ML System Design Interview Guide for FAANG 2026

You know that annoying feeling when you can train a model, explain transformers, and ship notebooks, but the moment someone says “design TikTok recommendations” or “build a fraud detection system,” your brain opens 47 browser tabs at once.

That is the ML system design interview.

And for FAANG-level roles in 2026, it is no longer a cute bonus round. At Meta, Google, Amazon, Apple, Netflix, OpenAI, Uber, Airbnb, and similar companies, ML system design is often the round that separates “strong ML engineer” from “can own production ML at scale.”

If you are aiming for Machine Learning Engineer, Applied Scientist, Data Scientist with ML scope, AI Engineer, Search/Recs Engineer, or Staff-level ML roles, this guide will help you prepare without turning your life into a 300-page PhD syllabus.

What the ML System Design Interview Tests in 2026#

ML system design is not just “draw a model architecture.”

The interviewer wants to see if you can build an end-to-end ML product that works in the real world, with messy data, latency limits, cost pressure, bad users, model drift, legal constraints, and angry product managers.

In 2026, expect the interview to test:

1. Product thinking

   • What problem are we solving?
   • Who is the user?
   • What metric actually matters?

2. Data judgment

   • What data do we need?
   • How do we label it?
   • What can go wrong with bias, leakage, or privacy?

3. Modeling choices

   • Simple baseline or deep model?
   • Ranking, classification, retrieval, generation, forecasting?
   • Batch or real-time?

4. System design

   • APIs, queues, feature stores, storage, serving layers
   • Latency and throughput
   • Online and offline processing

5. Evaluation

   • Offline metrics
   • Online A/B testing
   • Guardrails and safety metrics

6. Production ML

   • Monitoring
   • Retraining
   • Drift detection
   • Rollbacks
   • Cost control

The best candidates do not sound like they memorized a template. They sound like someone who has watched a model quietly ruin a dashboard at 2 AM.

Typical FAANG ML System Design Questions#

You will see a few question families again and again.

Recommendation Systems

These are extremely common at Meta, Netflix, Amazon, YouTube, TikTok, Spotify, and Airbnb.

Examples:

1. Design Instagram Reels recommendations.
2. Design Netflix movie recommendations.
3. Design Amazon product recommendations.
4. Design LinkedIn “People You May Know.”
5. Design Spotify Discover Weekly.
6. Design Airbnb search ranking.

Expected topics:

• Candidate generation
• Ranking
• Personalization
• Cold start
• Diversity
• Freshness
• Exploration vs exploitation
• Online metrics like CTR, watch time, conversion, retention

Search and Ranking

Common at Google, Amazon, Apple, Pinterest, LinkedIn, and marketplace companies.

Examples:

1. Design product search for Amazon.
2. Design job search for LinkedIn.
3. Design local business search for Google Maps.
4. Design search autocomplete.
5. Design ranking for app store results.

Expected topics:

• Query understanding
• Embeddings
• Lexical retrieval like BM25
• Neural retrieval
• Learning-to-rank
• Latency budgets
• Relevance metrics like NDCG and MRR

Ads and Monetization

Very common at Meta, Google, Amazon, TikTok, Snap, Reddit, and Pinterest.

Examples:

1. Design an ad click prediction system.
2. Design ad auction ranking.
3. Design conversion prediction.
4. Detect low-quality ads.
5. Build frequency capping for ads.

Expected topics:

• CTR and CVR prediction
• Calibration
• Auction mechanics
• User privacy
• Delayed labels
• Fraud
• Revenue vs user experience

Trust, Safety, and Fraud

Common at Apple, Uber, Airbnb, PayPal, Stripe, Coinbase, Meta, and Amazon.

Examples:

1. Detect fraudulent transactions.
2. Detect fake accounts.
3. Detect spam reviews.
4. Detect policy-violating content.
5. Detect bots on a social network.

