Google Data Engineer Interview Questions

I would explicitly design for lateness by using windows, backfill logic, or reprocessing rules rather than pretending every event arrives on time. The pipeline should know how long it can wait, how far back it can correct, and how to keep downstream metrics consistent. Without that, batch and streaming numbers will drift.

I would narrow the problem to input, transformation, or output, then inspect logs, metrics, and recent data changes. I would look for schema mismatches, permission issues, data skew, or bad assumptions about volume and freshness. A good debugging approach is systematic and starts with the narrowest reproducible failure.

I would ensure that rerunning the same logical input produces the same final result. That usually means using deduplication keys, atomic writes, merge semantics, and checkpoints or watermarks where appropriate. Idempotency matters because retries and backfills are normal in production data systems.

I would choose a deterministic key, define the deduplication window, and decide whether to keep the first, latest, or canonical record. The important part is that the rule is documented and consistent across the pipeline. If deduplication is ambiguous, the metric will not be trustworthy.

Common causes include scanning too much data, bad join order, missing filters, skewed joins, and unnecessary subquery complexity. I would also check whether the data layout supports the query pattern, since even good SQL can be slow on poorly organized data. Performance usually improves when the query and the storage model work together.