The phrase is most common in clusters, databases, storage systems, and distributed control planes. The danger is not only downtime. The larger danger is conflicting decisions made by isolated sides that later need to be reconciled. If both sides accepted writes, sent orders, or changed configuration, recovery can become manual and risky.
A concrete example is a two-node cluster with a flaky private network. The nodes stop seeing each other. Each node still sees some clients and believes the other node is dead. Both become primary. For a few minutes everything appears partially alive, then data, locks, or external side effects disagree. The incident report no longer asks “why was the service down?” but “which truth do we keep?”
Preventing split-brain usually requires quorum, fencing, odd-numbered voters, clear promotion rules, and failure drills that include network partitions rather than only clean shutdowns. The safest design makes inaction preferable to unsafe double ownership during real incidents.