Raft gets a cluster to agree on an ordered log of commands: elect a leader (randomized timeouts), replicate entries to followers, commit once a majority acknowledges. Designed as the teachable alternative to Paxos, it runs etcd, Consul and most modern coordination layers. Rule of thumb: consensus for the small critical state (locks, leaders, membership), not the data path.
Worked example: Raft elects one leader that owns writes; followers replicate its log and an entry commits once a majority (3 of 5) hold it, so the cluster tolerates a minority failing. A missed heartbeat triggers a new election. Gotcha: writes need a majority, so a partition without quorum halts writes — availability is deliberately traded for consistency.