A dashboard has shown the same numbers for eleven days. The pipeline reports success every night. What monitoring was missing?
Show the full answer Hide the answer
What the interviewer is testing
Whether you understand that data pipelines fail silently and successfully, and which specific signals catch it.
What happened
The job ran and succeeded. It processed whatever arrived — which was nothing, or the same file again, or a file with zero rows. Success meant "the code executed without throwing", which is what almost all pipeline monitoring actually measures.
Common causes: the source stopped delivering and nobody was told; a credential expired and the extraction silently returned empty; a partition filter began matching nothing after a source schema change; or the same file was reprocessed because the watermark did not advance.
The missing signals
Freshness. The maximum timestamp in the output table, monitored against an expected lag. This alone would have alerted on day one and is the single highest-value data monitor.
Volume. Row count per load, compared against an expected range with seasonality. Zero rows and ten times the usual are both alerts.
Absence. Alert when an expected load does not occur. This is the one that requires deliberate design, because every check that inspects arrived data is blind to data that never arrived.
Distribution. Have the values shifted in a way suggesting a semantic change upstream — a unit change, a new category, a null rate jump. This finds the genuinely dangerous failures and generates the most false alarms, so it needs tuning.
Lineage, so when something is flagged you can scope who is affected in minutes.
The design fix beyond monitoring
Expose freshness as a queryable value beside the data, so a dashboard displays its own staleness. Eleven days of unchanged numbers presented as current is a failure of the interface as much as of the pipeline.
What a strong answer adds
Routing and ownership. An anomaly detected and sent to a shared inbox is not detection — it must reach the dataset's owner with lineage attached, or the eleven days repeat with better instrumentation.
Common weak answers
Adding a job failure alert, which already existed and did not fire. Manual daily checks.