Expected topics:

• Class imbalance
• Precision and recall tradeoffs
• Human review queues
• Adversarial behavior
• Real-time scoring
• Explainability
• Appeals and false positives

GenAI and LLM Systems

In 2026, you should assume this can show up.

Examples:

1. Design a customer support chatbot.
2. Design a code assistant.
3. Design semantic search over company documents.
4. Design an AI writing assistant.
5. Design an LLM-based moderation system.
6. Design RAG for enterprise knowledge search.

Expected topics:

• Prompting
• Retrieval augmented generation
• Vector databases
• Embedding refresh
• Evaluation of generated answers
• Hallucination control
• Latency and cost
• Safety filters
• Human feedback

How FAANG Interviewers Grade You#

The interviewer is usually not checking if your diagram matches their secret answer. They are checking how you think.

A strong answer usually shows:

1. Clarification before solution

   • You ask who the users are.
   • You define success metrics.
   • You identify constraints.

2. A reasonable baseline

   • You do not jump straight to a giant transformer for everything.
   • You can start with rules, logistic regression, gradient boosting, or embeddings if appropriate.

3. End-to-end structure

   • Data collection
   • Labels
   • Features
   • Model training
   • Serving
   • Evaluation
   • Monitoring

4. Tradeoff awareness

   • Latency vs accuracy
   • Freshness vs cost
   • Recall vs precision
   • Personalization vs privacy
   • Revenue vs user trust

5. Production maturity

   • You discuss failure modes.
   • You mention model drift.
   • You include logging and alerting.
   • You plan rollback and retraining.

Weak answers often sound like:

• “I would train a neural network.”
• “I would use deep learning.”
• “I would use all user data.”
• “Accuracy is the metric.”
• “Then we deploy it.”

That is not enough for Meta E5, Google L5, Amazon L6, Netflix Senior MLE, or Apple ICT4 interviews.

Salary Context: Why This Round Matters#

This interview is worth preparing for because the money difference is real.

Approximate 2026 compensation ranges, depending on location and level:

1. Meta Machine Learning Engineer

   • US: $220k to $450k total compensation
   • UK: £120k to £250k total compensation
   • Germany: €110k to €220k total compensation

2. Google ML Engineer or Research Engineer

   • US: $210k to $430k total compensation
   • Ireland: €100k to €210k total compensation
   • Switzerland: CHF 180k to CHF 320k total compensation

3. Amazon Applied Scientist or ML Engineer

   • US: $190k to $380k total compensation
   • Luxembourg: €100k to €200k total compensation
   • UK: £105k to £220k total compensation

4. Apple Machine Learning Engineer

   • US: $200k to $420k total compensation
   • Germany: €100k to €210k total compensation
   • UK: £110k to £230k total compensation

5. Netflix ML Engineer

   • US: $300k to $650k total compensation
   • Some senior roles can be cash-heavy with fewer bonus games.

At startups and scaleups, you might see lower cash but higher equity upside. For example, Mistral AI in Paris, Anthropic in San Francisco, Databricks in Amsterdam or London, and Stripe in Dublin can all pay very serious ML compensation.

So yes, spending 30 focused hours on this interview can be worth more than learning yet another Kaggle trick.

The Best Framework for ML System Design Answers#

Use this structure in almost every ML system design interview.

Do not announce it like a robot. Just guide the conversation this way.

1. Clarify the Product Goal

Start with questions.

For example, if the prompt is “Design a recommendation system for YouTube Shorts,” ask:

1. Are we optimizing watch time, likes, shares, retention, or creator revenue?
2. Is this for new users, existing users, or both?
3. Are we designing the full system or just ranking?
4. What latency do we need?
5. Are there safety constraints, like age-sensitive content?
6. Are recommendations personalized by user history, context, or location?

Then propose a goal.

Example:

“Let’s optimize long-term user engagement, using watch time and return rate as primary metrics, with guardrails for content quality, diversity, and policy violations.”

