GitLab Bookclub - Hands On Machine Learning, 8 Mar 2022

Previous Meeting

Next Meeting

⏯

youtube image

►

From YouTube: BookClub - Hands On Machine Learning [AMER KICKOFF]

Description

BookClub Issue: https://gitlab.com/gitlab-com/book-clubs/-/issues/40
Discussions Repository: https://gitlab.com/gitlab-org/bookclub-hands-on-machine-learning

A

When it's recorded, okay.

B

um Hello, everyone and welcome uh this is gonna, be the second kickoff, which is uh sounds a little bit weird, but uh yeah. We had a small discussion uh last week for the folks that are more on the immediacy on the on the main side, and now I want you to deal with spaceship if you want more of the american time zone so yeah. This is why we're here. Thank you for the interest on the book.

B

It's a very scary book once you receive it and you start clicking through it and that's fine, but I think it's very exciting. For me, it's very exciting that I see people from all over the business. However, like every almost every stage has one representative uh on this. uh This handle on this book club and that's very that's, always a lot of fun.

B

C

A

B

Did a little bit of uh just a small introduction uh for everyone? uh We have it over here, but I, if you guys, want to put it over there uh on the kickoff document itself or if you feel like you want to channel a bit um for me. I would like to know your experience with machine learning. If this is the first time ever, you get uh exposed to it or if, uh if it's already already, some courses like, for example, the coursera machine learning course, which is a classic.

B

I don't know tell me.

D

Sure I can go first, um so um I'm a technical account manager, um which basically means I spend my time trying to help the customers use get lab as as as well as possible and solve their particular problems.

D

Most of my customers are u.s public sector like different government organizations, and a lot of them are trying to use machine learning to like deliver better public services more efficiently and for just a variety of different things and, like usual here for the government in the us they're behind everyone else um and trying to catch up in a lot of cases.

D

So um I get a lot of questions about data science and machine learning, and things like that and I'd rather be able to answer them myself. Rather than going and bothering people on slack about it. So I'm trying to at least understand the fundamentals and the sort of tools they work with and uh yeah to see, see how I can be helpful. So I was a software developer before I was a tam, but I have not done a lot with machine learning so.

B

E

Okay, okay, so um I've never done anything with machine learning. In fact I uh so before here I was at apple, where I was in the siri team um part of I so I was in ops in the siri group before siri got launched and um at the at that time, siri was not what people thought.

E

It was every almost everything that siri replied with was pre-thought out, like somebody had to sit down and think about everything and sort of pre-populate stuff, um so they didn't switch to ml until a while, later really kind of around the time that I left the siri team.

E

So I was always um based on the people that I knew who were implementing the ml with siri. I was pretty skeptical about.

F

How well it was.

D

E

Work- and I don't know they seem to have done pretty well anyway, um so I basically have zero experience with ml in any way other than like knowing people who work on it. um So I my interest in it here is so I'm in the support team at gitlab, and it seems like a rich environment where ml could be useful if we could figure out a way. So I don't know we'll see.

G

I'm zeph I'm a senior sct in the quality department, working mostly with the runner team. um I have some.

G

Very brief exposure to machine learning, actually more deep learning, neural networks um from a previous job. I've got a rather humorous article from mashable on the project. I was working on that in the chat we were working on this particular robot before we spun it off and it was. It was fun, but I was working mostly on like the front end side of things, because we were leveraging users to actually validate data and kind of boost our machine learning, so it I just got a taste of it and it really interested me.

G

So I kind of wanted to dive into it a little bit more. I'm definitely interested in the second half of the book, but I'll I'll definitely I'll have to get through the for the first half first, because I'm just not there.

B

Oh weird uh yeah, uh as involved because people were in another zoo call like there were five between the defensive call and then there were twos and calls happening over this same thing. So thanks uh uh thanks. uh Can you to take a look at it to reach out to me? I don't know why this happened, but well. It happened.

F

Yeah, I think there were two invites in the meeting and we all were tweeting our thoughts, I'll look up. Who was in that meeting and ping them on that slack thread. I just pinged funny.

