Idempotency

Pipelines that append on every run cannot be safely retried. Making each run replaceable rather than additive is what makes…

When a client retries a request because it timed out, the server cannot tell whether the original request actually succeeded,…

Networks fail after the request arrives but before the response returns. Without idempotency, your retry logic is a duplicate-charge generator.

At-least-once delivery is the practical guarantee almost every queue actually provides, and a job handler that assumes exactly-once will eventually…