Data scientist interview questions and how to answer them
Questions on the skills open data scientist listings mention and on the core work of the role. For each: what the interviewer is checking, and a STAR answer outline with placeholders for your own example.
Based on 114 open data scientist listings on ROLIVA as of . Each question below is about a term from ROLIVA’s skills vocabulary that appears in at least two of these listings, most-mentioned first.
01 Tell me about something you built or automated in Python that other people relied on. What would you do differently now?
Why this question: Python is mentioned in 101 of 114 listings.
What the interviewer is checking. Whether you write Python that works beyond your own laptop: structure, error handling, tests or checks, and whether you can judge your earlier work honestly.
Answer outline (STAR)
Situation.[The task or problem, and who depended on the script, notebook or service].
Task.[What it had to do and any constraint, e.g. data size, run frequency or a deadline].
Action.[The approach and libraries you used, how you handled failures or bad input, and how you tested it].
Result.[What it replaced or made possible, and the one thing you would change today]. If there is no number you can support, say what changed as a result or what you learned.
02 Tell me about a machine learning model you worked on that was used for a real decision. How did you know it was good enough?
Why this question: Machine learning is mentioned in 75 of 114 listings.
What the interviewer is checking. Whether you frame the problem, choose a baseline and metrics that match the decision, and check for leakage and drift rather than reporting a single score.
Answer outline (STAR)
Situation.[The decision the model supported and the data available].
Task.[Your part: framing, features, modelling, evaluation or deployment].
Action.[The baseline, the model, how you evaluated it, e.g. a held-out set or an online test, and the risks you checked].
Result.[How it was used and how it performed after release]. If there is no number you can support, say what changed as a result or what you learned.
03 Walk me through a SQL query you wrote that answered a real question. How did you know the result was right?
Why this question: SQL is mentioned in 74 of 114 listings.
What the interviewer is checking. Whether you can turn a question into joins, filters and aggregations, and whether you check your own output instead of trusting the first number the query returns.
Answer outline (STAR)
Situation.[The question someone needed answered and the tables or data sources involved, e.g. orders joined to customers].
Task.[What you had to produce, for whom, and by when].
Action.[How you built the query: the joins, filters and grouping you chose, and how you checked it, e.g. row counts, a hand-checked sample or a comparison with a known total].
Result.[What the answer showed and what was decided because of it]. If there is no number you can support, say what changed as a result or what you learned.
04 Walk me through a Spark job you built or tuned. What made it slow or expensive, and how did you fix it?
Why this question: Spark is mentioned in 28 of 114 listings.
What the interviewer is checking. Whether you understand distributed processing (partitions, shuffles, skew, memory) and can tune by evidence rather than by guesswork.
Answer outline (STAR)
Situation.[The job, its input size and what consumed its output].
Task.[The problem, e.g. a job missing its window, failing on memory or costing too much].
Action.[How you diagnosed it, e.g. the Spark UI or stage metrics, and what you changed, e.g. partitioning, a broadcast join or handling skew].
Result.[Run time, reliability or cost before and after]. If there is no number you can support, say what changed as a result or what you learned.
05 Tell me about work you did with large language models as an engineering component. How did you evaluate the output and handle its failures?
Why this question: LLM is mentioned in 22 of 114 listings.
What the interviewer is checking. Whether you treat a language model as an unreliable component that needs evaluation, guardrails, cost control and fallbacks, not as a finished feature.
Answer outline (STAR)
Situation.[The feature or system and the role the model played in it].
Task.[What you were responsible for, e.g. evaluation, integration or reliability].
Action.[How you built an evaluation set, measured quality, handled wrong or unsafe output, and controlled cost and latency].
Result.[What you shipped or decided not to ship, and why]. If there is no number you can support, say what changed as a result or what you learned.
06 Describe something you built or ran on AWS. Which services did you choose, and what went wrong at some point?
Why this question: AWS is mentioned in 20 of 114 listings.
What the interviewer is checking. Whether you understand the trade-offs behind the services you used, including security, cost and failure modes, rather than just naming them.
Answer outline (STAR)
Situation.[The system and what it did, e.g. an API, a data pipeline or a batch job].
Task.[Your part in building or operating it].
Action.[The services you used and why, how you handled permissions and cost, and how you dealt with the failure].
Result.[How the system ran afterwards and what you changed to prevent a repeat]. If there is no number you can support, say what changed as a result or what you learned.
07 Walk me through a model you trained in PyTorch. How did you debug it when training did not behave as expected?
Why this question: PyTorch is mentioned in 19 of 114 listings.
