The short answer
Stay calm, stop the bleed, understand what committed, then repair with an idempotent replay.
extract OK → transform FAIL → load not started
or
load started writing day=2026-09-05 → crash mid-writeStep-by-step response
- 1Check orchestrator state: which task failed, error, attempt count.
- 2Assess partial writes: is the target incomplete or uncommitted?
- 3Do not mark success or silently continue downstream.
- 4Fix root cause (bad file, OOM, permissions, poison row).
- 5Replay safely: overwrite the partition / merge from last checkpoint.
- 6Validate: row counts, null checks, compare to source.
- 7Communicate: freshness impact to BI owners if is at risk.
Design that makes this easy
Stage-then-publish, partitioned overwrites, durable raw zone, clear runbooks.
Interview tip: Interviewers want process + idempotent recovery, not "I would rerun the " alone. Say what you check before hitting retry.