Your résumé mentions a data pipeline you built. Walk me through its failure modes and how you monitored it.

How to answer

The interviewer wants how your pipeline fails, and which failures you would notice before a customer did.

1. Give the shape in two sentences, with numbers. Sources, schedule, volume, destination and reader: “Three hospital feeds, pulled hourly, landing in a table the billing team reconciles every morning,” then the daily row count and how you measured it. Say which parts you built and which you inherited.
2. Walk the failures in the order the data flows. Upstream: a feed arrives late, empty or with a renamed column. Transform: a retry writes a batch twice, a timezone shifts rows into the next day, or a rerun gives different output because it reads the current time or a lookup table that has changed since. Load: a partial write, or a backfill that overwrites corrected data. For each, say what caught it.
3. Separate loud failures from silent ones. A crashed job pages someone. A green run with rows missing pages nobody unless you check the data: freshness of the newest event, row count against the same weekday last week, null rate on join keys, duplicate keys, a total reconciled against the source. Give each threshold, who received the alert, and one check you rejected and why.
4. Tell one real incident. Which check fired, or why none did; who noticed first; the root cause in mechanical terms; what you added. If the customer noticed first, say so.
5. Close with what you would change, such as the alert people learned to ignore, and what breaks first at ten times the volume, such as a full reload that no longer fits its window.

The trap is the tool tour. “We monitored it in Datadog” answers nothing; “an alert when the newest row was over two hours old, sent to on-call” does. Expect the interviewer to pick one check and ask how you chose its threshold.