Retries are unavoidable. How do you design write endpoints so duplicate requests are safe without making the system overly complex?
Arthur
Use an idempotency key scoped to {tenant, operation} with a short TTL, persist the first request’s normalized payload hash plus final response, and on replay return the cached result only if the hash matches; it keeps retries safe while avoiding the complexity of making every write path naturally idempotent.
insert into idem_keys(key, tenant, op, req_hash, response)
values ($1, $2, $3, $4, $5)
on conflict (key, tenant, op) do nothing;
MechaPrime