B

Thing that this is very probabilistic that have people just arrived at one meeting and the other half of another different meeting, yeah.

B

So yeah and for the two that were here also.

F

B

Different backgrounds- and I like what uh diana said about like okay siri started without first just very, uh like rock heuristics, and for me, that's the right way to start with machine learning, because you first have to test the use case and starting with machine learning, is very, very slow and it takes a lot of resources and testing out. That way is great, even though the users they might not keep the best answer or the best outcome, but at least it sets up for success later.

B

But within your areas right now, do you see what kind of like, if you think of machine learning, what kind of use cases you see over there.

B

That's exactly before, knowing anything about machine learning, because I want to see this before and after as well.

F

Oh, um I actually, when I was reading chapter, one wrote a bunch of notes for places where it might be interesting for get lab products. So I don't know what I would have said before, but I could tell you what little notes I wrote on the side of my book about it.

A

F

That certain page, where I was like uh classifying incidents summarizing incident timelines- I was thinking about these for the ops section, which I'm.

D

F

With uh they, like, they specifically said, detecting credit card fraud but pipeline anomalies and anomaly alerts for like metrics and alerting and then recommended and similar issues. I thought about for great use cases there's like a page about example. Applications that I wrote on the site of.

G

Yeah for me, from the quality side, it was more interested in in trying to analyze our failures and figure out exactly or if we can categorize these things and see. If we can point to other other sources, we have lots of possible infrastructure related failures, but it's we're talking about flaky tests and we can't tell where the flakiness is coming from kind of situation.

G

um So it would just be helpful to be able to have something go through some, sometimes the massive number of issues that we can have and uh that maybe even simplify our uh our rotation process we have for for watching all the pipeline jobs for deployments.

D

Beyond um what I was mentioning earlier about just that, I get a lot of questions from my customers about their own data science and machine learning type projects and how get lab can support them.

D

We're also just starting to look at um using the data we've been collecting to help us uh work more effectively, so the data team is currently helping us build a model to predict which customers are most likely to either downgrade from ultimate to premium or reduce the number of seats they're using, and I think that's going to be the beginning of a lot more work in that area. So I've been just sort of you know, providing uh support to them as a tam, giving them data and stuff like that.

D

But I can't really be of any use uh beyond that at the moment, because I don't have a lot of background in this, but I'd like to be able to be a better partner to the data team and others that are helping us make better use of what we have and understanding the fundamentals here, I think, would help with that.

E

From the support team perspective,.

E

We get lab, as you know, is a very complicated product and we get all kinds of crazy things from customers, and I mean every customer might explain something completely differently. So you have all basically all of these twisty passages all alike, and it it's it's too big of a a surface of knowledge for any one person to keep in their head and a.

D

Lot of times, we'll.

F

H

E

Come up like oh with 14.8, there's this problem, and and how do you keep track of all of this? And so um I don't know my idea, zendesk, which is what we use, I think, has some kind of ml that they use to build a knowledge base that customers can look at which we're not using.

E

But I was thinking of it in terms more of being able to help those support engineers to easily sort of get some direction, because I mean one of the bigger problems that I see frequently is everybody has you know somebody will take a ticket and they'll go down a path and it's completely the wrong path and it it really becomes a rat hole that never ends, um and so, if we could prevent even just that with some kind of a direction up front. I think that would be helpful for us.

B

We lost kenny already yeah he's trying to figure out this one.

B

Okay, um I'm back okay, it's interesting! It's nice to see like those potential use gifts around the flat coming from the edges, not from like the central place, thinking about cases where potential machine learning or children can be used, but from the edges from the use cases coming and hi uh alan. uh Finally, we get. We got sorry about the neo and everyone who's joining. uh Now it was a small part.

B

I don't know how, when that, with two different invites uh or two different links, but well I left it annoying so, uh like I said before uh today, folks we had as well. uh We had another kickoff last week for email and I want to give this opportunity to to american times as well, so that we can discuss a bit uh the process uh and uh why we are here in general, getting to know you as well, because we're gonna try we're gonna, try to do this, uh mostly async, rather than sync, um so we're.

