text stringlengths 0 2.35k |
|---|
**Mike Eastham:** But I usually do like a capital M, capital L, capital O, and then the rest lower-case. I don't know what the specifics are on that, but I think some people do it all-caps, and there's a bit of a disagreement there. |
**Johnny Boursiquot:** Any other way, then it's just Emlops... \[laughter\] Or mlops. |
**Mike Eastham:** Yeah, it could be confusing. |
**Johnny Boursiquot:** "What is this mlops thing I keep seeing everywhere?" \[laughs\] "Is that a new product?" |
**Natalie Pistunovich:** I'm happy to say that very few people actually write with a space, but I guess that kind of makes sense, because you already write devops as one word, so it kind of continues that one. But if anybody was curious about that, here's all the information you need, and the terminology. |
**Johnny Boursiquot:** I will say, I am curious where Go fits into this wonderful world of mlops. \[laughs\] |
**Mike Eastham:** \[27:58\] Yeah, great question. So I can kind of give you some perspective on how we're using it at my company, and then maybe a little bit on how I think it's going in the outside ecosystem... Although I'm obviously more familiar with what we're doing. |
So I've mentioned before we have this online surveying interface for the system where low-latency is one of the big requirements... And so that's primarily where we're applying Go at Tecton. |
So we have basically a server that we deploy for every customer, that is taking in online requests for feature values, and we basically have a bit of a hybrid between pre-materializing the data and then -- so we have some of it prematerialized, but we do some kind of final aggregation steps at serving time, and estimat... |
So we basically have a bunch of prematerialized data if we get this request coming in, and we do final aggregations, and then return that result to the user. So it's a pretty straightforward thing it's doing, but latency is really important, both keeping it low and then also keeping it consistent, because typically, pe... |
The reason we chose Go for this use case - for me, the biggest thing came down to the tools that Go gives you, kind of during performance tuning, and being able to dig into performance problems; things like the built-in tracing and profiling we've used really extensively, in order to meet our performance targets for th... |
**Natalie Pistunovich:** Was Go the first choice for this, or did you start with something and switched? |
**Mike Eastham:** Go was the first thing we used for this component, yeah. I'm trying to remember, it's been a while now... I don't think we super-seriously considered other options. We kind of by necessity have a large part of the system written in Python, because our SDK that we give to data scientists - they're kind... |
Interestingly, we actually do have a Python interpreter embedded in this Go server. We allow people to do -- we call them on-demand transformations. They're able to put in transformations that run basically during serving time, and you might wanna use that functionality, for instance, to compute a feature which is like... |
So you might wanna have a feature which is like -- a silly example would be what the difference in the length of the query and the last query was, or something like that. And you would do that inside the Python interpreter on the feature server. So it is a little bit of a hybrid, but 95% of it is all written in Go. |
**Johnny Boursiquot:** Now you have me curious about this thing you've embedded inside of the -- so this is not like you're getting a raw Python string literal and then you're shelling out to some Python runtime, and then getting results back. This is like you're actually interpreting the Python code and translating th... |
**Mike Eastham:** So we are running a full Python interpreter; we're not translating the Python into some -- |
**Johnny Boursiquot:** Okay. |
**Mike Eastham:** Yeah, that would be pretty cool, but we're not doing that. |
**Johnny Boursiquot:** I'm like, "You need to open source that thing." \[laughs\] |
**Mike Eastham:** I have actually kind of looked into this. There are projects floating around that will let you basically compile Python down to LVM bytecode or something, and then you could potentially link that into a Go binary. We haven't really seriously investigated that. |
Most of these transformations are doing quite simple things, they're not super performance-intensive, so we haven't really seen a need... But it's something we could look at doing, potentially. |
