 Hi guys. Hi, I'm Manu. I work for the software architect as from flip card today. I'm going to talk about decisions in platform platform. We have built for scaling models. We have had great success. I want to share our success story with you. So what are the challenges people in flip card for working on the business and but are doing analysis of business data facing right. So firstly, the users spend a lot of time in data plumbing. They download large data sets on their laptop. They run. They run our scripts. They run Python scripts sometimes doing data analysis on top of large amount of data on their local laptops or even in a remote machine might not even be possible because of the amount of data involved. So what kind of models are we talking about? Here we are talking about some simple formula, maybe some linear regression model to what we do in demand planning specific use case that we are working on that uses feed forward neural network. So how do you scale something as complex as feed forward neural network to simple formula and give very basic constructs to your business users and data analysts and scale their use case. So the problem statement that our platform faced is something like this or users are very different. So we have data scientists, people who typically come from programming background. And we also have business analysts, people who come from an MBA. They started rating Excel macros and now they have graduated to this programming language are so the language and run times and libraries. They're like very diverse and the user persona is also very diverse. So target users, they have a widespread requirements. So we have like running optimizer for supply chain network where we'll run integer linear programming to find out what are the best routes between cities to send goods. So that's something that we work on and also working for it also trying to compute price elasticity of a product that is, you know, like given a price, how much will this product sell. So we came up with very minimalistic constructs. We didn't want to build a ML platform or anything. So the scale problem is there in all of these use cases. We have this use case of running business logic and machine learning models on large amount of data. So the solution that we have built right now it is called DSP decision sense platform. So what we have built is it's a workflow as a service essentially a workflow which is language agnostic and it's runtime agnostic. So think of a workflow where you know it has let's say three steps and step one runs in python step two runs in our and step three again runs in python. So each step of the workflow has its own container and hence its own runtime and it is linearly horizontally scalable and also the data scatter gather happens through the platform. So let's say step one runs for let's say all users and now let's say you want to run step two for specific segments of user. The platform actually takes care of data scatter at step one and data gather at step two. The interfacing is something that we wanted to be as simple as possible. So we even even as of now we don't have a web based UI for our platform. So we have given a command line utility. It's literally a jar and that jar actually runs out of users laptop. So the reason we took this approach is we want users to be as comfortable within their ID. So people are comfortable with pycham. That's an ID for python. People are comfortable with our studio. That's an ID for our. So we want them to be within their comfort zone but also encapsulate all the higher level constructs within the platform. So the client which is a command line tool which is a thin client that runs out of users laptop. It encapsulates get so the client pushes the data into cloud and it executes the script on the cloud and makes the data available for user. So these are the frameworks that we use. So misos is at the center of all this. So misos does the container management views docker to specify the container which language and which packages that we would need in the container. So Hive is used for data preprocessing and post processing and then we have Hadoop as a runtime. So this is the architecture. We're out of time. Thanks for listening. So this is the architecture. So we have a client that is in running out of users laptop that talks to a web based service. The MySQL holds all the metadata. So how many workflows are onboarded and what are the actions on those workflows. And we use us command for our workflow management as command is a linden developed framework. Then we use like I said apache misos which does the container management. So misos has this thing called roles. So you can specify I now want a container. I that contains two cores and these many GBs of RAM and we would run that script on that spec of container. So the actual script is read from GitHub. So that is the enterprise software that manages Git repositories for Flipkart. The script execution happens within the container.