What the interviewer is checking. Whether you understand the training loop, data pipeline and common failure modes, and debug in a methodical way.
Answer outline (STAR)
Situation.[The task, the data and the model architecture].
Task.[What you were responsible for and the problem, e.g. a loss that would not fall or results that would not reproduce].
Action.[How you debugged it, e.g. overfitting a small batch, checking data and labels, learning rate or seeding, and the fix].
Result.[How the model performed afterwards on the evaluation you trusted]. If there is no number you can support, say what changed as a result or what you learned.
08 Walk me through an A/B test you designed or analysed. What would have made you distrust the result?
Why this question: A/B testing is mentioned in 13 of 114 listings.
What the interviewer is checking. Whether you understand experiment design: a clear hypothesis, a primary metric chosen in advance, enough sample, and checks for problems such as uneven groups or peeking.
Answer outline (STAR)
Situation.[The change being tested and why].
Task.[Your role: design, implementation or analysis].
Action.[The hypothesis, metric, sample size reasoning and the checks you ran on the data].
Result.[What the test showed and what was decided, including a null or negative result]. If there is no number you can support, say what changed as a result or what you learned.
Questions about the core work of a data scientist
These follow from the work the role title describes, not from a count of listings.
09 Tell me about a model or analysis you took from a business question to something that was used. What happened after it was delivered?
What the interviewer is checking. End-to-end ownership: problem framing, choice of method, evaluation, and whether the work had a real effect rather than ending in a notebook.
Answer outline (STAR)
Situation. Describe the business problem: [the team, the decision, and why a model or statistical analysis was needed].
Task. State what you were responsible for: [e.g. framing, data preparation, modelling and evaluation].
Action. Explain your approach [the baseline you compared against, the method you chose and why, how you evaluated it, and how it was handed over or deployed].
Result. Give a result you can support [e.g. the measured improvement over the baseline, the decision it informed]. If it was not adopted, say why and what you learned.
10 Describe a time your results looked too good. What did you do?
What the interviewer is checking. Scientific scepticism: spotting data leakage, bias or flawed evaluation before others rely on the results.
Answer outline (STAR)
Situation. Describe the result and why it seemed suspicious: [the metric and what you expected instead].
Task. Say what was at stake: [e.g. a presentation to stakeholders or a deployment decision].
Action. Explain your investigation [the checks you ran, e.g. for leakage, sample bias or test-set contamination; what you found; how you corrected the evaluation].
Result. Share the result you can support [the corrected performance and how you reported it]. If the result held up, say how the checks increased confidence.
11 How do you explain uncertainty in your findings to a non-technical decision maker?
What the interviewer is checking. Communication of uncertainty without hiding it or overwhelming the audience, and whether you help them decide anyway.
Answer outline (STAR)
Situation. Describe the finding and the audience: [what you found, who needed to act on it, and how uncertain it was].
Task. State what they needed from you: [e.g. a recommendation with a clear level of confidence].
Action. Explain how you communicated it [e.g. ranges rather than single numbers, plain-language framing, what would change your conclusion].
Result. Give the result you can support [the decision they made]. If they misread the uncertainty, say how you adjusted your communication.
12 Tell me about a time you chose a simpler method over a more advanced one.
What the interviewer is checking. Judgement: matching the method to the problem, the data and the people who will maintain it, rather than chasing complexity.
Answer outline (STAR)
Situation. Describe the problem and the options: [the simple approach and the more complex one you considered].
Task. Say what constraints mattered: [e.g. data volume, explainability, time or maintenance].
Action. Explain your decision [how you compared the options, e.g. a baseline test, and why the simpler one was enough].
Result. Share a result you can support [the performance and how it was used]. If you later needed the complex method, say what triggered that change.
Using the outlines
One real example per answer. Replace every [bracketed] part with something you did and can talk about in detail.
Keep the situation short. Spend most of the answer on what you did and why.
Say “I” for your part. Name the team’s work as the team’s, and your own decisions as yours.
Results you can support. If there is no number, say what changed or what you learned. Never estimate a figure you cannot back up.
How these questions are chosen
Each open data scientist listing’s description and stated skills are checked, on whole words, against the fixed skills vocabulary ROLIVA uses on its job pages, resume examples and monthly skills reports. A listing counts once per term. A mention is not a requirement, and the counts change as employers open and close roles. The questions and outlines are preparation prompts written for this role. They are not a record of what any employer has asked, and no employer’s process is described here.
Interview Studio prepares questions and STAR outlines for a real job from its description and the experience you have confirmed. No generative AI: when a detail is missing, it asks you instead of inventing one.