text stringlengths 0 1.49k |
|---|
**Joe Doliner:** Yeah, absolutely. It's really taken off just because we've gotten more people working on it, we've refined a lot of the concepts... Open source is all about having these very tight iteration loops, so what we have now is the result of basically two years of iterating with our community as fast as we po... |
**Brian Ketelsen:** What kind of users do you have out in the open source world? What's the average Pachyderm user look like? |
**Joe Doliner:** I get that question a lot, and what's interesting is it's very hard to nail down an average user, because these tools can sort of apply across a vast spectrum of use cases. One of our most notable users - and they are actually customers, so we can use that word - is a fusion company called General Fusi... |
Another big case where we've seen a lot of customers and users coming is in machine learning. You guys have talked a lot about that in your previous jobs. One of the things that actually happens - and I think the last two months or so it's been really good for us - is the EU actually just passed a law where consumers h... |
If you're in the EU and the bank says, "Sorry, you're not getting this loan", you can turn around and ask them why; what data went into making this decision? As you can imagine, having a system that just tracks all of your provenance for you makes that incredibly easy. You can just look at the provenance of your machin... |
**Brian Ketelsen:** Yeah, that's really big. |
**Erik St. Martin:** The hard part, if people imagine that problem, is this data is continuously changing every day. Sometimes data will be bad and have to be purged from the system, and the dataset is constantly changing that's being used as part of the scoring model so to be able to go back in time and say, "This is ... |
**Joe Doliner:** \[07:50\] Yeah, absolutely. And the other thing that we really like about provenance is that it really allows the scientific method. Because if you've got a data lake running in the company - you're a data scientist, you come in, you see some result there that was computed by some huge chain of steps t... |
**Brian Ketelsen:** That's really amazing. Can you talk about where Go fits into this? I know there's a lot of containerization, but do you have daemons that are written in Go? What's the glue that holds all this together? |
**Joe Doliner:** Yeah, absolutely. The entire system is written in Go. That was a choice made a) because the things that we're interacting with are all written in Go, and so their nicest client libraries are gonna be in Go, too. Those are Docker, Kubernetes, Etcd... It's basically an entire Go stack that we deploy. We ... |
It's also just a really great language for this type of system stuff. Our main thing that we deploy is the Pachyderm daemon, which we just call Pachd. That's just a Go server; it's using gRPC. gRPC just makes it super simple to stand up this structured API and get going writing the actual code that goes into it. That's... |
We also have a frontend command line interface tool called Pach Control - that's all written in Go. It uses Cobra, which is the Steve Francis command line interface-making library that we found to be really good. That about covers it, to be honest. It's a pretty simple architecture. |
**Erik St. Martin:** So the main motivation for exploring the use of Go was mainly just because all the other components you were looking at to build the system were already in Go? |
**Joe Doliner:** Yeah, absolutely. And we knew the main use case that the Go team cares about is our use case. This is what people are using Go for internally at Google - they're using it to write services and they're using it to write command line interface tools, so we knew... It's always best to be on the main line ... |
**Erik St. Martin:** That's awesome. How about the data layer inside the container? Do you offer anything as part of Pachyderm to help people build their own components to process the data, or you pretty much leave everybody to their own devices inside the container and you handle the workloads? |
**Joe Doliner:** Yeah, we handle everything that could be termed "data orchestration" - which data gets processed, where does it get processed inside of a container? The user handles everything about how it gets processed. For example, in Pachyderm we think of things in terms of repos. We sort of mimic the language of ... |
The code that you will write is just a binary, and you put it in a container, you give it to Pachyderm, you tell it "Here's my container, here's the command I want you to run in the container" and when that command runs inside of the file system that it sees in the container, there will be a small chunk of the log mess... |
\[11:58\] After that code runs, Pachyderm knows to slurp up that data, put it back into the file system and trigger downstream processes in exactly the same way. We orchestrate all of that pipelining, we orchestrate all of where the data goes, and you just orchestrate what specifically the actual analysis is. |
This can get pretty complicated too with how you're joining data sets. You can do vast multi-way joins of different data sets and we'll distribute all of that, we'll orchestrate all of that so that your code gets run... But again, you're just writing the "Once I've got data on disk, what do I do with it?" That's all yo... |
