text stringlengths 0 1.82k |
|---|
• Comparison between Go's garbage collection and Java's, with Go being chosen for its simplicity and speed |
• Adoption of Go in other projects at Netflix, including Chaos Monkey |
• Performance techniques used in Rend, such as avoiding external dependencies and using standard lib Go |
• Use of metrics library and integration with Prometheus for collecting metrics and logging information |
• Explanation of the architecture of EVCache and the role of Rend within it |
• Moneta project aims to store cold data on disk while keeping hot data in RAM for faster access |
• Rend is an on-box memcached proxy that allows wire-compatible interaction with the existing Java client |
• The system has two layers: a caching layer (memcached) and a storage layer (RocksDB) |
• Performance considerations led to unconventional design choices, such as using mutexes instead of channels for concurrency |
• Mutexes were found to be more efficient in certain cases, especially when dealing with high-concurrency scenarios like Netflix's scale |
• Channels are not interchangeable with mutexes and should be used judiciously; Rob Pike's video on Concurrency Design In Go is recommended for guidance |
• Support for new memcached commands and adding features to an open-source project |
• Wire compatibility with memcached, and the decision not to support all its commands |
• The use of EVCache Java client at Netflix and its limitations |
• The "How To Block Forever In Go" blog post and its evolution as a list of ways to create deadlocks in code |
• Static analysis tools and their potential to catch deadlock patterns |
• Netflix's open-source projects, including those related to Go |
• GoKit reaches 0.1.0 milestone |
• Discussion of API stability and confidence tagging in GoKit release |
• Use of GoKit's logging package by multiple attendees |
• Alternative approach to logging using metrics instead of logs, as implemented at Netflix |
• Updates to Vim Go and Hugo projects |
• Review of Francesc's "Go Tooling In Action" video and discussion of related topics |
• Discussing external dependencies and work-life balance |
• Introducing the project "Iris" which claims to be 20 times faster than other web frameworks |
• Debate on router performance and whether new frameworks are needed in Go |
• Discussion of vendoring tools (Govendor) and dependency management in Go projects |
• Consensus on vendor folder best practices for libraries vs. commands |
• Mention of the survey sent by Ed Muller from Heroku to gauge usage of Go and its libraries |
• Criticism and dislike for Maven |
• Discussion on dependency management and the need for consensus among developers |
• Updates on the Go team's involvement in facilitating discussion on dependency management |
• Announcements about upcoming episodes, including Beyond Code Season 3 featuring GopherCon 2015 interviews |
• Discussion on the growth of the Go community and its influence on the computing industry |
• Humorously mentioned "hacking" onto the show and participation from listeners |
• Jessie Frazelle's dotfiles: a comprehensive collection of configuration files and scripts |
• Network Programming With Go: an open-source book by Jan Newmarch |
• Go standard library: enabling deeper understanding of programming concepts |
• Radare2: a reverse engineering framework with Go bindings |
• Importance of exploring and using alternative tools and projects in software development |
**Erik St. Martin:** Alright everybody, welcome back for another episode of GoTime. Today we have a special guest with us, but first we'll go through who the typical hosts are. I'm Erik St. Martin, we also have Brian Ketelsen, which does not sound like Brian Ketelsen today, but I assure you it is the real Brian Ketelse... |
**Brian Ketelsen:** Today the part of Brian will be played by somebody with a very scratchy voice. |
**Erik St. Martin:** \[laughs\] We also have Carlisia Thompson... |
**Carlisia Thompson:** Glad to be here, hello. |
**Erik St. Martin:** And our special guest today is Scott Mansfield who we've talked about a couple times on the show, about Rend and a couple of other posts we've read. Welcome to the show, Scott. |
**Scott Mansfield:** Hello, everybody. |
**Erik St. Martin:** So you decided to heckle last show, so I guess the punishment is you have to be on the show now. \[laughter\] |
**Scott Mansfield:** Yeah... Like I said, I need to learn to keep my mouth shut sometimes. |
**Erik St. Martin:** So Scott, you work on some products at Netflix