Company interview guide
Spotify Machine Learning Interview Guide (2026)
Practical Python and SQL, music-flavoured modeling cases and recommendation design, with a real values round.
Spotify's machine learning interview loop has 8 reported stages, from recruiter call to research scientist loop; typical length: 4-6 weeks.
Roles covered: ML Engineer, Data Scientist, Applied Scientist and Research Scientist. The question bank holds 14 questions reported from Spotify interviews, 1 of them free to solve in your browser.
- Interview process length
- 4-6 weeks
- Stages
- 8
- Reports linked to bank questions
- 18
- Bank report dates
- Apr 2020–Aug 2026
- Questions in bank
- 14 (1 free)
- Last reviewed
- October 2026
How the Spotify interview loop works
A recruiter call, one live technical screen, a conversation with the hiring manager, then a final round of four or five interviews with people across the team, usually held remotely. Data science loops add a take-home case that you present at the final round. Recent ML engineer finals split into data coding, ML depth, a modeling case, ML system design and values. Scheduling is often slow and follow-up can lag, so expect gaps between stages and chase your recruiter politely.
The Spotify interview process, stage by stage
01Recruiter call
About 30 min by phone or video
A conversation about your background, the role and why Spotify. Some recruiters add a few technical or behavioral questions, and level is often set here: one 2025 candidate was moved from senior to a mid level after this call.
What it tests
- Why Spotify and why this team
- How your past ML or analytics work matches the role
- Deployment and experimentation experience, briefly
- Logistics: salary range, location, relocation, level
How to prepare
- Have a two-minute summary of one shipped model or analysis with its business impact.
- Ask which rounds the final includes and what each covers; candidates say Spotify shares a short summary of each interview in advance.
- If you are aiming for a senior level, make the scope of your past work explicit now.
A recruiter call opens the process in 26 candidate reports across data science, ML and research roles, six of them from 2024-2026; the company's hiring page lists it as the first step.
02Technical screen
One live interview of about 60 min in CoderPad.
Data science: Python or pandas plus SQL, often a stats or A/B question. ML engineer: one or two easy-to-medium algorithm problems, sometimes with ML questions
The first technical round is run by one or two team members. For data roles it is practical data manipulation rather than algorithm puzzles, and code runs in a real SQL dialect. ML engineer screens lean on a classic coding problem plus some discussion of ML practice; a few older screens were ML fundamentals only.
What it tests
- Filtering, joining and aggregating tables in both SQL and pandas
- Reading an A/B result and saying what you need before calling it significant
- Basic probability and statistics questions
- An easy-to-medium coding problem: binary search, BFS, dynamic programming, a two-pointer sum
- Explaining your code and its complexity while you write it
How to prepare
- Practise the same question in SQL and in pandas; several screens asked for both.
- Write SQL that actually runs, since syntax errors cost time in a compiled pad.
- When an A/B question gives only two numbers, ask for sample sizes and variance rather than guessing.
- Keep algorithm practice to clean easy and medium problems; reports describe nothing harder.
One live technical interview after the recruiter in 36 reports: 23 data science (Python or pandas plus SQL, often stats) and 13 ML or research. The hiring page describes a team interview in CoderPad.
03Hiring manager and values
One conversation with the hiring manager, sometimes joined by an engineer; values can also be a separate final-round interview
A behavioral conversation about past projects, teamwork and fit with Spotify's values. In data science loops it often sits between the screen and the take-home and can include a mini product case. In recent ML engineer loops the hiring manager and values rounds are part of the final day.
What it tests
- Ownership and pride in your work, with concrete results
- Collaboration with product managers, researchers and other teams
- Adapting to change and ambiguity
- Depth on one project: baseline, model choice, imbalance, what you would change
- A short product question, such as investigating a drop in sign-ups
How to prepare
- Prepare two or three stories that show impact you can measure, and rehearse the follow-up questions.
- Know your strongest project well enough to defend a different model choice on the spot.
- Treat a hiring manager call as partly technical; several turned into a small case.
A hiring manager or values round in 25 reports across all tracks, placed before the take-home or inside the final loop. The hiring page says values alignment, ownership and collaboration are assessed.
04Take-home case and presentation (data science)
About a week to analyse a provided dataset or prompt, then a presentation (one report says 15 min) followed by questions
The signature data science step. A frequent prompt asks what makes a playlist successful; you choose metrics, analyse data and present findings and recommendations to the panel that then runs the rest of the final round. Candidates warn it can take more than four hours.
