went through the NVIDIA data engineer loop earlier this year, sharing notes while they're still fresh.
recruiter screen was pretty standard. 30 min, light on technical, mostly confirming I knew what data engineering was. then a technical phone screen with one of the team leads. that one was 45 min and heavier than I expected.
the SQL round: they gave me a HackerRank-style environment but the questions were more analytical than typical leetcode SQL. one was a multi-join aggregation across a fact-dimension schema tracking GPU telemetry. they wanted window functions, specifically LAG/LEAD over a time-series partition. I've been doing data eng for 7 years and even I had to think for a minute. not 'can you write a GROUP BY' territory.
pipeline design: they asked me to architect a data pipeline that could ingest high-throughput sensor data (they never said what kind, but it felt GPU-adjacent), process it with low latency, and make it queryable by analysts. I talked through Kafka ingestion, Spark streaming for transformation, and landing to Delta Lake or Iceberg for the query layer. they pushed on fault tolerance and exactly-once delivery. good conversation, not gotcha-y.
the onsite (4 rounds): coding: SQL-heavy again plus one Python question about pandas vs. writing raw SQL for large joins. I said raw SQL, they seemed to like that. systems design: design a data platform for ML training pipeline observability. basically: how do you track model training runs, log metrics, handle data lineage. MLflow came up, so did custom solutions. behavioral: three full STAR questions. one on cross-functional conflict, one on owning a project that failed, one on handling ambiguity. a 'reverse technical' where the hiring manager walked me through what the team actually does and I asked questions. felt like a cultural fit round but also technical.
took about 3.5 weeks from first recruiter call to offer. comp was competitive with FAANG bands for a data eng role at their Santa Clara office, though NVIDIA's base skews a bit lower and they lean heavy into RSUs.
overall: harder than my last two DE loops. they actually care about data quality tooling (Great Expectations came up twice), not just moving data from A to B.