Cloud Native Computing Foundation EnvoyCon North America 2021, 30 Oct 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Lightning Talk: State of Go Control Plane: What Does the Future Behold? - Alec Holmes, greymatter.io

Description

Don’t miss out! Join us at our next event: KubeCon + CloudNativeCon Europe 2022 in Valencia, Spain from May 17-20. Learn more at https://kubecon.io The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects.

Lightning Talk: State of Go Control Plane: What Does the Future Behold? - Alec Holmes, greymatter.io

In the recent years of Envoy’s lifecycle, the service mesh landscape has matured drastically. With the industry growing and utilizing the xDS APIs, the products coming to market and the various solutions designed around the complexities of envoys discovery mechanisms, control plane bifurcation and disparity are more relevant than ever. Go Control Plane has become a targeted part of the Envoy ecosystem, and this talk will walk through what's planned in the pipeline, and what has been merged in this year.

A

Well, uh thank you guys for coming to my talk. uh My name is alec. I am a core product engineer at gray matter, um and also one of the active maintainers of envoy's go control plane, and uh I want to take this little. This little 10 minute talk to bring everyone up to speed on what exactly has actually happened this past year um and where we, where we intend to take uh this project as it's become a pretty critical part of the core infrastructure for the envelope ecosystem.

A

So, let's first start off with a little recap of: what's actually happened so if you're, following the xds lifecycle, v2 has been pulled out so we're pretty much strictly on v3 and focusing more on the ud which is v4 and forthcoming.

A

So in 0.9.9 I believe we pulled out v2. So if you still are dependent upon that you're going to need to go ahead and pin back to an old version in june, we introduced ecds.

A

So as the project's growing, more xds services are coming out more discovery and we are continually integrating those into go control plane, as there are quite a lot of projects that do do rely on us, so uh that was that came in in june.

A

um This I'll talk more on in a little bit, but uh profiling has become a big target for go control plane. We know that again. This project is required in a lot of places upstream, but none of us really know what actually is happening under the covers. We've, never really profiled, we've, never really benchmarked. We know it works, but we don't know how. Well in august, we actually reworked the snapshot api.

A

So I don't know if you guys are familiar with the codebase, but this is pretty much the core interface with the cache and uh we found that it was really difficult to continually update for those rolling discovery services that kept coming out.

A

uh So we readjusted, we changed the api and we haven't reached 1 0 yet, but I'm going to talk about that in a little bit, so um we just merged those in and we called it another minor release incremental went in also that was a pretty big milestone, but I won't talk too much upon that since last year.

A

I gave a talk on this, so if you guys want to go learn more about that, I would recommend checking that talk out and then we just actually merged in the scoped route discovery service, and that was a pretty big change to the core internal logic, because it obviously changed a little bit how we handled route associations to listeners.

A

So that's pretty much what happened in 2021 um and then I'll go ahead and talk about where we want to go with the repo.

A

So what are some goals for go control? Plane? um We've had a lot of requests for a 1.0 release. Cut we've had some pretty major projects now become dependent upon gcp, so, like grpc go their xcs implementation requires on us big projects like istio, kuma, et cetera, et cetera. Those are all using go control plane, um so we can't just continually break if we just keep breaking it's going to become a huge nightmare and a big pain for all those control plants that are dependent upon our code.

A

um So what does that entail? The 1.0 stuff kind of involves a lot of protobuf upgrades and also a little bit changes to the bazel system that envoy uses.

A

uh We want to focus on benchmarking, so, like I said in the beginning, we don't really know how this is performing. There could be optimizations out there, we don't understand, and we want to investigate into that and we're going to continue to bug fix again. This is a pretty popular project. A lot of people use it, so we want to make it as stable as possible and then a reference implementation.

A

So I will save that. But that is a big big goal for go control point in the coming year.

A

So for the 1.0 we we decided to go with a multi-module release plan. So this means that uh if you're familiar with the code, we have the generated proto stubs from the actual data plane uh in the codebase, as well as the actual reference implementation of the xds server and cache. So we have to kind of be able to version those independently and asynchronously, so by doing that, we're going to treat that the proto stubs version separately and then etc, etc.

A

So those uh we have a plan, I I'd recommend downloading the slides and checking out that proposal to see how that's actually going to be versioned. It's a pretty uh complicated process, but we are working to get that done.

A

We have. We are dependent upon some deprecated libraries. We want to move to the latest, which is what envoys pointed at. So that's going to require again some of those bazel build upgrades, but that's pretty in-depth and we we are lacking a little bit of bandwidth on there. So uh if anybody wants to help out with that, please please make an issue and we would definitely reach out and like to collaborate.

A

So, for benchmarking uh this past year we integrated pprof, um which is google's, go profiler, uh where we started investigating how the code's actually performing and optimizations that we could make.

A

So the reason this was done is because incremental was merged in so we wanted to see the differences in actual compute load for the xds types and we are actually working on building a full end-to-end throughput benchmark- that's not done, but it is something that I think a lot of people will benefit from as they'll be able to see what the actual benefits of incremental from state of the world are.

A

So the the reference implementation uh previously go control plane uh was not really looked at as something that a lot of people would really use or depend on so with the ecosystem growing and maturing so quickly. uh There's been a lot of kind of issues that have occurred around uh replicating common functionality across multiple projects.

A

uh Matt actually wrote a really good blog post last on last year's envoy con um about like greenfield deployment and the code replication et cetera, and I would recommend you guys go read that, but that replicated code is a problem. People are solving the same solution over and over again. So how do we fix that?

A

Well, we bring it all upstream and uh because go control point is an envoy sponsored project. um It's a perfect place to do it, so uh we break down some common functionality across these projects. We uh we bring them up and maybe and go control plane is a really good place to do that. uh We could even potentially have something like a reference service discovery mechanism, because what do all these projects do? They're all tackling the same problem, they're all performing service discovery, they're all doing health checking circuit breaking et cetera, et cetera.

A

So is there really a need to replicate this over and over again.

A

Yeah, so that's basically, where go control point is going. That's our thoughts and ideas on the project and again we would love for people to help out and also maintain this project. uh We have a pretty small bandwidth for the maintainers.

A

um I think there's only four of us, so the more help we can get the better but yeah. Thank you guys and I will be in the slack if anybody has any questions.