A partition looks like an outage from inside

When region B stops answering, region A cannot tell why. B may have failed, the link may be down, B may be alive but too loaded to reply in time, or A's own network may be the broken part. From A these are one observation: silence.

Gorton makes the point for a single request: a crashed server, an overloaded one, and a lost reply look the same to the caller, which cannot know whether an answer is coming without waiting indefinitely. So systems decide by timeout. MongoDB members exchange heartbeats every two seconds by default and call an election after ten seconds without hearing from the primary. A timeout turns silence into a verdict, which is why Failure detectors can only suspect.

Treating unreachable as dead will sometimes promote a replica while the primary still serves users who can reach it, and then both sides take writes, the situation CAP describes.

Better detection does not remove the ambiguity; a rule about who may act works around it. A MongoDB primary stranded in a minority partition steps down, so only a majority elects a leader, and a third location such as a witness gives one side that majority.