Adobe · Primly Community

Adobe data scientist interview: SQL, case questions, stats, and what actually matters

ds_dmitri · 4 replies

went through the Adobe DS interview loop for their Digital Media analytics team. five years in DS now, fourth company loop, so I have context for comparison. here's what I found.

the structure phone screen with recruiter, then a technical screen, then a full virtual onsite with four rounds.

technical phone screen 30-45 minutes. two SQL questions and one light stats question. the SQL was medium difficulty: one multi-table join with aggregation, one window function question (rank, lag, something like that). the stats question was about A/B testing: how do you determine sample size, what's statistical power, what does it mean to commit a Type II error.

tip: the A/B testing stats question comes up every time. at Adobe specifically they care about it because they run enormous experiments across Creative Cloud. know how to explain p-values without the usual hand-wavy stuff.

onsite rounds SQL deep dive (45 min): harder queries. I had a question about computing 7-day rolling retention across cohorts. also a question about deduplicating events from a raw clickstream table. these are realistic business problems, not trick questions. stats and probability (45 min): Bayesian vs frequentist framing (they asked which I prefer and why), confidence intervals, a brain teaser about probability that wasn't super hard but tested if I could explain my thinking out loud. also a question about how to handle imbalanced classes in a model. product case (45 min): given this dataset about Acrobat DC user behavior, what question would you ask and how would you structure the analysis? I had to pick a metric to improve and justify it. felt like a DS/PM hybrid round. behavioral (45 min): same as the PM and SWE side. Adobe-wide questions about collaboration, ambiguity, impact.

what they care about most honestly: SQL fluency and communication. a lot of DS candidates have the stats knowledge but fall apart explaining it to a non-technical stakeholder. the product case round explicitly tests that. "how would you explain this finding to the marketing team" came up twice.

difficulty vs other DS loops easier than Netflix or Airbnb on stats depth. harder than most mid-size companies on SQL specifics. realistic overall.

4 replies

analyst_ana

the 7-day rolling retention SQL question is a great one to prep. I've seen variations at four companies now. the tricky part is usually handling the cohort date properly and not double-counting users who return on day 3 and day 7. did they care about the exact syntax or more the approach?

de_derek

clickstream deduplication is my least-favorite interview question because the answer depends so heavily on what your event schema looks like. did they give you a schema upfront or was part of the exercise figuring out what columns you needed?

ds_dmitri

they gave a simplified schema: eventid, userid, eventtype, timestamp, sessionid. the dedup was by (userid, eventtype, timestamp) within a session with a 1-second tolerance window. I used a ROW_NUMBER partition approach. they said there were multiple valid approaches.

ml_mike

imbalanced classes question in a DS loop is interesting. do they want SMOTE, class weights, threshold tuning, or are they just testing if you know the problem exists? feels like a shibboleth question sometimes.