How do you make a write endpoint safe to retry?

Give the client an idempotency key, store it alongside the result of the first successful write, and return that stored result on any repeat. Enforce uniqueness at the database level so concurrent duplicates collide instead of double writing. Scope keys per customer, expire them after a sensible window, and use the natural business key such as an order id as the deduplication anchor whenever one exists.

The client generates an idempotency key, usually a UUID, and sends it as a header. Server side I keep a table keyed on customer id plus that key with a unique constraint, and I write the key and the response in the same transaction as the business write. A repeat then hits the unique constraint or finds the stored row and just replays the original response, so retries are free. Two details matter. First, the uniqueness has to live in the database, because if you do a select then an insert, two concurrent retries will both pass the check. Second, you have to decide what to do when the second request arrives while the first is still running. We returned 409 with a Retry-After rather than blocking, which kept our connection pool healthy. I also store a hash of the request body so that reusing a key with different content is rejected instead of silently returning the wrong answer.