipped instead of, so you can make this depend on something else. So this example is something that we use quite commonly when we connect to remote service sort of thing. So basically first drive the developers running it on their own local host, if not, see if they're in the office, try connecting to the server on that network sort of thing. But if NIDR is there, then whatever, let's go and skip it. Combining this with the session scope, for example, means that you don't keep doing that again and again because that might be slow operation or something. Another thing to note here is that when you call pi test.skip like this with a message, basically pi test or skip will raise an exception in your code. So for the control flow of your fixture, you have to realise when you execute pi test or skip, you basically raise an exception in there and pi test will interpret that as. But no code will be executed afterwards anymore. Just like you got pi test or skip, there's a pi test or fail will do exactly the same but failure test for whatever reason if you want to do that. And this is sort of a sort of slide side step. Talking about introducing marks a little bit. Hopefully you've already encountered marks but pi test has a very flexible marking system. So it's basically just a decorator, pi test.mark. And then a name you choose yourself sort of thing. And you can just apply that decorator to your test functions and then your test functions will be marked. That on its own doesn't provide you very much. You can sort of use that on the command line to select your marks but it doesn't provide you that much yet. I'll get back to that on the next slide basically. But one thing to note here is that you can mark multiple marks. You can obviously apply multiple marks onto your single test function or anything like that. But another thing is that because marks are so flexible, they're sort of, they're allowed to be made available on the fly basically. So one sort of side effect of that is that if you make a typo in a marker, you may not notice that and they may kind of hurt you later on. And that's why you sort of have two camps of people I think. Some people like me prefer to use the minus, minus strict option to pilot test so that you get caught out and you have to basically declare your markers up front and you get notified of any mistakes you're doing. And the obvious way to just always enforce that is to write in your configuration file. So that's basically the example of how you write in your configuration file. Adopt basically always adds the command line option when you invoke it and then you declare your markers. In this example now, if you try to run this, it would fail because only one of the markers has been declared. So that's sort of markers. So markers are like, you know, you use by plug-ins etc. as well if you've used plug-ins. They often make use of that. And one way to make use of that is inside, basically detect them in your fixtures. So in here the test function wants a Mongo client. It's a little bit contrived because PyMongo doesn't quite work like that, but almost. So basically it needs a Mongo client. And we're also declaring basically the markers, basically declaring the database to put in the URI to be used at connection time. And then when you actually look in the fixture, you can basically what it tries to do is like, it tries to look if there is a marker, yes, then I'm going to use that as the database name. If not, I'm just going to use a default database name and then create the client and return it. And to actually get this marker information, you use request.node. So request.node itself is basically the test node, which is an internal part of the test representation of your test itself. So one of the attributes will actually be the function that you're actually testing. And it has quite a few methods. I think you can look it up in the documentation, at least part of them are documented. But getMarker is basically how you get a whole of a marker. The only thing is like, you don't basically get this sort of marker object. And in the case, so either you get this marker object or none, if you get none, you just use the default one here. But in the case of a marker object, the way it passes, because as you see, the mark takes an argument here. And the way it passes on the arguments that you've specified, it just passes these on as this arcs and Kw arcs attributes on your marker object, which is a list and a dictionary, which is a very common representation. But the problem with that is that it's not how Python signatures sort of work. So I use this little helper function, which is like a call API phone in this case. And it doesn't do anything useful other than it uses Python itself to pass a signature for me. So it means that if someone, if I wrote the decorator here differently in MongoDB, DB equals users, it would still work because Python would just pass a signature and return my database back. So that's how I call that function with star arcs and Kw arcs. So that's sort of a little trick to get native Python signatures working there. I have to add that in the future that I think probably one of the next releases, that there is going to be a slightly different way or a new way introduced of declaring markers, which will get around this little hacky signature passing stuff. So that will improve. But this is sort of the way we do things currently, and it works very nicely. Another commonly used thing is fixtures can also be automatically used basically. So normally fixtures are always dependency injected, so your test function requests the fixture by name it wants to have. But sometimes that might not be suitable. So there's this auto use equals true argument to the fixture decorator. In that case, it will basically be called for every test function automatically, whether it's been requested or not. So this is kind of a lot closer if you're used to the unit test kind of things to set up and tear down because that's just called every single time. One of the nice things here is this well that you can combine this with the scope argument as well. So if you do this, if you create an auto use fixture with a scope of the session, that basically means you only actually write the beginning of your test session, you'll be doing set up and at the end you'll be doing some tear down. The auto use fixtures, in this case, I sort of use it with an underscore sort of indicating that I don't expect it to be used immediately or directly. But there's nothing stopping you from returning a value as well and explicitly requesting it. So if someone explicitly requests it, you can just mix those two together as well. So in this case, it's sort of usually to create a kind of custom skipping logic for something that's only supposed to work on Linux. There's various ways of doing that, I guess, but it's an example. Then parameterising fixtures is sort of another very powerful feature that you can do with fixtures. So in this example, we have a test function, the first test function there, TXN, which we have a test function that just uses basically your I to connect to a database and say the problem here is I want to ensure that whatever this works with PostgreSQL Oracle and SQLite, for example, at the same time. So instead of having to write three tests or write tests with different fixtures or something, you can basically parameterise your fixture itself. So by parameterising this fixture, Pyta test will create, will call my test function three times once with each parameter. And the way you know which parameter, so the arguments in the list to the params keyword in your fixture decorator can be anything, it's up to you. So you can use direct values there if you want to, or in this case, are you simple strings which then I'll probably use later with an if statement or something. But you get access to the parameter being passed in by this request.param attribute on the request's fixture. And that's how you access them. Another sort of building on top of that is if you have multiple fixtures with parameters, you can combine them, which is basically what the last function does. So in this case, because both fixtures are parameterised, Pyta test will basically call the last function six times because it requests both of these and you will get each combination tested automatically. So in this case, the example suggests that I want to test the connection of both IPv4 and IPv6, but you automatically get all combinations of your fixtures. But a slightly building on top of that is you can optionally mark your parameters again in your parameterised list. So you can, if there is a fixture, so in this example, I don't assume that every developer has their Oracle DBI PI installed. So I just try and put Oracle if not. And I sort of introduced two things at once here. So we've already seen the markers basically. So this is Pyta test, mark skip if is a built-in marker that Pyta test provides. But instead of using it directly on something, I just assign it to a variable now. And that's just to make my line a bit shorter later on basically, or maybe I want to use it more than once or something like that. So you can just assign your marker to a variable and now you have that marker object there, which you can use later on. In this case, I apply it basically on because my parameter is not a function, so I can't really use decorator here. So I just manually apply that onto my parameter. And it means that if I don't have Oracle installed here, despite that the test function should be parameterised, only one version of them will be run and the second will be skipped. This is sort of an example taken roughly from an early version of the Pyta's Django plugin. But it shows kind of how you can see basically the thing here is the requested fiction names part of the fixtures. And it allows you to see all the other fixtures that are being requested by the current test function. So you can, in this case, I'm just stopping two fixtures that are mutually exclusive being used at the same time by simply calling a PyTest or Phelon if I detect that case. But you can sort of peek into what the test function is asking and have a look at the other fixtures around, which sometimes could be a useful thing. Then sort of a very another sort of step aside, I guess. You've probably, if you've been using PyTest, seen this ConfTest.py files. So this is sort of typical directory like you'll get. At a simple level, ConfTest.py, so you can declare your fixtures in there and then you'll see them in their entire directory basically. But ConfTest.py is basically, from PyTest's point of view, it's just another plugin just like any other PyTest plugin you can install. But it's a per project kind of plugin. And essentially, when you start building a big test suite, you're going to be building a per project plugin kind of thing. The plugins work with this hook system. So basically, in your plugin, you find a function with your hook. PyTest will at certain times call these hooks then if you defined one. These are just a few common ones. There's a much longer list in the documentation. I'll be showing like basically the add option one here, which is sort of adding a new command line option. The parser object you get in that hook is basically just an odd pass command line parser. So you can just use that documentation to add your options here. I'm here adding an option saying, so basically the idea here is that if my, for my script that invokes PyTest on my CI server, I'll be just passing in that minus minus CI option so I know if my test suite is running on the CI server. And just very quickly, you can get to it, to command line options both from inside fixtures as well as test functions. Again, via this request object which has a config, which is basically PyTest config representation. So you can get access to both command line options or configuration file options in there. And basically the both request.config and PyTest config are basically the same instance. They're different ways of getting to them. And those two, combining all of that together sort of, I can, this is sort of extending on an example I used earlier. So again, so here I have this external service, in this case a ready service, that before I checked out is it on my developers laptop, is it on my, is it in the office or something. And I was just skipping those. But in this case, by using request.config and checking that option, I can make sure that when I'm running on my CI server, it really is going to execute. And if not, I actually get a hard failure so I can see. So if the server is down, I don't just suddenly stop testing part of my test suite sort of thing. So that's another thing that we use quite a lot and it's pretty handy. That's basically what I had so far. So, yeah, thanks for listening. I hope it was useful. Okay, so the question is, is there a way to get to the value of that feature? Okay, so the question is, I showed how to get the names, the other fiction names that are being requested by the test function. Is there also a way of getting the value of them? The answer is sort of twofold. Definitely yes, obviously by simply requesting it yourself. But that's not dynamic. There is an escape hatch to do that dynamic, although it's discouraged because basically on the request fixture itself, there is a getfunkarg value call, which you can then use to call another value. So that's how you could get to it. But generally that's discouraged because if you use that on one that's already requested, it's probably okay. But if you use that on one that hasn't been requested yet, Python test sort of loses its view of all the combinations of things and it won't be able to do its caching and scoping properly as well, et cetera. So you have to be a little bit careful with using that escape hatch, I would call it. I wasn't familiar with the parameterisation that you could do with testers. Does that play well with the other parameterisation that you can apply to testers directly? Yes, so the question is, does the parameterisation on fixtures that you can do on fixtures play well with the parameterisation you can do on testers directly? So on testers are actually in case you don't know. It's basically not a marker, so you can decorate the test at pytest.mark.parameterised and then give it a list of its parameters. Yes, those two basically work together. When you do, so when you're requesting, so when you use the decorator at parameterised, you have to use the request and request.param again to get hold of the value that's being parameterised. So when you're requesting request, which is already going to be parameterised and all the parameterised fixtures, you're just going to get all the combinations again. So, yeah. So, the question is, when pytest and coverage are used together, if you import stuff at the comftest.py in the comftest.py file, it tends to be skipped by coverage and someone is nodding yes. The problem is plugging, is it, for a problem, is a coverage plug-in. What I'm just doing is I use a coverage group to run pytest and then there is no problem at the import level. Because the problem is that if you do something in a fixture, then that plug-in coverage bring you to the plug-in. So, yeah, the plug-in itself can't catch that as well. Okay, so the new PyT coverage plug-in should be fixed. So, it should just work then basically. All right, well, if we're not made, thank you again for us.
