The full picture
| Protocol | Leader | Distinctive trait | Ships in |
|---|---|---|---|
| Basic Paxos | none (any proposer) | agrees on a single value; dueling proposers can livelock | textbooks; parts of Chubby |
| Multi-Paxos | stable distinguished proposer | 1 RTT steady state; under-specified in the paper | Chubby, Spanner, older Google stack |
| Raft | strong, mandatory | understandability as an explicit design goal | etcd, Consul, CockroachDB, TiKV, KRaft |
| ZAB | strong, with sync phase | primary order + recovery-driven follower sync | ZooKeeper |
| Viewstamped Replication | strong | the 1988 original; Raft rediscovered much of it | academic lineage |
- The one real behavioural difference worth citing: leaderless Paxos permits dueling proposers — two proposers repeatedly out-preparing each other and making no progress. Raft's randomized timeouts make that vanishingly unlikely. This is a liveness difference, never a safety one.
- All of them are safe under partition and all of them stop accepting writes without a majority. If an interviewer is fishing for 'which is more available', the correct answer is that they're identical on that axis, and the real availability differences come from failover speed (timeout tuning) and from client behaviour during elections.
- EPaxos / Flexible Paxos are where the genuine innovation is: EPaxos is leaderless with commutative fast paths (great for geo-distribution, no leader bottleneck); Flexible Paxos observes that the election and replication quorums need only intersect each other, not both be majorities — which lets you trade write latency against failover cost.
- The architect's actual decision is almost never 'which protocol' but: how many replicas, spread across how many failure domains, with what election timeout, and — most importantly — what belongs in the consensus system at all. Consensus is for small, critical, low-volume state.
- Cost of coordination scales badly with distance: a 3-region majority means every write waits for the second-closest region. That single fact drives more architecture (regional sharding, single-region write masters, async cross-region replication) than any protocol choice.