A

B

Very few uh of this uh session, so most of it will be uh async studies, but it's nice and get to know each other a little bit before this starts. So for the folks that are joining now. I'd like to know a little bit uh your roles within gitlab and your experience with with the machine learning so far, so that we can a little bit divide the groups that that are studying.

F

I guess I can start my name's kenny, I'm a product leader. I cover the ops section. I uh have rudimentary programming experience uh and actually the thing that made me feel more comfortable about joining this book club is that my rudimentary experience is all actually in python. So hopefully that will be of use here um and uh yeah. My experience with machine learning in get lab is none, um but you know I am interested in learning what it's like to be a developer.

F

Who is interacting with both models in code simultaneously, as well as think through ways? We can use this tool to solve user problems in gitlab the product itself.

C

Hi everyone, I'm neil, I'm a front-end engineering manager within our security department. I haven't been I've, been with skidlab for two years now rarely hands on um it's it's difficult to dive into product and feature work. So when I saw this book club come up, it's an opportunity for me to parallelize in an area I'm interested in. I don't have experience with with machine learning, but I figured a guided experience. Working with other peers sounds fun.

H

um My name is alan cook: I'm a backing engineer of release and uh been a software developer for a while, and machine learning is just an area. I don't really have a lot of experience with, so I really wanted to dive in and get some hands-on uh knowledge about. What actually I can do with this stuff.

A

Brett walker, with uh plan. um The only experience I have is uh watching stan's uh little labeling engine run, um which is not much um but yeah. I'm I'm going to see how uh one uh learn the technology a little bit and see how maybe it'll apply to plan and issues and stuff.

B

Cool um so yeah, like I said for other excites.

B

And this is really fun, because this is where the defense I mean, and especially since we work with remote, we are usually just on a double of the people we work normally with, so we only talk from stage.

B

And this is a little bit of cross pollination of knowledge between the different areas, so I find this very really fun. So what we did last week was come up with a small process. uh I added this to the agenda.

B

um We tried to find a way uh to do this as much as basic as possible, uh and I want to hear your feedback if you have any other idea. Other ideas you can put on top of it. So the basic assumption is that we will not finish the book. We don't expect to finish the book. I think this is setting up ourselves for failure.

B

Really um it's just too high of expectation what we decided, instead of doing really a book club more a chapter club almost so we choose which we go through the book choose which chapters we want to go through uh and we want to want to study, and then we created a project where each chapter has its own issue and we go there and we assign ourselves the issue for that specific chapter and write down uh that you are reading that book.

B

That issue becomes the form where we ask questions where we discuss uh use cases or discuss code for that specific chapter or anything. Any question that you might have so this way is also rated.

B

If we want to do a further uh uh book club on this specific book, we, like all other people, want to study uh our notes or anything they have this like uh breadcrumbs and paper trail of what we did and then once most people or a few people have finished that chapter we can get together again for the chapter and I don't know, write a blog post or a small discussion of what each one learned possible use cases within their uh school and I don't know whatever the group wants to learn.

B

uh Last but not least, I I think it would be nice if we could set ourselves to read one chapter per week, but again, there's no police that will just knock on their door, and I arrest you because you don't read it. um So no there's no time pressure, it's better that you read two or three chapters.

B

Well, then, you try to push one per week and then you burn out you like burn out of the book really by the second week, and you don't read anything also if you notice that this is not adding uh to your to your expectations of what the book would be, don't be afraid to drop out.

B

I don't know it's not worth to spend the time on something that you're not really interested about. It's better just choose the time to play video games, or I don't know whatever trying to push yourself on it right.

B

So that's the the the tlvr of what we discussed last week. If you have anything, any suggestions that you give, uh let's discuss it.

F

um I do have a process c question, so um I have added my name to the first three chapters but like, as I read a chapter, I should kind of put my notes about what I thought of that chapter in the issue and then others might respond to it or I guess how. How will we facilitate a discussion about it if we're all kind of reading at various different times.

B

What I thought about was sid the the issue chapter, uh the chapter issue with the questions on the back of the chapter, so that we can discuss. If you want to put your notes, that's great because then we can uh other people can discuss the notes. If you have questions it's a forum, you can really add anything you might want uh over there. It's a free-for-all. We're I've never ran a book club this large, um and I really don't know if this format will be uh successful. I I have high hopes for it.

