Two regions cannot outvote each other

With voters in two regions, one region always holds the majority or the two tie. Losing the majority side, or either side of a tie, leaves the survivors unable to elect a leader or commit a write. No split of voters across two regions survives the loss of either one.

MongoDB's documentation walks through the case: a set with three members in one data center and two in another stays writable if the smaller one fails and becomes read-only if the larger one does. Its warning is general, that a deployment in only two data centers can stop taking writes when the survivor cannot form a majority. CockroachDB states the same limit as a rule, requiring at least three database regions to survive losing one.

Many two-region designs claiming tolerance only tolerate losing the smaller region. Failing over without a majority reopens split brain. The fix is a third location that can vote, one that need not serve traffic or hold a full copy: A third region can be a witness.

The limit reaches past databases. Locks and leader election for other components usually rest on etcd, ZooKeeper or Consul, which are consensus stores underneath and inherit the same arithmetic.