Amazon RDS for PostgreSQL 17.4

a year ago

PostgreSQL uses MVCC for transaction isolation, with 'Read Uncommitted' and 'Read Committed' both acting as Read Committed, 'Repeatable Read' providing Snapshot Isolation, and 'Serializable' ensuring Serializability.
Amazon RDS for PostgreSQL is a managed service automating tasks like provisioning, replication, and backups, with Multi-AZ deployments enhancing fault tolerance through synchronous replication.
RDS provides primary and reader endpoints; the primary supports all isolation levels, while secondaries do not support Serializable, limiting the strongest consistent level to Snapshot Isolation.
Testing involved a Jepsen-adapted workload on RDS clusters, focusing on transactions over integer lists to detect isolation anomalies via dataflow dependencies.
Amazon RDS for PostgreSQL exhibited G-nonadjacent cycles (violating Snapshot Isolation), including Long Fork anomalies, across all tested versions (13.15 to 17.4).
The anomalies suggest RDS may offer Parallel Snapshot Isolation, weaker than single-node PostgreSQL's Strong Snapshot Isolation, potentially affecting transaction consistency.
Users are advised to scrutinize transaction structures for Long Fork risks or restrict safety-critical operations to the writer endpoint to mitigate anomalies.
Jepsen's findings are experimental, demonstrating bugs but not proving correctness, with no compensation received for the independent analysis.

Hasty Briefsbeta