B

I think we really got some like cool, setups, very detailed uh setup if you will uh yeah, but we're still figuring out the process and really if the process doesn't match what your learning style, your study style feel free to just do your own thing and use uh create a different issue that you can ask questions or or whatever it's just that we need a more easy, better place to ask questions. I think then I.

F

Rely on the scenes I like it and I think the idea of adding yeah the prompt for questions would be great, sounds good.

B

Any other comments, any other ideas.

D

No, I mean, I think, what what's in place, looks like a great first iteration and we won't know whether or not it works until we try it and then we'll adjust as needed.

B

One one concern that I have is how some people are really uh self-driven and they can learn this, but some people like when somebody nudges them to go through uh some chapter, that I have what ideas would we like? How could we nudge these people? How could we should we tag everyone like? Should I go over in two weeks and tag everyone for the status update or I would like some ideas on that.

E

I mean, I think that once you add yourself to an issue on a specific chapter, then you're going to get pinged on. You know anything that anybody else adds to that issue. So I think that in itself is a nudge um which may be enough.

G

So we're talking about um adding ourselves as a signees to the issue.

E

I added my so on the issues that I looked at. There was in the description. People were just adding their name as I'm signing up to read this chapter. I think, and so yeah. That's what I did.

D

If we add ourselves as a signees, we could stick a due date on each of these in the roughly like one a week starting like next week, and then people would get a your issues almost do email from get lab, but that might be more. We have to decide the right amount of pushy to your point. Eduardo like we want to. We want to nudge, but we don't want to shove people or make them feel like guilty about it.

B

Oh okay, let's do the following. If you want me to invite you send me a dm and then I'll write down that I should guide you after x amount of time, and then I will do this on private so that I start like public, shaming or anything, um but I agree with diana. I think that will give like when the problem happens. People forget so if we have a constant emails because we are assigned over there and put our names, it might help like drive this.

B

uh It's like bringing the food from the back of the fridge to the front when we get that email and we end up, not forgetting I put myself. As I say, I think it's kind of like a cool setup that okay, this is assigned to myself and when I finish, I cannot assign it's also on my list of to do stuff that I need to do the following weeks as well.

B

So I like that, um but initially it was just put your name uh on the on the thing on the read on the description, but I think it might be a good setup to put yourself to the setting, because then you know, if you open your issues, you will see that over there that you, you need to read or you want. You sign yourself up to reading that book.

B

Anything else on this: uh anyone who wants to give a suggestion tell a joke or anything any comments, or I don't know.

D

I'll share one quick thing I ran into if you bought it on kindle and you and you try to use the kindle web viewer. It won't. Let you copy text out of it, which is really super annoying because oftentimes, when you're walking through a tutorial, and they give you a big chunk of code- you don't want to have to manually type it. So if you, google, you'll, find um the repo with all the jupiter notebooks for all the chapters.

D

I think it's on that other git hosting website that we don't like, um and you can. You can look at the at those and then incorporate them into your notebook. For you know whatever chunk of code, it is but don't plan to copy and paste out of the ebook if you've got that.

A

Can you can you post a link to those jupiter chapters on the project readme.

H

B

A little bit below over here, it's it's on the on the issue of the it's on both the reading of the repository of the project and also on the issue of the book club.

A

Yeah, I see it now.

B

It's over there.

D

Already done cool so yep uh now I found that um repo to be very handy.

B

Okay, so what I'm gonna do as well? I'm not gonna see every question on the on the issues. I'm gonna post one question every two days so that it's always happening for the people that are essentially issues and reminding them I'm gonna beat it. Will you go out uh for everyone? I guess you forget your spanish lessons.

B

So, okay, uh if anybody has anything else to say uh otherwise,.

D

Yes, thank you so much for organizing and all this um I'm it looks like you got more interest than you thought you did, which means more management than you thought you were going to have to do and- uh and we appreciate it.

