Samsung · Primly Community

Samsung data scientist interview (SQL + case + stats): what the loop looked like for a mid-senior DS role

analyst_ana · 4 replies

went through a Samsung DS interview loop in Q1 2026 for a mid-senior role on their SmartThings analytics team. the loop was 4 rounds. here's the breakdown by type.

SQL round (60 min): this was harder than most SQL rounds i've done. not just joins and aggregates. they gave me a schema representing user device events (deviceid, eventtype, timestamp, session_id) and asked: find users who interacted with 3+ distinct device types in a 7-day window rank devices by engagement rate where engagement = events per active session write a query to identify sessions that ended without a user-initiated action (inferring from event type sequence)

the last one required window functions and some careful reasoning about event ordering. i wrote it in postgres syntax (they said any dialect was fine).

statistics + case round (60 min): two interviewers. first half: stats questions. they asked about A/B test design for a new SmartThings feature (how to handle devices that appear in both test and control via a household), then asked me to explain p-value in plain english and describe a scenario where a statistically significant result shouldn't influence a product decision. second half: a brief product case. 'a new Samsung TV feature has 20% adoption in the first 30 days but drops to 5% by day 90. what happened and what would you investigate?' they wanted both analytical and product intuition.

ML/modeling round (60 min): walked through a past project in detail (they reviewed my resume first and picked a specific project). lots of follow-ups on model choice rationale, how i handled class imbalance, what i'd do differently now. then a short whiteboard exercise: design a model to predict whether a smart device will go offline in the next 72 hours. they cared about features, label definition, and evaluation metric choice (precision vs recall tradeoff).

behavioral (45 min): standard. stakeholder communication, ambiguous problem framing, disagreeing with a PM decision.

hardest part overall: the SQL window function question under time pressure. if you're rusty there, practice before you go in.

4 replies

analyst_ana

the household contamination problem in A/B tests is such a good question. i never thought about that for device companies but it makes total sense: one person in a household gets the feature, all their devices are affected. how did you answer it?

sdr_sky

i said randomize at the household or account level rather than device level, then acknowledged that makes your effective N smaller. mentioned that you can also do difference-in-differences if you have pre-exposure data. they seemed to want to see that i understood the clustering issue, not necessarily that i had a perfect answer.

de_derek

the event sequence SQL question is brutal under time. 'session ended without user-initiated action' requires you to define what the last event in a session looks like and whether the absence of a certain event type within a time window counts. that's a data modeling judgment call as much as a SQL question.

ml_mike

device churn prediction as a modeling exercise is actually a really clean domain for an interview. you have time-series sensor data, clear failure events as labels, and real precision/recall tradeoffs (false negatives = device goes offline unexpectedly, false positives = unnecessary maintenance requests). good choice by them.