**Brian Ketelsen:** Oh my god, I just wanna stop what I'm doing now and go play with it. |
**Joe Doliner:** We would love for you to play with it, but don't stop right now, because I think we need to do the rest of the show. \[laughter\] |
**Brian Ketelsen:** We've proven in the past that the show will go on without me, so I'm just gonna step out and start downloading Docker containers. I'll be back. \[laughter\] |
**Joe Doliner:** Okay. |
**Erik St. Martin:** We'll bring him back in at the end, and he can tell us what he learned. |
**Joe Doliner:** Okay, hopefully he'll have a good experience. I'm crossing my fingers right now. There's a users channel if you run into any problems. |
**Brian Ketelsen:** It really sounds amazing. |
**Erik St. Martin:** It might make for a fun episode where you assign somebody on the show a mission and then come back to him at the end. |
**Joe Doliner:** I think that would be a great radio version of the open source experience. You could see us live, providing support for an open source project. Unfortunately, that might get really boring for the listeners, because I'm like "Alright, do you have the permissions right? Do you have the user path right?" ... |
**Brian Ketelsen:** Yeah, my Docker volumes aren't working. |
**Joe Doliner:** Oh, no... Not the Docker volumes. |
**Brian Ketelsen:** I ran into that two or three days ago... I installed a new Linux distribution, because it's distro of the month for me, it always is. I went back to Arch, but the newest Arch Linux Docker packages don't work unless you're using overlay file system, but the Arch packages didn't bundle whatever lib ov... |
**Joe Doliner:** That sounds like the quintessential Linux experience right there. Somewhere there is some little message that if you Google it, it will tell you exactly what you need to do, but until you Google that, you're completely lost. |
**Brian Ketelsen:** Right, that's exactly true. That's what we do, and that's why this kind of mission would not play well in a time-boxed show. \[laughter\] You never know whether it's gonna be a ten-minute mission or a ten-day mission. |
**Joe Doliner:** The worst is when you come across a forum post for something on Linux, and someone's got this question and then nobody answers it, and then ten days later they're like, "Found the answer, thanks guys", but they don't post what the answer was, so you're just as screwed as you were before, but you know s... |
**Erik St. Martin:** I know somebody in the Go channel will probably link this, but there's an xkcd episode where they do that, and it's like "Who were you so-and-so? What did you see?" because you come back to it years later. |
**Brian Ketelsen:** That's right. Honestly, I think worse than having somebody solve it but not tell you what the solution is is you going out on the internet and searching for the solution and finding your own damn post explaining the fix for it. I don't know how many times I've done that. |
**Joe Doliner:** Yeah, I've done that, too. |
**Brian Ketelsen:** Could I not possibly have learned at the first time? |
**Erik St. Martin:** I wanna point out how awesome the channel is, because I made the joke that somebody in the channel would do it. In less than 30 seconds, somebody had it linked. |
**Brian Ketelsen:** Nice! |
**Erik St. Martin:** So for anybody who's listening, it's number 979. |
**Brian Ketelsen:** DenverCoder9. Who were you? \[laughter\] That's awesome. |
**Erik St. Martin:** I tend not to post my question, though. I usually reach out privately to people or talk in channels, so I don't really come across my own question. |
**Brian Ketelsen:** \[16:05\] Mine are usually blog posts. I wrote this blog post about how to do X, Y and Z, and then tried to do it again two years later and I find my own blog post answering my own questions. I hate it. |
**Erik St. Martin:** Your own blog posts... |
**Brian Ketelsen:** Yeah, it's embarrassing. So the Pachyderm orchestration system being written in Go, what nice features of Go made that work well for you, and how is that working with a community of Go developers on such a large scale orchestration system? |
**Joe Doliner:** The features of Go that really make it work well are a) just having all the built-in libraries that you need; having an HTTP library that's there and it's really good... Before this, I was programming in C++ on RethinkDB, and there I remember writing our own HTTP libraries, writing our own HTTP server,... |
I really think that the single best feature of Go is how batteries included it is. You have just good things in the standard libraries that you can use. Other than that, having goroutines is to me the best way to do concurrency. |
When we were making RethinkDB, one of the first things we did was make a call routine library for C++ so that we could have that style of concurrency, and it was nowhere near as clean and elegant as Go's is, because it just has a few primitives and you can basically do everything. Other than that, we're using all the s... |
**Brian Ketelsen:** Right. At least not successfully. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.