One-liner: Under a network partition a distributed store must choose consistency or availability; PACELC adds that without partitions you still trade latency vs consistency.
[A] <--X--> [B] partition between A and BCP: minority side returns errors / blocks until healed -> consistent, not availableAP: both sides accept writes, diverge, merge later -> available, not consistent
Partitions are not optional in a real network, so P is a given; the choice is C vs A during a partition.
CP: leader needs majority acks; nodes that can't reach quorum stop serving writes (and often linearizable reads).
AP: each side serves; conflicts resolved with last-write-wins, vector clocks, or CRDTs; anti-entropy repairs.
PACELC in normal operation: synchronous replication = higher latency, lower lag (PC/EC, e.g. Spanner, VoltDB); async = lower latency, stale reads (PA/EL, e.g. Cassandra, DynamoDB default).
CAP "consistency" ≠ ACID consistency. CAP's A ≠ "high availability" in the SLA sense; a CP system can still be 99.99% available in practice.