Aurora

This mostly isn't a decision made on Aurora's own merits — match your team's existing engine and tooling. But one specific mechanism is worth knowing before you assume parity: Aurora PostgreSQL doesn't run vanilla vacuum behavior unmodified. Adaptive autovacuum (rds.adaptive_autovacuum, on by default) automatically makes autovacuum more aggressive as a table's transaction age approaches autovacuum_freeze_max_age, which vanilla PostgreSQL doesn't do on its own.

Aurora trades per-operation I/O billing (on Standard storage) and a higher instance rate for AZ-level durability and reader scaling that shares the writer's storage layer instead of replaying a binary log. You're not paying for "faster" — you're paying for a different storage architecture, and that architecture has a different cost shape than RDS's flat EBS-backed billing.

I/O-Optimized isn't a general upgrade — it's a tax on compute and storage in exchange for removing I/O charges entirely. Storage runs roughly 2.25x (about +125%) the Standard rate, and instance/compute pricing runs roughly 30% higher, in exchange for zero per-operation billing.

An Aurora cluster trades per-replica cost (up to 15 Aurora Replicas) for read scaling that shares the writer's storage layer — but the default reader endpoint trades isolation for simplicity, load-balancing all read traffic across all replicas undifferentiated.

Serverless v2 trades a per-ACU-hour premium (its rate is set so flexibility costs more than provisioned capacity at comparable steady utilization) for scaling in 0.5 ACU increments instead of whole instance classes — real elasticity for spiky or idle-most-of-the-day workloads.

Aurora's fast failover trades a same-AZ storage architecture for an application-layer requirement — your connection/retry logic still has to reconnect via the cluster's DNS endpoint, because "fast" at the database layer doesn't automatically mean invisible at the application layer.

Backtrack (MySQL-compatible only) trades storage cost for change records — billed per million change records per hour — for a fast in-place rewind that avoids a full restore-to-new-instance. It's not a backup; it's a precision undo tool with its own cost and its own limits.