B

It's I have a lot of fun, organizing this kind of stuff. I think it's really fun. I like getting people from different places of the business together to talk about this kind of thing. uh This is what writes innovation uh for me, the most people that don't share the same day-to-day process or knowledge get together and study and forget it's when new ideas come up, so I'm enjoying it as well.

B

I won't stop the recording uh now.

G

uh Actually eduardo, I got yes, one thing I was adding it sorry, adding it to the agenda. um I noticed the book doesn't have a whole lot of guidance on like getting your system set up and all of those things, I'm I'm curious really from uh from a gitlab perspective.

G

How are we doing it uh in your group, for example, managing your system or using uh paying pi emv virtual env, whatever uh to manage all of all of your things, um so it it would probably be useful. I know I get lost in the weeds when I start going through some of those. Some of that tooling.

B

Yeah, uh no, it's uh it's we're still learning. I think we should add. This is a really good question that I can spend a lot of time writing about. um I think we could add this as a question to the issues uh create the issue specific for this question, because then we can write it down over there and discuss, and then I have other people that are not understand exactly even type the people from the applied ml team that because they they could answer this a lot better.

B

We have people also from verify now from uh using machine learning. We have a small uh kind of like meet my learning meetup as well, where they discuss setup uh around how they do the setup for their machine learning models to discuss evolution of it as well. So, let's add a question over there and then we can. uh I can change uh in there and then it's open for more people eventually as well and the public as well.

G

um Since the chapters are are specific, more to actually kind of ml theory,.

F

G

Not any one of them actually address kind of the the back end technical things should we have a separate issue for it, so we know where to look for it. I think.

B

This is an issue, that's the question that might get a lot of. uh We can either create an issue for random questions that we don't think we don't know if it fits in there or not, or we can create an issue specific for this question, and I don't know what would be best I'll. Let you decide when you create the question: okay,.

E

I think I think, having a specific issue, for the setup in particular is a good idea um and then maybe there's another one for random questions, or maybe maybe, as people come up with random questions, they decide on the fly. Whether or not it's worth creating an issue.

B

Exactly I think I think the second chapter is about end to end uh machine learning, project or something it could be over there. It's a little bit about the setup about how to you that question could come up on the second chapter as well.

D

That's where I ran into it, because the first chapter is all introduction. The second chapter was the point where I hit that same question, and I was like okay. How do I actually get this thing working and then I monkeyed around with it for a while um and eventually got something that was good enough, but I definitely want to hear how you're actually supposed to do it from people who know what they're talking about.

B

You'll be surprised at how not classic it is like, uh but.

A

Yeah, I know we.

B

Can discuss a little bit better, but I usually for me most. uh We have now. uh The problem now has set up a full pipeline for machine learning, lab pipelines which, which is really cool, to see- and I have some examples over there also of different powerpoints that I set up uh with gitlab. So we can share over there.

G

Yeah, I just haven't gotten to chapter two, yet there's there's some details in there, but um but if there's something like like, we do in within gitlab teams that is kind of more standard. I'd love to hear that too. There's no.

B

Standard game, so the there's no standard is very young.

B

um The team is very young, so I don't know six months now I have been in the company for I'm not on the machine learning, I'm not on the model. Upstage, I'm on the incubation engineering department, which is kind of a different thing, but I also joined six months ago. So most of the use cases are very new to the company, we're still building the culture uh around. uh Actually I mean the team is doing a great demo. Quite a mellow is doing a great job at that as well.

B

So uh we can share everything and it's nice, because now we can see this evolution because our our customers are going through this evolution themselves right. So we are kind of like dog foodie, uh not only the product but building up the culture, which is a really cool experience.

B

Okay, anything else otherwise, there'll be five. Four, three, two! Okay! Thank you all folks. uh It's a pleasure! I'm really really excited about this book club is the first time running here that it's quite large, I'm quite surprised, but I think it's gonna be a lot of fun. um Let's do this async and then we will figure out the next steps over time cool.

B

Thank you all for joining and sorry for this mode, but in the beginning I have absolutely no idea how that happened, but I'm glad everyone joined this one afterwards, thanks kenny for pointing that out. That was weird.

F

Awesome. Thank you. Eduardo thanks.

G