We had duplicate payouts twice last year. Both times: client retried after a timeout, server had already pr…

We had duplicate payouts twice last year. Both times: client retried after a timeout, server had already processed the first call. What fixed it for good: 1. Every mutating API takes an Idempotency-Key header (client generates UUID). 2. We store (key, request_hash, response, status) in Postgres with a unique index on key. 3. First request inserts IN_PROGRESS inside the same transaction as the business write. 4. Retries with same key + same hash get the stored response. Same key + different hash = 422. 5. Keys expire after 24h via a nightly job. The important part is #3: if the idempotency row and the business write aren't in one transaction, you've just moved the race somewhere else. Happy to share the filter code if people are interested.