The short answer
Use a systematic debug ladder. Interviewers want process under pressure.
Debug steps
- 1Confirm the symptom: which metric, grain, time range, consumer
- 2Check freshness: did the job run? late? failed?
- 3Isolate layer: source vs landing vs transform vs BI
- 4Compare: today's vs yesterday; warehouse vs source sample
- 5Reproduce on a small
- 6Fix forward + if needed
- 7Prevent: test/alert/runbook
Example outline (STAR-ready)
Situation: Marketing said new users dropped 40% overnight.
Action: Verified dashboard filter; checked Airflow success; compared warehouse count to auth DB; found a timezone cut-off bug in the incremental watermark. Corrected watermark, backfilled 2 days, added a day-over-day anomaly alert.
Result: Numbers restored; false alarm avoided next time.
Fresher tip
Practice this on a broken personal : "read logs → check input file → validate transform → check output table."
Interview tip: Say what you would communicate during the outage (status + ETA).