Meta Data Scientist interview questions

489 Data Scientists have shared notes from their Meta interviews, with the questions they were asked in each round. That is 21% of the 2,377 Meta interview notes on Screna AI.

489 Data Scientist interview notes

Meta·Data Scientist·Technical Phone Screen

Jul 2026
Meta DS interview focused entirely on a single sprawling experiment design question about network interference and spillover effects. Dense, technical, and pretty humbling if you haven't thought hard about SUTVA violations before.
  • A new social feature might cause network spillovers (like hashtag following or invite prompts spreading between users). Design an experiment that can separately measure direct effects on treated users and indirect spillover effects on their neighbors, covering the full design, estimands, power calculations, analysis approach, guardrails, and what happens when the graph changes mid-experiment.

“This was basically six questions wrapped in one.”

View Post

Meta·Data Scientist·Technical Phone Screen

Jul 2026
Meta DS technical screen, basically one big SQL problem broken into three parts. The schema wasn't too scary but the overlap logic and P90 requirement made it genuinely hard to keep in one clean query. Left feeling like I got maybe 70% of the way there.
  • Given a call participation schema with overlapping time intervals, compute peak concurrent participants per call while excluding test users and users from a specific email domain. A call qualifies as a 'group call' if its peak concurrency reaches at least 3.
  • For each calendar day in a given date range, return the total calls started, number of group calls, and the 90th percentile of peak concurrency, filtering to calls where group calling was enabled.
  • Return the top 3 calls by peak concurrency that started on a specific day, including call metadata and the first timestamp within the first 10 minutes of the call when concurrency first hit 3. Return NULL if it never reached 3 in that window.

“The filtering part was fine, just a join on users with a couple of WHERE conditions.”

View Post

Meta·Data Scientist·Technical Phone Screen

Jul 2026
Meta data scientist interview with a stats-heavy question about evaluating a marketing campaign's CTR. The whole thing was basically a hypothesis testing deep dive, which I expected but still found myself fumbling through parts of it.
  • You're told a marketing campaign has a 4.2% click-through rate. Leadership wants to know if that's good or bad. How do you formally set up the hypothesis test, choose the right statistical test, and figure out what additional data you need? Walk through computing a p-value and confidence interval, and explain both statistical and practical significance.

“This question has a lot of layers and I didn't pace myself well.”

View Post

Meta·Data Scientist·Technical Phone Screen

Jul 2026
SQL-heavy technical screen for a Data Scientist role at Meta. Three questions, all SQL, progressively nastier. The last one about growth proportion had enough edge cases to make me second-guess my own logic twice.
  • Write a query to compute daily revenue broken out by creation source, ordered by date and creation source. Only include dates that actually appear in the ads table.
  • Find the 10 countries with the fewest distinct active advertisers over the last 30 days, where active means at least one ad with revenue greater than zero in that window. Countries with zero active advertisers should still appear in the results. Break ties alphabetically.
  • For each creation source, compute the proportion of advertisers who grew their spend by at least 1000 between the first 8 months of 2024 and the same period in 2025. The denominator is all advertisers with any spend in either period for that source. Advertisers missing from one period should be treated as having zero spend for that period.

“Straightforward group-by aggregation.”

View Post

Meta·Data Scientist·Technical Phone Screen

Jul 2026
SQL-heavy technical screen for a DS role at Meta. Two questions, both around aggregations and joins, but the second one had enough edge cases to slow me down a bit.
  • Given a table of buyer-seller listing interactions and a product dimension table, write SQL to find how many sellers have more than 3 distinct products where that product's total interaction count (summed across all rows for that seller-product pair) is greater than 1.
  • Among products created in the US within the last 7 days, what percentage of total listing interactions comes from the 'vehicle' category? Return the result as a value between 0 and 100.

“The definition of 'multiple interactions' tripped me up at first.”