using Go. Do you wanna give everybody a little background? |
**Scott Mansfield:** Sure. So the project is called Rend, and it's a memcached proxy and server that's written in Go. It was an interesting choice of language, because Netflix is pretty much an all-Java shop, and we needed something that was more performant, more productive and less - not bloated, but doesn't have any ... |
We also wanted to have some sort of performance, because the service that I work on is actually called EVCache. That's a distributed, charted memcached. And we're very latency-sensitive, so having an 80-ms GC in Java would be great for a lot of people, but for us that would be horrendous, so we picked Go. |
**Brian Ketelsen:** That's interesting, if you don't mind me interrupting. A lot of people are still very twitchy about Go's garbage collection. Can you elaborate a little bit more on why Go's garbage collection worked out for you when Java's couldn't? |
**Scott Mansfield:** There's a variety of reasons. Partially, the Go memory model itself is simpler and has less indirection, which allows the garbage collections to be faster, but really it was sort of a "Let's just create the program in Go and we'll see how it works." It didn't really start out as a work project; it ... |
**Erik St. Martin:** So is this the first project within Netflix to adopt Go, or is there other ones? |
**Scott Mansfield:** I actually don't know when the other ones started, there's a bunch here. I don't know if you guys are familiar with the Chaos Monkey system. |
**Carlisia Thompson:** Yeah, yeah. |
**Scott Mansfield:** \[03:58\] So there is actually a new version of the Chaos Monkey coming out. It's not open source yet, but the whole backend of the Chaos Monkey has been rewritten in Go, and it's actually in production right now, striking fear into everybody's hearts here. I actually spoke to the developer before ... |
**Erik St. Martin:** That's awesome, because until now I love everything Go does, and now it's wreaking chaos on my system right? \[laughter\] It is now both evil and good. |
**Carlisia Thompson:** I wanted to ask you, with the project, the Rend library being so in need of performance... Did you apply specific techniques, did you apply any design concepts, did you use specific libraries to make it as performant as possible? Or did you just apply good Go idioms and it came out performing wel... |
**Scott Mansfield:** Interesting enough, there's... Well, I'm trying to parse the whole question, there's a lot of pieces there. Sometimes the good Go idioms can be less performant than if you tried doing something lower level, such as ownership of data with the goroutine and then sending messages back and forth for th... |
One of the things that's actually a good example is our metrics library, because there's all kinds of metrics libraries out there where you have a counter struct and then you go and increment it and it still does atomic increments in the background, but it really didn't fit our use case quite as much. |
**Carlisia Thompson:** Is your metrics library also in Go? |
**Scott Mansfield:** Yeah, the whole thing... It's all in the same repository. |
**Carlisia Thompson:** Because I was thinking to ask you as well, I'm not sure if it applies what I wanna ask anyway... If you're using a Prometheus to collect metrics and log information. |
**Scott Mansfield:** Yeah, so our deployment is actually quite interesting. On the same server we're running Rend, memcached, our L2 disk-backed solution called Mnemonic which reuses part of the Rend code, and a Java-based Sidecar process that's actually the hook into the rest of the ecosystem. So in Netflix we have a ... |
**Erik St. Martin:** And now, the mnemonic part of it was the RocksDB portion of Rend, right? |
**Scott Mansfield:** Yeah. |
**Erik St. Martin:** We should probably back up a little bit too, because we discussed it a little bit in another show, but we might wanna kind of talk about what Rend is, and the components of that and what you're using it for, to kind of give a better understanding of what it does and why performance was so critical. |
**Scott Mansfield:** \[08:02\] Okay, yeah. So earlier, as I mentioned, I work on EVCache, which is a distributed charted memcached. It's the second or third highest volume system that we have here. It's a cache that fronts pretty much everything - I mean, not everything-everything, but quite a lot. It's in all three of... |
**Erik St. Martin:** Wow. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.