During a partition, a replica answers stale or not at all.
A write reached G1; G2 heard nothing. If G2 answers 0, consistency breaks. If it waits or errors, availability breaks.
Pick any two of three?
Bad: Choose CA
You can’t opt out of partitions. One side still stops or diverges.
Good: Partition? Then C or A
Only while it lasts, give up fresh answers or give up answering.
A server can’t tell a cut link from a slow one. It waits on a replica; the timer fires.
The trick
CAP is a timeout decision: cancel and lose availability, or proceed and risk consistency.
Same partition: CP refuses the minority, AP accepts and merges later.
CP isn’t “down”; the majority keeps working. AP answers everyone, then pays in conflicts that must be merged.
10 ms
of round trip per 1,000 km of fibre
PACELC’s else. With no partition at all, every replica you wait for still costs a round trip. Mumbai to Singapore is at least 40 ms.
Per operation, is a wrong answer worse than no answer?
Money or a double-sold berth? Choose consistency. A blank feed or a frozen map? Choose availability, and plan the merge.
The label belongs to a setting, not a product
- Cassandra at ONE behaves PA/EL
- At QUORUM, the minority side fails, near PC/EC
- Leader + async replicas: neither CAP-consistent nor CAP-available
Remember
During a partition, each operation answers stale or not at all. Else, it trades latency for consistency. The ledger is PC/EC; likes are AP/EL.