What it tests
- Defining success metrics for a product feature such as playlists
- Turning exploratory analysis into a clear recommendation
- Proposing an experiment to test what you found
- Presenting to a mixed audience and handling challenges to your choices
How to prepare
- Timebox the analysis and spend real effort on the story and slides.
- State metric trade-offs explicitly, for example engagement versus retention versus follower growth.
- End with the next experiment you would run and how you would read it.
A take-home case presented to a panel in 16 data science reports from 2021-2026 and one 2016 ML report; no ML engineer report since then mentions a take-home.
05Final round: data coding and analytics
45-60 min rounds. ML engineers: one SQL or small data-pipeline round on listening data. Data scientists: product sense, metric movement and experiment rounds with stakeholders
Both recent ML engineer finals included a data round on streaming logs: join, deduplicate and rank, then talk about scale. Data science finals instead hold several conversations with product managers and other data scientists about metrics, experiments and visualisation.
What it tests
- Window-function ranking per group (top users or artists per country)
- Deduplicating retried events before aggregating
- What changes when the table is too large: partitioning, sharding, incremental jobs
- Explaining a metric movement and scoping an open-ended analysis
- Forecasting a product number such as monthly active users
How to prepare
- Be fluent with ROW_NUMBER versus RANK versus DENSE_RANK and say which tie rule you chose.
- Ask whether event ids can repeat before you sum anything.
- For forecasting, separate the existing-user base from new acquisition and state your assumptions.
SQL, pipeline or analytics rounds in 3 ML-role final loops (two from 2025-2026); 4 data science reports, mostly 2021-2023, describe product, metric and experiment rounds with stakeholders.
06Final round: ML depth and modeling case
One or two discussion rounds, no live coding reported
One round goes deep on a family of methods you should know in detail, such as unsupervised clustering. Another is an open modeling case on a Spotify problem, for example finding free users likely to convert to premium, with questions on labels, features, embeddings, loss and evaluation. Pure theory questions are rare.
What it tests
- Comparing clustering methods: assumptions, hyperparameters, scale, evaluation
- Framing a propensity model: label definition without leakage, prediction horizon
- Feature and embedding choices from listening behaviour
- Loss functions, class imbalance and offline evaluation
- Defending modeling choices from a past project
How to prepare
- Walk every modeling case in the same order: objective, labels, data, features, model, loss, evaluation, launch.
- Prepare one method family in real depth; interviewers followed up on details of each model.
- Have a ready answer for how you would pick hyperparameters and build a baseline.
Free reported questions for this stage
- Free-to-Premium Conversion Propensity ModelML System DesignHardSolve
Modeling-case or ML depth rounds in 8 ML-role reports; both 2025-2026 final loops had them, while the other six date from 2016-2022 and give few details.
07Final round: ML system design
One open design discussion, on a shared whiteboard such as Mural
Design a production ML system around a Spotify surface: the backend for a song recommendation model, a pipeline that writes podcast recaps with language models, or a system that tracks listens. Interviewers push on serving, streaming ingestion and infrastructure, not only the model, and some expect a particular approach.
What it tests
- Candidate generation, ranking and serving for recommendations under a latency budget
- Batch versus streaming ingestion for new content
- LLM summarisation pipelines: transcription, chunking, evaluation, cost
- Logging, feedback loops and online evaluation
- Cold start for new users and new tracks
How to prepare
- Cover data flow and serving early; one 2026 candidate felt the round wanted more infrastructure than modeling.
- Prepare one recommendation blueprint and one LLM pipeline blueprint for audio or podcast content.
- Check in on the interviewer's expected direction before going deep on an alternative.
ML system design in 7 ML-role reports, including both 2025-2026 final loops (recommendation backend, podcast recaps); older ones cover listen tracking, podcast search and playlist models.
08Research scientist loop
Research track: two technical screens, then a research talk and themed interviews (research agenda, product, programming, ML, values), often 7-9 interviews
Research scientist candidates describe longer loops: a talk on their own work, a discussion of the research agenda they would pursue, and brainstorming on real problems from Spotify's recommendation and product teams. All of our research reports predate 2024.
What it tests
- Presenting and defending your own research
- Proposing a research agenda that fits the team
- Working through an open recommendation or product problem together
- ML fundamentals such as clustering and regularization
How to prepare
- Prepare a talk that ends with how your work could apply to listening or podcast data.
- Expect the process to take two to three months and plan your other interviews around it.
Five research scientist reports from 2017-2023 describe a research talk and themed rounds; nothing recent confirms the current format.
