text stringlengths 0 2.35k |
|---|
**Johnny Boursiquot:** Now, how granular -- you gave an example about having a deterministic way of getting somebody's age, giving your data model, your schema, your data store, wherever you're storing your user information. Is the point to have these things very granular, to ease their reuse? Because they are very sma... |
**Mike Eastham:** Yeah, I think some of that depends a bit on what types of techniques you're using to build the model. So from the perspective of someone who's building a feature store, we kind of just wanna make sure that we support whatever types of features that people wanna encode in the system... So you can have ... |
**Johnny Boursiquot:** So a feature then is part of the set of inputs that you're feeding into your model? |
**Mike Eastham:** Yeah, so you use it both for training, so you'll provide basically a bunch of previous examples of feature values, as well as the outcome. An example would be I might have a bunch of features about a prospective shopper, like let's say their age, maybe the last ten things they had searched for would b... |
And then during model training, you provide all these examples, and then a model is the result of that. And then later on, you can basically use that model to predict the outcome from just the inputs. |
An example I gave before - if you're trying to predict while they're shopping what it is that they might be interested in buying, you have just those feature values as inputs; you feed that into your model serving system, and then you get out a predicted outcome. |
**Johnny Boursiquot:** \[11:50\] So I'm trying to figure out where features fit in the greater set of tools that you would use; if we're talking personalization here, for example, I do know in having used some commercial personalization products that there's different algorithms that are used and different models that ... |
So these kinds of recommender systems - it sounds like these are a kind of problem that the features that you're creating would feed into sort of helping these kinds of problems. So I'm trying to imagine in my head where the layer of things needed to have a fully-operational recommendation engine, what unit -- where's ... |
**Mike Eastham:** Yeah, so I think it is one of the first things you would start thinking about, but I think another kind of important aspect of MLOps in comparison to DevOps is that you have to go through a lot of cycles. So you might start thinking about features at the beginning, but to really effectively build ML s... |
So I think it is one of the first things you would think about, but you also kind of have to be continuously monitoring what the values of your features are looking like, if you need to change the definitions, if you need that nuance throughout the lifetime of the model. |
**Johnny Boursiquot:** So what makes features sophisticated enough or complex enough to require their own data store? |
**Mike Eastham:** So there's a couple of interesting requirements from a systems engineering perspective when it comes to feature data. One of the biggest ones is that I mentioned that you kind of end up using feature values in two main places; you use them once when you're building the model or training the model, as ... |
Generally, when you're making a prediction, in the context of most recommender systems, for instance - you have a user generally that's kind of sitting there at the other end of the connection, waiting for search results to be ranked, or something like that. You obviously want that to happen quickly; so in that kind of... |
Whereas when you're training the model, the latency is not super-important. I mean, you obviously don't want people sitting around for hours, waiting for training to happen, but the throughput is more important in that case. Generally, the more examples you can put into training, the better results you're gonna get. So... |
In addition to the performance requirements being different, in the online case you're generally looking for the freshest data, so you want the most recent version of the feature value. That doesn't make sense with the age example I've been using, but if you're talking about a feature that's like the last product that ... |
\[16:07\] So you want those to be fresh, and have the most recent copy, whereas in the training or offline contexts, you want to get the correct value for like a historical point in time. So you wanna be retrieving the feature value as of the example that you're computing features for. So the interface looks a little b... |
**Natalie Pistunovich:** Would you say that the reason that feature stores became a thing rather recently is that we expect faster everything? Like you said, fast search results, and everything? For example, feature stores were not really around five years ago. |
**Mike Eastham:** Yeah, I think that's a big part of it, the performance requirements. I also just think that people are trying to deploy these machine learning systems in a much larger variety of businesses and products than they were maybe 5-10 years ago. I think really large tech companies have had some form of what... |
**Johnny Boursiquot:** So given that this falls in the category of MLOps, as you've been referring to it, is this meant for the people that operate these systems? I'm trying not to conflate the people actually building and training those models with the people actually running those things. So there's a seam there, but... |
**Mike Eastham:** Yeah, that's a good question. The answer is that both of those categories are users of feature store systems typically, and really, one of the most important things that a feature store can do is provide a clean seam between those two different roles. |
I briefly mentioned before, but before people were kind of integrating this feature store concept into their workflows, oftentimes the way a feature would make it into production would be a data scientist would kind of do the exploration in the data, come upon insights and figure out what feature values they wanted, an... |