That sounds senior.

2. Define Inputs, Outputs, and Constraints

Make the system concrete.

For a feed recommendation system:

• Input: user ID, session context, device, location, time, recent interactions
• Output: ranked list of videos
• Latency: under 200 ms for ranking after candidates are fetched
• Scale: millions or billions of users, millions of items
• Freshness: new content should be eligible quickly
• Constraints: safety, privacy, fairness, creator diversity

This helps the interviewer see that you are not floating around in theory land.

3. Pick North Star and Guardrail Metrics

Do not say “accuracy.” Please. Somewhere, a senior staff engineer just spilled coffee.

For recommendations, use:

• Watch time
• CTR
• Completion rate
• Like or share rate
• Day 7 retention
• Creator follow rate
• Long-term satisfaction survey score

Guardrails:

• Hide or report rate
• Policy violation rate
• Repetitiveness
• Latency
• Diversity
• User churn
• Ad load if monetized

For fraud:

• Fraud loss prevented
• Precision
• Recall
• False positive rate
• Manual review rate
• User friction
• Latency at checkout

For search:

• NDCG@K
• MRR
• Query reformulation rate
• Click satisfaction
• Conversion rate
• Zero-result rate
• Latency

4. Design the Data Pipeline

Talk about events.

For a marketplace ranking system like Airbnb:

Collect:

1. Search query
2. User filters
3. Listing impressions
4. Clicks
5. Saves
6. Messages to host
7. Bookings
8. Cancellations
9. Reviews
10. Price and availability changes

Then explain storage:

• Raw event logs in S3, GCS, or HDFS
• Stream processing with Kafka, Flink, or Pub/Sub
• Batch processing with Spark or Beam
• Feature storage in online and offline feature stores
• Training data generated from logged impressions

Mention label leakage.

For example:

“If we use booking as a label, we need features that existed at ranking time. We cannot include future availability, future review score, or post-booking signals.”

That one sentence can save your whole answer.

5. Build a Baseline First

FAANG interviewers like ambition, but they love sanity.

For most systems, propose a baseline:

• Recommendations: popularity plus user-category affinity
• Fraud: rules plus logistic regression or gradient boosted trees
• Search: BM25 plus business rules
• Ads: logistic regression or GBDT for CTR
• GenAI support bot: keyword retrieval plus FAQ matching before RAG

Then improve.

You can say:

“I would start with a simple baseline to establish logging, evaluation, and serving. Then I would add more advanced models once the feedback loop is reliable.”

This signals production experience.

6. Choose the Model Architecture

Now you talk models, but only after product, data, and metrics.

Examples:

Recommendation System

A common two-stage architecture:

1. Candidate generation

   • Retrieve a few hundred or few thousand items.
   • Use collaborative filtering, two-tower embeddings, approximate nearest neighbor search, popularity, trending, social graph, or content similarity.

2. Ranking

   • Rank candidates using a richer model.
   • Use GBDT, wide and deep model, deep ranking model, or transformer-based sequence model.
   • Include user, item, context, and cross features.

3. Re-ranking

   • Apply diversity, freshness, safety, business rules, creator caps, or exploration.

Fraud Detection

A common architecture:

1. Rule-based first layer for obvious fraud.
2. Real-time ML model for transaction risk score.
3. Graph features for connected accounts, devices, cards, and IPs.
4. Human review for uncertain cases.
5. Feedback loop from chargebacks, disputes, and confirmed fraud.

Models:

• Logistic regression for interpretability
• XGBoost or LightGBM for tabular fraud signals
• Graph neural networks for entity networks
• Sequence models for behavior over time

LLM RAG System

A common architecture:

1. Ingest documents.
2. Chunk documents.
3. Generate embeddings.
4. Store in vector database like Pinecone, Weaviate, Milvus, Vespa, or pgvector.
5. Retrieve top chunks.
6. Rerank with cross-encoder or LLM.
7. Generate answer with citations.
8. Apply safety and policy checks.
9. Log feedback.

