This is an akka based implementation of the Raft consensus algorithm. It is generic enough that you can build your own replicated state machines on top of it (with raft keeping track of the consensus part of it).
This implementation is akka-cluster aware, so it can be easily deployed on multiple machines. Implementation wise, all parts of the raft whitepaper are covered:
- Leader election
- Dynamic membership changes (with transitioning periods implemented as Joint Consensus)
- Deployment across multiple nodes
- Log compaction, via snapshotting
This project is still work in progress and has not been stress tested (athough it is tested on multiple nodes already)
Raft is a distributed consensus algorithm, much like Paxos (but simpler). This implementation is fully akka (and akka-cluster) based, and can be used to deploy a replicated state machine on top of akka clusters.
THIS API IS STILL SUBJECT TO CHANGE
class WordConcatRaftActor extends RaftActor {
type Command = Cmnd
var words = Vector[String]()
/**
* Called when a command is determined by Raft to be safe to apply;
* Application results are sent back to the client issuing the command.
*/
def apply = {
case AppendWord(word) =>
words +: word
log.info("Applied command [{}], full words is: {}", command, words)
word // will be sent back to original actor, who sent the AppendWord command
case GetWords =>
val res = words.toList
log.info("Replying with {}", res)
res
}
}
// ...
val members = (1 to 3) map { i => system.actorOf(Props[WordConcatRaftActor], name = s"raft-member-$i") }
val clusterConfiguration = ClusterConfiguration(raftConfiguration.members + additionalActor) // 0, 1
members foreach { _ ! ChangeConfiguration(clusterConfiguration)
// todo implement re-routing if you send to a non-leader
// then send messages to it; the state machine will only be applied when consensus has been reached about a value
leader ! ClientRequest(AppendWord("I"))
leader ! ClientRequest(AppendWord("like"))
leader ! ClientRequest(AppendWord("capybaras"))
// ... after some time
leader ! GetWords
expectMsg(List("I", "like", "capybaras"))
And if you want to enable snapshotting support it's as simple as implementing one method and matching for InstallSnapshot
in your Actor:
class SnapshottingWordConcatRaftActor extends RaftActor {
type Command = Cmnd
var words = Vector[String]()
def apply = {
case AppendWord(word) =>
words +: word
word
case GetWords =>
val res = words.toList
log.info("Replying with {}", res)
res
case InstallSnapshot(snapshot) =>
words = snapshot.data.asInstanceOf[Vector[String]]
}
override def prepareSnapshot(meta: RaftSnapshotMetadata) =
Future.successful(Some(RaftSnapshot(meta, words)))
}In the above examples, the client implementation is very naive, and assumes you have some way
of finding out who the current Leader is (as this is a requirement to interact with any Raft cluster).
Thankfully, you can use the provided RaftClientActor, which works like a proxy that forwards all your messages
to the current Leader, or stashes them if the cluster has no Leader at the moment (is undergoing an election) and
sends the messages once the Leader becomes available.
Simply: Apache 2.0
Issues, Pull Requests as well as Tweets and Emails are more than welcome!
- Raft - In Search of an Understandable Consensus Algorithm whitepaper
- See other implementations (many lanugages) on raftconsensus.github.io
- Docs on akka-cluster in 2.3.0-RC2.
We have discussed this paper both in Kraków and London, on these awesome reading clubs (drop by if you're into CS papers!):
