Module 09 Summary
What this module established
Classify failures before retrying - a 400 retried three times is three times the cost and a delayed error. Retry only where something could plausibly differ next time.
Carry forward
- Exponential backoff with jitter and a hard ceiling. Jitter prevents a thousand clients retrying at the same instant.
- A timeout is ambiguous, not failed. Retry with an idempotency key, because the call may already have done the thing.
- Build the middle rungs of the degradation ladder. Retrieval survives a model outage, and source text with a citation is a real answer.
Before moving on
Move on when a timeout retries safely with a key, and your five error messages are distinguishable and free of internals.