|
Floris Bruynooghe - Advanced Uses of py.test Fixtures One unique and powerful feature of py.test is the dependency injection of test fixtures using function arguments. This talk aims to walk through py.test's fixture mechanism gradually introducing more complex uses and features. This should lead to an understanding of the power of the fixture system and how to build complex but easily-managed test suites using them. ----- This talks will assume some basic familiarity with the py.test testing framework and explore only the fixture mechanism. It will build up more complex examples which will lead up to touching on other plugin features of py.test. It is expected people will be familiar with python features like functions as first-class objects, closures etc.
|
10.5446/19971 (DOI)
|
Okay, welcome to the second talk in this session, given by FlavioPairCoco on OpenStack. Hello, everyone. Today's talk is about system integration. It's not actually 100% related to OpenStack. It's mostly related to integrating systems to each other. And I'll use OpenStack as an example because most of the methods I will present to you are being used by OpenStack itself to integrate all the services that we're using. So that's me. That's my Twitter handler. Pretty much everything you want to know about me is out there in the Internet. Something I want you to know. I work for Red Hat and I'm part of the RDO community. RDO is a community of a bunch of really great people working together to make OpenStack amazing on real-based distributions. Other things about me, I'm not going to go through those. The one I will mention is I'm a Google Summary of Code and OPW Mentor. And I wanted to mention this because I really believe it's very important. And if you have spare time in your day and you want to mentor people, please sign up. We need more mentors there. So let's get to it. Before we go through the methods that you would use to integrate systems, let's first define a little bit what system integration means. System integration is basically what you do when you have a set of subsystems and you want to make them work together towards a common goal or scope, right? So all the methods and technologies and strategies that you would use to make those systems work together and to get towards that goal is what we call system integration. This is put in a very simple way. There are a bunch of different definitions of system integration. You can integrate systems. Systems are not necessarily software. You could use hardware as a system. You could use many other things as a system. So system is a very generic term that you would use to say that you have a set of subsystems working together for a single cost, so to speak. There are many different generic strategies to integrate systems. These are the three that I will present very briefly, and we will dive a little bit more on the last one. So vertical integration, vertical integration means or basically looks like a small graph up here. It looks like stars. Basically, you have a set of systems and you will talk to the system. The system that I will be talking to is done based on each subsystem features and what you need from them. So you have a web service that integrates to your database and then you have two systems that are working together. Or you have your authentication service and then you have your other services below it. And you will make this service with the real features to your authentication service and you are integrating those two services vertically. The star integration, well, it's called star integration because it's supposed to be like a star, but it's more like spaghetti integration because all services know what other services do and they all talk together and they do that in a case-by-case basis. So you have service A that needs something from service B, but before doing that it talks to service C because it needs something from service C before getting to service B. So it's quite a mess. Use cases for this, there are plenty of them, but it's very risky and it's very, very error-prone. And there's for sure a very high risk of not having a contract when you're talking, when those services are talking together. Not having a contract basically means that you don't know what you're going to get back and you don't know when something's going wrong, when you get something from service C to talk to service B, but service B is expecting something different and it turns out that service C was updated. And the other method that we're going to dive a little bit more today is horizontal integration. And it's based on a service bus and service bus is, I call it communication bus. I don't like to use the term messaging bus here because it doesn't matter, I mean it's not about messaging itself, but about making those services communicate together through the same bus. So you have service A, service B, and service C and they all communicate through these communication bus, sending either messages or just a data asset that will make the whole feature work through this communication bus. So diving a little bit more on horizontal integration from an application's point of view. So how would you make all these, so imagine that you have a set of applications that you want to make them work together. You need to have these communication bus. So you have to come up with an idea, with a technology that you would use to make them talk together. What I'm going to do now is I will present like four different methods, four different methods to make those applications talk together. And these are not new methods. They are actually been around for a long time. Many people use them and they actually don't know that they are basically integrating a system and they actually don't know what the whole thing they're doing means. And each one of these cases are good for very, sorry, each one of these methods are good for very specific cases. Some of them are more generic and others are more specific. And the first one is, sorry, and the first one is files. Files is probably, it's probably the oldest way to integrate different services talking together. For a long time people used to open a file, get a file descriptor, put something in there and have another application within the same piece of hardware, reading out of it a data asset that it will use to do something. So some people would use files as a messaging bus, like you would use a RabbitMQ right now, so a messaging broker. It's good for very specific cases. Try not to use it. It's good for cases like embedded systems. And at the embedded system you won't have RabbitMQ running for sure. So if you have very limited hardware and very limited processor and memory, you probably want to use something that's really cheap, files are cheap. Definitely accessing the file system has a cost. It has a high risk in terms of security and reliability. But it works very well for embedded systems. We used to use these files, we used to use files in OpenStack to have some kind of in-server distributor lock for some time. Many things went wrong with that, so don't do it. We're now working on another way to have distributor locks. But that's one of the cases where, for example, in OpenStack we used files and we moved away from them. Like I said, they're very good for hardware that has very limited resources. But if you can afford something more expensive, you probably want to do that. Databases. Databases are probably one of the most common. By the way, all these statements are based on my own experience. I don't have really actual data that proves that this is the most common or the files are the most oldest. This is all based on my experience and my own research. Databases are probably the most common way to integrate services. They are asynchronous, data-wise. What that means is that when you have a message and you want another service to get that message, you would just store it in the database and you are done with it. The producer stores the message in the database, the producer is done with the message and then the consumer eventually will get the data out of the database and will do something with it. Databases are really great for starting states and I'm saying this is probably the most common one because most of the web services out there, like I couldn't think about a web service that does not rely on a database. If you want to scale your web service, you definitely have or most probably have a single database for your whole thing and you have several services talking to that database and getting data out of there. They are great, great, great for states. The way we use this in OpenStack is in OpenStack, most of the services are probably the biggest services have been split in several smaller services. In Nova, for example, how many of you know OpenStack or have heard of it? Awesome. So Nova is the service that is responsible to spawn new instances, virtual machines, so it probably is like EC2 for AWS. And Nova has three sub-services, well, it has many more than that, but the main services that you need from Nova are like three or four services and you have the API service, you have the compute node and you have the scheduler and you have conductor that gets messages and stores everything on the database. So when a request for a new instance comes into the Nova API service, a new record will be creating the database and then a message will be sent to the scheduler that will then talk to the Nova compute node to spawn the new virtual machine. So what Nova compute does is it gets the data of these new instances that were requested out of the database, it spawns the new virtual machine and when it is running, it will update the state of the virtual machine saying that, hey, the virtual machine is running and it will update the data. So that system integration in that really small scale. And that's a way you could use databases to integrate systems. So that's why you say that they're probably the most common way to integrate systems and probably many people don't know that they're actually integrating systems by using databases. Stuck. So LibreOffice is, there you go. I hate your LibreOffice. So does any of you have any questions so far? Feel free to interrupt me if you have questions. So messaging. What I mean by messaging here is not a broker, is not an NQP and is not the specific technology that allows you to send a message from point A to point B. What I mean about messaging here is the message itself, like the message as you need to send data from point A to point B. Whatever method you use to send that message from point A to point B, the benefits of using messaging is that it's loosely coupled and it adds way more complexity because by being loosely coupled it means that you don't have a contract on the message so the service A can send a message to service B. But service B has a hypothetical idea of what it's going to get and what it wants to do with that message. It adds more complexity because if you don't know what the message may look like, you probably will have some parsing errors, type errors or whatever depending on the language and what you want to do with that message. Some benefits though, it's like being loosely coupled, you can say I will send this message and whoever gets this message can do whatever it wants with it. One of the places where we use this kind of messaging or loosely coupled contract is in Cilometer. So Cilometer would