Company interview guide

Netflix Machine Learning Interview Guide (2026)

Rapid-fire ML and LLM fundamentals, practical coding, team-specific design and a heavy culture check.

Netflix's machine learning interview loop has 8 reported stages, from recruiter & hiring-manager calls to culture & hiring manager; typical length: 3-6 weeks.

Roles covered: ML Engineer, MLOps Engineer, Applied Scientist and Data Scientist. The question bank holds 24 questions reported from Netflix interviews, 7 of them free to solve in your browser.

Try a free question

Interview process length
3-6 weeks
Stages
8
Reports linked to bank questions
30
Bank report dates
Mar 2015–Sep 2026
Questions in bank
24 (7 free)
Last reviewed
October 2026

How the Netflix interview loop works

A recruiter call and usually a hiring-manager call about your background, an online assessment for some roles (mostly internships), then a technical screen of 30-60 minutes. The onsite is often split over two days and mixes coding, ML system design, a partnership-style conversation with a stakeholder or partner-team leader, and culture conversations with the hiring manager or directors; data scientist loops replace much of this with experimentation, statistics and SQL. The hiring manager drives the decision. Several candidates report long silences or abrupt rejections, so keep other processes moving.

The Netflix interview process, stage by stage

01Recruiter & hiring-manager calls

Recruiter or HR call of about 30 min, then a call with the hiring manager (from under 15 to about 30 min)

Netflix often reaches out first, by recruiter or directly from a hiring manager, and sometimes the hiring manager call comes before the recruiter. The recruiter checks background, motivation and which Netflix values you relate to; the hiring-manager call goes through your resume, the team's work and why you want to leave your current job.

What it tests

  • Why Netflix, and which of its values you relate to
  • How your past projects map to the team's domain
  • A clear reason for changing jobs
  • Logistics: level, location, timeline

How to prepare

  • Read the culture memo before the first call; values questions can start here.
  • Do not lead with compensation as your reason for applying.
  • Ask which team you are interviewing for and what each later round covers; loops differ by team.

A recruiter or HR call in 17 candidate reports and a hiring-manager call in 15, mostly data science candidates; the four ML-role reports that describe the start follow the same pattern.

02Online assessment (some roles)

Online test, reported at about 90 min: a few multiple-choice ML questions plus coding or short ML problems

Reported for internships and some ML engineer pipelines, not for every loop. A 2026 intern test had four basic multiple-choice items (Bayes rule, forward-pass arithmetic, nearest neighbours, transformer facts) and two coding problems, one a sparse matrix product. Older tests mixed multiple choice with two ML questions or basic coding and ML items.

What it tests

  • Quick Bayes-rule and forward-pass calculations by hand
  • Basic concept checks on KNN and attention
  • Exploiting sparsity in a matrix product
  • Correct, complete code under a timer

How to prepare

  • Bank time on the multiple-choice items; they are meant to be fast.
  • For sparse products, iterate only over nonzero entries and still return the full dense result shape.

Free reported questions for this stage

Four ML-role reports (two internships, two full-time, 2021-2026) mention an online test; Netflix's intern and new-grad pages also describe a take-home assessment before interviews. No data science report mentions one.

03Technical screen: ML fundamentals and coding

30-60 min video call: often a project deep dive, rapid-fire ML questions, then a short coding task; data scientist screens focus on experimentation and SQL instead

Recent ML engineer screens open with 25-30 minutes on one of your projects, then move to quick conceptual questions, mostly on transformers and fine-tuning. The coding part is short and dictionary-based, such as nesting flat checkpoint names or counting words from a stream; one 2026 screen swapped coding for a live review of a feature-pipeline pull request. Older screens, including data scientist screens with ML content, probed classical topics such as bias and variance, L1 versus L2 regularization, tree models and logistic regression.

What it tests

  • Self-attention mechanics, the square-root scaling and the KV cache
  • Tokenizers and how LLaMA-style blocks differ from the original transformer (RMSNorm, RoPE, MoE)
  • LoRA, why it saves memory, and Adam versus SGD
  • Loss and preprocessing choices: MSE versus cross-entropy, L1 versus L2 normalization
  • Clean Python with dicts and classes, plus spotting leakage and train/serve skew in pipeline code

How to prepare

  • Answer fundamentals precisely and briefly; candidates have been rejected for fumbling MSE versus cross-entropy, LoRA, and L1 versus L2 regularization.
  • Prepare a 25-minute walk-through of one project that covers data, modelling choices, evaluation and impact.
  • Expect the conversation to follow your resume: research-heavy candidates were asked to extend their own work to new problems.
  • When reviewing pipeline code with passing tests, look for point-in-time joins, label leakage and features computed differently offline and online.

