A distributed data store can guarantee at most two of consistency, availability, and partition tolerance. Pick two on the triangle — the third is what you give up. See what that costs you when the network splits, and where real databases land.
click two vertices to choose a pairing · click a third to see why you can't
The fine print. CAP describes behavior only during a network partition — when there's no partition, a system can be both consistent and available. "CA" systems don't escape the theorem; they simply assume partitions won't happen, which for anything truly distributed is a bet that eventually loses. Most modern databases are also tunable per-request (e.g. quorum levels), so the same product can behave CP or AP depending on configuration. The richer model, PACELC, adds: else (no partition), trade Latency vs Consistency.