hesela.devSplit brain
Split brain is a clustered-system failure state in which nodes unable to coordinate independently operate a resource that requires coordinated access.
Scope: A network partition describes lost communication; split brain describes a resource-control failure that can accompany it. Coordinated multi-writer operation is not itself split brain. [1]
Recovery: DRBD distinguishes a partition from disconnected primary roles and divergent data histories. A reconnection handshake can detect split brain and drop replication. Recovery policies may select a surviving copy and discard changes on the other side, not merge application updates. [2]
Policy selection: The after-sb-0pri, after-sb-1pri and after-sb-2pri options distinguish primary-role counts at detection. They do not count the nodes that accepted writes during the disconnection. Therefore, a current Secondary role does not by itself establish that discarding that copy loses no updates. [2]
Relations
Limitations
- Successful fencing prevents further conflicting access within its enforced scope but does not reconcile writes already accepted. [1, 2]
Citations
- ClusterLabs: Pacemaker Explained 3.0, Fencing, sections 8.1-8.3 and 8.11
- LINBIT: DRBD 9 User Guide, split brain notification, recovery policies and manual recovery
Sources & caveats
- DRBD 9 guide sections 2.12, 5.19.2 and 7.5 describe implementation-specific detection and recovery, not a universal definition or a tested failover procedure.