plug into the notification stream of OpenStack and it will get all the notifications of what's happening in your infrastructure if you spawn a new virtual machine, a new notification will be sent so Cilometer gets it, parses it and does something with it, creates new events, creates stats and allows you to build users based on what you've done. And one thing about messaging is that it may depend on message routers and transformations. So when you use messaging and you want to send a message from point B to point C but you have to go first to point B, you will need in point B some kind of logic or technology that will allow you to route that message to point C and you will do that based on the message information itself so you have to know what to do with it and you have to try to parse it and get information out of it to know where the message has to go. And this is something that Nova scheduler for example does. It doesn't get a notification, it gets an RPC message and we will go to that later. But it gets a new message, it parses it and it tries to get a Nova compute node that will do the work for it and it will send that message to Nova compute and it does that by using some filter logics and availability and all that important information. But let's not get to that. They are very easy, they are very cheap but they add complexity to your system. And the last method that I want to present today is RPC. RPC, it stands for remote procedure calls. It was probably introduced pretty much by the enterprise war when system integrators wanted to integrate systems for the customers and they would go and use RPC calls to do that. And RPC calls, the way it works is you will send up formatted messages so you have a counter-agundant message to point B. Point B will do that and it's called remote procedure call because you are basically calling a remote function just by sending a message. You will say call this function and pass this argument to that function and give me the result back. It's the most used method throughout OpenStack and I do have numbers for this. The message channel may vary, you can use databases, message brokers, so like I said, I'm not talking about the method you would use to send a message from point A to point B. In the OpenStack case, we use message brokers to do this, to do RPCs. And one of the drawbacks but it's actually something required for RPC is that it's tightly coupled so you have a protocol, you have to invent something, you have to agree on a contract when you send a message from point A to point B because you want to call a function that you know it exists in point B and you have to pass some arguments to that function and you want to get a result back. So you have to know what are you going to get back and you have to know what you have to send to the point B to call that function. So it's really tightly coupled, you will need to design your own protocol to this. But it's really common, it's very useful for doing that kind of remote calling function thing. But you have your benefits and your drawbacks taken from this. So in the OpenStack case, this is pretty much a high-level overview of how it works in terms of system integration. It's based on shared nothing architecture. If you don't know what shared nothing architecture is, it's basically a very simple way. Services working together but not sharing anything. And by not sharing anything, I mean they don't share memory space on your box, they don't share processes, PIDs and other resources. They can live together on the same box but they won't share the same resources. They don't have their own space within that box. So every service knows very few things about other services. And with that, we managed to get and to keep all those services very isolated from each other, which is really good if you want to integrate systems together. You want your services to be independent, you want your services to be isolated from each other. And if something happens to one of your services, you would see, you definitely want your other services to still be alive and being able to work on top of other services in your system. So we used databases for interservice communication. Like I said, Nova API will store a new instance record with booting state and then Nova Compute will update that state. And we use RPC for interservice communication. When Nova API gets a new instance request, it will send an RPC message to the scheduler and the scheduler will get that message and then it will send another RPC message to the compute node that will then boot the virtual machine. And we use messaging for cross-service communication. And I already mentioned this. The way it works is that services, when something happens with our Pest Act, services will generate notification. Then they will send it to some specific topic in the broker that other services can just plot into and get messages out of it. And they can do something with those messages. So since OpenStack relies a lot on brokers and it's probably right now one of the common tools to integrate services in many deployments, I would like to say a few things about brokers and the technology that you could use or how you could do integration based on protocols like AMQP or just using technologies like message brokers. So the first thing I want to say is that scaling brokers is really hard. You may have read or heard something like broker scaling is already fixed and you can scale Rabri MQ. I'm sorry, that's a lie. That doesn't work that way. There's a lot of documentation. Yes. There's some explanations how you can do it. Yes. There are some demos that people have done it yet. When you get it to big scales, it doesn't work that way. So scaling brokers is hard because synchronizing messages between different nodes of your broker that is heavily read and heavily written on, it's really hard and it doesn't work that way. Another thing is that brokers need a lot of memory. It all depends on your use case. If you don't have many messages traveling around your system, you probably won't use a lot of memory. But if you have a big deployment, your broker is definitely going to use a lot of memory. It really depends on how fast you write to it and how fast you read from it. If you write really fast and you read as fast as you read, if you read as fast as you write, your broker will probably use less memory. The memory footprint will be pretty linear and stable. But if you have more writes than reads, your broker will use a lot of memory. Brokers need a lot of storage. And if you want to have durable queues and you have your messages to stick around if something bad happens, you probably will use a durable queue. If you use a durable queue, your broker will have to write everything on this. Because if the broker goes down, it has to start from somewhere. So you will read all your messages out of whatever database or storage system it is using and it will make those messages available again. So again, if you have a lot of writes and not as many reads as you have writes, your broker will use a lot of storage. So I was looking at the time and it says nine minutes because LibreOffice went down and I was like, oh, I'm already done. So like since I've been ranting about brokers for a bit, I would like to say something about those. If you are going to use brokers or any messaging technology, prefer federation instead of centralization. What I mean by that is if you have a centralized broker and you want to scale the broker and that broker goes down, your system is off. You will have HA and all that you want. You want to scale the broker and you want to have it replicated and all those kind of things. But if you prefer federation instead of centralization, you will have a whole bunch of nodes that are lightweight workers and if they go down, you will probably set up a new one and you won't rely on a single broker that is in the middle of your system processing all your messages. And one way to do that is relying on NQP 1.0. NQP 1.0, I'm pretty sure most of you are familiar with NQP itself. Current latest version of the NQP protocol being used by Ruby NQ and most of the brokers is NQP 0.10. NQP 0.10 is not a standard and many brokers have implemented it in different ways. Whereas NQP 1.0 is actually a standard and it detects how messages will go from point A to point B. So how you can send messages from between two peers. NQP 1.0 is peer based in a message basis. What I mean by that, it explains how a message will travel from a point to another point. But it also, in the specification, there's also an explanation how you would do that with an intermediate broker. So it doesn't say that you have to have a completely federated system. You could also have a broker in the middle that is capable to speak NQP 1.0. So NQP 1.0 is all about messages and how those messages will go from point A to point B. And if you want to scale it and have more routing intelligence, so to speak, in your system, you could use something like QP dispatch that will allow you to create new rules to send those messages between your services as you would do with using routing keys in NQP 1.10. So in NQP 1.0, you don't have exchanges, you don't have Qs, you don't have binding rules, and you don't have routing keys. In NQP 1.0, you just have messages, and you have links, and every link is basically a connection to one of the peers in your system. So after having said all that about methods to integrate systems and technologies that you could use and protocols and all that stuff, I would like to give you some tips and tricks about system integrations. This is mostly based on our experience in the open stack community. First and foremost, transmission protocol matters. But transmission protocol, I'm not talking at the lowest level, like I'm not talking about UDP against TCP. I'm talking about a probably a higher level, like whether you want to use a protocol that's TCP wide or you want to use HTTP or you want to use some other RPC protocol whatsoever. Transmission protocols matter. Depending on the protocol you choose, you have some extra cost on your messages and transmission of your messages. So be aware of that. Depending on your use case, make sure you choose the best protocol for your use case. Use versions for your wide protocol. If you choose to use RPC to integrate your systems, you probably will have to agree on a protocol, and you probably have to define that protocol by yourself. Something that has been around in open stack for a long time is the version of those protocols. So when you define your protocol, you probably will say, my protocol is a dictionary that I will send between services. And that dictionary has a key that is called function, and that function is the key, the value of that key is actually a function name. And then I will have R and K keywords in the dictionary, and that will be the value to pass to the arguments and keyword arguments of my function. But then you want to update that protocol. You say, I want to also specify the return type I want from that function. And if you have your system deployed and you want to make a change to your protocol, you can do that. But if you don't have versioning, you will probably have to tear all your services down and then up again once you update the protocol. Because if some service gets a message, an RPC message with an RPC format, it doesn't recognize it will probably