Hacker News Reader

Comment by tedsuo | Hacker News Reader

CoreOs (etcd team) is currently rewriting their raft implementation because it had a lot of problems. I don't believe there is a "good" implementation out there yet, because it's all very new. And should an implantation be "good", AKA implement the spec without introducing incidental errors, will it still have edge cases because the spec has edge cases? This is all currently unknown. It's precisely these assumptions that I'm pointing out.

A quick google search brings up libpaxos and other open paxos implementations, so it appears that they exist.

View on HN

The new etcd raft implementation is working well and has a solid set of deterministic tests. We recently released an alpha of etcd that utilizes this improved raft implementation: https://github.com/coreos/etcd/releases.

The core raft state machine comes in under 600 LOC last time I checked and the docs are here: https://godoc.org/github.com/coreos/etcd/raft. There are a few non-etcd projects that have started working with it, most notably the cockroachdb folks have started to use it to implement the "multi-raft" that they need for their DB. You can find the discussion and work on github.com/cockroachdb.

It turns out getting a raft implementation that is easy to audit with a copy of the paper in hand and easy to test in a fast and consistent way requires a significant engineering effort. I hope that based on the lessons learned from goraft that this implementation is something that a number of Go projects can use and leverage. Particularly since we carefully separated the WAL, RPC and state machine from each other.

If you are looking for a good one. Checkout ours https://github.com/cloud-software-foundation/c5-replicator

Wow - this looks like a great implementation. Who is cloud-software-foundation? Is this the software that powers WANdisco?

This does not power wandisco. WanDISCO uses DConE.

Interesting - thanks for the libpaxos link. Is it any good?

I'm curious: what are these edge cases you're talking about? I've worked on a Raft implementation, and I don't remember any particular edge cases (though it's been a while).

There's definitely some interesting choices for when to replace nodes in the face of failure, but I don't think that's a Raft algorithm shortcoming.

Reply on news.ycombinator.com