What a partition is
The US data centre sends a message to the EU and waits for the reply. None comes. Any of three things may have happened:
- The link is cut, and the message is lost.
- The EU data centre has crashed.
- The EU data centre is slow, and its reply will come late.
From the US side all three look the same: silence, and then its , the limit after which it stops waiting. A timeout says that no reply came in time. It does not say why.
The servers are split into groups that keep running and keep getting requests, while no message gets from one group to another in time.
Of these three, only the cut is a fault of the network, but no server is told which one it has met. It sees a reply that does not come, and must decide without knowing the cause. So the US side must treat any reply that misses its timeout as a cut, whatever the cause: a cut link, a crashed EU and a slow one look the same from inside.
CAP names three properties: consistency, availability and partition tolerance. The P is not a feature a has. It is an assumption about the network: messages may be lost at any time, and for any length of time.
A partition leaves both sides running but unable to reach each other, looks from inside like silence, and CAP's P assumes one can happen at any time.