Comparison
FailovervsSplit brain
Failover
the primary stopped answering and something promoted a replica, and now you find out how much of the last few seconds you lost.
Promoting a replica to leader when the current leader is deemed dead. Under asynchronous replication it loses any unreplicated writes; under synchronous it may refuse to proceed. The hard part is the deeming: too eager and you failover on a network blip, too slow and the outage is the timeout.
Full entry →Split brain
the network partitioned, both sides decided they were the leader, and both accepted writes to the same data.
Two nodes simultaneously believing they hold exclusive leadership, usually after a partition or a slow heartbeat. The result is divergent state that has to be reconciled or discarded by hand. Preventing it is what majority quorums, fencing tokens and STONITH exist for; detecting it after the fact is expensive and manual.
Full entry →