Metrics:

• Answer correctness
• Citation accuracy
• Hallucination rate
• Retrieval recall
• Latency
• Cost per query
• Escalation rate to human support

A Full Example: Design a TikTok or Reels Recommendation System#

Let’s walk through a realistic answer.

Clarify the Goal

You might say:

“I’ll assume we are designing a short-form video feed for existing and new users. The primary goal is long-term engagement, measured by watch time, completion rate, and day 7 retention. Guardrails include content safety, diversity, creator fairness, and latency.”

Good. You sound calm.

System Requirements

Functional requirements:

1. Generate personalized feed.
2. Support new users and new videos.
3. Refresh recommendations quickly.
4. Learn from user behavior.
5. Filter unsafe content.

Non-functional requirements:

1. Low latency, ideally under 300 ms end-to-end.
2. High availability.
3. Real-time event logging.
4. Scalable to millions of QPS at peak for a FAANG-scale app.
5. Privacy and policy compliance.

Data

User events:

• Impression
• Watch time
• Completion
• Skip
• Replay
• Like
• Share
• Comment
• Follow
• Report
• Hide
• Session length

Video features:

• Creator ID
• Upload time
• Caption
• Audio
• Visual embeddings
• Topic
• Language
• Safety label
• Historical engagement

Context features:

• Device
• Network
• Time of day
• Country
• Session depth
• Recent viewed topics

Candidate Generation

Use multiple candidate sources:

1. Videos similar to what the user watched recently.
2. Videos from followed creators.
3. Trending videos by region or language.
4. New videos for exploration.
5. Social graph signals, if available.
6. Embedding retrieval from a two-tower model.
7. Content-based retrieval for cold start.

You can mention ANN search using FAISS, ScaNN, Milvus, or similar systems.

The goal is to reduce millions of possible videos to maybe 1,000 candidates quickly.

Ranking Model

Use a ranking model that predicts multiple outcomes:

• Probability of watch
• Expected watch time
• Completion probability
• Like probability
• Share probability
• Hide probability
• Report probability

Then combine them into a utility score.

Example:

“Score could combine expected watch time, completion, and positive engagement, while subtracting predicted hide or report probability.”

Model options:

• GBDT baseline
• Wide and deep model
• Deep neural ranking model
• Sequence model using recent user behavior
• Multi-task learning model

For senior roles, mention calibration.

If the model predicts probabilities, they should be calibrated enough for ranking and tradeoff decisions.

Re-ranking

After ranking, apply:

1. Safety filters
2. Diversity across topics and creators
3. Freshness boost
4. Repetition limits
5. Exploration bucket
6. Business constraints, if needed
7. Age-appropriate filtering

This is important because the highest scoring 20 videos might all be from the same creator or topic. Users say they want relevance, but they also get bored fast.

Evaluation

Offline:

• AUC for classification tasks
• NDCG@K
• Recall@K
• Calibration error
• Counterfactual evaluation, if possible
• Slice metrics by country, language, age group, device

Online:

• Watch time
• Completion rate
• Session length
• Day 1 and day 7 retention
• Hide or report rate
• Creator diversity
• Latency
• Crash rate
• Long-term satisfaction surveys

Say this:

“I would not ship based only on offline AUC, because a ranking model can improve AUC and still hurt feed quality.”

That is the kind of sentence interviewers remember.

Common Follow-Up Questions and Strong Answers#

Interviewers will poke holes. That is normal.

“How do you handle cold start?”

For new users:

1. Ask onboarding interests.
2. Use country, language, device, and context.
3. Start with popular and diverse content.
4. Explore quickly and learn from skips, watch time, and likes.
5. Avoid over-personalizing after one click.

For new items:

1. Use content embeddings from text, image, audio, or metadata.
2. Give controlled exploration traffic.
3. Use creator history.
4. Compare against similar items.
5. Promote fresh content only within safety limits.

