Google Data Engineer Interview Questions

I would build the pipeline so historical partitions or time windows can be recomputed without corrupting current data. That means clear partitioning, deterministic transformations, and a way to replace or reconcile historical outputs. Safe backfills are crucial because real data systems need repair paths, not just forward flow.

I would use streaming when freshness has direct business value and the system can support the operational complexity. I would use batch when latency is less important and simplicity, cost, or reliability matter more. The decision is really about the trade-off between immediacy and operational burden.

I would validate schema, volume, freshness, null rates, uniqueness, and business-rule consistency at key checkpoints. Data quality should be monitored continuously so issues are detected before they affect dashboards or downstream models. I would also keep alerting actionable so the team can tell whether the problem is ingestion, logic, or source data.

A warehouse is optimized for structured analytics and governed reporting, while a lake is often more flexible for raw or semi-structured data. In practice, many modern systems combine both ideas, so the real question is how much structure and governance the workload needs. I would choose based on query patterns, reliability requirements, and the team’s ability to manage data discipline.

I would compare query plans, scanned bytes, changed join paths, and recent data growth. Cost spikes often come from a change in query shape, partition misuse, or a larger-than-expected input rather than from the platform itself. The right fix depends on whether the cost increase is structural or temporary.