Break the delivery.
Inspect the outcome.
Small, repeatable experiments for a deceptively difficult question: did the message fail, or did its acknowledgement disappear?
Three experiments to start with
| Experiment | Parameters | Expected observation |
|---|---|---|
| Repeated delivery | At least once; 3 duplicates; 1000ms timeout; idempotency on | 4 envelopes, 1 effect, 3 duplicates ignored. |
| Lost acknowledgement | At least once; 0 duplicates; 2 retries; 200ms timeout; idempotency off | 3 effects and a dead letter. Exhaustion does not mean absence. |
| No recovery budget | Once; failure on attempt 1 | 1 failed attempt, no effect, 1 dead letter. |
A clock you can reproduce
The simulator uses virtual milliseconds. A new effect takes 400ms; a duplicate lookup takes 50ms. An injected failure happens before the effect at the selected attempt, once per delivery envelope. Backoff doubles from 250ms. The next envelope arrives 100ms after the previous envelope finishes.
Once mode permits one envelope and one attempt. At-least-once mode permits the original envelope plus your selected duplicates, with a separate retry budget for each envelope. This is a bounded teaching model: its name does not promise eventual successful processing.
The selected timeout is an acknowledgement deadline. Below 400ms, the new effect still commits but the attempt is marked failed. Processing is sequential, so the simulated retry starts after the original handler finishes and the backoff elapses.
What this model leaves out
No requests leave your browser. The lab has no real provider, queue, database, payment system or game session. It does not simulate concurrency, jitter, network partitions, authentication, storage failures, retention expiry or process crashes. Reloading clears the in-memory result.
Use the trace to develop a hypothesis, then test that hypothesis at your own transaction boundary. Start with the idempotency guide for concurrency tests and callbacks and retries for a delivery failure taxonomy.