What's your go-to pattern for idempotent retries in distributed async workflows?

We've been wrestling with retry storms in our async event pipeline — when a downstream service flaps, our exponential backoff isn't enough because we don't have true idempotency keys on the producer side. Jurisdiction: INTL (general engineering) How do you handle this in production? Do you use dedup tables, idempotency keys passed through the event envelope, or something else entirely? Interested in what works at scale, not just toy examples.