just finished the Netflix DS loop a few weeks ago for a senior data scientist role on the content analytics team. sharing specifics because it's not well documented.
the loop (4 rounds total after phone screen):
round 1 - SQL + metrics. this was 60 minutes. the SQL was legitimately hard. not 'join two tables and group by' stuff. i got a multi-step query involving window functions, a self-join, and an aggregation with filtering at each level. second half was metrics definition: how would you measure whether a new UI feature is improving engagement, what's your north star, what could go wrong with that metric. they pushed on the 'what could go wrong' part extensively.
round 2 - experiment design / A/B testing. expected this but not the depth. they gave me a scenario: Netflix is testing a new content recommendation surface for a small user segment. design the experiment. the basic experiment structure took maybe 10 minutes. the rest was: how do you handle novelty effects, network effects in a social content context, what do you do if your primary metric goes up but engagement duration goes down. very netflix-specific problems because their recommendations actually affect global watch behavior.
round 3 - statistics + ML concepts. no coding, but conceptual depth. i got asked about: interpreting p-values in the presence of multiple testing, bias-variance tradeoff in the context of a real recommendation model, and how i'd approach building a churn prediction model from scratch (what features, what evaluation metric, why). they went deep on the last one - about 20 minutes on that single question.
round 4 - culture contribution. same format as the SWE loop. two Netflix values, real stories, multiple layers of follow-up.
what i'd prep: get really comfortable with window functions and experiment design edge cases. those two areas were where the bar felt highest. stats was more conceptual than computational. the culture round is real - do not skip it.