← Meta Interview Insights

Meta·Machine Learning Engineer·Onsite - System Design / Architecture·Senior

SeniorPrefer not to say
Jun 2026

Summary

Meta ML Engineer system design round, basically one big question about recommender systems that kept branching into sub-topics for the whole session. Solid prep on two-tower models helped but I underestimated how deep they'd go on the serving and feedback loop side.

Questions Asked (6)

Q1

Design an end-to-end recommender system for a content or product feed, covering candidate retrieval, ranking, re-ranking, and online serving.

System DesignTechnical Trade-offs
Author's notes

This was the whole interview, basically.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the problem scope, including the type of feed (content vs. product), scale, and business objectives. Then walk through the system architecture in stages: candidate retrieval, ranking, re-ranking, and online serving, highlighting trade-offs and design choices at each stage. Emphasize how you would evaluate and iterate on the system using offline and online metrics.

Pro tip: Show awareness of the two-tower model for retrieval and the importance of feature freshness and embedding updates in online serving. Also, discuss how you would handle cold-start and diversity in re-ranking, as these are common challenges at Meta's scale.

1. Clarify Requirements and Scope

Ask questions to understand the feed type, scale (users, items), latency constraints, and business goals (e.g., engagement, revenue). Define success metrics and constraints.

2. Design Candidate Retrieval

Propose multiple retrieval sources (e.g., collaborative filtering, content-based, trending) and describe how to generate candidates efficiently, possibly using two-tower models or ANN search.

3. Design Ranking Models

Outline a multi-stage ranking approach: a lightweight ranker to prune candidates, then a heavier model (e.g., deep neural network) to predict engagement. Discuss features, model architecture, and training.

4. Design Re-ranking and Business Logic

Explain how to apply re-ranking to enforce diversity, freshness, and business rules (e.g., deduplication, policy constraints). Mention techniques like MMR or rule-based filters.

5. Design Online Serving and Evaluation

Describe the serving architecture (e.g., pre-computation, caching, real-time inference), latency considerations, and how to monitor and A/B test the system. Discuss feedback loops and retraining.

Key Points to Mention

  • Two-tower models for candidate retrieval and embedding-based ANN search
  • Multi-stage ranking with a cascade of models (e.g., lightweight to heavy)
  • Feature engineering: user, item, context features, and real-time features
  • Re-ranking for diversity, freshness, and business rules
  • Online serving: low-latency inference, caching, and pre-computation
  • Evaluation: offline metrics (recall, NDCG) and online A/B testing

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

How would you construct training data for the ranker, and how do you handle delayed or implicit feedback signals?

Data ModelingTechnical Trade-offs
Author's notes

I went straight to impressions as negatives and engagement as positives, which they seemed fine with.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by outlining the end-to-end process of constructing training data for a ranking model, from logging user interactions to generating labels. Then, discuss the challenges of delayed and implicit feedback, and propose methods to mitigate them, such as using surrogate labels or modeling delays. Emphasize trade-offs between data freshness, label accuracy, and model performance.

Pro tip: Demonstrate awareness of production constraints by mentioning how you would validate the training data pipeline and monitor for feedback loop biases, ensuring the model doesn't reinforce existing biases.

1. Data Collection and Logging

Describe how to collect user interaction data (clicks, views, purchases) and log relevant features and context. Highlight the importance of unbiased logging and handling missing data.

2. Label Generation for Ranking

Explain how to construct labels from implicit feedback, such as using clicks as relevance signals, and techniques like negative sampling or counterfactual logging to address position bias.

3. Handling Delayed Feedback

Discuss strategies for delayed feedback, such as waiting windows, importance weighting, or modeling the delay distribution to incorporate delayed conversions into training.

4. Training Data Pipeline and Validation

Outline the pipeline for generating training data, including feature engineering, splitting, and validation. Mention the need for temporal splits to avoid leakage and monitoring for distribution shifts.

5. Trade-offs and Iteration

Summarize key trade-offs (e.g., freshness vs. accuracy, complexity vs. scalability) and how to iterate based on offline metrics and online A/B tests.

Key Points to Mention

  • Position bias and how to correct it (e.g., using inverse propensity scoring)
  • Negative sampling strategies for implicit feedback
  • Delayed feedback handling: waiting windows, importance weighting, or survival models
  • Temporal validation to prevent data leakage
  • Trade-offs between using immediate vs. delayed signals
  • Monitoring and mitigating feedback loops and biases in production

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

How do you handle cold start for both new users and new items in a recommender system?

