Backend systems & reliability
Dispatch
A Python and SQLite service that separates accepting an event from delivering it. Retry timing, crash recovery and duplicate acceptance are inspectable rather than hidden behind a success flag.
Persist before delivery
The intake API stores an event with an idempotency key. Workers claim stored work using expiring leases and record an outcome for each attempt. The database holds both the queue and its history, keeping the first deployment small enough to inspect.
The ambiguous acknowledgement
A receiver can accept an event while Dispatch never stores its response. Delivery is at least once: retrying may repeat an accepted request. A stable event ID lets a cooperating receiver deduplicate its own effects; Dispatch cannot promise exactly-once effects in another system.
A real process boundary
The crash demonstration pauses its own worker after local HTTPS acceptance, terminates that process, then starts a replacement. It waits for actual lease expiry and retry backoff. The recorded run observes two acceptances and one simulated receiver effect. This is one deliberately injected boundary, not a host or disk failure test.
Measured, within limits
Three repeated baseline/retry pairs each use twenty synthetic events and one worker against local HTTPS. Raw timing observations and environment details accompany the report. These small shared-desktop runs establish reproducible behavior, not production capacity. The service is not publicly deployed and its loopback intake API is unauthenticated.