Split-brain, consensus, and quorum

What's really being asked

Whether you understand the failure mode of dual leadership under partitions and how majority-based consensus structurally prevents divergent committed state.

Why it matters

Split-brain occurs when a network partition divides a cluster and each side independently believes it holds leadership. If both sides accept writes, they diverge irreconcilably, and merging later may silently drop or corrupt data. This is the worst outcome for a system of record.

The full answer

Consensus algorithms like Raft and Paxos prevent split-brain by requiring a majority quorum, more than half of all nodes, for two critical actions: electing a leader and committing a log entry. A candidate becomes leader only after receiving votes from a majority, and in any partition at most one side can hold a majority. The minority side cannot elect a leader nor commit writes, so it correctly refuses to make progress and serves at most stale reads. Raft adds monotonically increasing terms so stale leaders are detected and stepped down.

The mistakes people make

Claiming heartbeats or timestamps alone prevent split-brain, allowing the minority partition to keep writing, or confusing a quorum with simply any two nodes regardless of cluster size.

What usually comes next

Why odd cluster sizes are preferred; what happens with an even split; how fencing tokens stop a deposed leader's late writes from being applied.

A concrete example

A five-node cluster partitions three-to-two. The three-node side retains a majority, elects a leader, and continues committing. The two-node side cannot reach the quorum of three, so it stops accepting writes. When the partition heals, the two stale nodes replay the leader's log and rejoin cleanly with no divergence.