Aurora DSQL: Scalable, Multi-Region OLTP

AWS paper describing the architecture of Aurora DSQL, a serverless, PostgreSQL-compatible OLTP database with active-active multi-region capabilities.

The design is fully disaggregated into three horizontally-scalable layers: compute (query processors in Firecracker MicroVMs, stateless), transaction coordination (distributed adjudicators + a Journal replication system), and storage. Query processors are PostgreSQL-compatible and carry no local state, so they can be scaled to zero.

The key latency trick is deferring all coordination to commit time. Multiversion concurrency control with precise timestamps lets reads proceed coordination-free, while writes use optimistic concurrency control resolved by adjudicators at commit. This means cross-region coordination is needed only for commits — not for every statement — keeping p50 read-write latency low even across regions.

The system offers strong consistency, full ACID transactions, and continuous availability through zone and region failures, scaling from zero to millions of transactions per second.