text stringlengths 0 2.08k |
|---|
**Mikeal Rogers:** Our focus on today's episode with Andrew and Arfon was around open source metrics and how to interpret data around dependencies and usage. We talked about what we currently can and cannot measure in today's open source ecosystem. |
**Nadia Eghbal:** We also got into individual project metrics. We talked with Andrew and Arfon about how we can measure success, what maintainers should be paying attention to and whether stars really matter. |
Andrew, I'll start with you. What made you wanna build Libraries.io? How was that informed by your GitHub Explore experiences, if at all? |
**Andrew Nesbitt:** I got a little bit frustrated working at GitHub on the Explore stuff. It was me kind of deprioritized whilst I was there, and my approach of libraries, rather than just build the same thing again outside of GitHub, was to use a different data source, which started at the package management level, an... |
**Mikeal Rogers:** Right. So tell me a little bit about that. So you pull out the whole dependency graph data - do you go into the kind of deep dependencies, or do you sort of stay at more of a top layer of first-order dependency data? |
**Andrew Nesbitt:** So for each project, it only pulls out the direct dependency. But as it picks up every project, because every time it finds anything that depends on anything else, it will go investigate that as well. It ends up having the full dependency tree, but right now I don't have it stored in a way that make... |
**Mikeal Rogers:** Interesting. Okay. So you said that this is a much more interesting way to go about this in the GitHub data. What's something that you found when you started working with the dependency data that you never had in GitHub Explore, or just to the GitHub data? |
**Andrew Nesbitt:** GitHub stars don't really give you a good indication of actual usage, and GitHub download data is only really accessible if you are a maintainer of a project, rather than just someone who's looking at the project from a regular browser's perspective. If you actually look at the dependency data and n... |
\[04:21\] But you can actually use package management data to connect the dots between GitHub repositories as well. You can say, "Oh, well given this GitHub repository, how many other GitHub repositories depend on it through npm or through RubyGems. |
**Mikeal Rogers:** It's good to hear that stars are useless, because I've also thought that. \[laughs\] That's been my assessment, as well. |
**Andrew Nesbitt:** Yeah, I've \[unintelligible 00:04:46.03\] over how you shouldn't judge a project by its GitHub stars. There's one particular project that's a great example of that, it's called Volkswagen. It is essentially a monkey patch for your CI to make sure it always passes. I think it's got something like 5,0... |
**Mikeal Rogers:** Yeah, that's by Thomas Watson. It was a joke when VW had that scandal where they were just passing all their tests, so he wrote a module called Volkswagen that just made all your tests pass, no matter what. \[laughs\] It's brilliant... But yeah, utterly useless in terms of actual usage. |
**Andrew Nesbitt:** Yeah, and if you actually look at the stars... Of course, people have contributed to it, but even looking at contributed data doesn't give you a good indication of actually is this a useful thing, a real thing, and should I care about it? I always look at GitHub stars as a way of... It's kind of lik... |
**Mikeal Rogers:** I think stars are an indication of attention at some point in time, and that is all we can say about them. So if you look at stars versus pageviews on a given repo, they correlate very well. So in defense of stars, we shouldn't use them as "This is what people are using", but they're a good measure o... |
**Andrew Nesbitt:** Yeah. I saw someone actually build a package manager; I think it was only a prototype, but I really hope it never actually became a thing, where it would pick the right GitHub repository if you just gave it the name rather than the owner and the name, by the thing that had the most stars, which soun... |
**Mikeal Rogers:** Yeah, that doesn't sound like a good idea. You mentioned something interesting, which was you can understand how people use it in terms of just it being depended on. Recently GitHub did this new BigQuery thing, and one of the results is that you can do - RegEx has done the actual file content of a lo... |
**Andrew Nesbitt:** Yeah, so just to refresh the data that we put into BigQuery, it's basically not only the event data that comes out of the GitHub API, which is just "Something happened on this public repo" - and that's what the GitHub archive has been collecting for a long time - this is actually in addition to that... |
\[08:15\] So that allows you to do things like if there's a particularly - maybe a method call in your public API that you wanna try and measure the use of, then you can now actually go and look for people using that explicitly. So currently really complex kind of RegEx stuff on GitHub searches is pretty hard; in fact,... |
For me I think the exciting thing about that... I think that's really complementary to things like libraries that go and look at package managers - that's incredibly useful, but I think not every language has a strong convention for the package management that they use, unfortunately, some of us forget that; I think we... |
**Arfon Smith:** Yeah, for languages like C, that's pretty much the only way to do it. There's just no convention there, other than the language itself. And then for some other package managers, you actually have to execute a file to be able to work out the things that it depends upon, which I avoid doing because I don... |
**Mikeal Rogers:** Well, in the NodeJS project we've been trying forever to really figure out how are people using some of these methods, because if we wanna, say, deprecate something, we'd really like to know how many people are using that in the wild and to which level is it depended on. But we've had several project... |
**Andrew Nesbitt:** Yeah, it kind of makes me sad that we've made everyone write crazy RegExes, but sorry about that. Hopefully, that will be useful. \[laughs\] Hopefully a bunch of good stuff can be done; people are gonna have to level up their RegEx skills, I think. |
**Nadia Eghbal:** Just for people who are newer to metrics world, why should they care to be blunt about this dataset being open and being on BigQuery? What are some things that you expect the world to be able to do with this data? Even outside of people like Mikeal with Node, but policy makers or researchers or anyone... |
**Arfon Smith:** One of the things I think is incredibly difficult right now for some people is to measure how much people are using their stuff. For a maintainer of an open source project maybe that's not a huge problem, because you can go and look at things like libraries and see how many people are including your li... |
\[12:26\] It's currently kind of hard to do that if you have your software only on GitHub and you don't have any other way of measuring when people use the library. You don't have any direct ways of doing that, other than just looking at the graphs that you have as the owner of the software on GitHub. So I'm excited ab... |
