Paxos Made Practical

Paxos [3] is a simple protocol that a group of machines in a distributed system can use to agree on a value proposed by a member of the group. If it terminates, the protocol reaches consensus even if the network was unreliable and multiple machines simultaneously tried to propose different values. The basic idea is that each proposal has a unique number. Higher numbered proposals override lower-numbered ones. However, a “proposer” machine must notify the group of its proposal number before proposing a particular value. If, after hearing from a majority of the group, the proposer learns one or more values from previous proposals, it must re-use the same value as the highest-numbered previous proposal. Otherwise, the proposer can select any value to propose. The protocol has three rounds. In the first round, the proposer selects a proposal number, n > 0. n’s low-order bits should contain a unique identifier for the proposer machine, so that two different machines never select the same n. The proposer then broadcasts the message prepare(n). Each group member either rejects this message if it has already seen a prepare message greater than n, replies with prepare-result(n, v) if the highest numbered proposal it has seen is n < n for value v, or replies with prepare-result(0,nil) if it has not yet seen any value proposed. If at least a majority of the group (including the proposer) accepts the prepare message, the proposer moves to the second round. It sets v to the value in the highest-numbered prepareresult it received. If v is nil, it selects any value it wishes for v. The proposer then broadcasts the message propose(n, v). Again, each group member rejects this message if it has seen a prepare(n) message with n > n. Otherwise, it indicates acceptance in its reply to the proposer.