Gnurdle / raft

a clojure raft implementation

raft

A implementation of the Raft consensus algorithm written in Clojure. This implementation has a heavy emphasis on providing an algorithm disentangled as much as possible from concerns like network transport.

This implementation provides a function com.manigfeald.raft/raft that creates a value that describes the state of the algorithm, and a function com.manigfeald.raft/run-one that steps a given state of the algorithm to the next state.

The state of the raft algorithm is described in five records:

1. ImplementationState
   • The overall container of the rest of the state
2. Timer
   • a number representing now
   • a number representing the next timeout
   • the period between timeouts
3. IO
   • a possible message received
   • a PersistentQueue of messages to send out
4. RaftState
   • raft state common to all nodes
5. RaftLeaderState
   • raft state used by the leader node

Why is this so weird?

You can think of it more as a model of a Raft implementation, but the model is in clojure so it is easy enough to just wire the model in to some transports and run it. Like wiring in to a spreadsheet model, but without the instinctive gag reflex of wiring in to a spreadsheet.

Usage

[com.manigfeald/raft 0.1.0]

call com.manigfeald.raft/raft to get an initial state, call com.manigfeald.raft/run-one to advance the state. if you want the node you are modeling to receive a message, assoc the message in to [:io :message] in the state, [:io :out-queue] will contain messages the node you are modeling wants to send out. Update the value of [:timer :now] to model the passage of time.

the tests have an example using core.async channels as the transport

Generated Docs

http://ce2144dc-f7c9-4f54-8fb6-7321a4c318db.s3.amazonaws.com/raft/index.html

See also

https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf

What is missing

The raft paper also has snapshot generation to avoid logs growing too large and an interesting scheme for changing cluster membership, both of which are unimplemented.

The raft paper also talks about providing a fast path for reads that avoids the entire log and commit process, I have left that out over fears of screwing it up, so reads and writes both go through the entire consensus process

License

Copyright © 2014 Kevin Downey

Distributed under the Eclipse Public License either version 1.0 or (at your option) any later version.