A pipeline step occasionally fails with a transient service error. Repeating it is safe because its outputs are idempotent. Which approach best fits?

1. Use a date-scoped idempotent backfill strategy.
2. Use bounded retries with backoff and observable failure handling.
3. Declare explicit dependency and success conditions in the orchestration workflow.
4. Validate expected data freshness and volume in addition to job exit status.