Amazon RDS for PostgreSQL 17.4

Summary

Background PostgreSQL is a popular open source general-purpose SQL database. It uses multiversion concurrency control (MVCC) to provide three levels of transaction isolation. PostgreSQL’s “Read Uncommitted” and “Read Committed” are both Read Committed. The “Repeatable Read” level actually provides Snapshot Isolation, not Repeatable Read. “Serializable” provides Serializability. Amazon RDS for PostgreSQL is an AWS service which provides managed PostgreSQL clusters. RDS automates provisioning, storage management, replication, backups, upgrades, and more. Multi-AZ deployments distribute database nodes across multiple availability zones, reducing the probability of correlated failure. RDS uses synchronous replication to ensure that transactions are durable both on primary and (at least one) secondary instances before acknowledging. From a user perspective, Amazon RDS for PostgreSQL provides a pair of URLs which speak the PostgreSQL wire protocol: a primary endpoint for read-write transactions, and a reader endpoint for read-only transactions. The primary endpoint supports all PostgreSQL isolation levels, while secondaries do not support Serializable. The strongest level supported across all nodes is therefore Snapshot Isolation (which PostgreSQL terms “Repeatable Read”). Test Design We adapted Jepsen’s test library for PostgreSQL for use with Amazon RDS for PostgreSQL with a small wrapper program. For each round of tests, we provisioned an RDS cluster using AWS’s CreateDBCluster API, using gp3 storage and db.m6id.large instances. We then launched a single EC2 node to run our tests, and provided it with the main and read-only endpoints of the RDS cluster. We performed no fault injection, and triggered no failovers. As in our previous work on PostgreSQL, our primary workload consisted of transactions over lists of unique integers. We stored each list in a single row, encoded as a TEXT field of comma-separated values. Transactions could either read a list by primary key, o...