Retries repeat attempts; idempotency preserves the identity of one intended operation

If a creation request times out, the server may already have created the resource. A second attempt with a fresh identity can create a second resource. Backoff changes timing; it does not resolve that ambiguity.

AWS describes a caller-provided request identifier as an explicit expression of intent. Repeated calls by the same caller with the same identifier represent the same operation. Recording that identifier and applying the mutation must be coordinated atomically; otherwise either record can exist without its counterpart.

Illustrative design: generate one token for “create this job,” persist it with the pending operation, and reuse it after a timeout. Generate a different token for a genuinely new job. Do not silently reuse a token with changed parameters. Check the service's retention period, parameter-mismatch behavior, and handling of late retries.

Choose bounded retries only after classifying the error and understanding that contract. Where no such guarantee exists, reconcile the remote state before repeating a side effect.

Evidence: source-based distributed-systems guidance; no service was fault-injected here.