Starbucks · Primly Community

Starbucks data scientist interview (SQL + case + stats), full breakdown

analyst_ana · 5 replies

went through the Starbucks DS interview loop for a data scientist II role on the personalization team (Seattle-based, some remote flexibility). took notes throughout because the process was more involved than i expected for a retail/CPG company.

full loop timeline: recruiter screen -> technical screen -> final panel (4 hours)

technical phone screen (60 min): two SQL problems and a stats/probability question. the SQL was on a HackerRank-style interface, Postgres flavor. one was a straightforward aggregation with a window function (rank within a partition). the second was a multi-table join to answer something like: which reward tiers have the highest variance in visit frequency over 90 days. not hard, but you need to be fast and clean.

the stats question was a classic: explain p-value to a non-technical stakeholder, then: if we run an A/B test on the loyalty program and see p=0.04, should we ship? they wanted to see that i'd ask about: power, sample size, primary vs secondary metrics, and whether the test was designed properly before i just said "yes p<0.05 ship it." that trap is a real filter.

final panel:

case round (90 min): a data-heavy case. scenario: engagement with the Starbucks Rewards program dropped 8% YoY in a specific demographic. walk me through how you'd diagnose it. this is a classic DS case and they went deep. expect to be asked: how you'd segment the data, what SQL you'd write to validate a hypothesis, and what the business action would be even before you fully confirm the root cause.

ML/modeling round (60 min): how would you build a propensity model to predict which users are at risk of churning from Rewards? feature engineering, model choice, how you'd evaluate offline vs online, how you'd set a threshold for intervention. pretty standard ML interview content but they pushed on how you'd tie model outputs to actual marketing/ops actions.

behavioral round (30 min): mostly about cross-functional collaboration with marketing and finance. they run data science very collaboratively with stakeholders here, not as a pure research function.

comp (from my offer): base $145k, 10% target bonus, some RSUs. Seattle, data scientist II level, 3 YOE in DS. not Google-tier but honestly solid for the role scope.

5 replies

analyst_ana

the p-value trap is so real. i've heard that question in various forms at like six companies and people still say "yes, p<0.05, ship" without thinking. glad they're actually filtering for it.

ux_uma

was the SQL on a live shared environment or did you code on your own and share screen? and did they run your code or just read it?

firsttime_mgr

HackerRank-style, they could see my screen. they ran my code, yes. output mattered. make sure your joins don't drop rows you need, that's where i got tripped up on the second problem initially.

ml_mike

for the churn propensity question: did they want you to go toward logistic regression and XGBoost type answers or were they open to more complex approaches? also curious if they care about model interpretability given it's a marketing use case.

sec_sasha

they were pretty open to model choice. i started with logistic regression as a baseline, mentioned gradient boosting, and they seemed most interested in the interpretability piece. marketing stakeholders need to explain why someone got an offer, so SHAP values or feature importance matter. go deep on that angle if you interview there.