CAP is about partitions, not outages

CAP says something narrow. When the network between parts of a replicated system fails, the system must either refuse some requests or let its copies diverge until the partition heals. While the network is healthy it can be both consistent and available.

It is usually quoted as pick two of three, which makes it sound like a permanent property. Daniel Abadi argues that only an actual partition forces the trade, that partitions are comparatively rare, and that systems relaxing consistency in normal running are responding to latency instead. His PACELC names both trades: under a partition, availability or consistency; else, latency or consistency, the everyday form of the synchronous replication trade. Newman adds that the choice need not be system-wide. One service can tolerate a stale balance on display yet insist on consistency when the balance changes.

For regional failover the useful reading follows from that. A region that has truly failed accepts no writes, so nothing diverges; the risk lies in a region that only looks gone while still serving users. Choosing availability there leads to divergence. Choosing consistency means refusing writes somewhere, and Read-only mode is a legitimate answer is the least painful way to do it.