Proof
Live Consistency Lab
The two guarantees that make Aurora DSQL the right database for safety-critical decisions, run them yourself, against the live cluster.
No stale-safe read, across regions
Issue a recall on Region A’s endpoint, then read it back from Region B’s endpoint, immediately.
us-east-1
writes
SAFE
epoch 0
us-east-2
reads
SAFE
epoch 0
Commits a recall through the us-east-1 endpoint, then reads the same row from the us-east-2 endpoint. With DSQL strong consistency, Region B sees it immediately, no replication-lag window.Commits a recall through the us-east-1 endpoint, then reads the same row from the us-east-2 endpoint. With DSQL strong consistency, Region B sees it immediately, no replication-lag window.
Optimistic concurrency, on the guard row
Two transactions write the same model’s safety guard at once. DSQL lets one commit and rejects the other, the conflict our retry wrapper handles.
Transaction Awaiting
Transaction Bwaiting
Runs two real transactions in parallel that both write the same guard row. DSQL's optimistic concurrency lets one commit and rejects the other with SQLSTATE 40001. The winner varies each run, a genuine race.Runs two real transactions in parallel that both write the same guard row. DSQL's optimistic concurrency lets one commit and rejects the other with SQLSTATE 40001. The winner varies each run, a genuine race.
Correctness under load
Fire 100 concurrent sale attempts at a recalled unit, against the live multi-region cluster. Every one must be blocked, no amount of concurrency may let a recalled unit sell.
Fires 100 concurrent sale attempts at a recalled unit against the live cluster, then reports throughput and latency. Every attempt must be blocked: zero recalled units may sell under load.Fires 100 concurrent sale attempts at a recalled unit against the live cluster, then reports throughput and latency. Every attempt must be blocked: zero recalled units may sell under load.