fail. Instead, if you have versioning, you can do rolling updates on your system by restarting services one at a time and updating those services so you don't have any downtime. Versioning is not just useful for upgrades. It's also useful for backward compatibility. If you do a change, and that change turns out to be really bad, you can go back to your previous version and you still have your services that used to work with that version. Keep everything explicit. I have a really nice quote that I got from Jeff Hodge's talk at the RIAC conference. He basically said, in a distributed system, having implicit things is the best way to fuck yourself. That's really true. If you have implicit things happening in your system, like you send a message and you don't agree on the contract for that message, you will probably face some several issues that you didn't expect to happen. So keep everything explicit. Even if it is more verbose, even if you need more code, even if you need more nodes running, that's fine. Just keep everything explicit. Because when something bad happens, you will know what it is. You will know how to debug it and how to fix it. Just at the time. Can you step to the microphone? I can repeat it. He is asking for an example of an implicit, oh, I can get one of the open stack issues. For a long time, FartCo correctly, it gets messages out of the notification stream in open stack. There were some implicit fields being sent by some services and those fields were sent by other services. So Cilimir didn't know about that. There was a case where it failed when you got those messages. A good thing is that it was before the release. So it could be fixed. But anyway, you don't need to, like, something that you want to get implicit is how you distribute your system. How your nodes are running and what nodes can run alongside with other nodes. You don't want to have all nodes running on the same server. So if you keep your architecture and distribution very explicit and even in the way you use separate services, it will be easier for you to estimate the scale and how to distribute those. A good example of this is Nova itself. Again, Nova has a Nova API service and it has a Nova scheduler service. So if you are getting a lot of API requests, you will get a lot of messages going to your scheduler. If your scheduler is under a lot of pressure, you can add more schedulers to it. You can scale them horizontally very easily. So the way you distribute your services in terms of code, like you have an API service, a scheduler service, a conductor service, and a compute service, it's another way to be explicit in how your distributed system should look like. Designed by contract. I've been using the word contract a lot today. If you design by contract, you don't have to, like, you know what service B is expecting you to send and service B is expecting you to send something. So service B can run, let's say, a set of aspirations before running anything and it will be replied back if some of those requirements are not met. So it's pretty another way when you integrate your system and you want two services to talk to each other, you have a contract between them. Like, pretty much like your account manager and yourself. You have a contract with him and you know what he's expecting you to do when you pay something and he wants you to get all the receipts and give those receipts to him. And you know when you give the receipts to him, he will do something with it and you will pay him for that service and he will expect you to pay for his service, right? So you have a contract with him. The same thing happens with services. When you send up requests to service B, service B is expecting something from you. You know what the service is expecting, so you will send that. If you don't met those requirements, he will reply back with an error. And if you send all the requirements, you are going to expect something back from him. If you don't get what you're expecting, you can just call back again and say, hey, this was not what I was expecting, so please give me what I want. This design by counter was probably known by most of you. It was introduced by AFL, the programming language, and it's basically part of the coding style of the language itself. Keep services isolated as much as possible. Like I said, share nothing architecture is very useful when you want to keep your distributed system safe from failures. And it's not completely safe from failures, but if one of your services goes down and it's isolated from all your other services, you can probably run another one somewhere else and make it talk to you. Keep them isolated. Keep your services very stupid, if you can. And I'm not talking about microservice architecture and having thousands and thousands of microservice doing just one little function thing. But keep them isolated and very focused in context on what they have to do. Avoid dependency in cycles between services. I wouldn't recommend using the star integration method. It's really messy and when something goes wrong, it's very difficult to debug it. So avoid having dependency in cycles between your services. If you can have a service policy, you can send messages through it. Make sure you don't depend on it. Both services don't depend on each other to get something done. Mock is not testing. If you have a distributed system, you probably want to test it. If you want to test it, you would say, hey, the easiest way to test it is by mocking what I'm expecting from the other service. Yeah, that works and it probably will succeed every time, but that's not testing. If you want to test your distributed system, get it installed, run everything live, and that's a way to test it. That's how you would know when something is working and it's not working. We have mocks in OpenStack, but we also run everything live for every single patch. This is very important. Like, many bugs we have found in OpenStack and that are related to how services are distributed, we're not tested live, and we have mocks for those tests. So mock is not testing. And before closing, this is a Pythoner interconference. Here are three libraries for doing integrations, like Combo is for sending messages. Combo is a library that's actually used by Celery, and it supports transport, and every transport is basically a messaging technology you could use. You can use RabbitMQ among the Redis and some other sort of technologies that it supports. Celery is a distributed task manager. Well, there was a presentation before mine about it. Basically, it allows you to have distributed workers doing something based on messages, and Celery itself uses RPC implicitly to tell workers what they have to do. And also messaging is an RPC library, and that's what we use in OpenStack to send RPC messages through services. And it also supports, well, it has the architecture to support many brokers. It just supports RabbitMQ and QPIT for now, and we're working on the MQP 1.0 support for it. And these are some messaging technologies that you could use. You probably already know them, like Halfcam, Acony, DreamMQ, RabbitMQ, and the QPIT family. You have the QPD, which is the broker, and it supports 0.10. And, well, it actually supports 1.0 as well. You could use QPD Pro, which is fully MQP 1.0 and QPD is for routing messages throughout your system. And that's pretty much it. Any questions? Please come to the microphones if you want to ask questions. Hi. Thanks for your talk. I was really curious about how do you do your systems integration testing? Do you have some automated system integration testing of setting up a cluster with all the services and so on? What tools do you use? So in a second, we use Garret for code review. Every time you submit a patch, there's tool which is our testing integration tool. It basically gets notification from Garret, and it runs a Jenkins job every time it gets a new patch. And those Jenkins job will install OpenStack completely in a single node, and it will test. We have live tests that call APIs and it will send messages throughout the whole system and simulate a live environment, like spawning new virtual machines, taking it down, creating new volumes, deleting volumes, creating new images and deleting images and all that kind of things. So it's been tested live. We do have automated tools. Jenkins is basically the one that does everything. And we use DevTag to install pretty much all the tools in those Jenkins jobs. All right. Thank you. You're welcome. I have a question. You didn't talk about security if you run this messaging infrastructure. How do you secure it? Sure. So right now, in OpenStack, security is pretty much done by binding everything on your private network in this layer. We have some work going on on signing messages and encrypting messages before sending them through the whole pipe, so to speak. There was a talk about Marconi that was done yesterday where one of the things, the good things about Marconi that was presented is that it is good when the message broker is not good enough. One case is especially security. We have guest agents running in virtual machines. We don't want those agents to talk to the central broker. So Marconi would be good for that use case where you can just set up a new service that doesn't have to take a high load of messages in your infrastructure and you will isolate everything from your message broker. So the security that is done in OpenStack right now is just by binding everything on the private network and we don't allow anyone to talk to that except for the services running in the OpenStack deployment. And like I said, we have some work going on to sign messages and encrypt messages before sending them through the wire. Great. Yes, I have another question. Do you have a way to make the dependencies between your services visible? Because when I see this communication bus, it looks very clear and simple. You just put a message on the bus and somebody else will get it. But in the end, it's just a way for the services to communicate with each other. And you can easily build a spaghetti dependency system by just using a very clean bus. So how do you prevent this? Logically. Like we don't have any assertion between services that say like, hey, we can depend on each other. It's just done logically when the design decisions are being taken. We cannot make service A to depend on service B and service B to depend on service A. So let's try to figure out a way to do that, which basically means create a service C, unfortunately. But yeah, it's done logically. Cycle dependencies, in my opinion, are bad, but they are not always bad. Like everything in software. But we try to avoid them as much as possible. It's all done logically. We have everything explicit. So since we know which service depend on each other, logically speaking, or function wise or feature wise, we know that we cannot create cycles in some of the services. Or we try not to. Can you use the mic, sorry? And that's explicitly written down somewhere in the code? No, it's not somewhere in the code. It is in the code, definitely, but it also has documentation about it, written on the weak pages and the documentation of each other's service. And the operations book, obviously, because you have to know how to play the whole thing. If there are no further questions, I'd like to thank the speaker again. Thanks. And thanks for attending. Thank you.
