What does the CAP theorem actually say?

Consistency in CAP means linearizability: every read sees the most recent completed write. Availability means every request to a non-failed node receives a non-error response. Partition tolerance means the system keeps operating despite dropped messages between nodes.

Networks partition. If a node cannot reach its peers, it must either reject some requests to avoid returning stale data, or answer with possibly stale data. That is the real decision. A system can be consistent and available when the network is healthy, which is why CAP alone is a coarse tool.

CP and AP in practice

• CP during a partition: minority replicas refuse writes, so clients see errors or higher latency instead of divergence
• AP during a partition: replicas keep accepting writes locally and reconcile later, so reads may be stale
• Leader-based systems are usually CP for writes; leaderless quorum systems can be tuned toward either side
• CAP says nothing about latency when the network is healthy, which is often the more important trade-off

Healthy network:  C and A can both hold
Partition:        choose C or A for the affected operations
PACELC:           if Partition, trade A vs C
                  else trade Latency vs C

PACELC extends the idea: when there is no partition, the trade-off is latency versus consistency. A quorum read that waits for a slow replica is choosing consistency over latency, even though no partition exists.