A deep technical analysis of the Aurora DSQL architecture paper by a former AWS team member who worked on the project. The post breaks down DSQL's core design philosophy of 'exploding the monolith' into independent horizontally scalable services: Query Processors, Storage Nodes, Adjudicators, Journals, and Crossbars. Key architectural bets include relying on synchronized physical clocks (AWS TimeSync) for coordination-free reads, Optimistic Concurrency Control paired with MVCC under Snapshot Isolation, strong linearizable consistency, hard transaction size caps (3,000 rows / 10MiB), and a Warp-inspired single-leader 2PC optimization. The payoff is 0-RTT consistent reads, 1-RTT commits, independent scalability of compute/storage/commit layers, and no pessimistic lock contention. Tradeoffs include write-write conflicts on hot keys under OCC, WAN latency penalties before conflict detection in multi-region scenarios, and friction with foreign key constraints and high-locality sequences. The author also reflects on how reusing existing AWS infrastructure (Journals, PostgreSQL engine) and strong upfront design made building a novel global database surprisingly tractable.
