Make failure understandable

Clients need to distinguish a missing resource from a temporary outage or a rejected request. Clear status codes and consistent error shapes give both users and operators a better next step.

A fallback is appropriate only when it preserves meaning. Returning an empty result for a failed request can look successful while quietly hiding important content.

Bound the waiting

Timeouts put a limit on how long a dependency can hold a request open. Pair them with measured retries for transient failures, and avoid retrying operations that may have already caused a side effect.

Logs and metrics should retain enough context to diagnose the failure without exposing credentials or private user data.