Meta Data Scientist Interview Loop
The Meta data scientist interview has three stages: a recruiter call, a 45-minute technical screen, and a loop of four back-to-back 45-minute interviews. Most people prepare for the wrong thing. They expect a lot of coding, and they get a loop where product judgment decides the outcome. The SQL is the minimum. Meta wants to see whether you can turn a query into a decision.
What decides it
Product thinking. Meta's own guidance to interviewers puts it above mathematical sophistication, and candidate reports back that up. If you're a strong coder with weaker product instincts, spend your time on Round 3.
Practice by round
SQL basics on product tables like users, posts and events, with a product question attached. The questions are easy and the round is timed, so speed is what gets scored. Nobody fails here on a difficult question. They fail by being slow on a simple one.
Medium and hard SQL, well past the screen, and the follow-ups matter most. You'll query retention or engagement. Then you'll be asked how you'd measure what you just computed, what could confound it, and what you'd do if it moved.
Analytical reasoning on an open product problem. Two dashboards disagree and you find out why. A new notification system needs a launch decision. You judge how healthy a Group is, or what makes a friend a Close Friend. There's no right answer. There's a structured way of thinking, and that's what gets scored. Say it out loud: define the metric, list the causes, and pick what to check first.
| Case study | Difficulty |
|---|---|
| Investigate the Discrepancy | Medium |
| Close Friends Metrics | Medium |
| Ad Success Metrics | Medium |
| Best Friend | Medium |
| Health of Groups | Medium |
| New Notification System | Medium |
| Clicks on Search Result | Medium |
| Teenager's Parents on Facebook | Medium |
| Defining a Product Success Metric | Medium |
| Launch or No-Launch Decision Framework | Medium |
Statistics, probability and experiment design. Expect A/B testing and sample size, conversion rate, serving ads within Newsfeed, and what happens to a test when network effects connect the two groups. The bar is fluency rather than depth. Explain why you'd pick a metric, as well as computing a p-value.
| Question | Difficulty |
|---|---|
| Conversion Rate | Easy |
| Negative R Squared | Easy |
| Different Card | Medium |
| Two Cards Same Suite | Medium |
| Expectation of Variance | Medium |
| Mean, Median Age in Mexico | Medium |
| Used by Different Groups | Medium |
| Serving ads within Newsfeed | Medium |
| Sample Size | Medium |
| A/B Testing | Medium |
| Propensity Matching for an Opt-In Feature | Medium |