|
Flavio Percoco - Systems Integration: The OpenStack success story OpenStack is a huge, open-source cloud provider. One of the main tenets of OpenStack is the (Shared Nothing Architecture) to which all modules stick very closely. In order to do that, services within OpenStack have adopted different strategies to integrate themselves and share data without sacrificing performance nor moving away from SNA. This strategies are not applicable just to OpenStack but to any distributed system. Sharing data, regardless what that data is, is a must-have requirement of any successful cloud service. This talk will present some of the existing integration strategies that are applicable to cloud infrastructures and enterprise services. The talk will be based on the strategies that have helped OpenStack to be successful and most importantly, scalable. Details ====== Along the lines of what I've described in the abstract, the presentation will walk the audience through the state of the art of existing system integration solutions, the ones that have been adopted by OpenStack and the benefits of those solutions. At the end of the talk, a set of solutions under development, ideas and improvements to the existing ones will be presented. The presentation is oriented to distributed services, fault-tolerance and replica determinism. It's based on a software completely written in python and running successfully on several production environments. The presentation will be split in 3 main topics: Distributed System integration ----------------------------------- * What's it ? * Why is it essential for cloud infrastructures? * Existing methods and strategies OpenStack success story ---------------------------- * Which methods did OpenStack adopt? * How / Why do they work? * What else could be done? Coming Next --------------- * Some issues of existing solutions * What are we doing to improve that? * Other solutions coming up
|
10.5446/19969 (DOI)
|
Now with us is Federico Marani talking about Ansible and DevOps. Hello everybody, my name is Federico Marani and I'm going to talk about Ansible. Ansible is a DevOps tool and we've been using it for a while now, we've been using it for around a year and a half, kind of started with like, version 1.2 I think, and then we kind of went through like some experiments and then we kind of turned our infrastructure into Ansible scripts and then we kind of like worked a lot on that idea and then we kind of extended it to do like many other things like such as school deployments and a lot of like kind of server setup. I've been trying some other tools like those tools and Ansible is a nice compromise because it's quite simple and like it just works in the way like it. Okay, so just to give a bit of introduction, I'm a coder so I've been coding Python for a long time and I've been involved with some open source projects but like many company work. I code with like many languages and Python, did some scarlet, did some PHP in the past and also like always like the kind of the sysami side of things because it's quite nice when you know both sides, you can code, you can deploy, you can configure server. So I always had an interest in kind of sysami first and then moving towards DevOps. Our first company called Tri-Reese, it's a startup, we are based in London and I had a very good engineering there. Before that, they work with other companies, some of you may recognize some names. Okay, so what is the thing about? Like, you know, obviously we like talk about Ansible but like what is the real problem behind it? The problem behind it is DevOps and the problem behind DevOps is system administration really. You can do DevOps if you don't, if you don't understand sysami, obviously you can use Ansible but like Ansible is just a tool that helps you to do the same sysami work but like you still need to know sysami in order to do that. Like it just kind of simplifies some of the things but like certainly it doesn't tell you how to configure in Ginect, it doesn't tell you how to configure Sudo, it doesn't tell you how to configure any other tools that you might use in your stock. So, yeah, I mean basically this failure is lots of the complexity of system administration. So, you really need to know how these systems work and then you can use Ansible to simplify your workflow, basically. So, what is DevOps? DevOps is like basically like one simple concept and is having infrastructure as code. So, like just saying historically, you had like this figure called sysami which like go on the server and then do this, all this manual thing and like nobody, nobody knew what they were doing, they were like just doing magic on the server and somehow the server got into a state it was working and then suddenly these people, I leave the company and nobody knows like what they've done to the server and like we basically lose track of like all these changes. So, like what changed some time ago, especially when kind of DevOps became a big thing, basically now every change you do on the system is going through this process of coding and then you basically have the coding as a first step and then these deployments of infrastructure changes like you basically roll them out on one server or many servers like you know whatever number of servers you have. And what DevOps is about is about automation as well because I mean we are engineers so we like to, without when things are automated like this, this less, it's just pretty nice, you don't have to think about these things anymore, you just have this automation in place. People leave companies, so that's another good point, like you don't want knowledge to escape the companies, you want to keep the knowledge within the company and anybody can read a code, anybody can understand how the system works. Another point I feel strongly about and there's a lot of DevOps tool out there which like they borrow too many ideas from programming languages. So I mean I love coding, I've done a lot of it but I don't feel necessarily the connection between like doing DevOps and doing coding. So I don't think DevOps should not require primary experience. This tool is like, you know, Chef is really kind of Ruby based for example and the Papa does like, you know, entire its own language. Yeah, I think that this, you know, I was looking for a tool that kind of made this distinction a bit clear and that's when I came to Ansible. Ansible is a really nice tool, it's really quick to get started, it's really easy, it builds on top of like tools that we all should know like Python, SSH, YAML files, it's written in Python, it's extensible with Python, you can write plugins in Python, you can write plugins in other languages I think. Like differently from other systems, it's based on the idea of like pushing updates to the server instead of the server having the responsibility of pulling updates. Obviously this behavior can be tweaked but the idea is like you run Ansible on your machine or a management machine and it basically connects to every server and pushes the operation that you want to run on the server. This operation may be like installing packages or maybe configuring packages, you know, whatever you chose to run. Yes, this push approach. The nice thing about this is it does require you to install agents on the servers, it does require you to have a central repository of configuration files. Yeah, I mean it's just a different way. It's based on the idea of playbooks, so you can use Ansible in a simpler way but like I'm going to talk about playbooks and playbooks are basically a list of tasks, so there is like a task section I'm going to describe and there's another file called inventory file, so those are like the two basic files you need to set up. Okay, you can use Ansible for many things. I generally like tend to distinguish in two big groups. One is configuration management and if you want you can use Ansible for code deployments but like that's quite a separate thing. Configuration management, that's like the traditional way to use these tools. So when I say configuration management, I mean like installing software on the server, bring like this software, making sure all the demos are started, making sure all the network interfaces are up, making sure all the firewalls are set, you know, these type of operations are basically like the traditional way to use these tools. Okay, so this is a playbook, this is like how you kind of the basic file for using Ansible. It's divided in three sections, you got like cross section, task section and the section. So really the most important part here is the task section. In this playbook there are three tasks, each one with a name and each one with an action. Action is in this case is APT template and service. So I mean this file is really easy to put this file down, specifically set up on GNX server on your machine, on many machines. So the very first thing you would do is make sure GNX is installed and it's the latest version. So to do that you basically need one line, APT line, which is like there, you specify the package name and you specify the state you want the package to be. When Ansible runs this action, what it will do is basically check if GNX is already installed, if it's already installed and it's already the latest version, it won't do anything, it won't cause any change in the server. If it's not installed or if it's out to date, it will be upgraded. Okay, so second action, like some say you probably do like for most of the software install, you have to use your own configuration files because I mean for instance actually you need to set up like you have a file to like configure the website so like you need to upload this. Template is basically a copy but like with some pre-processing down to it, just templating. Ansible is using a GINJA, GINJA is templating engine, the same one used with Flask and same autoflask. What the template action does is takes this GINJA file and you like feed it to the template engine with a bunch of variables that are available in Ansible and you can specify also variables in playbooks and that the output of this kind of templated file is then copied on the server to that destination path there. Yeah, that's basically it. Third action here is the service action. Service action basically is an interface to Winning scripts and in this case it's just about like making sure GINJA is running. So if it's not running, we call the InScript to start GINJA. If it's running, it won't do anything. So then there's another important section here called endless. So the basic idea around endless is basically you list your endless and you can execute those on demand like at the end of your task list and you can execute them only if there's been a change like kind of associated to the, in this case, to the template action. So if the template action caused a change in the target server, it basically triggers the notify which is then connected to the endless. So in this case, if the configuration file changed on the server, you want to restart GINJAX. So the notify will be triggered and then the endless will be executed. It's just like basically they are action you execute only when a notify has been called. So the last thing here is the host section. It's actually a section you need, you always need and basically like the group of servers you want to apply this playbook to. All right. So task order is important. That's how Ansible works. That's not how other tools work. It's really kind of typical of Ansible. And I mean, to be honest, like it really makes sense, like from my perspective, when you set up machines, when you set up servers, like you would do things in order. So like it's quite convenient for me to, you know, to reuse the same order when I define when I define this task, it's really kind of a nice way to think about problems like in steps. So like really kind of say the thing as it's an same party programming. So it's not like, you know, the order is quite important. That's how you define dependencies between tasks. That the thing is task card is important. Meaning that like you can execute the playbook as many times as you want and you won't try to install it twice. You won't try to override the configuration file if the configuration file is already on the server. That's a nice feature. And basically like you won't change the system. You won't try to change the system if the system is already in the state you want that to be. I already can describe the endos a bit. Basically there are commands fly for later execution. So we execute only if there has been a change in the system. Typical case reloading a demo. Okay. So the last type of file that you need to set up is the inventory file. That would be like much simpler than a playbook. It's just a list of domains or IP addresses. You can group them by OS group if it makes sense to you. Like, you know, like what we normally do is like list all the web servers in one section. List all the database servers in another section. We have monitoring servers. We have a trillion of type of servers. Like one thing you can do with inventory files that we found really helpful is you can define like variables per inventory per OS group. For example, like for all the web servers you want to declare that the build environment is production or staging or whatever environment you have. You may want to declare some database names in case you're running