Notesreplication🌱 seedling

Leaderless Replication

Also: Dynamo-Style Replication · Leaderless Quorum Replication

No replica is special: writes and reads go to all n replicas of a key in parallel and succeed once w (or r) respond. Stale replicas are fixed later by read repair and anti-entropy, not by a log.

Why it exists

Built for "always writeable" stores such as Amazon's shopping cart: there's no leader to fail, no failover, and a write succeeds as long as enough replicas are reachable.

How it works

  • Each key has n replicas. The client, or a coordinator node acting for it, sends each write to all n and waits for w acknowledgements.
  • Reads go to all n and wait for r responses; each carries a version, and the newest wins. See Quorum.
  • A replica that was down just misses writes; it catches up via Read Repair, hinted handoff or Anti-Entropy.
  • Concurrent writes need resolution: Last Write Wins, Version Vectors with siblings, or CRDTs.

Trade-offs

Gains Costs
no failover; tolerates n − w down for writes never linearizable, even with strict quorums
tunable latency vs freshness per request conflicts on concurrent writes
no single write bottleneck repair is an operational duty

In practice

  • Cassandra exposes the choice per query as a consistency level (ONE, QUORUM, LOCAL_QUORUM, ALL).
  • Amazon DynamoDB is not leaderless despite the name: it elects a leader per partition with Multi-Paxos.

Pitfalls

  • Assuming quorum reads are strongly consistent. See Quorum for the edge cases.

Recall

These cards are also in your review deck, scheduled with spaced repetition.