Anti-Entropy

Every node tracks a small set of {key, value, version}. Two nodes meeting:

1. Each side sends a digest (key → version).
2. Each requests keys it has older versions of.
3. Other side sends those updates.

Versions break ties — concurrent updates need conflict resolution. Common approaches:

• Last-Write-Wins (LWW): timestamp-based. Simple; loses concurrent updates.
• Vector clocks: per-node version counters. Detect concurrency, surface conflicts.
• CRDTs (Conflict-free Replicated Data Types): merge automatically (sets, counters).

Cassandra uses LWW with timestamps. Riak supports vector clocks. SoundCloud's Roshi uses CRDTs.

Two flavors:

• Push: I send my updates to a peer.
• Pull: I ask a peer for their updates.
• Push-pull: both. Most efficient.