There oftentimes were mistakes made in that process, people didn't quite understand what was going on in one version or the other... And if you have discrepancies between the training system and the online system, it really hurts the performance of your model, typically. |
So the whole idea with the feature store is that we can just have the data scientists define their features once in a way that they understand... So we support SQL, we also support defining them in Python, if people prefer that... But once they've done that, it already is meeting the performance requirements that are k... |
So we have one part of the interface that's focused towards the data scientists that are doing exploration, defining new features, and then another piece that is focused towards people that are operating the serving systems and integrating them into CI/CD pipelines, and things like that. |
**Natalie Pistunovich:** \[20:11\] Yeah. A company I used to work for in the past, they had exactly what you described. There was the data science/engineering team, who were building models and taking data and crunching and coming up with new ideas and new rules, kind of... You know, "if this, then perform this action.... |
**Johnny Boursiquot:** That's not error-prone at all... \[laughs\] |
**Natalie Pistunovich:** Yeah, even just translating from Python to Go, you would lose some of the precision that they found, or something interesting... |
**Mike Eastham:** Yeah... I mean, sometimes model training systems can even latch onto little quirks of behavior you wouldn't really think about when you're doing this translation... But even if there's some edge cases where certain numerical libraries they're using, or something behaved differently, it can really caus... |
**Natalie Pistunovich:** Yeah. |
**Break:** \[21:16\] |
**Natalie Pistunovich:** So we asked you to define some things, Mike, and I guess we've covered a little bit what was your definition of MLOps, but maybe you can say that for us, how do you see this, and how do you see this different from the devops and the infrastructure as a thing. You mentioned, for example, how man... |
**Mike Eastham:** Yeah. From my perspective, the most significant difference is that with devops generally you're talking about deploying servers, or some other programs into production... You've kind of got this pipeline that's going from some code in your Git repository or wherever it is, and then you're compiling it... |
\[24:04\] With MLOps, it's really kind of an extension of the same thing, it's just that that process of transforming the raw data into the final product, which is like a model in this case, is really quite a lot more complex. So it takes longer, either because there's manual steps that are involved, or because some of... |
You've got different kind of disciplines involved, with different stages of that pipeline... There's data scientists, data engineers, so multiple different stakeholders and people that are concerned with that pipeline. So really, there's a lot more complex pieces involved, is the main difference that I see. |
In addition to taking longer, it's also -- with a traditional continuous deployment pipeline it pretty much only goes one way. So you have the code, you produce the server, you deploy it, and then that's kind of it. But with MLOps, you always continuously have this loop back, because you're producing more data from the... |
So yeah, I think that's really the biggest difference... I think a couple other things are -- I mentioned this a little bit already, but you really have a lot of different kind of skills and roles involved with most ML systems than you do with a typical software system... So you have people that are just straight up so... |
So I think you kind of have to spend more time thinking about the different roles that are involved in making this system operate well, and make sure that you can kind of accommodate what they wanna do at different stages in the pipeline. |
**Natalie Pistunovich:** There's two names for this, right? There's MLOps and there's AIOps. |
**Johnny Boursiquot:** Ooh... |
**Natalie Pistunovich:** \[laughs\] |
**Johnny Boursiquot:** I wanna say, it sounds like we're splitting hairs, but I don't know enough to know for sure... \[laughs\] |
**Natalie Pistunovich:** So I did a tiny bit of research on Google trends, and things like this... AIOps is more common in the Americas; MLOps is more common in Europe and Asia. |
**Johnny Boursiquot:** But they really mean the same thing? |
**Natalie Pistunovich:** It kind of means the same thing. I mean, of course, you can go into splitting hairs, but this is kind of practically interchangeable. |
**Johnny Boursiquot:** Right. |
**Natalie Pistunovich:** I was also curious if you write that as one word, or two words. |
**Mike Eastham:** Well, I typically use MLOps, actually... |
**Natalie Pistunovich:** As one words, and you're based in the U.S. |
**Mike Eastham:** Yeah. |
**Natalie Pistunovich:** You're breaking the statistics. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.