Just finished my Databricks MLE loop. Applied for a senior MLE role (roughly L5 equivalent) on one of their MLflow-adjacent teams. Going to be direct about what surprised me.
What I expected: Heavy ML theory, maybe some coding, standard FAANG-style behavioral.
What I got: Much more systems and data engineering than I anticipated. Makes sense in retrospect since their whole value prop is the Lakehouse architecture. If you come in thinking it's a pure ML research interview you'll be caught off guard.
The loop (5 rounds):
Coding: Two rounds. Both were what I'd call "data-heavy coding" rather than pure LeetCode. One involved parsing and aggregating structured data. One was a graph problem applied to a dependency graph. Mediums. Nothing from the NLP or model-training track.
ML system design: This is where it got interesting. I was asked to design a feature store for a large e-commerce platform. We talked about point-in-time correctness, offline vs. online consistency, and how to handle training-serving skew. I got deep into Spark for batch computation and they followed me there, asking detailed follow-ups about partitioning strategies. This is genuinely a Databricks-flavored question.
ML depth: One round that was more classic ML. Bias-variance, regularization, how I'd debug a model that performs well offline but degrades in production. The interviewer pushed on the production part specifically. They want ML engineers who think about deployment, monitoring, and data drift, not just model accuracy.
Behavioral: Pretty standard. Impact stories. Cross-functional influence. One question about a time I had to help a non-technical partner understand a model failure.
My honest take: if your background is heavily research-leaning with limited production systems experience, spend time with Spark, MLflow, and the Lakehouse docs before interviewing. They're not looking for Databricks certification but they do want engineers who understand the space they operate in.