Why it matters
Gossip is what makes large clusters converge quickly on shared state without centralized coordination. Understanding it explains a lot about distributed system behavior.
The architecture
Basic gossip: each node picks K random peers per round and sends its state. Peer merges received state with local state, propagating updates.
Anti-entropy: periodic full state exchange between random peers. Ensures long-term convergence.
How it works end to end
Rumor spreading: nodes gossip only recent updates. Efficient for hot state; anti-entropy handles catch-up for stale nodes.
SWIM: gossip variant for membership + failure detection. Combines periodic pings, indirect pings (ask peer to ping suspect), and gossip about membership changes.
Convergence time: O(log N) rounds to reach every node with high probability. Rounds are typically 100-1000 ms.