PACELC is a way to compare distributed database trade-offs. CAP asks what happens when a network partition prevents nodes from communicating. PACELC adds a question for normal operation: how much latency is a system willing to accept for consistency? section 6.2 section 6.3
The name describes two separate decisions:
- If a partition occurs (P), favor availability (A) or consistency (C). An available node may keep responding with outdated data. section 6.3 section 6.2
- Else (E), when there is no partition, favor low latency (L) or consistency (C). section 6.3
Combining the decisions gives four possible designs:
- PA/EL favors availability during partitions and low latency otherwise; Amazon DynamoDB is an example. section 6.3
- PC/EC favors consistency during partitions and in normal operation; CockroachDB is an example. section 6.3
- PA/EC favors availability during partitions but consistency when the network is functioning normally. section 6.3
- PC/EL favors consistency during partitions but low latency in normal operation. section 6.3
Why response time changes the choice
Suppose a request eventually receives a non-error response, but only after thirty minutes. It can meet CAP's definition of availability while being unusable to a person waiting for it. PACELC makes response time explicit; a practical latency target might be 500 ms. section 6.10.1 section 6.3
Ask both questions when evaluating a database: what does it prioritize when nodes cannot communicate, and what does it prioritize when they can? The second answer matters even if partitions are rare. section 6.3


