← Meta Interview Insights

Meta·Software Engineer·Onsite - System Design / Architecture·Senior

Senior
Jul 2026

Summary

Meta MLE onsite covering recommendation system design. The focus was on label quality issues and model selection trade-offs, with a lot of open-ended back-and-forth rather than clean textbook answers.

Questions Asked (2)

Q1

How would you design a recommendation system, and how do you handle label quality issues that arise during training?

System DesignTechnical Trade-offsProduct Analytics & Metrics
Author's notes

This is where most of the time went.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the product context and scale, then outline a high-level recommendation system architecture covering data, model, and serving. Dedicate significant time to label quality issues, discussing sources, detection, and mitigation strategies, and tie everything back to business metrics.

Pro tip: Emphasize that label quality is often a data pipeline problem—invest in logging and validation early, and consider weak supervision or human-in-the-loop systems to continuously improve labels.

1. Clarify Requirements and Scope

Ask questions to understand the product (e.g., what is being recommended, user scale, latency requirements) and define success metrics (e.g., CTR, engagement, revenue).

2. High-Level System Design

Outline the end-to-end architecture: data collection (user interactions, item features), feature engineering, model training (candidate generation, ranking), and serving (real-time inference, A/B testing).

3. Label Quality Issues: Sources and Detection

Identify common label quality problems (noise, bias, missing labels) and discuss detection methods (e.g., data validation, outlier analysis, cross-validation with held-out sets).

4. Mitigation Strategies for Label Quality

Propose techniques to handle noisy labels: robust loss functions, noise modeling, data cleaning, semi-supervised learning, and human-in-the-loop verification.

5. Evaluation and Iteration

Explain how to evaluate the system (offline metrics, online A/B tests) and set up feedback loops to continuously monitor and improve label quality and model performance.

Key Points to Mention

  • Two-stage architecture: candidate generation (e.g., collaborative filtering, embeddings) and ranking (e.g., deep learning models).
  • Importance of feature engineering and real-time features for recommendations.
  • Sources of label noise: implicit feedback (clicks, views) vs. explicit feedback (ratings), position bias, and missing data.
  • Techniques for handling noisy labels: noise-robust loss functions (e.g., bootstrapping, co-teaching), data cleaning, and re-weighting.
  • Use of human-in-the-loop and active learning to improve label quality.
  • Evaluation metrics: offline (precision@k, NDCG) and online (CTR, dwell time, conversion rate), and how to design A/B tests.

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

Q2

When would you choose a gradient boosting model like XGBoost over a neural approach in a recommendation setting, and what are the downsides?

Technical Trade-offsSystem Design
Author's notes

Felt okay on this one.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the decision around data characteristics, latency, and infrastructure constraints, then contrast XGBoost's strengths (tabular data, fast training, interpretability) with neural approaches' strengths (representation learning, scalability). Conclude by discussing the downsides of XGBoost in recommendation settings, such as limited expressiveness and cold-start handling.

Pro tip: Emphasize that in production, the choice often depends on the existing feature pipeline and serving infrastructure—XGBoost integrates easily with traditional feature stores, while neural models require embedding tables and GPU serving. Mention that hybrid approaches (e.g., using XGBoost for ranking after neural retrieval) are common at Meta.

1. Clarify the recommendation task

Identify whether the problem is candidate generation, ranking, or re-ranking, as different stages have different latency and data requirements.

2. Compare data and feature characteristics

Discuss when tabular, dense features with limited interactions favor XGBoost, while sparse, high-cardinality categorical features and sequential behavior favor neural models.

3. Evaluate latency and scalability constraints

Consider training and inference latency: XGBoost is fast to train and serve on CPU, while neural models may require GPUs and more complex serving infrastructure.

4. Assess interpretability and debugging needs

Highlight that XGBoost offers feature importance and easier debugging, which can be critical in early-stage or regulated settings.

5. Discuss downsides and trade-offs

Acknowledge XGBoost's limitations: difficulty capturing high-order feature interactions, poor performance on raw unstructured data, and challenges with online learning and cold-start.

Key Points to Mention

  • XGBoost excels on tabular data with engineered features, while neural models learn representations automatically.
  • Neural approaches scale better with massive datasets and can incorporate diverse signals (text, images, sequences).
  • XGBoost is easier to deploy and interpret, but may underperform on sparse, high-cardinality categorical features common in recommendations.
  • Latency and infrastructure: XGBoost serves well on CPU with low latency; neural models often need GPUs and embedding tables.
  • Cold-start and online learning: neural models can leverage content features and incremental updates more naturally.
  • Hybrid architectures are common: neural retrieval followed by XGBoost ranking, or vice versa.

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