Partition Tolerance
The ability to keep operating when the network drops or delays messages between nodes — not a choice, but a property of any system spanning more than one machine.
The most misread letter in CAP. Partition tolerance is frequently discussed as though it were optional — "we chose CA" — but a partition is a network event, not a design decision. If your system spans two machines, the link between them can fail, and something must happen when it does.
So the real statement of CAP is narrower and more useful: during a partition, choose availability or consistency. "CA" describes a single-node system, or a system that has not thought about it.
What counts as a partition is also broader than a cut cable. A garbage-collection pause long enough to miss heartbeats, a saturated network link, an overloaded node that stops responding, or a misconfigured firewall rule all present identically to the rest of the cluster: messages stop arriving. Which is why partition handling gets exercised far more often than the phrase "network partition" suggests.
The practical implication: every distributed system needs a written answer to what does this do when a node cannot be reached, and how do we tell that apart from the node being dead?