System DesignProduct Sense & Ideation
Author's notes

Felt okay on new items since content embeddings are a pretty natural answer.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the problem scope: distinguish between user cold start (new users with no interaction history) and item cold start (new items with no engagement data). Then propose a multi-pronged strategy that leverages side information, content-based methods, and exploration techniques, and discuss how to transition to collaborative filtering as data accumulates.

Pro tip: Emphasize the importance of a feedback loop: use exploration to gather data on cold-start entities, then quickly incorporate that data into the model. Also, mention that at Meta scale, you'd likely use a hybrid approach with embeddings from content and graph-based features.

1. Clarify the problem and constraints

Ask about the specific recommender system context (e.g., feed, marketplace), the definition of cold start (zero interactions vs. few), and business metrics (e.g., CTR, engagement).

2. Leverage side information and content-based methods

For new users, use demographic, contextual, and declared interest data; for new items, use content features (text, images, metadata) to compute similarity and make initial recommendations.

3. Employ exploration and bandit algorithms

Use multi-armed bandits or Thompson sampling to balance exploitation of known preferences with exploration of new items/users, gathering feedback efficiently.

4. Transition to collaborative filtering

As interactions accumulate, gradually blend in collaborative signals (e.g., matrix factorization, deep learning models) to improve personalization and item ranking.

5. Evaluate and iterate

Define offline and online metrics (e.g., coverage, diversity, CTR) and set up A/B tests to measure the cold-start strategy's effectiveness, iterating based on results.

Key Points to Mention

  • Content-based filtering using item metadata and user profiles
  • Hybrid models that combine content and collaborative signals
  • Exploration-exploitation trade-off via bandits or reinforcement learning
  • Use of graph neural networks to propagate information from connected entities
  • Transfer learning from related domains or tasks
  • Cold-start specific evaluation metrics like coverage and novelty

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q4

What metrics would you use to evaluate the recommender system offline versus online, and how do you decide when offline metrics are sufficient?

A/B Testing & ExperimentationProduct Analytics & Metrics
Author's notes

Talked through AUC and recall at K offline, then CTR and retention online.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by contrasting offline and online metrics for recommender systems, emphasizing that offline metrics measure predictive accuracy on historical data while online metrics capture real user engagement and business impact. Then explain that offline metrics are sufficient only when they correlate strongly with online outcomes and the system changes are low-risk; otherwise, online A/B testing is necessary. Conclude by discussing how to validate offline metrics through correlation studies and when to escalate to online testing.

Pro tip: Mention that at Meta, offline metrics are often used for rapid iteration, but online metrics are the ultimate arbiter; highlight the importance of guardrail metrics to catch regressions in user experience or system health.

1. Define offline metrics

List common offline metrics such as precision@k, recall@k, NDCG, MAP, and AUC, and explain how they evaluate the recommender's ranking quality on historical data.

2. Define online metrics

Describe online metrics like click-through rate (CTR), conversion rate, dwell time, and user engagement metrics, as well as business metrics like revenue and retention, measured through A/B tests.

3. Establish correlation

Explain that offline metrics are only useful if they correlate with online metrics; discuss methods like offline-online correlation studies and counterfactual evaluation to validate this relationship.

4. Decide sufficiency

Outline criteria for when offline metrics are sufficient: low-risk changes, strong historical correlation, and limited resources; otherwise, recommend online testing.

5. Consider trade-offs

Discuss the trade-offs between offline and online evaluation, including speed, cost, and potential biases in offline data, and how to combine both for robust decision-making.

Key Points to Mention

  • Offline metrics: precision@k, recall@k, NDCG, MAP, AUC
  • Online metrics: CTR, conversion rate, dwell time, revenue, retention
  • Correlation between offline and online metrics (e.g., through A/B tests)
  • When offline metrics are sufficient: low-risk changes, strong correlation, resource constraints
  • Guardrail metrics to monitor for regressions
  • Counterfactual evaluation and off-policy evaluation techniques

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q5

How would you design the online serving infrastructure to support real-time inference at high query volume, including feature consistency between training and serving?

