It Ran Twice — Or Not At All

The failure these share

There is no delivery setting that neither loses nor duplicates. A broker offers at-most-once or at-least-once, and every one of these is a measured case of a codebase assuming a third option that does not exist.

• Celery acks_late Verdict — measured: a killed worker ran the task TWICE with `acks_late=True`, and made it VANISH under Celery's default.
• BullMQ Stalled Job Verdict — measured: a job added with `attempts:1` ran on two workers. One completion, no failure, the only trace on a channel nobody subscribes to.
• Queue Lease Semantics — lease held, redelivery cause, poison message: the decisions an at-least-once consumer actually has to make.
• Idempotency Key Manager — a key containing a timestamp never matches on retry. It protects nothing while appearing to.
• Webhook Fulfillment Kit — an order fulfilled twice, or a paid order never fulfilled at all.
• Retry Budget Guard — retries are added to improve reliability and are the standard way to turn a partial outage into a full one.

Why buy them together

Idempotency is the only real remedy, and it has to hold at every layer that can re-deliver. Making the worker idempotent while the webhook handler and the retry loop are not simply moves the duplicate.

Each ships with its own tests and is sold separately. This is the set, below the sum.