Google Data Engineer Interview Questions

I would deduplicate using a stable business key or request identifier before aggregation or downstream writes. If the source is at-least-once, I would assume duplicates will happen and design the pipeline so the same event does not create double counts. The key is to make deduplication explicit, not implicit.

I would ingest events through a durable messaging layer, process them with a streaming compute engine, and write them into an analytical store optimized for reporting. The design should handle late data, retries, and backpressure while keeping latency predictable. I would also separate raw ingestion from serving layers so failures do not corrupt downstream analytics.

I would compare the source queries, timezone handling, deduplication logic, and event definitions used by each dashboard. Often the mismatch comes from different filtering rules or boundary conditions around day changes. The goal is to identify the semantic difference before assuming one system is broken.

I would model based on the questions the business needs to answer, not just on source system structure. That often means building fact and dimension layers, defining grain clearly, and keeping metrics consistent across teams. Good modeling makes queries simpler, cheaper, and less error-prone.

I would include event identifiers, timestamps, user or session keys, event type, and any context needed for later aggregation or debugging. I would also plan for schema evolution because event payloads tend to change over time. The schema should be flexible enough for growth but strict enough to support trustworthy analysis.