multiple versions of the same website on one machine. You can do the same thing like for database servers. Yeah, that's the feature we use a lot, basically. Okay. So the important things here are like OS groups. Like really kind of trying to understand how OS groups work. I mean, it's actually quite easy. But there are some little things that you need to know. There's a feature called roles in Ansible. We found it like really helpful, especially because basically defines like a common convention like to include files within your playbook. So if you have like a long list of tasks you need to run on the server, you may want to split this task into multiple files. I mean, the same idea behind the programming language is you basically split your file into multiple files. That's why there are includes and you should use them. And always try to use the actual inside effects. Meaning if you don't need to template a file, you can just use copy. Basically, there's less chances to trigger a change in the server and the change you don't really want to trigger. Okay. So this is how do you find includes. These are actually like a real snippet of production code we use. So you can basically see the operations you do on the server on a more logical level. Obviously to store what packages there are like many packages or many configuration files you need to apply. Just try to see that from a kind of higher ground. So you basically store all the web packages and then we configure on GNX, then we configure on a supervisor because we're using supervisor. Another thing we do is have a day restore backups. Everybody should do like backup testing. But like we restore it from production to staging. So you can do conditional includes if you want. And all the tasks in this include will be included only if that variable is, you know, if that is evaluating the truth. I'm going to come back to conditional except to be more later. You can tag operation, tags meaning you can just write tag SQL and any keyword. That's quite a nice thing because sometimes you only need to execute parts of your Ansible scripts or you may want to ignore some tags and that's a nice way to do it. Okay. So that was like kind of the basics basically. We introduced a playbook, we introduced inventory files, we did some other things. But like the operations you do on configuration management, they are quite similar. Unless you do, you know, like lots of Java and demos to start or anything crazy like that. The other thing we use Ansible for are code deployments. The problem with code deployments is that they can be really custom. Like the really like one, like, you know, when you write things like configure servers, they are always, you know, they normally quite standard. Code deployments, like it's really personalized like for your, for environment. We have like, you know, a ton of Python. We have many based on Django. So, you know, this is actually like some playbooks that work well for us. But, you know, just to describe, you know, the basic things we do, we basically create virtual environments. We store dependencies in the virtual environments. We use Bower, we use like Node.js and we use Grunt, like to do compilation of facets like server side. And Ansible has some support for these tools, especially in NPM. It doesn't support Bower, but I can always run shell commands so we can run Bower store with the shell command. We use Grunt, so we just trigger a shell command to run the ground compilation server side. And there's some, you know, standard operation you do with Django, like you basically collect all the static files, you run the migrations, you want to run the migrations only in the case the migrations are not already applied. And, yeah, I mean, there is some setup to do that. We use a whiskey, so when we finish all this, we are starting whiskey, we're starting celery, we're starting everything we need. Okay, so basically in this extra code that we have, like, it gets a bit trickier because there's like, you know, there's many things that we need to add which are not standard. They already introduced the conditioners. Conditioners are basically can be applied on any task. And it's just like an extra line saying one and then an expression. Expression is all the way to using Jinja, so we got all the power of Jinja for free. In this case, like, we don't want, like, when we deploy something, we need to know what environment we're deploying. So we first failure if the app environment is not defined. Yeah, I mean, that's quite easy. We just want to know the environment. Okay. It's another operation that we found quite useful. We used this in a few places. It's a register operation. The idea behind register is that, I mean, you can register a variable name. And that variable name at the end of the execution of that task will contain some information about the task. So the problem we had is we want, like, I want to deploy production only the version of the website which are being tagged with the version. Because I know if they've been tagged with the version, they are stable. So they might be deployed. There's no task support in the snow magnetic model to do tasks, to read the git tags. But you can still use a share command. So we run git tag. We put the output of this command in the git tags variable. And then we can reuse that variable in a conditional later on. So in this case, like, the conditional is a production and the tag is not in the git tags list. In that case, we need to fail. Like, I don't want to deploy the production version that they don't be tested. Yeah. And one thing to add is git tags basically contains many properties. And one of these properties is the standard output. But I also contains X equals, like the time this command took, many other things. Yeah. Just go and check the website that's written, everything there. One thing we use a lot in a lot of places is with items. So sometimes you want to run the same action multiple times on, like, many packages or you're going to store, like, for example, like many packages or many debut packages. What you can do is repeat the same action, copy and pasting it, like, for every package. And it might make sense. This is certainly like a nicer way to do it. So you basically run the same action. It's basically a loop. You run the same action on many items. In this case, you want to store virtual element supervisor with pip. Another nice thing that there is in Ansible is something called facts. Facts are basically data coming from the server, from the current server. Ansible facts may be like the hostname, for example, or the IP addresses that is machinous or the mount points or a distribution name, distribution version, and data about CPU, about disks. And you may need to do some of these, some of the information, when you, for example, like, you know, write template files or you directly in Ansible playbook. In this case, like, we as a company use Ipset, and we, like, I want to let everybody know that, like, something is in the ploy on a particular server with Ansible as a model called Ipset. It's already done for you, so you don't need to, you know, write any Python code to talk to Ipset. You specify the room, you specify the message. The message is trained that can be a template. And basically what happens is this action will be run for every server that is in your playbook, and this action is run, you will get, like, a different message. So, deploy to, W-W-W, deploy to whatever your hostname is, and you get that as many times as you have servers. Okay. So, there's a lot of packages in Ansible. I mean, Ipset is just one of the many. And there are some which are being more standard than others. Some of them are really specific on, with EC2 or with this solution or, like, something, or, like, interface to backtracking system. The actions or modules that we normally use are APT because we're using, like, Ubuntu everywhere. Service, if you know, like, kind of interface with Inescript. PIP packages, if you're in store, PIP packages. We use Git. The Pro with Git is quite limited. You can only check our repositories. You can do any other Git operation, which are many. There's a file module if you want to check the presence of directories, if you want to check the presence of files or links. And some more modules specific to the Python world, like supervisor, for example, is one, or jungle manage their interfaces to run jungle management commands or supervisor commands. As I said, there's many more. So just to give you, like, an idea of the size, like, we have more than a thousand lines of playbooks. We do a lot of things with them. We have, like, around, we have four environments, some of them production environments, some of them staging environments. We actually have more than the production machines. I just added a couple. We were on, like, Posca Square, Neo4j, NGINX, Solar Machines. Basically, like, the way you set up all of them is quite similar. There's a bit of extra setup for Solar and Neo4j because they're based on the JVM. We have, like, that machines, like, everyone of the team has, like, both like local machine and the dissolution box so they can deploy anytime. We have some brass machines. We run, like, on multiple cloud providers, like AWS dissolution. Every fragrant box is set up with sensible. So basically, it's quite nice because you do a big up, up, and then you run step provision automatically. So you actually get, like, the final server. It takes a while, but, like, you actually get it. Yeah. I mean, that's, I'm cutting to the end of the talk. You know, like a few suggestions. Just try to keep server stateless if you have, like, you know, especially when you scale, like, you really need to, like, not, for example, I store a file on a particular server and not on another. Because that file will become, like, a state and then that's kind of the thing that will stop you when you have to scale and you have to have more than one web server or more than, like, you know, many web servers. And the nice thing about DevOps is it kind of allows you to do things in the right place. You can do, like, IP geolocation, for example, both in code or in server, like infrastructure level, just like models for engine X to do that. So you might want to configure engine X in a way that does geolocation for you or you can do it in the code. Yeah. I mean, I think that's probably it. Thank you very much. Thank you. Yes. Do you have any questions? Okay. So the question is I was reading the question. Do you have any experience with that? We, like, we use load balance, sorry. So you were asking about load balance certainly for use, like load balance and modules when we do deployments. So you're talking about a specific type of load balance or load balance or load balance. Well, yeah, my company really gets five, but basically questions rather be of an experience with as a load load. Okay. We use load balancer for some of the, actually one of the environments we have. Well, like, we still do it manually pretty much. But like this, the sort of support like for is it true load balancer and other type of load balancer. But yeah, I mean, we don't do it. We kind of do that process manually. So we take the machine off the load balancer and then we deploy to that specific machine and then do this thing manually basically. Okay. So the question was about if there is a specific convention about like kind of where you put files and, you know, as a rule, you are the thing that in kind of forces you a lot of this is a lot of convention. So especially when you use Ansible, Ansible rules, basically automatically gives you like a folder structure you need to, you need to follow when you kind of declared various sections. The thing besides that you can pretty much come out with a structure you want. Yeah. Hi. How do you control who gets to configure the structure? So how do I get to control who figures who like kind of does the infrastructure, sets up the infrastructure. You can like because everything is like kind of committed to repository can always use the repository to do that level of control. Who gets to deploy this? I mean, if you have the kind of SSAs permission on the machine basically means you can run Ansible. So like the control is really kind of built on top of SSAs. I mean, there are tools that you can put on top of Ansible if you know, especially when you use management servers and this tool they release on Ansible Tower. I don't think it's free or maybe it's free for like some like a limited amount of servers. But like as it stands, basically control is on SSH. So if you have the SSH key you can configure SSH. Sorry, how do you copy SSH key? The way we do it now is basically like I have the permission to like to kind of apply this configuration files, scripts on all machines. Setting up like a master server now to do that. Sorry, if you have an error in the playbook. So the problem is if there is an error in the playbook, let's just say for example like you are copying a configuration file and then like an error happens so then this didn't run like how do you do that? Yeah, that's really annoying. And what I normally do is you can either like force a change in the file or like I'm sure there are like, you know, more clever ways to do it. But yeah, I mean the store, I don't have a purpose to wish of it. Okay. That was it. Thank you very much again. Thank you. Thank you.