“How do you avoid filter bubbles?”

You can say:

1. Add diversity constraints in re-ranking.
2. Reserve slots for exploration.
3. Track topic concentration.
4. Penalize repetitive content.
5. Use long-term satisfaction, not just short-term clicks.
6. Run experiments on retention and survey quality.

“What about delayed feedback?”

For ads, purchases, fraud, and job applications, labels may arrive late.

Handle it by:

1. Training with delayed-label correction.
2. Using proxy labels like clicks or add-to-cart.
3. Separating short-term and long-term objectives.
4. Backfilling labels when final outcomes arrive.
5. Monitoring label freshness and missingness.

“How do you monitor the model?”

Track:

1. Prediction distribution
2. Feature distribution
3. Missing feature rates
4. Data pipeline delays
5. Online metrics
6. Segment-level performance
7. Drift by geography, device, and user cohort
8. Latency and error rates
9. Cost per prediction
10. Model version comparison

Also mention alerts and rollback.

“If key metrics move outside expected bounds, we can automatically reduce traffic to the new model or revert to the previous stable version.”

Nice. Very production-y.

GenAI ML System Design in 2026#

You need to be ready for LLM system design, even if your title is “Machine Learning Engineer” and not “AI Engineer.”

Companies like Google, Meta, Apple, Amazon, Microsoft, OpenAI, Anthropic, Salesforce, Databricks, and Shopify all ask versions of these questions now.

Example: Design a Customer Support Chatbot

Start with product goal:

“Reduce support ticket volume while maintaining answer quality and user trust.”

Metrics:

• Resolution rate
• Deflection rate
• Customer satisfaction
• Escalation rate
• Hallucination rate
• Average handle time
• Cost per conversation
• Latency

Architecture:

1. User asks question.
2. Classify intent and risk level.
3. Retrieve relevant documents from knowledge base.
4. Rerank retrieved chunks.
5. Generate answer with citations.
6. Apply safety checks.
7. Escalate to human for sensitive cases.
8. Log conversation and feedback.
9. Update knowledge base and evaluation set.

Important details:

• Use RAG for company-specific knowledge.
• Do not let the model answer policy-sensitive questions from memory.
• Use citations so users and agents can verify.
• Cache common answers.
• Use smaller models for simple queries and larger models for complex ones.
• Add human review for refunds, legal, medical, financial, or account security cases.

Evaluation:

1. Golden dataset with known answers
2. Human review
3. LLM-as-judge with caution
4. Retrieval recall
5. Hallucination tests
6. Red-team prompts
7. Online A/B test

Cost control:

• Prompt compression
• Caching
• Smaller model routing
• Batch embedding updates
• Token limits
• Early escalation for impossible requests

If you mention cost, interviewers usually smile, because real LLM systems can get expensive very quickly.

The Mistakes That Make Candidates Look Junior#

Here are the big ones.

1. Starting With the Model

Bad:

“I’d use a transformer.”

Better:

“I’d first clarify the goal, success metrics, constraints, and available data. Then I’d start with a baseline and improve from there.”

2. Ignoring Labels

A model is only as good as the labels.

Always explain:

• What the label is
• When it becomes available
• Whether it is biased
• Whether it can be gamed
• Whether it creates feedback loops

For example, clicks are not always satisfaction. People click ragebait too.

3. Forgetting Negative Feedback

For recommendations, negative signals matter:

• Skip
• Hide
• Report
• Short dwell time
• Unfollow
• Not interested
• Search reformulation
• App close after impression

A ranking model that only learns from likes will become weird very fast.

4. Not Discussing Scale

You do not need exact numbers, but you need reasonable thinking.

Say things like:

• “We cannot score every item for every user in real time.”
• “We need candidate generation before ranking.”
• “Popular features can be precomputed.”
• “User session features may need real-time updates.”
• “We should separate offline training from online serving.”