Free reported questions for this stage

Eight ML-role reports (six since 2025) plus three DS screens describe this; Netflix's intern and new-grad pages say interviews test technical and role-specific skills and advise preparing to discuss resume projects.

04Onsite: practical coding

About 60 min of live coding; one recent loop built a small stateful service

The most detailed recent report describes building a service that picks a title for the homepage billboard: titles have relevance scores that can be updated, the highest score wins, but the same title should not be returned twice in a row when alternatives exist. Other ML engineer reports describe one coding round of medium to hard algorithm questions, with follow-ups on Python internals such as how dictionaries work. Older data scientist onsites had whiteboard Python plus SQL.

What it tests

  • Designing a small class with clear APIs (upsert, get top)
  • Keeping a priority structure correct when scores change
  • Encoding an extra business rule without breaking the ordering
  • Edge cases: a single title, repeated updates to the same id

How to prepare

  • Talk through the data structure choice first: a heap with lazy deletion or a sorted structure keyed by score.
  • Write a few calls by hand to show the rotation rule works before optimizing.
  • Know Python dict and set behaviour well; one L5 candidate was asked how dictionaries work internally.

Three ML engineer reports from 2024-2026 describe a coding round after the screen, one in detail, plus three older DS whiteboard rounds; Netflix's early-career pages say interviews assess technical skills, without detail.

05Onsite: ML system design

About 60 min, discussion on a shared doc or whiteboard

The prompt follows the team. A 2026 ML engineer loop asked for a recommendation system design; an ML infrastructure loop asked how to allocate CPUs and GPUs across several clusters for submitted ML jobs. An older data science onsite asked how to detect fraud in unlabeled login and payment data.

What it tests

  • Recommendation pipelines: candidate generation, ranking, logging and evaluation
  • Scheduling and resource allocation for training and inference jobs
  • Feature freshness, feature stores and online versus batch training
  • Working without labels: anomaly signals and how to validate them

How to prepare

  • Ask early which team's domain the design should target, and anchor on Netflix-like scale and surfaces.
  • Cover monitoring and rollback, not only the model; senior candidates are pushed on operating models at scale.

Two recent ML engineer reports (recommendation design, multi-cluster job scheduling), one 2013 phone round on building a Netflix-like recommender and one older data science fraud case.

06Stats & experimentation (DS and applied tracks)

30-45 min rounds on A/B testing, metrics and probability, usually with SQL; one senior loop added a take-home presentation

Data scientist loops at Netflix centre on experimentation: testing a personalized ranking change, choosing short-term proxies for retention, designing metrics for ranked lists, and p-values, multiple testing and variance reduction. Most rounds also include SQL of easy to medium difficulty, and product-sense cases appear in screens and onsites. No first-hand ML engineer report describes a separate experimentation round.

What it tests

  • Designing an A/B test for a ranking or recommendation change
  • Metric design for ranked lists and short-term proxies for long-term outcomes
  • Hypothesis testing, p-values and multiple-testing pitfalls
  • Bayesian updating and MLE versus MAP
  • Sizing the business impact of growth versus retention work

How to prepare

  • Practise explaining interference between rows and why retention needs long tests or proxies.
  • Be ready to compute a posterior or a covariance by hand; the questions are short but exact.

Fifteen data science reports from 2014-2026 describe experimentation, statistics or metric rounds, most with SQL; ML engineer reports do not mention this stage.

07Onsite: partnership round

About 45 min with a future stakeholder (for example a data scientist) or a director from a partner team

A Netflix-style round where someone you would work with tests how you collaborate. In a 2026 ML infrastructure loop a data scientist asked how you would handle slow model iteration, a launch that underperforms and data blocked by approvals; a 2024 L5 loop used a director from a partner team. It is half technical, half behavioral.

What it tests

  • Understanding the partner's pain points before proposing fixes
  • Diagnosing iteration bottlenecks and launch regressions
  • Improving process, such as data access, without ignoring governance
  • Speaking the partner's language rather than pure infrastructure

How to prepare

  • Do not promise a perfect fix; the interviewer pushes back on suggestions that sound too clean.
  • Show that you understand modelling challenges, not only systems.

Two ML engineer onsite reports (2024, 2026) describe a partner or stakeholder interview; data science loops mention peer-team interviewers without this framing.

08Culture & hiring manager

One or more conversations with the hiring manager, directors or senior leaders, mostly inside the onsite

Culture fit weighs heavily; reports say the hiring manager makes the final call with the team's support. Expect questions on candor, feedback, judgment, conflict and failure tied to the culture memo, plus a walk through your past experience. Senior leaders may also test how you explain technical work to non-technical people.

