Split brain is two primaries accepting writes

Split brain is what failover can produce when the old primary was cut off rather than dead: two nodes each believe they lead, and both accept writes. The result is two histories that disagree, not merely a gap.

It begins because A partition looks like an outage from inside, so a replica on the far side gets promoted while the original still serves the clients that can reach it. Gorton describes a Redis Cluster leader stranded in a minority partition that keeps taking writes until a timeout; those writes are lost when the partition heals. Newman's general point is that writes accepted on both sides must later be resynchronized, and the longer the partition, the harder that gets.

GitHub's October 2018 incident shows how little it takes. Connectivity returned after 43 seconds, but by then failover had begun directing writes to the West Coast, each data center held writes the other lacked, and service stayed degraded for over 24 hours. Automatic failover can make an outage worse picks up that thread.

The defences make a replaced primary unable to act: storage rejects its writes, as fencing tokens arrange, or leadership needs a majority, which consensus provides.