Split-Brain Problem
Subject: Computer and Technology noun /ˈsplɪt ˌbreɪn ˈprɑː.bləm/ 1 definition
- distributed-systems
- fault-tolerance
- high-availability
- data-consistency
-
Definition (noun)
Entry #356 · UpdatedThe Split-Brain Problem is a failure condition in clustered or replicated systems in which a communication breakdown divides nodes into isolated subgroups that each operate independently under the false assumption that peer nodes have failed. When heartbeat messages or replication links are severed by a network partition, congested switch, or asymmetric fault, nodes on multiple sides of the break can trigger failover elections or promote themselves to primary status simultaneously. Because partitioned clients may continue reading from and writing to conflicting active primaries, the separated halves accept incompatible updates, overwrite shared storage, or grant duplicate resource locks, producing silent data corruption and state divergence when connectivity is restored. Preventing split-brain behavior establishes a critical safety boundary in high-availability architectures, forcing systems to trade minority-side availability for consistency through majority quorums, fencing tokens, and node-isolation mechanisms rather than relying solely on naive peer-reachability timeouts.
Common Forms
Dual-Primary Replication Split: Occurs in leader-follower database or storage clusters when a partitioned primary continues serving clients while a promoted standby accepts new writes, creating diverging transaction histories.
Shared-Storage Corruption: Arises in clustered filesystems or storage area networks when two isolated compute nodes concurrently mount and modify the same shared volume without SCSI persistent-reservation fencing, destroying metadata integrity.
Symmetric Two-Node Tie: Happens in two-node high-availability pairs lacking an external witness or quorum arbitrator, where a broken interconnect causes each host to assume the other has crashed and claim primary ownership.
Asymmetric Partition Divergence: Emerges during partial network faults—such as unidirectional packet loss or segregated management and client VLANs—that block inter-node cluster heartbeats while allowing external clients to mutate multiple partitions.
Example
In a two-node PostgreSQL cluster managed by Pacemaker and Corosync with
stonith-enabled=falseandno-quorum-policy=ignore, a link failure on the10.10.0.0/24interconnect cuts Corosync heartbeats betweendb-1anddb-2. Each node assumes its peer is offline, sodb-2binds the virtual IP192.168.50.10via gratuitous ARP and invokespg_ctl promotewhiledb-1keeps accepting writes over existing TCP connections, causing both primaries to commit conflicting rows toordersacross divergent WAL timelines.