Went through the Palantir data engineer loop in late 2025, got to final round, ultimately didn't accept (comp didn't work out for me). Sharing the technical details because the DE-specific info online is almost nonexistent.
Technical rounds I saw:
SQL round. More involved than the typical 'write a GROUP BY' gauntlet. They gave me a schema that resembled a real operational dataset -- event logs, user actions, some hierarchical data. Then asked me to write queries that answered business questions. One involved window functions and I needed to handle some tricky null semantics. The schema had gaps (no explicit FK in one case) and they expected me to notice and ask about it. Read your query explain plans if you can; they asked about index usage.
Pipeline design round. This was more like the system design round for SWEs but scoped to data infrastructure. I got something like: 'Design an ingestion system that pulls from 5 different source systems with different formats, normalizes them, and makes the data available for analysis within 15 minutes.' Trade-offs they cared about: idempotency, late-arriving data, schema drift. Know what CDC (change data capture) is and why it exists. Know the difference between micro-batch and true streaming and when each is appropriate.
Python/coding round. One problem on data transformation in Python. Less 'LeetCode' and more 'write a function that parses this messy JSON structure and outputs a clean schema.' Messiness was intentional. Edge cases were the point.
Behavioral. Same as what others have said. High-stakes, ownership, mission alignment.
Prep specifically: Spark vs. Flink trade-offs, Airflow DAG patterns, schema evolution strategies, dbt familiarity is nice to have but not required. They don't use a generic stack -- they use Foundry -- but they want to see if your mental models transfer.
The loop is genuinely technical. If you've been doing mostly dashboard work and not pipeline engineering, it'll show. Be honest about your depth in the screen.