Consensus

Where it shows up in real systems

• Distributed SQL

  Each write is committed through a consensus protocol (Raft or Paxos), so it is never lost and every read is up to date, even across regions. The price is latency: a write must reach a majority of replicas, and across regions that means tens of milliseconds.

  In production: Google Cloud Spanner, CockroachDB, YugabyteDB, TiDB, Amazon Aurora DSQL

• Coordination service

  Runs Raft or ZAB on an odd number of nodes: three nodes survive one failure, five survive two. Use it for small, critical data, never as a general database.

  In production: etcd, Apache ZooKeeper, Consul KV