Questions and themes reported most often
- Python or pandas plus SQL data manipulation in the screen (14 reports, technical screen)
- Easy-to-medium algorithm problem (binary search, BFS, DP, two-pointer sum) (7 reports, technical screen)
- Is this A/B result significant, and how would you design the test (7 reports, technical screen)
- Deep dive on a past project and the choices behind it (6 reports, hiring manager and values)
- Take-home: what makes a playlist successful (4 reports, take-home case and presentation (data science))
- Design a song recommendation system or its backend (4 reports, final round: ml system design)
- Explain a metric movement or a drop in sign-ups (3 reports, final round: data coding and analytics)
- Why Spotify and what you want next (3 reports, recruiter call)
What Spotify asks about
- SQL & data coding20%
- ML modeling cases & concepts20%
- ML / recommendation system design18%
- Stats, A/B testing & product cases17%
- Python & algorithm coding13%
- Behavioral & values12%
Spotify interview prep checklist
- Solved a top-N-per-group query with deduplication in both SQL and pandas
- Can explain what you need before calling an A/B result significant
- Presented a playlist-success style analysis in 15 minutes with a clear recommendation
- Walked a free-to-premium propensity case from labels to evaluation
- Sketched a song recommendation backend and a podcast recap pipeline with serving and metrics
- Prepared three values stories and one project deep dive
Tips
- Drill SQL window functions and pandas groupby on listening-style tables; data coding shows up for both data science and ML roles.
- Frame every case around a Spotify surface: playlists, home recommendations, podcasts, premium conversion.
- For data science roles, budget well over four hours for the take-home and rehearse the presentation out loud.
- For ML roles, prepare a recommendation system design and an LLM content pipeline design, each with serving and evaluation.
- Pick one ML method family and know it to the level of hyperparameters, complexity and failure modes.
- Prepare values stories on ownership, adaptability, collaboration and drive, the traits the hiring page lists.
- Ask the recruiter for the round summaries and for feedback after the final; a 2026 candidate received detailed feedback.
- Expect slow scheduling and occasional silence; follow up with your recruiter after about a week.
See the rest of the Spotify loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Spotify loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Spotify interview FAQ
How many interview rounds does Spotify have for ML roles?
Candidates report 8 stages: Recruiter call, Technical screen, Hiring manager and values, Take-home case and presentation (data science), Final round: data coding and analytics, Final round: ML depth and modeling case, Final round: ML system design and Research scientist loop. A recruiter call, one live technical screen, a conversation with the hiring manager, then a final round of four or five interviews with people across the team, usually held remotely. Data science loops add a take-home case that you present at the final round. Recent ML engineer finals split into data coding, ML depth, a modeling case, ML system design and values. Scheduling is often slow and follow-up can lag, so expect gaps between stages and chase your recruiter politely.
How long does the Spotify ML interview process take?
Reported timelines run 4-6 weeks (About 4-6 weeks; some loops ran 2-3 months).
What does Spotify ask in ML interviews?
By reported question mix: SQL & data coding 20%; ML modeling cases & concepts 20%; ML / recommendation system design 18%; Stats, A/B testing & product cases 17%; Python & algorithm coding 13%; Behavioral & values 12%. The most repeated themes: Python or pandas plus SQL data manipulation in the screen; Easy-to-medium algorithm problem (binary search, BFS, DP, two-pointer sum); Is this A/B result significant, and how would you design the test.
How should I prepare for the Spotify ML interview?
Drill SQL window functions and pandas groupby on listening-style tables; data coding shows up for both data science and ML roles. Frame every case around a Spotify surface: playlists, home recommendations, podcasts, premium conversion. For data science roles, budget well over four hours for the take-home and rehearse the presentation out loud.
Are there free Spotify interview questions?
Yes. LeetML has 14 questions reported from Spotify interviews, and 1 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- NetflixRapid-fire ML and LLM fundamentals, practical coding, team-specific design and a heavy culture check.
- PinterestRapid-fire ML fundamentals, edge-case-heavy coding and embedding-centric recommendation design.
- LinkedInProbability-flavoured coding, ML fundamentals and recommender design built around LinkedIn's own products.
- AirbnbMarketplace ranking design, a hard look at your past models, one practical coding round and a values check.
- UberGrid and graph coding, ML from scratch and marketplace system design built around ETA, pricing and Uber Eats.
- TikTokHour-long mixed rounds: a resume grilling, ML fundamentals and an algorithm problem, repeated two to four times.