A task may fail transiently after writing some output. What must retry configuration consider?

1. Unlimited retries with a new output identity each time
2. Delete all source data after the first error
3. Bounded retries and idempotent or transactional output handling
4. Report every exception as success