**Johnny Boursiquot:** \[32:10\] So this was done really to solve not really a technical problem, but more of a meeting your internal users, where they are, kind of thing... Because if Python is what they know and that's their tool, that's the thing that they know best, supporting that to some degree was sort of the go... |
**Mike Eastham:** Yeah, that was definitely the main thing, that our customers writing these are primarily familiar with Python. There's also a bit of a nice thing technically with Python, in that you can just take the function definition as a string and just send it to the server, so you don't have to worry about link... |
I know there are other systems, like -- I may be misspeaking here, but I believe Airbnb has published, they have an internal feature store product where I think they... I think they have an equivalent concept where they have people write the transformations in Rust maybe, and they can make this whole thing super-fast..... |
**Natalie Pistunovich:** Would you say that Go is generally a good choice for machine learning ops? MLOps? Mlops? |
**Johnny Boursiquot:** Mlops... \[laughs\] |
**Mike Eastham:** Yeah, I think so... I think you'll find at the various layers of the MLops stack there are always gonna be these components that are pretty latency-sensitive. Across most of the customers I've talked to, generally, pretty low latency for the end-to-end system is a pretty important requirement. And I r... |
**Johnny Boursiquot:** And that is worth the price of admission, my friend. |
**Mike Eastham:** Yeah, definitely. I would say the other thing too is that there's -- I mean, in the broader MLOps ecosystem I think a lot of people are using Kubernetes as part of the solutions that they're deploying, and I've found that just like being familiar with Go and the ecosystem is really helpful in that spa... |
**Natalie Pistunovich:** You mentioned the tooling is a thing that is super-useful for you in your world as a developer, in the world of mlops. Interesting. I would really expect that you would also mention something like the speed of the language, because it is really fast. |
**Mike Eastham:** It is certainly fast compared to something like Python... In my previous life as a software engineer at Google we were mostly using C++, so compared to that, it's not intrinsically faster. For me, I guess, given my background, I just look at the tooling and the ease of making everything work as the ma... |
**Johnny Boursiquot:** It's all relative. |
**Mike Eastham:** Yeah. |
**Break:** \[36:01\] |
**Natalie Pistunovich:** When did you start writing Go? What got you into this? |
**Mike Eastham:** Yeah, so it was at Google... I was trying to remember before I came on here... I think it was probably around 2015 was the first time... And it was not the primary thing I did there; like I said, it was mostly C++, but I worked a lot on the basically experiments system for web search at Google, and a ... |
But yeah, at the time I was working on it everything was in Go. That system had -- there were a bunch of different teams at Google that would contribute their own metric to it. So the image search team would have metrics having to do with images, there would be some that were common across all the different search prop... |
**Natalie Pistunovich:** And that's also, would you say, the main language for you now, right? Some Python, but mostly this. |
**Mike Eastham:** Yeah, we've got definitely a mix. All of our online serving stuff is written in Go; there's the Python for the customer-facing stuff... And then just for variety we have Kotlin that's involved in some of the more management and orchestration systems that are kind of off to the side. |
**Natalie Pistunovich:** And feast is the open source project, right? |
**Mike Eastham:** Yes. |
**Natalie Pistunovich:** And it has some Go. |
**Mike Eastham:** It has some Go. And frankly, I'm less familiar with the exact composition of everything on the Feast side... They're a lot more Python-focused, because they basically run a lot more of the feature store system inside the client, whereas Tecton's architected with more things living on the backend... Bu... |
**Natalie Pistunovich:** \[40:12\] Okay. So anybody who's listening and wants to dive into this a little bit by contributing code, can do this in the repo of Feast. |
**Mike Eastham:** Yeah. Feast is a good place to start. |
**Natalie Pistunovich:** I love the way he alalyzed the documentation, I have to say, of Feast. It's a great job at that. |
**Mike Eastham:** Yeah, they've done a really good job with all that stuff. |
**Natalie Pistunovich:** Cool. Okay, and Johnny, as we laughed earlier, quickly skimming through everything, and now he's becoming the expert... \[laughs\] |
**Johnny Boursiquot:** Ingesting data, yeah. Next week interview me on mlops. I'll have all the answers. \[laughter\] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.