finished the uber DS loop about 6 weeks ago. writing this because the information out there is mostly for SWE and i had to piece things together from scattered posts.
the loop structure for DS (L4/L5 IC): sql round (45 min) product case round (45 min) statistics/ml round (45 min) behavioral round (45 min) sometimes a HM round (i had one, not everyone does)
sql round: this was legitimately hard. not like 'write a GROUP BY' hard. more like multi-CTE window function with a business logic wrinkle you have to figure out. my problem involved session data and calculating some metric over rolling time windows. if you're comfortable with rank(), lag(), lead(), dense_rank(), and nested CTEs you're probably ok. they cared a lot about my reasoning, not just getting the right answer.
they ran my code during the round which was mildly terrifying. it worked first try, small error on edge case, they asked about it, i fixed it. seemed fine.
product case round: this is where DS at uber diverges from pure analytical DS elsewhere. they want you to think like a PM. prompt was something like 'how would you measure success for [Uber Eats feature]?' you have to define metrics, pick the right ones, explain trade-offs, design an experiment if applicable. classic north star metric vs guardrail metric framing helped me a lot.
stats/ml round: at L4 level this was more stats than ml. A/B testing, hypothesis testing, p-values and their limits. one question about an experiment that showed positive short-term lift but they were worried about novelty effect. how do you test for it? what's your recommendation to ship or not ship? that kind of thing.
if you're expecting a 'implement gradient boosting from scratch' round, that's more ml engineering. DS at uber is more applied stats and product thinking.
overall vibe: rigorous but fair. interviewers were engaged. my sql interviewer pushed me pretty hard but it felt like intellectual sparring, not hazing.