scale.bythebay.io: Stepan Pushkarev Interview
Recording: scale.bythebay.io: Stepan Pushkarev Interview
[Music] I'm Stephan I'm CEO of hydrosphere IO we are mostly focused on the production izing machine learning pipelines so it's pretty new fields field it's pretty wild and besides engineering engineering research that we are doing constantly how to optimize them how to provide better user user experience for data scientists and machine learning engineers there are other challenges like for socializing the ideas and getting a support from the thought leaders of the community here yeah just because it's creating you and right now everybody is focused on me mostly on training on training machine learning models and educating themselves how to how to even apply the machine learning and as far as well like jumping a little bit ahead of the of the crowd yeah it's very challenging even to find the right the right research articles in academia and yeah we just we're crying crunching a very low-level stuff to do that to make things right so for scaling or scaling I would say I would say so just to scale just an ordinary traffic it's it's not a big deal because company like delivering a high load software for years but scaling traffic for machine learning it's a lot of different challenge yeah we could use a best practices from that world but it should be somehow rethought and rebuilt for the data intensive and machine learning intensive applications okay so for example GPU versus CPU so the GPU is not widely used in production okay which it's used for training but it's not like a best practice right now and how to how to how to simplify how to make it cost-effective to use it in introduction it might be like any interesting interesting challenge for cloud providers because it's it's definitely doesn't make sense to to do it yourself yeah so it's probably something that that will come pretty pretty soon probably in AWS conference in two weeks so the the challenge probe also the challenge is in scaling the data flows when you you you can compare so do you have a java application it receives like a HTTP request and it's pretty tiny it just process some apply some business logics and return a result so it's pretty stateless and so there are best practices how to do how to deal with that but with the machine learning if you receive like a image Qwest that just like some binary data so it's they the the data is much more is it's bigger bottleneck rather than computing power how to help how to provide these coaches so the did yeah as I mentioned it it's it's challenging to scale the data at the data flow rather than computing power behind that so it's a business like the chip the challenge by IC and there is no exact solution for that so well we all like so we all like like building and researcher like you know like in a flat both buffers in a in other like inter process communication so how to build the how to build the micro services and minimize the the latency and data movement between these micro services because in a data intensive application it becomes a real bottleneck and for scaling data-intensive applications so yeah [Music] yeah so functional programming of course Italy of course so we use functional programming you know you know when we when we develop our systems we develop our programs but it really helps and change the the thinking about different machine learning problems in a functional way so for example for how to explain data scientist what should you deliver to the to developers in order to get make a deploy to production for example it could be if the data scientist could deliver script notebook or something else in a program just a program and if we tell a design data scientist can you provide a function a pure function without any side effects it just has an input and output so we can control what happens with that function introduction and we can optimize it as accordingly it it really helps even in communication level in how to guess it's in intuitively clear for everybody what function is everybody like learned that in the university and so providing such kind of functional framework for everybody in the team not even a data scientist and in a day for for data engineers in order to process that data the the data stream or data batch you need to write a function but whatever and you don't need to and the less side effects you will have an in that function the less queries you will send from that function to these 2d like external databases the easiest you it will be easier to support and maintain maintain that code and in the future so that's that's a very everything is based on a functional paradigm in our world so yeah this is like geekiest community in bay area I would say I would say for sure pirate and I hang out in multiple meetups here I talked to the people and sometimes even if we have ten people showing up on a meet-up these meetups are much more interesting than road shows with the keynote speakers like strata or whatever so this like a narrow and very focal focused community is really driving driving things [Music]