PACELC Theorem

The PACELC theorem, proposed by Daniel Abadi (2012), extends CAP by addressing what happens in the absence of a network partition. CAP only describes behavior during partitions; PACELC recognizes that even under normal operation, distributed systems face a fundamental trade-off between Latency (L) and Consistency (C). The full statement: if a Partition occurs, choose between Availability and Consistency (PAC); Else (normal operation), choose between Latency and Consistency (ELC). This makes PACELC more practically useful than CAP — systems like DynamoDB (PA/EL), Cassandra (PA/EL), and Google Spanner (PC/EC) can be precisely categorized and their trade-offs predicted.

Key Points

• PACELC notation: PA/EL (DynamoDB, Cassandra — available during partitions, low latency normally), PC/EC (Google Spanner, Zookeeper — consistent during partitions, consistent normally).
• Google Spanner achieves PC/EC using TrueTime (GPS + atomic clocks) to minimize uncertainty windows, achieving external consistency with ~7 ms commit latency globally.
• Cassandra's consistency level tuning (ONE vs. QUORUM vs. ALL) is a direct PACELC slider: ONE is PA/EL, QUORUM is PA/EC/EL compromise, ALL is PA/EC.
• The ELC trade-off is why synchronous replication adds write latency — waiting for the replica ACK adds one network round-trip (≥1 ms intra-AZ, ≥70 ms cross-region).
• PostgreSQL with synchronous_commit=on is PC/EC; with synchronous_commit=off it becomes PA/EL, accepting possible data loss on crash for lower write latency.
• PACELC helps explain why geo-distributed systems feel slower — strong consistency (EC) across continents requires consensus rounds that accumulate speed-of-light latency.
• Most real-world databases are tunable on the ELC axis: eventual → session → monotonic read → monotonic write → strong consistency — each level adds coordination overhead.
• For interview depth: CAP is sufficient for "distributed systems 101"; PACELC demonstrates mastery of the nuances that distinguish principal-level engineering thinking.