Meta data scientist interviews: how they run, what they ask, how to answer
Two stages. First a 45-minute technical screen: a product case, SQL and an A/B test. Then an onsite of four rounds, one of them SQL. Built from interviews candidates and AirCareer users shared with us, and a prep session Meta ran for candidates.
| Stage | Length and format | What it covers |
|---|---|---|
| Technical screen, also called the phone screen | 45 minutes, video call and CoderPad | A product case, two SQL questions, and a measurement question with and without an A/B test. |
| Onsite | Four rounds | One round is SQL. We have no recording of the other three, so we do not know exactly what each one covers. |
The short version
- Meta says the product case is not a quiz on its products. If you do not use a feature, ask how it works.
- Four question types come up again and again: why a metric changed, what a product is worth, whether to launch, and how to design it.
- Hypotheses are not enough. For each one, say exactly what data you would pull.
- The coding rounds are SQL on product data, not algorithm puzzles: joins, aggregations and group by.
How the interviews run
Data scientists at Meta sit inside product teams. They work next to product managers, designers and engineers, and help the team decide what to build and when to launch. The interviews test exactly that.
The technical screen
- A product case. The recorded one was about ads.
- Two SQL questions.
- Measuring a product, with and without an A/B test.
The onsite rounds
Four rounds, and one of them is SQL. Candidates often call the other rounds product sense and analytical execution. Each loop is a round, and the onsite is the final round, even when it runs over video.
Is the coding separate from the rest?
Candidate: “Is the coding separate, like LeetCode questions?”
Meta data scientist: “Coding is covered in the initial interview for part of it, and one of the four loops in the final round is dedicated to coding.”
The four question types Meta asks
Meta’s own data scientists listed four types in their prep session. Meta gave candidates one minute per question, so read the example, think, then compare your answer with Meta’s.
1. Why did this metric change?
Example: Likes per user on Facebook dropped 10% over the past year. What analysis would you do? Interviewers want hypotheses, and for each one, a way to check it in the data.
How Meta’s data scientists answered:
- Should we even worry? Is 10% unusual for this metric? Sudden or gradual? All users or some? Are other metrics falling too?
- Hypotheses: a product launch (a sharp drop on one date), competition (a slow decline), only some groups affected, people switching to comments or reactions, or different content in the feed.
- Check each one. For a launch, plot likes per user, mobile vs web. For switching, plot likes, comments and reactions together, and check that time spent stayed flat.
- Recommend a fix, personalization for a struggling group, or tracking likes plus reactions as one metric.
2. What is this product worth?
Example: Messenger now lets businesses chat with customers. Why is that valuable, and how would you measure it? Interviewers want value for every side: users, businesses and Meta, and the risks.
How Meta’s data scientists answered:
- Value: customers buy more when questions get fast answers. It is cheaper for a business than a contact center.
- Risks: businesses could spam people, and brands might stop buying ads if they can message customers directly.
- Success: message threads where the customer replied. That beats messages sent, because spam gets no replies. Add businesses using it, and store visits or sales.
- Counter and ecosystem metrics come next, in the metrics section below.
3. Should we launch it?
Example: A new gaming tab makes time spent go up, but friend interactions go down. Is that a success? Interviewers want a success metric, a counter metric and your rule for trading one against the other.
- There is often no single right answer. Interviewers want to see how you weigh the trade-off.
- Name every side: players, game publishers, other users and Meta.
- Check that the tab is not taking time from the rest of the app.
- Then say how you would test it, and which result would make you say launch.
4. How should we design it?
Example: How should Facebook show Marketplace product suggestions in the feed? Interviewers want pure product sense: every type of user, and how the feature fits the rest of the product.
- Ask follow-ups until you know how the product works today.
- Cover buyers, sellers, people who do not shop, and Meta.
- Say how it fits next to everything else in the feed, including ads.
In the recorded screen, one case mixed types 1 and 2: are advertisers really spending less on brand ads, and what are brand ads worth?
How to structure your answer
Meta gave candidates a four-step framework. It works for every question type above.
- Frame it. Ask who the change affects and what matters to each of them. Then repeat the goal back in one sentence.
- Form hypotheses. Say what could be going on, list a few, then rank them. Use the interviewer’s answers to rule some out.
- Make it testable. For each hypothesis, describe an analysis clear enough that the interviewer could repeat it. Name the metric, the split and the chart.
- Recommend. Sum it up, state the caveats, and end with a decision: fix, launch, personalize or change the metric.