View Post

Meta·Data Scientist·Technical Phone Screen

Jul 2026
SQL-heavy technical screen for a DS role on an ads platform. Two questions, both multi-part, both requiring you to actually think about the data model rather than just write a join and call it a day. Left feeling okay about Q1 but Q2 had a wrinkle that tripped me up a bit.
  • Given an ads platform schema with advertisers, ads, and daily spend tables, identify advertisers whose total 2023 spend exceeded $1,000. Then, for that cohort, calculate what percentage of total 2024 platform spend came from them. Return the eligible advertiser count, cohort 2024 spend, total 2024 platform spend, and the cohort share.
  • Write a query returning monthly 2024 spend broken down by ad creation source, excluding certain advertiser types. Also include, within each month, the spend from advertisers where AI-assisted ad spend went up year-over-year while their manual ad spend went down, to surface potential cannibalization between creation sources.

“The cohort definition part is pretty clean, just a subquery or CTE filtering on 2023 spend.”

View Post

Meta·Data Scientist·Technical Phone Screen

Jul 2026
Meta DS interview with a product health angle, basically one meaty SQL question dressed up as a real business scenario. The Oculus usage framing was a nice touch, made it feel less like a leetcode grind and more like actual work.
  • You're given a table of Oculus headset session data. First, define what 'unhealthy' usage looks like as a SQL-expressible rule (e.g. more than 120 continuous minutes per day for at least 3 consecutive days in the past month). Then write SQL to return the count of unhealthy users, total active users, and the percentage of unhealthy users over the last 30 days.

“The open-ended part tripped me up more than the SQL itself.”

View Post

Meta·Data Scientist·Technical Phone Screen

Jul 2026
Meta data scientist interview with a heavy stats focus. Four questions, all technical, no behavioral fluff. The kind of session where you either know your Bayesian priors or you don't.
  • Given an A/B test with 5 interim looks already done, compute the two-sided p-value and 95% confidence interval for the difference in conversion rates between control and treatment, then explain how O'Brien-Fleming alpha spending would affect whether you'd call the result significant.
  • You have five metrics with sorted p-values of 0.001, 0.012, 0.019, 0.070, and 0.300. Apply Benjamini-Hochberg at a 5% FDR and identify which are discoveries. Then compare to what Bonferroni would give you.
  • Your bot detection model has a 2% false positive rate and 10% false negative rate. Bot prevalence is 5%. If a user gets flagged, what's the actual probability they're a bot? Walk through the Bayes calculation.
  • Average Order Value is right-skewed with a heavy tail. Compare IQR-based outlier removal, z-scores, and MAD or Huber M-estimators for handling this. Then explain a log-transform approach and how you'd report the effect back in original dollar terms using a smearing estimator.

“This is the question I spent the most time on.”

View Post

Meta·Data Scientist·Technical Phone Screen

Jul 2026
Meta DS technical screen, all SQL, no fluff. Three questions off one schema and they just let you cook. The queries got progressively nastier and I definitely fumbled the percentile one before recovering.
  • Given a schema with users, posts, and comments, write a query to return the top 3 posts by distinct commenter count within a 7-day window, excluding deleted comments. Break ties by newer post date, then lower post_id. Include a count of distinct commenter countries.
  • For each post and day within the same 7-day window, compute the P50 and P90 of comment text length over non-deleted comments using percentile_cont. Output one row per post per day, only for days that have at least one qualifying comment.
  • Identify users who commented on their own posts within the 7-day window. Return user_id, the count of such self-comments, and the timestamp of the most recent one. Order by self-comment count descending, then user_id ascending.

“The tie-breaking tripped me up a bit.”

View Post

Meta·Data Scientist·Onsite - Cross-functional / Panel

Jul 2026
Interviewed for a Data Scientist role at Meta and got hit with a meaty stakeholder scenario about bad causal inference from sales leadership. One question but it had five sub-parts and basically ate the whole session.
  • Sales leadership observed a correlation between call volume and win rate and wants to double required calls starting next week, claiming it will increase wins. The analysis is confounded by deal stage and rep mix. How do you push back diplomatically, propose a low-cost pilot, set timeline and data quality expectations, align stakeholders, and handle pressure to publish early directional results before the pilot has enough data?

“This one was brutal in the best way.”

View Post