Paxos Made Simple
Leslie Lamport2001ACM SIGACT News 2001
Read it on lamport.azurewebsites.net(opens in a new tab)Why this one
Read this after In Search of an Understandable Consensus Algorithm. Raft gives you the engineer's shape of replicated logs, leaders, terms, and commitment. Paxos gives you the smaller, stranger core underneath: how a group can choose one value even when messages vanish, machines pause, and nobody can tell whether a peer is dead or slow. What people get wrong is treating Paxos as a protocol you copy into production. The paper is more useful as a proof-shaped mental model. It teaches you which guarantees come from quorums, which come from promises, and why the hard part is not electing a leader but preserving safety across every half-finished attempt. If you build storage, queues, schedulers, config systems, or anything that says highly available, this is an evening well spent because it changes what you distrust in your own designs.
What to take away
- Majorities are used so any two successful decisions must share at least one acceptor.
- The prepare promise prevents an older proposal from overwriting a value that may already be chosen.
- Paxos separates safety from liveness, which is why it can be correct while still making no progress.
Reads with
- Impossibility of Distributed Consensus with One Faulty Process
explains the asynchronous fault model paxos is designed to survive safely
- In Search of an Understandable Consensus Algorithm
recasts the paxos safety core as an engineer friendly replicated log
- Spanner: Google's Globally-Distributed Database
shows paxos embedded under a real global transactional database