I think it's actually a hard problem to solve, but I know there are some environments - I'm trying to think of some large institutional compute facilities, big HPC centers... People have done some work, doing some reporting on when something's being installed or run, and actually Homebrew I think have started doing tha... |
**Nadia Eghbal:** So for you coming from the academia world, have you gotten requests from people from the scientific community around using this type of data? Did those experiences help inform the genesis of this project at all? |
**Arfon Smith:** Yeah, a little bit. Very early on when I joined GitHub I got some enquiries from people saying, "We'd love to get really, really rich metrics on how much stuff is being downloaded, where people are downloading from..." - all this stuff that you needed if you had to report and you wanted really rich met... |
The other things is - and I think this is a kind of a pretty sane standpoint for us to take - we take very seriously user support, so if somebody comes to me with a data request, it may be ethically possible for me to service that, and it might be technically possible for me to service that. But if it takes two weeks o... |
\[15:58\] I think the next step for me personally with this data being published is to start to kind of show some examples of how it can be used to answer common support questions that we see. I think that's kind of the obvious next step from my standpoint. |
**Mikeal Rogers:** And Andrew, you're in a position where you're actually taking a bunch of public data that's out there in all these different public ecosystems and then kind of mashing it together, so you're like your own customer for this data. What are some of the interesting things that you've been looking at? Wha... |
**Arfon Smith:** Unfortunately I didn't have access to the BigQuery earlier, so I've been collecting it manually via the GitHub API for the past year and a bit, which takes a lot longer, but it also picks up all of the repositories that don't have a license, which I guess often it's probably best not to pull people's c... |
Some of the things that I've been able to pull out and have been quite interesting is looking at not only the usage of package managers across different repositories, but the amount of repositories that use more than one package manager, or that use Bower and npm, or RubyGems and npm, and then looking at the total coun... |
Coming from a time working with Rails before Bundler, it was incredibly painful sharing projects or coming back to projects and trying to reinstall the set of dependencies that all worked, given the transitive dependencies that move around all the time with new versions. And it looks like the Ruby community is pretty m... |
**Nadia Eghbal:** I think we're heading into the break right now... When we come back we'll talk about the open source ecosystem. |
**Break:** \[18:49\] |
**Nadia Eghbal:** \[19:38\] We're back with Andrew from Libraries.io and Arfon from GitHub. In this segment I wanna talk about the broader open source ecosystem and the types of metrics that are and aren't available to people, because I've heard a lot of confusion about "Well, why can't we measure what is being measure... |
**Arfon Smith:** I don't think so. I mean, I don't know what myths there might be. I would love to hear things that you've heard that you would love to know if they're true. I don't know of any kind of whisperings of what GitHub might be doing, so I'm happy to respond to questions. |
**Nadia Eghbal:** I hear a lot around just download data, and whether GitHub actually has the data and isn't sharing enough of it, why not use download data in addition to stars as something that people can see... |
**Arfon Smith:** Sure... Yeah, okay. So there is a difference between what you as a project owner can see about a GitHub project and you as a potential user of that software. So there are graphs with things like number of clones of the software, which is I think a good metric, there are graphs for showing how many page... |
**Andrew Nesbitt:** I seem to remember the metrics weren't exposed because of some of the referrer data potentially leaking people's internal CI systems. |
**Arfon Smith:** Yeah, that might be possible. I'm not hugely familiar with exactly why the data isn't exposed right now. I think it's important to remember that we take user privacy very seriously, so the thing here is you wanna be on the right side of people's expectations of privacy. There are things that GitHub cou... |
\[23:47\] So yeah, we capture more information than we expose, but that's just a routine part of running a web application and having a good engineering culture around measuring lots of things. The decision about what to further expose to the broad open source community or the public at large is largely one based on ma... |
**Mikeal Rogers:** I think also any metric has to be qualified, right? A lot of this talk about stars is that stars is not an indication of quality, it's an indication of popularity at a point in time, like you said, but people take it as that because it's the only data that they have. |
An example is in NodeJS we have metrics for which operating system people are using, so we always put out two data points. One is the operating systems that have pulled downloads of Node, either the tarballs or the installers of some kind, and then we also have the actual market share for the NodeJS website, visitors t... |
One metric that is huge on the people end is Windows, and incredibly small on the actual computer end is Windows. But we do a lot to qualify those before we put them out, to set people's expectations about them. |
**Andrew Nesbitt:** Yeah, and there's another thing... I think the Python package index has a similar - like a badge you can put on your profile. And you see this, people will put it, the number of downloads last month from the Python package index, and it's exactly the same problem. For a fast-moving project where the... |
**Mikeal Rogers:** Yeah, the problem with download metrics on packages too is that you also get into the dependency graph stuff, right? Downloads are really good at looking at the difference in popularity between something like Lodash and Request. They're both very popular, but the difference in downloads gives you som... |
**Andrew Nesbitt:** Yeah, I have one of those, [base62](https://github.com/andrew/base62.js). I don't think there are many projects that use it, but it gets like one and a half million downloads a month because React transitively depends upon it, so it's downloaded by everyone all the time. But it never changes, it's n... |
**Mikeal Rogers:** That's funny. There's a lot of packages like that. The whole Leftpad debacle was people did not know that this was used by a thing that used a thing that used a thing. It wasn't that popular of like a first-order dependency, it just happened to be in the graph of a couple really popular things. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.