Raft: In Search of an Understandable Consensus Algorithm - Stanford, 2014

• [Paper] as it appears in ATC’14 [Mirror]
• [Extended Paper] with notes on implementation
• [ATC’14] [Video]
• [Notes] by MIT’s Distributed Systems Reading Group with some interesting thoughts
• Here is a great visualization / tutorial on the basics of Raft
• The goal of Raft is to easy to understand (or at least easier than Paxos)
  • I did find it to be pretty easy to understand
• They note that replicated state machines are implemented with a replicated log
• Raft is based around a replicated log with leader election
• Leader election is controlled by the epoch (they call it term) and the log length.
• Each server votes for no more than one candidate per epoch (term)
• Split votes are mitigated using a randomized timeout to start an election
  • i.e. the duration a member waits before assuming the leader is unavailable is randomly chosen
  • if a split vote does occur than the participants wait for their individual (re-randomzied) timeouts
  • there is no increasing back-off as in TCP
• Notes from MIT’s DSRG:
  • One of the members wrote an implementation and said it “just worked”
  • Once you add all the features you need for a production system, they feel it is approximately the same complexity as Paxos