System DesignTechnical Trade-offs
Author's notes

Feature store architecture came up here and I talked through offline versus online feature consistency, which is a classic training-serving skew problem.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying requirements: QPS, latency SLA, model size, and feature freshness. Then propose a layered architecture covering request handling, feature retrieval, model inference, and caching, with a strong focus on feature consistency via a feature store and logging for training-serving parity. Finally, discuss trade-offs between latency, cost, and consistency, and how to monitor and mitigate skew.

Pro tip: Emphasize that feature consistency is not just about using the same code, but about ensuring the same data and transformations are applied in both training and serving pipelines, often via a feature store with point-in-time correctness. Also, mention the importance of logging served features and predictions to detect and debug skew in production.

1. Clarify Requirements and Constraints

Ask about expected QPS, latency SLA (e.g., p99 < 100ms), model size, feature types (real-time vs batch), and consistency requirements. This shapes the entire design.

2. Design Serving Architecture

Propose a scalable serving layer: load balancer, stateless inference servers (e.g., using Triton, TF Serving), autoscaling, and caching. Consider model partitioning, batching, and hardware acceleration (GPU/CPU).

3. Address Feature Consistency

Introduce a feature store (e.g., Feast, Tecton) that provides offline and online stores with point-in-time correct features. Ensure transformations are defined once and reused in both training and serving.

4. Handle Real-Time Feature Retrieval

For real-time features, use low-latency stores (e.g., Redis, DynamoDB) and precompute aggregations where possible. For batch features, use a fast key-value store. Discuss caching and fallback strategies.

5. Monitor and Mitigate Skew

Log served features and predictions, compare with training data distributions, and set up alerts for skew. Use shadow deployment and A/B testing to validate consistency.

Key Points to Mention

  • Feature store with online/offline consistency and point-in-time correctness
  • Low-latency feature retrieval using in-memory stores and caching
  • Model serving optimizations: batching, quantization, hardware acceleration
  • Logging and monitoring for training-serving skew detection
  • Trade-offs between consistency, latency, and cost
  • Scalability patterns: autoscaling, load balancing, and stateless services

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q6

How do you think about feedback loops and filter bubbles in a deployed recommender system, and what can you do to mitigate them?

Product StrategyTechnical Trade-offs
Author's notes

Honestly the question I was least prepared for.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Acknowledge that feedback loops and filter bubbles are inherent challenges in recommender systems, and demonstrate a balanced understanding of their causes and consequences. Structure your answer by first defining the problem, then discussing mitigation strategies across data, modeling, and evaluation, and finally emphasizing the importance of trade-offs and continuous monitoring.

Pro tip: Show that you think about both user and system perspectives: filter bubbles can harm user experience and long-term engagement, while feedback loops can degrade model performance. Mention that mitigation often involves trade-offs between short-term metrics and long-term health, and that Meta values proactive solutions.

1. Define and Diagnose

Clearly define feedback loops and filter bubbles in the context of recommender systems, and explain how they manifest (e.g., popularity bias, homogenization).

2. Assess Impact

Discuss the potential negative impacts on users (e.g., reduced diversity, echo chambers) and on the system (e.g., model drift, reduced long-term engagement).

3. Mitigation Strategies

Outline concrete strategies across data, modeling, and evaluation: diversify training data, use debiasing techniques, incorporate exploration, and design metrics that capture diversity and long-term value.

4. Trade-offs and Evaluation

Explain how to balance mitigation with other objectives (e.g., relevance, engagement) and how to evaluate effectiveness through A/B tests and long-term studies.

5. Monitoring and Iteration

Emphasize the need for continuous monitoring, feedback mechanisms, and iterative improvements to adapt to evolving user behavior and system dynamics.

Key Points to Mention

  • Feedback loops: self-reinforcing cycles where model predictions influence future data, leading to bias amplification.
  • Filter bubbles: users are exposed to increasingly narrow content, reducing diversity and potentially polarizing opinions.
  • Mitigation via data: ensure diverse and representative training data, use inverse propensity weighting, and incorporate negative sampling.
  • Mitigation via modeling: add exploration (e.g., epsilon-greedy, Thompson sampling), use regularization, and consider multi-objective optimization.
  • Evaluation metrics: beyond accuracy, measure diversity, serendipity, and long-term user satisfaction; use counterfactual evaluation.
  • Trade-offs: balancing short-term engagement with long-term user well-being and platform health; consider business and ethical implications.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.