|
Federico Marani - Scaling with Ansible Ansible is a powerful DevOps swiss-army knife tool, very easy to configure and with many extensions built-in. This talk will quickly introduce the basics of Ansible, then some real-life experience tips on how to use this tool, from setting up dev VMs to multi-server setups. ----- Infrastructure/Scaling is a topic really close to me, I'd like to have the chance to talk about how we set this up in the company I work for. Our infrastructure is around 10-15 servers, provisioned on different cloud providers, so a good size infrastructure. Presentation is going to be divided in 3 parts, first part is going to be focused on comparing sysadmin and devops, then there will be an introduction to the basic concepts of Ansible. I want to spend most on the last part, which is going to give some tips based on our experience with it. Many ideas will come from this presentation which i gave at DJUGL in London, with a longer session I will have more chances to delve into more detail, especially on how we use it, from vagrant boxes setup to AWS and DigitalOcean boxes, network configuration, software configurations, etc... I want to offer as many real-life tips as possible, without going too much offtopic as far as Ansible is concerned
|
10.5446/19968 (DOI)
|
Okay, Eric is going to tell us all about Bitbucket. With a focus on git, Eric? Is it judging by your shirt? Just a bit. Yeah, with a focus on git. All right. Well, it's a lot of faces. More than I think I've ever seen in one room staring at me. We'll see how this goes. So I'm Eric and I am with Atlassian. And I work on Bitbucket. I'm one of the more back-end developers on Bitbucket. And I'm going to tell you all about Bitbucket's architecture and infrastructure. Or at least as much as I can in 30 minutes. Before I do that, though, I want to share with you this photo. And those who don't instantly recognize the rocket here, this is a Saturn V rocket. It's the rocket from the Apollo program. That's the moon rocket. So that's the one that, I mean, not that particular one, but got Armstrong to the moon and back. And I want to show it to you because it's like the whole Apollo program is, I find, like a fascinating piece of history. And I'm sure I'm not alone here. It's like this rocket when they built this. And I guess the program around it. We're really sort of the pinnacle of innovation and engineering at the time. And a goal that they set out to achieve was so ridiculously ambitious in the 60s, sending a man to the moon and bringing him back. When, I guess, the state of the art was the Russians who had just flung a chunk of metal into orbit. There's quite something. Enormous undertaking. I think at some point, like 500,000 people were working on it. Ridiculously large. Billions of dollars. But it worked. And so you must have assumed that only really the smartest people worked on that and were able to pull this off. So quite literally, rocket science. And I'm a bit of a nerd. And earlier this year, I actually went to Florida and visited the Kennedy Space Center in Cape Canaveral. And they've got one of these things on permanent display. So there's a, here it is, an actual remaining Saturn V rocket that they've taken apart into the separate rocket stages. So you can see it up close. You can see sort of what's inside, right? And what struck me when I was there and I looked at this. The first time I'd ever seen this stuff is that it looked sort of, I don't know, simple. Maybe that's not the right word, but rudimentary perhaps. As in it was very functional. Look at this thing. It's like a sheet of rolled up metal around a, I guess, a massive gas tank. I mean, it's really not much more there. I mean, there's some plumbing, but even that is limited. And I guess I never really considered what would be inside of a rocket like that to be able to do the things that it did. But I guess I sort of expected something more complex, more ingenious. I don't know. It's a similar story at the back or the bottom. It ends here. There's a flat surface and we'll just bolt some engines at the bottom. If you're there, you can actually see the engine mounts, like the screws and everything. It's not really polished. You see bolts protruding everywhere. And now I don't mean to disrespect the Apollo program, by the way. I mean, it's still as amazing as I thought it was. But it's sort of seeing this stuff up close, sort of, I don't know, made it more approachable. It brought it down to earth, if you will. And I think that is representative of how we tend to perceive technology that we have in high regard, but we don't really know much about. We tend to assume that things are more complicated than they really are and that the people working on it are, by definition, much smarter than we are. The whole grass is greener thing. And it is that potential perception that I want to debunk today by laying out the architecture behind BitBucket. And also, at the same time, share some anecdotes and, I guess, some of the instances where we screwed up. So if you are a little bit like me and you tend to assume that other people are smarter than you, then you'll be glad to hear that there's really no rocket science behind BitBucket and everything that is running now is built around the same tools that you will use yourself. You sort of try to break it down a little bit. This is roughly the architecture of BitBucket. I've separated into three logical areas. So there's the web layer, which is responsible for load balancing, high availability, that kind of stuff. Then there's the application layer. That's where our code is, that's where all the Python stuff is. BitBucket is almost exclusively written in Python. And then lastly, the storage layer where we keep our repository data and all that. So talk about each layer individually and time permitting. I'll share some anecdotes. So the first layer, the web layer, consists really of two machines only. So BitBucket is all, there's no virtualization in BitBucket. We run real hardware. We manage them ourselves. We have a data center in the U.S. and we have two load balancer machines. They own the two IP addresses that you see when you resolve BitBucket.org. These machines basically run Nginx and HAProxy. Web traffic that comes into the load balancer first hits Nginx. Nginx is a, for those who don't know it, it's an open source web server. It's pretty good at SSL. It can also be used really well for reverse proxying. And that's what we do here on this layer. So when a request comes in, it is encrypted. Everything on BitBucket is always encrypted. So the first thing we do is strip off the encryption. And that's done using Nginx. And then once it's decrypted, we forward it on to HAProxy, which runs on the same machine. HAProxy also an open source reverse proxy server, but it's really good at doing load balancing and failover. We have a whole bunch of backend servers. And so that HAProxy inspects the request and based on some properties decides how to forward it on. And ultimately it will forward it on to one of our many actual application servers. And on there, there is another Nginx instance. So this Nginx instance is also just a reverse proxy server. It's not our actual web server. It takes care of things like request logging, compression, response compression, and asynchronous response and request buffering. And that's why logically it's part of the web layer because it doesn't actually process the request. And then ultimately that forwards it on to the real Python web server on the application server. Now so that's HTTPS. We also do SSH. And SSH takes a bit of a different path. SSH is a different protocol. We can't easily decrypt it first. And so we do need load balance it. So it does go through HAProxy. Just as a TCP connection and HAProxy now forwards it on to the least loaded back end server. So that path is a lot simpler. But make no mistakes. It's not necessarily easier to run that reliably as we found out really just recently when users started to complain about SSH connections dropping out sometimes. Like users would say that they get hung up on. And looking at the error messages it seemed like that was indicative of a capacity problem. Like we had not enough capacity on the server side basically to handle the request rate. But our monitoring tools told us a different story that we had plenty of capacity. So we're stunned for a little while until we started analyzing the network traffic on the load balancers. In particular we looked at the frequency of SIN packets that were arriving. So SIN packets are part of TCP and it marks the start of a new TCP connection. And so time stamping those, each single one of them gives you a really good accurate view of the incoming traffic. You see that here. So what you see here is an interval of 16 minutes over which we captured every SIN packet. And you can see right away that it is ridiculously spiky. And these spikes are, aside from being very high and very thin, are also very evenly spaced. If you count them you'll see that there are 16 spikes in an interval of 16 minutes. In the SNO coincidence, these spikes occur at the start of every minute. Like precisely at the start of every minute. They last about one to two seconds only. But you can see that the rate at that point is ridiculously high. It's like three to four times higher than our average load. Our working theory behind this is that this is the result of thousands of continuous integration servers all around the world that are configured to periodically pull their BitBucket repos. That in combination with NTP, everybody I guess these days uses NTP and clocks are really accurate. This is what you get. And that was a bit of a problem because even though we have enough capacity for the average rate, during these spikes we actually don't have enough capacity. Now solving this, it's not, we can't really quadruple our SSH infrastructure to be able to deal with the large spikes. So we went back into the web layer into ACA proxy where we basically have a hook into that traffic that comes in. We configured ACA proxy to never forward traffic at a rate higher than what we knew our capacity could take. But then don't make any changes to the ingress side. And so during these spikes ACA proxy will happily accept all the incoming traffic, but it won't actually connect or forward all of the connections at once. And so it sort of spreads it out over a few seconds. And now this graph on the application servers is a lot smoother than it is on the load balancer side. You could probably see it if you have a cron job and you started at that very start of the minute every time versus any other second. It'll probably have a few seconds less lag. So it's a bit of a funny problem. Never really considered it until it crapped up. Probably won't have it really with websites that have humans click on links. But if you operate like a public API that is very popular and people script against it, you might see similar issues. So the application layer done. This is where all the magic happens sort of. This is where the website runs. And this layer is distributed across many tens of servers, real servers. They all run a whole bunch of stuff. They run the website. The website is a fairly standard Django app, really. Like Bitbucket started out as a pretty much 100% Django app. And it's still very important. We run that in Gunnycorn, the web server. Gunnycorn, Python web server that is relatively simple. We run it in perhaps the most basic configurations. We use a sync worker, meaning that our processes process one request at a time. So we have a whole bunch of processes and multi-processing to get concurrency. And then SSH. We handle SSH using really just the standard open SSHD server daemon, the same one that you all run in your Linux machines and laptops, with one difference. So we made a small change to it, a small patch that allows us to use the database, to look up some of the database for public keys. OpenSSHD looks on the file system. It's hardwired to look at the file system to find public keys. It's not practical for us, and so we have a little change to make that happen. Other than that, it is the standard openSSHD server. So we don't need to maintain that. We also do background processing. So any sort of job or process that we can't guarantee will respond in a few milliseconds. We dispatch off to our background system. That's comprised of a cluster of high available RebitMQ servers. RebitMQ is an open source, early implementation of AMQP broker. And to consume jobs, we use celery. So we have a whole farm of celery distributed across all these machines to process these jobs. The examples of that are if you fork a repo, for instance, there's actual copying of files involved so that it might not complete immediately. That's dispatched. So it looks very basic. At the end of the day, it's the same components that you all run, distributed statelessly across multiple servers. There's nothing really special about it. It's simple. It's usually good. We've had this set up for years. And BitBucket is now, I think, over 35 times bigger than it was when we started, when we acquired it, I should say. And this held up really well. However, you can still screw up as we do from time to time. And one of those examples was when we decided to upgrade our password hashes. So up until that time, we never stored passwords in our database, but we stored salted, shawan hashes. It's very common. So it means that if somebody, for some reason, gets a hold of our database, they only have hash values until they still don't have your