A failed job has several successful upstream tasks. What should be reviewed before repairing the run?

1. Delete all source data first
2. Which tasks rerun and whether their writes are safe to repeat
3. Only the job's display color
4. Assume every successful task is automatically rolled back