Backend systems · failure semantics

Transactional Outbox
Reliability Lab

See why “save the order, then publish an event” can lose messages—and how an outbox trades that failure for recoverable, at-least-once delivery.

● deterministic simulation

Run a failure scenario

Change the write pattern and crash point.

No external services. Each run uses one synthetic order and one crash boundary.

What happened

Naïve dual write · no crash

Orders DBWaiting
Outbox rowNot used by this strategy
Event brokerWaiting
Order committed—
Event intent durable—
Events delivered—
Lost / duplicate—
ReadyChoose a failure point and run the simulation to inspect the outcome.

Dual write

Two independent operations can split: DB success + publish failure loses the event.

Outbox pattern

Commit business data and event intent together; a relay retries delivery, and consumers deduplicate.

Reality check: an outbox does not provide magical exactly-once delivery. Real systems need durable relay state, idempotent consumers, monitoring, and operational recovery.