RAFT : In Search of an Understandable Consensus Algorithm

InSearchofanUnderstandableConsensusAlgorithm DiegoOngaroandJohnOusterhout StanfordUniversity (DraftofApril7,2013,undersubmissiontoSOSP) Abstract was our most important criterion in evaluating de- Raftisaconsensusalgorithmformanagingareplicated sign alternatives. We applied specific techniques to improveunderstandability,includingdecompositionlog. It produces a result equivalent to Paxos, and it is (Raft separates leader election, log replication, andas efficient as Paxos, but its structure is different from safety so that they can be understood relatively in-Paxos; this makes Raft more understandable than Paxos dependently)andstatespacereduction(Raftreducesand also providesa better foundationfor building practi- the degree of nondeterminism and the ways serverscal systems. In order to enhance understandability, Raft separates the key elements of consensus, such as leader canbeinconsistentwitheachother,inordertomake electionandlogreplication,anditenforcesastrongerde- iteasiertoreasonaboutthesystem). greeofcoherencytoreducethenumberofstatesthatmust • Strongleader: Raftdiffersfromotherconsensusal- be considered. Raft also includes a new mechanism for gorithms in that it employs a strong form of leader- changingtheclustermembership,whichusesoverlapping ship where only leaders (or would-be leaders) issue majorities to guarantee safety. Results from a user study requests; other servers are completely passive.