password. However, shawan hashes for passwords are slowly being phased out and replaced by more strong, secure hashing algorithms. The reason for that is that even though you can't decrypt a shawan hash to get a password, what you can do is think of a word that might be the password, compute the shawan, and then compare it with what's in the database. If you just think of enough words and try enough combinations, you might brute force the password. If you have a strong password, chances of anybody brute forcing that through shawan are going to be careful with some cryptographers maybe here in the room, but let's call it negligible. However, we have millions of users on BitBucket, and not everybody has a strong password. If you have a password that is a word in the dictionary, then it's a whole different story because there really aren't many words in the dictionary. Certainly not when it comes to a computer computing shawan values for it. So you're really at risk. Now, short of forcing people to not use simple passwords, another thing you can do is sort of upgrade to a stronger hash. Now, what these things do, nothing special really, but they're hash values or hash algorithms that are more expensive, deliberately more expensive by rehashing their hash value over and over again thousands of times, deliberately spending more CPU cycles. And that's what we wanted to upgrade to. And let me show you just how big that difference is. So we wanted to upgrade to Bcrypt hashes. Bcrypt is one of the sort of more modern iteration style cryptographic hash algorithms, and we compared it with shawan. So this script measures how many hash values you can generate in one second. And for Bcrypt, you can see that my laptop was able, so this uses Django's code, so Django's hashing algorithms, all that code, with the required or required, with the optional C extensions to make it as fast as you can. And my laptop that is amounts to three hashes per second for Bcrypt versus 160,000 for shawan. So it's five orders of magnitude more expensive, just CPU cycles. And so that is absolutely huge. And it's great because it means that even your weak password may stand a chance. But you have to realize that as a server, you have to incur that cost of that massively expensive calculation every single time somebody uses a password for authentication. We run a really popular high-volume API, and a lot of people use basic auth for authentication. It's all SSL, right? So it's not plain text passwords. But it means that we have to compute Bcrypt for every single request. And our API requests are relatively quick, like on average, in the tens of milliseconds. So you can imagine that if you add a 300 millisecond password check to every single one of these requests, you have a problem. And we did, because we naively rolled this out, and instantaneously the website went down. All cores and all the machines and all CPUs went to 100% and were calculating Bcrypt. Now, we realized our mistake fairly quickly, obviously not quickly enough, but fairly quickly. So we rolled it back and were able to keep the downtime minimal. But then we had a bit of a problem, because we still wanted to move away from SHA-1. Now, you can't really make Bcrypt cheaper. Actually you can, but you don't want to, because the whole point is to have an expensive algorithm. And so we could do, however, is do less of it. You know, when people use the API and write a client, they typically do more than one request in quick succession. And so they'd be using the same password over and over again, and we'd be computing the same Bcrypt over and over again. And so we decided to implement a sort of two-stage hashing system, where when a request comes in, instead of computing the expensive Bcrypt, we now compute an old-fashioned salted SHA-1 value, and then we use that to look up in a in-memory map dictionary the Bcrypt value. If that's empty in the beginning, we then compute the Bcrypt value ourselves, check it against the database to see if your password was correct, and then store that mapping, like SHA-1 versus Bcrypt, in the in-memory table. Then the next request that you make with the same password is able to look up the Bcrypt value from the in-memory cache. That way we're able to cut out, I'm guessing, 99% of all the Bcrypt calculations. But it's important to understand the ramification of this system, because you might be tempted to think that, well, you've now sort of weakened your Bcrypt authentication back down to SHA-1 strength. There's that whole story. The important thing, or the main thing, I guess, is that SHA-1 values never hit cold storage anymore. So the database is all Bcrypt. So if you get a hold of a database, you still only have Bcrypt. And even if you were able to somehow tap into our servers and get a hold of, like, copy memory access, then you'd get some SHA-1s. But you'd only get SHA-1s from the users that are active in that very moment, because these cache entries, that's essentially what they are, are expunged very, very quickly. So we're able to get this thing running in upgrade to Bcrypt. But even then, the remaining 1% of the time that we spent on Bcrypt is still very significant, very significant. Just look at that ratio, right? 160,000 versus three. And right now, today, if you look at one of our servers, and you run like a perf top or something, you can see that the Bcrypt Cypher method is the most expensive method that runs on that machine. And at any point in time, I think it eats like 12% CPU. So it's still hugely expensive. So in the future, I guess, we should probably be looking at migrating or offering something like an alternative to basic auth, maybe, you know, like, standard, relatively standard HTTP auth tokens, which are revocable and have a limited privilege set. And now let's move on to the storage layer. So here we keep track of your data, obviously. The biggest amount of data that we store, of course, is the contents of your repositories. There are millions and millions of repositories. We decided to keep the storage of that as simple as we could, like sort of in line with everything else that you've seen so far. And we decided to just store that stuff on file systems, just like you do on your local machines, right? Git and Macorial were designed for file systems that work really well. As a post, for instance, to modifying Git and Macorial to be able to talk to, like, a distributed cloud-based object store system of some kind, for instance, what Google Code does, we decided to keep it simple. The file systems that live on specialized appliances by NetApp, commercial company, are accessible from the application servers, simply using NFS. And then aside from that, we have other, so we have sort of no SQL storage, like distributed, like map systems. We have Redis and Memcached. We use Redis for your news feed, like the repository activity feed that you see. And we use Memcached for basically everything that is transient that we can lose. And then the data for the website is all stored in, or traditionally, just in SQL. So we use a PostPress SQL. The data is manipulated and accessed basically exclusively through the Django ORM, and that works pretty well. The only thing is that SQL databases, PostPress is no exception, are generally kind of hard to scale beyond a single machine, transparently, I should say. So unless you go implement application-level sharding to separate your data across multiple databases, transparently scaling an SQL database across multiple machines is an entirely trivial. So far, we've kept things simple. We are running a single PostPress database. It's a very, very big machine, and it has no trouble with the load at this point. And then for high availability, we have several real-time replicated hot slaves on standby. But yeah, in the future, should that thing ever sort of become a bottleneck, which hopefully it will, because that means service is popular, I guess we have to look into sharding. I could talk about this stuff all day long, and I wouldn't mind to do so either. But there's only 30 minutes. There's only a few minutes left at this point. So I want to leave it at this. If you have any questions, I'm happy to take some now. We don't have a lot of time, but I'll take some now. Otherwise, come chat to me afterwards. We also have a booth in the lower level, and so you can just find us there. And otherwise, we have a... I'd like to invite you for a drink tonight. We are hosting a drink up in a bar nearby, starting at 7 at Main Haasamzee. So I'd like to invite you over. Come have a drink on us, and you can talk all about this stuff. There's two of my colleagues, too. And we're also hiring, so if you want to talk about that, that is also possible. And with that, I want to thank you very much for listening, and I hope to see you all tonight. Thank you. Thank you, Eric. Would the next speaker like to come up and set his slides up if you'd like to take any questions over there? Yep. So any questions? Hi, question about the HA Proxy and NGINX in the beginning of the request. So HA Proxy actually has the SSL support. Have you tried that? Yeah, it does. So our setup on the WebLayer is a little convoluted, maybe. There are a lot of components, as you saw, and that's not strictly necessary. Part of that is sort of organic growth in historical. So HA Proxy hasn't always been very good at SSL, at least not in our experience. We've experimented with a ton of different SSL terminators. We've used Stunnel, a bunch of others, and at some point we found that NGINX was, at least for us, the most reliable. And so we've left it there, and I know the situation has definitely changed in HA Proxy, so it is something that we intend to revisit at some point in the future. So yes. Thank you. Thank you. Hi. So I noticed that Bitbucket uses quite a lot of JavaScript on the web server side, and I wonder if you use WebSockets, and if you do, what do you use for them on the server side? So do we use WebSockets? No, we don't currently use WebSockets. We've experimented with WebSockets quite a bit for things like real-time notifications in pull requests, for instance, those kind of things. But no, we're not currently using WebSockets. Okay, thank you. Anyone else? Yep. Yep. What's PgBouncer? What are you used for? PgBouncer, you said? Yeah. All right, so yeah, so I had a whole spiel about PgBouncer, but I wasn't no time to go into it, and there's not enough time to go into all of that right now either, so I invited to chat later on, but PgBouncer is a Postgres connection pooling daemon, and so we use Django. Django, by default, doesn't come with any connection pooling, and so getting Django to talk to a database efficiently is a bit of a challenge. It's not a challenge, but you need something else. So PgBouncer is part of the Postgres project, and basically what it does is it makes stateful connections, long-lived connections to the database, a limited amount, and then you configure Django to talk directly to PgBouncer. It acts like a database, and so then if Django open and closes connections at a very high rate because you're serving a lot of connections, that is a lot cheaper than open and closing actual database connections, and so it bridges between the two to make that more efficient and to also be able to limit the total number of connections that you end up having on your database. There's a lot more to it, by the way, so we actually, if you noticed, we have two layers of PgBouncer, and there's a good reason for that, but yeah, as I said, I'll have to talk about that afterwards because there's no time. Cool, thanks. Yeah, no worries. Hi. Hi. You talked about the machine with the database, like a little large machine. As far as I could understand, that's physical machine. So what happens if that goes down? Yeah, so if the physical machine goes down, we have several real-time replicated hot slaves, so we're streaming replication for Postgres to have a bunch of slaves. So if the machine goes down entirely, then we steal the IP, basically, and almost instantly, hopefully move over to the other. Yes, it never happened, by the way, but yes, it's configured that way. All right. Thanks a lot.
|
Erik van Zijst - The inner guts of Bitbucket Today Bitbucket is more than 30 times bigger than at the time of acquisition almost 4 years ago and serves repositories to over a million developers. This talk lays out its current architecture in great detail, from Gunicorn and Django to Celery and HA-Proxy to NFS. ----- This talk is about Bitbucket's architecture. Leaving no stone unturned, I'll be covering the entire infrastructure. Every component, from web servers to message brokers and load balancing to managing hundreds of terabytes of data. Since its inception in 2008, Bitbucket has grown from a standard, modest Django app into a large, complex stack that while still based around Django, has expanded into many more components. Today Bitbucket is more than 30 times bigger than at the time of acquisition almost 4 years ago and serves Git and Mercurial repos to over a million users and growing faster now than ever before. Our current architecture and infrastructure was shaped by rapid growth and has resulted in a large, mostly horizontally scalable system. What has not changed is that it's still nearly all Python based and could serve as inspiration or validation for other community members responsible for rapidly scaling their apps. This talk will layout the entire architecture and motivate our technology choices. From our Gunicorn to Celery and HA-Proxy to NFS.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.