text stringlengths 0 2.35k |
|---|
Outro: \[58:05\] to \[59:55\] |
• Introduction to a podcast discussing feature stores with Mike Eastham |
• Explanation of what Tecton is and its role in building systems for production machine learning applications |
• Background on Mike Eastham's experience working at Google on indexing and serving infrastructure for web search and his current role as tech lead at Tecton |
• Definition and explanation of a "feature store" and its purpose in managing feature definitions and data within an organization |
• Discussion on what constitutes a feature, including engineered inputs that go into machine learning models |
• Explanation of the objective of using feature stores, including making features reusable parts of the workflow for building multiple models |
• Features are inputs used for training and prediction in machine learning models |
• Features can be generic or specific to certain products and user combinations |
• Feature stores support various techniques for building features and provide a clean seam between data scientists and operators |
• Performance requirements differ between model training and prediction contexts, with latency being more important in online contexts |
• Feature stores address the need for faster everything, including fast search results and predictions |
• Both data scientists and operators use feature store systems |
• Feature store to manage features and ensure consistency between training and serving systems |
• MLOps vs AIOps: terms referring to the same concept of managing machine learning pipelines |
• MLOps pipeline complexity: longer process with multiple stakeholders and roles involved |
• Differences between MLOps and traditional DevOps: loopback from deployed model to retraining, more complex skills and roles required |
• Terminology debate: MLOps vs AIOps, one-word or two-word formatting, capitalization conventions |
• Use of Go in MLOPs: application in online surveying interface for low-latency feature value requests |
• Hybrid approach to data processing with some prematerialized data and final aggregations at serving time |
• Importance of low latency in data processing and consistency of tail latency |
• Use of Go for performance tuning and troubleshooting due to built-in tracing and profiling tools |
• Embedding Python interpreter within Go server for on-demand transformations |
• Consideration of compiling Python to LVM bytecode for linking into Go binary |
• Supporting internal users' familiarity with Python as a primary tool |
• Discussion of Go's suitability for Machine Learning Operations (MLOps) due to latency sensitivity and availability of tools for troubleshooting |
• The speaker Mike Eastham discusses his work with Google and the use of Go programming language in various projects. |
• He mentions that a metric pipeline was written in Go to process logs from thousands of experiments at once. |
• Feast is an open-source project that has some Go mixed in, but is more Python-focused. |
• Natalie Pistunovich mentions contributing code to the Feast repository for those interested. |
• Mike Eastham shares his unpopular opinion: he doesn't like maple syrup on pancakes or waffles. |
• The conversation then devolves into a humorous discussion about maple syrup and Canadians. |
**Natalie Pistunovich:** Hi-hi! I'm joined today by my co-host, Johnny. Hi, Johnny. How are you doing? |
**Johnny Boursiquot:** Hello, happy new year! This is like my first episode since the year started. It's the first episode of 2022. |
**Natalie Pistunovich:** Wait, is this really just the first episode of this podcast for the year? |
**Johnny Boursiquot:** No, I'm sure we released some already. |
**Natalie Pistunovich:** But this is the first one recorded in the new year. |
**Johnny Boursiquot:** For me, at least. |
**Natalie Pistunovich:** I'm not sure there were other episodes, but we will find out. Mystery. And we are joined today by Mike Eastham. Hi, Mike! It's great to have you. |
**Mike Eastham:** Yeah, it's great to be here. And you are joining us from the overseas side, or from Johnny's side of the overseas; so you two are based in the U.S, but you are based in Brooklyn... |
**Johnny Boursiquot:** Maryland. Not too far, actually. Both East-Coasters, yeah. |
**Natalie Pistunovich:** Is there a time difference between you two? |
**Johnny Boursiquot:** Nope. |
**Natalie Pistunovich:** Okay. Well, we have six hours apart to here, to Berlin, so that's 9 PM here. Fun way to start the year for me. \[laughter\] So Mike, you are a tech lead and a software engineer at Tecton. |
**Mike Eastham:** Yeah. |
**Natalie Pistunovich:** Or TectonAI, I should say, because the name of the product, as you told me, is Tecton. And Tecton is building systems to operate and manage feature pipelines and datasets for production machine learning applications for all sorts of customers that we all know... And you used to be a googler. At... |
**Mike Eastham:** \[04:08\] Yup. |
**Natalie Pistunovich:** Fun. So can you tell us a little bit on where you work and what you do there? |
**Mike Eastham:** Yeah, so Tecton is a company that was founded in 2019; the co-founders were all at Uber prior to starting the company, where they worked on a feature store product (or internal product) called Michelangelo. I think we'll get into it a little bit later in the podcast what it means to have a feature sto... |
**Natalie Pistunovich:** And I will pause you for one second to say that all the different names that you're mentioning, for example Project Michelangelo, we will include in the show notes. So if you wanna read more about this one product, you can find it in the show notes. |
**Mike Eastham:** Yeah, definitely. So yeah, the company was started around 2019; I joined in March of 2019, I was like the first engineer to join outside of the founders, so I've kind of worked on pretty much every part of the product since pretty near the beginning, when it was just very early prototypes, to where we... |
So the company is like 55 people now, around there. We've got a couple different sub-teams, and the engineering teams. So I guess we're kind of a mid-sized company now. Our main product is still this proprietary feature store that we've been working on since the beginning, but we've also started contributing to another... |
**Natalie Pistunovich:** So both the products, the enterprise and the open source one, are feature stores. |
**Mike Eastham:** That's right. |
**Natalie Pistunovich:** What's a feature store? |
**Mike Eastham:** I think it's a very buzzwordy term right now, so I think you'll probably see a lot of people trying to explain it in different ways, or having a different idea about what it is... |
**Johnny Boursiquot:** We just went through the whole serverless thing, so we're accustomed to -- |
**Natalie Pistunovich:** New year, new term. |
**Mike Eastham:** Yeah. I'll do my best to give my take on what it is. Feature stores - they're part of this broader concept called MLOps, which you mentioned before. It's all about kind of applying a devops style mindset to getting machine learning applications into production. Feature stores aim to solve one piece of... |
So you could kind of break down into a couple different components what a feature store is. I'd say the primary thing is that it acts as a source of truth for your feature definitions within your organization or your team, whoever is working on your model. Basically, the idea is that you just have a single place where ... |
In our case, we have data scientists and data engineers define those features generally as a SQL query, which is kind of a tool that those types of roles are typically familiar with. The idea is that they can do it in one place and not have to prototype it using SQL, and then later hand it over to another team that bui... |
**Natalie Pistunovich:** I will pause you just here to go one question deeper, and ask, for those who really focus on backend web servers, and that little world - what is a feature? |
**Mike Eastham:** Yeah, that's an important question. You can probably have a very mathematical definition of this, but generally, the features are sort of the engineered inputs that go into a machine learning model. So when you're in the process of making a machine learning model, you sort of have an algorithm that ta... |
\[08:16\] So if you have a bunch of, let's say raw data about users in a system, and maybe you're trying to predict a product they might be interested in buying, as someone who kind of knows a bit about or predicts what people might wanna buy, you might say, "Okay, I think their age might be an important determinant of... |
**Johnny Boursiquot:** Is the objective of a feature to make it a reusable part of the workflow for building these models? Or what is the objective of having a feature? |
**Mike Eastham:** Yeah, so I think definitely in an MLOps context you might say that you want it to be reusable... You can certainly build machine learning systems where all of the work is just done in a silo, and you don't reuse it across different models ever, but you're probably gonna have better outcomes for your o... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.