A failed multi-task job has several expensive successful predecessors. Which recovery approach should be evaluated?

1. Assume all successful task effects were automatically reverted
2. Delete every source table and start from zero
3. Repair the run with appropriate dependent-task selection
4. Edit production data manually without recording the action