5. Treating Offline Metrics as Final Truth

Offline metrics help, but production decides.

A model can improve NDCG and still reduce retention. A fraud model can improve recall and still destroy checkout conversion. A chatbot can reduce tickets and still make customers furious.

Always connect offline evaluation to online A/B testing.

A 30-Day Prep Plan for FAANG ML System Design#

You do not need six months. You need focused reps.

Week 1: Learn the Core Template

Study and practice these system types:

1. Recommendation system
2. Search ranking
3. Ads prediction
4. Fraud detection
5. LLM RAG system

For each one, write a one-page answer covering:

• Goal
• Metrics
• Data
• Labels
• Features
• Model
• Serving
• Evaluation
• Monitoring

Week 2: Practice Full Designs

Do one mock per day.

Prompts:

1. Design YouTube recommendations.
2. Design Amazon product search.
3. Design Uber ETA prediction.
4. Design Stripe fraud detection.
5. Design Meta ad CTR prediction.
6. Design LinkedIn job recommendations.
7. Design a customer support chatbot for Shopify.

Record yourself. Yes, it feels awkward. Do it anyway.

Listen for:

• Did you clarify?
• Did you ramble?
• Did you define metrics?
• Did you miss monitoring?
• Did you handle follow-ups?

Week 3: Add Depth

Pick weak areas.

If you are weak on recommendations, study:

• Two-tower models
• ANN retrieval
• Ranking losses
• Exploration
• Diversity
• Cold start

If you are weak on GenAI, study:

• RAG
• Embeddings
• Reranking
• Evaluation
• Safety
• Cost and latency

If you are weak on production, study:

• Feature stores
• Model serving
• Batch vs streaming
• Monitoring
• Rollbacks
• A/B testing

Week 4: Mock Like It Is Real

Do 4 to 6 full mocks.

Use a timer:

1. 5 minutes: clarify
2. 5 minutes: requirements and metrics
3. 10 minutes: data and labels
4. 10 minutes: model and system architecture
5. 5 minutes: evaluation and monitoring
6. 5 minutes: tradeoffs and follow-ups

After each mock, write down:

• One thing you did well
• One thing you missed
• One phrase to use next time
• One diagram to simplify

What to Say When You Get Stuck#

You will get stuck. Everyone does.

Use calm phrases like:

1. “Let me break this into offline training and online serving.”
2. “I’ll start with a simple baseline, then discuss improvements.”
3. “The main tradeoff here is latency versus ranking quality.”
4. “I want to make sure the label is available at training time and not leaking future data.”
5. “For scale, I would use candidate generation before a heavier ranker.”
6. “I’d evaluate offline first, but I would not ship without an A/B test.”
7. “For safety, I’d add guardrails before and after model scoring.”

These phrases buy you time and make you sound structured.

Final Interview Checklist#

Before the interview, make sure you can explain:

1. Candidate generation vs ranking
2. Batch vs streaming features
3. Offline vs online evaluation
4. Precision vs recall
5. Cold start
6. Feature leakage
7. Model drift
8. A/B testing
9. Human review loops
10. LLM RAG basics
11. Latency and cost tradeoffs
12. Monitoring and rollback

If you can cover those calmly, you are in good shape.

Final Takeaway#

The ML system design interview is not about proving you know every model ever invented. It is about showing that you can take an ML idea and make it useful, safe, measurable, scalable, and maintainable.

FAANG interviewers want to hire someone they can trust with a real product. So act like that person: ask good questions, define metrics, start simple, discuss tradeoffs, and remember that production ML is mostly about everything around the model.

And before you apply to Meta, Google, Amazon, Apple, Netflix, OpenAI, Anthropic, or any serious AI team, make sure your resume can actually get through the first filter. Run it through JobRise’s free ATS checker here: https://jobrise.io/en/free-ats-checker/