Meta said most candidates struggle at step three. They have good ideas but no way to check them.
Mistakes that cost candidates offers
- Hypotheses you cannot test. “Users stopped coming” is a guess. Say you would check daily active users and time spent.
- Running out of time. Long answers leave no time for the next questions. Be concise, then offer to go deeper.
- Success metrics only. Name what could go wrong, and how you would see it.
- Answering a different question. A question about a spending trend is not an A/B test question. A candidate lost about 3 minutes this way in the recorded screen.
- Integer division in SQL. Clicks divided by impressions, both integers, return 0 in PostgreSQL. Cast one side to a decimal.
Interview prep is included: when an interview comes up, AirCareer builds practice questions from the real job post.
Success, counter and ecosystem metrics
Almost every case ends with a question about how to measure the idea. Meta’s checklist for a good metric has five parts:
- Relevant: tied to the outcome you want.
- Measurable: Meta can actually log it.
- Specific: an average or a count, daily or weekly.
- Prioritized: a few best ones, not a list.
- Balanced: it shows the good and the bad effects.
Then sort your metrics into three groups. Success metrics show the product works. Ecosystem metrics show what happens to the rest of the app. Counter metrics catch the harm (many companies call them guardrail metrics), and they usually move the opposite way.
Meta’s point: most candidates name success metrics easily and forget counter metrics. Always name at least one.
Here is how Meta sorted the metrics for business messaging, from the second question type:
| Group | What it shows | Metrics |
|---|---|---|
| Success | The product works | Message threads where the customer replied; store visits or sales from a chat |
| Ecosystem | The rest of the app | Total ad spend on Facebook; monthly active businesses |
| Counter | The harm | Spam and fraud reports; messages customers never answered |
SQL and A/B tests
SQL rounds and queries
The technical screen has two SQL questions, and one of the four onsite rounds is SQL. You get tables and a business question, and you write the queries.
- Joins, aggregations and group by come up most.
- Talk about goals and metrics while you code.
- Python is allowed if you arrange it with your recruiter. Pandas is the usual pick.
A worked example in our own words: click-through rate (CTR) by campaign type for the last 7 days. It needs a join and a filter, and it shows the integer-division trap.
SELECT c.campaign_type,
COUNT(*) FILTER (WHERE e.event_type = 'click')::decimal
/ NULLIF(COUNT(*) FILTER (WHERE e.event_type = 'impression'), 0) AS ctr
FROM ad_events AS e
JOIN campaigns AS c ON c.campaign_id = e.campaign_id
WHERE e.event_date >= CURRENT_DATE - INTERVAL '7 days'
GROUP BY c.campaign_type
ORDER BY c.campaign_type;On a small test data set, the query returns 0.10 for brand and 0.25 for direct. The brand data had 1 click in 10 impressions, and the direct data had 5 clicks in 20. Without the cast, both return 0.
A/B test questions
A/B questions show up in the screen and in the onsite. These points separate strong answers:
- Say which metrics need a test and which do not. Adoption does not: launch and watch.
- Decide when a user counts as exposed: on opening the tab, or on clicking into it.
- Friends can leak the feature to the control group, so choose your unit with that in mind.
- State alpha at 5%, power at 80% and a minimum effect. Then give the sample size and the duration.
Traps to avoid:
- Comparing time in the new tab between groups. The control group has no tab.
- Tracking ten success metrics. More metrics mean more false positives.
- Assuming you can always test. The recorded screen removed the A/B test on purpose.
Meta data scientist interview questions
How many rounds do Meta data scientist interviews have?
Two stages. A 45-minute technical screen comes first. Candidates who pass get an onsite of four rounds, and one of the four is SQL.
Do Meta data scientist interviews have LeetCode-style questions?
Not in the interviews we have. The coding rounds use SQL on a product data set: joins, aggregations and group by, always tied to a business question.
Can candidates code in Python instead of SQL?
Yes. Arrange it with your recruiter before the interview. Meta said pandas is the most common choice.
Do candidates need to know Meta’s products well?
No. Meta told candidates that the product case does not test product knowledge. If you have not used a feature, ask how it works before you answer.
What does Meta look for most?
A clear structure and clear communication. Meta taught candidates a four-step framework for answers.
Practice from the real job post, before the interview
We match you to new data scientist jobs every day, tailor your resume to each one and apply for you on the employer’s form. You approve each one, or auto mode sends strong matches.
Read next
Practice questions from the real job post.
Get practice questions

