Data Scientist Interview Guide 2026
SQL, statistics, machine learning case studies, A/B testing, and product sense. everything you need for DS interviews at top tech companies.
What DS Interviews Actually Test
Data scientist interviews at big tech are broader than most candidates expect. They test five distinct skill areas: SQL and data manipulation, statistics and probability, machine learning theory and application, A/B testing and experimentation design, and product/business sense. Most companies test all five; the weighting varies. Meta and LinkedIn weight product sense heavily. Google and Amazon weight ML theory more. Stripe and Airbnb emphasize experimentation design.
Knowing which areas a specific company emphasizes is half the battle. LeakCode lets you filter DS interview reports by company and seniority so you can calibrate before investing weeks in prep.
SQL: The Baseline Everyone Must Clear
Nearly every DS interview includes at least one SQL round, and it's often an early-stage filter. The questions range from basic aggregations to window functions, self-joins, and multi-step analytical queries. You need to be comfortable with: GROUP BY and aggregation, HAVING vs WHERE, subqueries vs CTEs, window functions (ROW_NUMBER, RANK, LAG, LEAD, SUM OVER PARTITION BY), and query optimization basics.
The most common failure point is not knowing window functions well. If you can't write
SUM(revenue) OVER (PARTITION BY user_id ORDER BY date ROWS BETWEEN 6 PRECEDING AND CURRENT ROW)
fluently, prioritize this above all else.
Statistics and Probability
Statistics questions at DS interviews are often deliberately ambiguous to test whether you think carefully about assumptions. Core topics: probability (Bayes' theorem, conditional probability), hypothesis testing (p-values, Type I/II errors, statistical power), confidence intervals, distributions (normal, binomial, Poisson), and the Central Limit Theorem.
A frequently underestimated topic: variance and bias tradeoffs in ML models, and why a high-accuracy model on training data can fail in production. Understanding overfitting, regularization (L1/L2), and cross-validation is expected even for interviews with minimal ML focus.
A/B Testing and Experimentation
This is where DS interviews at product companies diverge from academia. You'll be asked to design experiments, identify flaws in existing experiment setups, and interpret results. Common pitfalls interviewers probe for: peeking (stopping early when results look significant), novelty effects, network effects violating SUTVA, incorrect randomization units (user vs cookie vs session), and Bonferroni corrections for multiple testing.
Practice walking through: "How would you measure whether feature X is working?" Starting with "what metric am I optimizing?" and ending with "what could go wrong with this experiment design?" is the structure interviewers want to see.
Machine Learning Case Studies
ML case studies test whether you can apply theory to real problems. A typical prompt: "Design a recommendation system for [feature]" or "How would you build a fraud detection model?" The interviewer is evaluating your problem decomposition, feature engineering instincts, model selection reasoning, and awareness of production concerns (latency, explainability, data drift).
Don't jump to deep learning for everything. Showing awareness of when a logistic regression or gradient-boosted tree is the right tool (and why) signals stronger ML maturity than defaulting to neural networks.
Statistics Depth Probes
Statistics rounds at DS interviews probe foundational understanding. Topics that interviewers consistently probe at Meta, Airbnb, Stripe, and Lyft: when to use t-test vs z-test (sample size threshold and known vs unknown variance), one-tailed vs two-tailed hypothesis tests (alternative hypothesis direction), Type I vs Type II error trade-offs (alpha vs beta, sample size relationship), confidence intervals vs prediction intervals (precision of mean vs precision of next observation).
Bayesian vs frequentist framing comes up at AI-lab DS interviews (Anthropic, OpenAI). Be able to articulate when Bayesian methods help (informative priors, sequential decision making, hierarchical models) vs when frequentist is appropriate (high-volume A/B tests at fixed sample sizes, regulatory contexts). Reports on LeakCode show this question framed as "you have prior data from previous experiments; how does that change your A/B test design?"
Metric Design and Trade-Off Round
Senior DS rounds increasingly test metric design judgment. Prompts: define the success metric for a new feature launch, propose a guardrail metric that catches negative externalities, design a north-star metric for an entire product line.
Strong candidates structure the answer: state the goal of the metric, define the metric precisely (numerator, denominator, time window, segmentation), justify the choice against alternatives, identify failure modes the metric does not capture, and propose at least one guardrail metric. Weak candidates pick a metric, defend it as the right choice, and miss that no single metric captures all the relevant dimensions. The discriminating skill is intellectual honesty about trade-offs.
Diagnostic and Causal Reasoning
Diagnostic rounds present an anomaly and ask you to investigate. "Daily Active Users dropped 8% last week. Walk me through your investigation." The structure that wins: ask clarifying questions (which platforms, which geos, sudden vs gradual), form hypotheses systematically (data-quality issue, experiment effect, external event, seasonality), design specific checks for each hypothesis, articulate what data you would request.
Causal inference techniques come up at senior+ DS interviews. Be able to articulate when correlation is sufficient (predictive modeling) vs when causal identification matters (policy decisions, feature attribution). Methods: randomized experiments (gold standard), difference-in-differences (when randomization is impossible but a comparable control exists), regression discontinuity (sharp policy cutoffs), instrumental variables (when confounders cannot be controlled directly).
DS vs Applied Scientist vs Research Scientist
Top tech companies have three distinct quant tracks with different interview loops. Data Scientist: product analytics, experimentation, SQL-heavy, focus on metric design and causal inference. Applied Scientist (Amazon, Microsoft, AWS): more ML modeling, less SQL, builds production models. Research Scientist (Google Brain/DeepMind, Meta FAIR, OpenAI, Anthropic): publishes papers, deep math/ML theory, less product framing.
Reports on LeakCode show candidates often confuse these tracks during recruiter screens, leading to mismatched loops. Clarify with the recruiter: what is the day-to-day work, what is the publication expectation, what is the SQL-to-modeling-to-engineering ratio. Mid-loop pivots are rarely successful; pick the right track upfront.
Browse Real DS Interview Questions by Company
Filter by Data Scientist role across thousands of companies. See what SQL, stats, and ML questions are actually being asked right now.
Browse DS Questions