Designing idempotent operations and safe retries

Goal

Let clients retry a failed or timed-out request safely, so that network faults do not create duplicate orders, articles or payments.

Prerequisites

A server that can store a small record per request key for a bounded time, and clients that generate unique keys per logical operation.

Steps

1. Use HTTP methods according to RFC 9110: GET, HEAD, PUT and DELETE are idempotent by definition; POST is not.
2. For POST operations that create resources, accept an Idempotency-Key header. Store the key together with a hash of the request body and the produced result, with an expiry.
3. On a repeated key with the same body, return the stored result and status; on a repeated key with a different body, return 409 or 422 and do not execute.
4. Execute the store-and-check inside the same transaction as the operation, or under a lock keyed by the idempotency key, so that two concurrent retries cannot both succeed.
5. Document the key format, the retention period and the conflict behaviour so clients can rely on them.

Expected result

A client that times out and retries with the same key gets exactly one created resource and the same response both times.

Limits and test basis

Idempotency covers the server's state, not side effects outside it (an email sent twice by a downstream system). Keys must be scoped per account to prevent cross-client collisions. The design mirrors the cited API documentation and this wiki's own implementation.