Sign inStart practicing
Start practicing Sign in

ML System Design Interview Questions: 30 Problems and What They Test

ML system design interview questions asked at Meta, Google, Amazon and startups, grouped by type, with what each tests and a worked answer to study.

Beginner 3 min read · Updated

These are the machine learning system design interview questions that come up most in ML engineer, applied scientist and data scientist loops at large tech companies and ML-heavy startups. They're grouped into families, because questions in a family share an architecture: once you can design one recommender well, the others are variations. For each question you'll find what it really tests and where to study or practise it. Answer any of them with the seven-step framework.

Ranking and recommendations account for a large share of ML design questions at ad-funded and consumer companies. If your time is short, master the recommendation funnel first, then ads, then integrity.

The question families at a glance

Video, feed, PYMKRecommendationsretrieve → rankWeb, product, localSearch rankingquery → LTRCTR, conversionAdspCTR → auctionFraud, harm, spamIntegrityclassify → reviewETA, demand, pricePredictionregressionFeatures, training, servingML platformsinfrastructureML design questions
ML system design questions fall into a handful of families. Each family reuses one core architecture, so studying one worked answer per family covers most of what you'll be asked.

Recommendation and personalisation

QuestionWhat it testsStudy / practise
Design YouTube's home page recommendationsMulti-stage funnel, multi-objective ranking, satisfaction vs clicksWorked answer · Round
Design a personalized news feed ranking (Facebook, LinkedIn)Value model over several actions, integrity, freshnessWorked answer · Round
Design People You May KnowGraph candidates, link prediction, batch precompute, privacyWorked answer · Round
Design Netflix or Spotify recommendationsSmall catalogue vs huge one, session signals, artwork personalisationRecommendation systems
Design an event or restaurant recommenderLocation and time context, cold start, availabilityCold start
Design similar listings for a rental marketplaceItem embeddings, two-sided marketplace constraintsEmbeddings
Design Instagram Explore or TikTok For YouUnconnected content, exploration, fast feedbackBandits
Design push notification rankingUplift vs open probability, volume caps, send time, long-term holdoutsFeedback loops · Round

Search and retrieval

QuestionWhat it testsStudy / practise
Design search ranking for an e-commerce siteQuery understanding, hybrid retrieval, LTR, position biasWorked answer · Round
Design YouTube video searchText plus engagement signals, multimodal content understandingLearning to rank
Design a visual search system (image to similar images)Contrastive embeddings, ANN at billions of vectorsWorked answer · Round
Design search autocompletePrefix retrieval, ranking suggestions, latencyTypeahead · Round
Design local business search for a maps appDistance and opening hours as features, relevance labelsSearch ranking

Ads

QuestionWhat it testsStudy / practise
Design ad click prediction for a social platformCalibration, sparse features, auction, frequent retrainingWorked answer · Round
Design ads ranking with conversion optimisationDelayed labels, multi-task click and conversion modelsContinual learning
Design an ad targeting / lookalike audience systemUser embeddings, similarity search, privacyEmbeddings
Design an ad click aggregator (data side)Streaming counts, deduplication, exactly-onceRound

Integrity, trust and safety

QuestionWhat it testsStudy / practise
Design real-time payment fraud detectionImbalanced and delayed labels, velocity features, cost thresholdsWorked answer · Round
Design harmful content detectionMultimodal multilabel models, review queues, prevalenceWorked answer · Round
Design spam or fake account detectionBehaviour and graph signals, adversariesFeedback loops
Design blurring of faces and plates in street imageryObject detection at scale, precision vs recall, batch pipelinesModel serving

Prediction and forecasting

QuestionWhat it testsStudy / practise
Design ETA prediction for food delivery or ridesStage decomposition, quantiles, real-time featuresWorked answer · Round
Design dynamic (surge) pricingDemand and supply forecasting, feedback effects, experimentsA/B testing · Round
Design demand forecasting for inventoryTime-series features, hierarchy, backtestingTraining data
Design churn prediction and interventionLabel definition, uplift vs propensity, actioningProblem framing

ML infrastructure and platforms

QuestionWhat it testsStudy / practise
Design a feature storePoint-in-time joins, online/offline parity, streaming featuresGuide · Round
Design a distributed training platformGPU scheduling, checkpointing, reproducibilityGuide · Round
Design a low-latency model serving platformBatching, autoscaling, canaries, multi-model hostingGuide · Round
Design an ML model monitoring systemDrift, delayed labels, alertingGuide · Round
Design an ML platform for many teamsSelf-service workflow, registry, governance, costGuide
Design an experimentation platformAssignment, metrics pipelines, guardrailsGuide · Round

Generative AI questions

Questions like "design ChatGPT", "design a RAG chatbot" or "design a customer support agent" are increasingly common, especially for AI engineer roles. They need a different framework; see generative AI system design interview questions and the AI engineer roadmap.

How to practise these

  1. Read one worked answer per family, then close it and redraw the architecture from memory.
  2. For a second question in the same family, write only what changes: label, features, metric, serving constraint.
  3. Do timed rounds out loud: 45 minutes, framework steps in order, with someone (or an AI interviewer) asking follow-ups.
  4. After each round, list the questions you couldn't answer and read the matching guide on the ML roadmap.

Frequently asked questions

What are the most common ML system design interview questions?

Design a video or feed recommendation system, ad click prediction, search ranking, fraud detection, harmful content detection, People You May Know, visual search, ETA prediction, and ML infrastructure such as a feature store or model serving platform.

How should I prepare for ML system design questions?

Study one worked answer per question family (recommendations, search, ads, integrity, prediction, platforms), redraw each from memory, then do timed practice rounds using a seven-step framework.

Further reading