What it tests

  • Giving and receiving candid feedback
  • Disagreement and conflict with a colleague
  • Judgment under ambiguity and owning decisions
  • Biggest failure and what changed afterwards
  • Explaining technical results to non-technical leaders

How to prepare

  • Know the culture memo well and map two or three of your stories to its values.
  • Prepare a concrete example of feedback you gave and one you received and acted on.
  • Directors from other teams may join; treat them as future partners.

Ten candidate reports, mostly data science, describe culture or behavioral rounds; Netflix's intern and new-grad pages also say behavioral skills are assessed and you may be asked which parts of the culture memo resonate.

Questions and themes reported most often

  • A/B test design and pitfalls for product or personalization changes (12 reports, stats & experimentation (ds and applied tracks))
  • SQL on members, signups and viewing data (11 reports, stats & experimentation (ds and applied tracks))
  • Culture memo values and feedback stories (7 reports, culture & hiring manager)
  • Classical ML fundamentals in screens: L1 versus L2 regularization, bias and variance, logistic regression (5 reports, technical screen: ml fundamentals and coding)
  • Rapid-fire transformer questions: attention, scaling, KV cache, tokenizers, modern LLM blocks (4 reports, technical screen: ml fundamentals and coding)
  • Project deep dive at the start of the screen (4 reports, technical screen: ml fundamentals and coding)
  • Design a recommendation system for a streaming product (3 reports, onsite: ml system design)
  • Metric to compare two rankings of shows (3 reports, stats & experimentation (ds and applied tracks))

What Netflix asks about

  • ML & LLM concepts30%
  • General coding20%
  • Culture, partnership & deep dive20%
  • ML system design15%
  • Stats & experimentation15%

Netflix interview prep checklist

  • Can explain attention, the KV cache, LoRA and RMSNorm/RoPE crisply
  • Know when to use MSE versus cross-entropy and L1 versus L2 normalization
  • Wrote a score-ranked service with upserts and a no-repeat rule in under an hour
  • Can sketch a streaming recommender and a GPU job scheduler end to end
  • Designed an A/B test for a ranking change with a short-term proxy for retention
  • Mapped three personal stories to Netflix's culture values
  • Rehearsed a partnership conversation about a slow model iteration loop

Tips

  • Ask your recruiter which team you are interviewing with; the domain decides the design round and much of the screen.
  • Drill transformer and fine-tuning fundamentals until you can answer each in two or three exact sentences.
  • Prepare one project for a 25-30 minute deep dive, including what failed and how you measured success.
  • Keep practical Python sharp: classes with clear APIs, dict-based bookkeeping, heaps with updates and how dicts work inside.
  • For design, have a recommendation-pipeline blueprint and an ML-infrastructure blueprint ready.
  • Read the culture memo and turn its values into concrete stories about feedback, judgment and conflict.
  • Treat the partnership round as a stakeholder conversation: listen, diagnose, and avoid overpromising.
  • Data scientist candidates should expect experimentation on personalization, metric design and SQL in most rounds.

See the rest of the Netflix loop

The later stages, the questions reported for them and how to prepare are in Pro.

Unlock with Pro

Practice the Netflix loop

Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.

Netflix interview FAQ

How many interview rounds does Netflix have for ML roles?

Candidates report 8 stages: Recruiter & hiring-manager calls, Online assessment (some roles), Technical screen: ML fundamentals and coding, Onsite: practical coding, Onsite: ML system design, Stats & experimentation (DS and applied tracks), Onsite: partnership round and Culture & hiring manager. A recruiter call and usually a hiring-manager call about your background, an online assessment for some roles (mostly internships), then a technical screen of 30-60 minutes. The onsite is often split over two days and mixes coding, ML system design, a partnership-style conversation with a stakeholder or partner-team leader, and culture conversations with the hiring manager or directors; data scientist loops replace much of this with experimentation, statistics and SQL. The hiring manager drives the decision. Several candidates report long silences or abrupt rejections, so keep other processes moving.

How long does the Netflix ML interview process take?

Reported timelines run 3-6 weeks.

What does Netflix ask in ML interviews?

By reported question mix: ML & LLM concepts 30%; General coding 20%; Culture, partnership & deep dive 20%; ML system design 15%; Stats & experimentation 15%. The most repeated themes: A/B test design and pitfalls for product or personalization changes; SQL on members, signups and viewing data; Culture memo values and feedback stories.

How should I prepare for the Netflix ML interview?

Ask your recruiter which team you are interviewing with; the domain decides the design round and much of the screen. Drill transformer and fine-tuning fundamentals until you can answer each in two or three exact sentences. Prepare one project for a 25-30 minute deep dive, including what failed and how you measured success.

Are there free Netflix interview questions?

Yes. LeetML has 24 questions reported from Netflix interviews, and 7 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources