Knowing the cluster
Every node needs a current view of which peers are alive and reachable. With thousands of nodes, a central registry becomes a bottleneck, so many systems use gossip instead.
How gossip spreads
Periodically, each node picks a few random peers and exchanges its membership view with them.
- New information spreads exponentially, reaching all nodes in roughly log many rounds.
- There is no single coordinator, so the protocol has no single point of failure.
- Bandwidth per node stays bounded because each node talks to only a few peers per round.
Convergence and scale
Gossip is eventually consistent: views may briefly disagree, but they converge quickly. This trade is worth it because gossip scales to very large clusters and tolerates message loss gracefully.
Key idea
Gossip spreads membership state through random peer exchanges, giving each node an eventually consistent view of the cluster without any central coordinator.