Gossip: each node periodically picks a random peer, exchanges state. In log(N) rounds, everyone knows everything. Scales beyond centralized approaches.
The mechanism
Every T seconds, each node picks K random peers, exchanges state. State converges. No central coordinator.
The mechanism
Every T seconds, each node picks K random peers, exchanges state. State converges. No central coordinator.
log(N) spread
Update spreads to all nodes in log(N) rounds. Fast at scale. Fault tolerant — even if messages drop, next round retries.
Anti-entropy vs rumor mongering
Anti-entropy: full state sync (correctness, expensive). Rumor mongering: only recent changes (efficiency). Combined for balance.
SWIM protocol
SWIM adds failure detection. Nodes exchange 'this node alive?' pings. Suspected failures gossip through. Used by Consul, Hashicorp Serf.
Eventual convergence
Under partitions or drops, may take longer but always converges when network heals. Perfect for eventual-consistency systems.