← Meta Interview Insights

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

SeniorPrefer not to say
Jun 2026

Summary

Meta Research Engineer system design round focused entirely on a real-time dwell prediction system for a TikTok-style image feed. Dense question with a lot of moving parts and I definitely didn't get to all of them cleanly.

Questions Asked (3)

Q1

Design an ML system for a TikTok-style image feed that predicts in real time how long a user will dwell on a post, then auto-scrolls to the next post once the predicted dwell time is exceeded.

System DesignTechnical Trade-offsA/B Testing & Experimentation
Author's notes

I spent too long debating regression vs classification upfront and the interviewer had to nudge me to move on.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying requirements and constraints, then outline a high-level architecture that separates offline model training from online real-time inference. Focus on the trade-offs between prediction accuracy and latency, and discuss how to handle cold-start and feedback loops. Finally, propose an A/B testing framework to validate the system's impact on user engagement.

Pro tip: Emphasize the importance of a fallback mechanism: if the model's prediction is uncertain or the system fails, default to a safe heuristic (e.g., average dwell time) to avoid degrading user experience. Also, highlight the need for continuous monitoring and retraining to adapt to changing user behavior.

1. Clarify Requirements and Constraints

Ask questions to understand scale (users, posts), latency requirements (real-time prediction), and success metrics (e.g., user engagement, time spent). Clarify if the auto-scroll is optional or forced, and how to handle edge cases like user interaction.

2. High-Level Architecture

Propose a system with offline training (using historical user interaction data) and online serving (real-time prediction). Components include feature store, model server, and a client-side component that triggers auto-scroll based on predictions.

3. Model Design and Features

Discuss model choice (e.g., gradient boosted trees or neural networks) and features: user demographics, past behavior, post content (image embeddings), time of day, etc. Address how to handle cold-start for new users/posts.

4. Real-Time Inference and Auto-Scroll Logic

Explain how predictions are served with low latency (e.g., using a cache or precomputed predictions). Describe the auto-scroll trigger: if actual dwell time exceeds predicted, scroll to next post. Include fallback logic for model failures.

5. Evaluation and A/B Testing

Outline offline evaluation metrics (MAE, RMSE) and online A/B testing to measure impact on user engagement. Discuss how to set up experiments, guardrail metrics, and iterate based on results.

Key Points to Mention

  • Latency constraints: real-time prediction requires low-latency serving, possibly with precomputed predictions or edge inference.
  • Feature engineering: incorporate user history, post content, and contextual signals; use embeddings for images.
  • Cold-start problem: use content-based features or fallback to global averages for new users/posts.
  • Feedback loops: auto-scroll changes user behavior, which affects future training data; need to account for this bias.
  • A/B testing: define control and treatment groups, measure dwell time and engagement, and ensure statistical significance.
  • Fallback and safety: if prediction is uncertain or system fails, default to a safe heuristic to avoid negative user experience.

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

Q2

How would you handle cold-start users in this dwell prediction system, and what privacy constraints affect your personalization strategy?

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

This came as a follow-up and I was already a bit scattered.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining cold-start users and the signals available for them, then propose a tiered personalization strategy that degrades gracefully from rich to minimal data. Explicitly address privacy constraints like GDPR/CCPA and Meta's data policies, and discuss trade-offs between personalization quality and privacy preservation.

Pro tip: Emphasize that privacy constraints are not just legal requirements but also user trust factors; propose privacy-preserving techniques like federated learning or differential privacy as part of the solution, not just as limitations.

1. Define cold-start and available signals

Clarify what constitutes a cold-start user (e.g., new user, no historical dwell data) and identify available signals such as demographics, context, or first-session behavior.

2. Propose a tiered personalization strategy

Outline a fallback approach: use global/popularity-based models for true cold-start, then quickly transition to lightweight personalization using early interactions, and finally to full personalization as data accumulates.

3. Address privacy constraints

Discuss relevant privacy regulations (GDPR, CCPA) and Meta's policies, and explain how they limit data collection, storage, and cross-user sharing, affecting personalization.

4. Integrate privacy-preserving techniques

Suggest methods like federated learning, differential privacy, or on-device personalization to balance personalization with privacy, and explain their trade-offs.

5. Evaluate and iterate

Propose metrics (e.g., dwell time, engagement) to measure the effectiveness of the cold-start strategy and privacy-preserving methods, and describe how to iterate based on results.

Key Points to Mention

  • Cold-start types: user cold-start, item cold-start, and system cold-start
  • Fallback strategies: popularity-based, content-based, and contextual bandits
  • Privacy regulations: GDPR, CCPA, and Meta's data use policies
  • Privacy-preserving ML: federated learning, differential privacy, secure multi-party computation
  • Trade-offs: personalization accuracy vs. privacy, latency vs. complexity
  • Metrics: dwell time prediction accuracy, user engagement, and privacy budget

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

Q3

What guardrail metrics would you set up alongside engagement metrics when running A/B tests for this auto-scroll feature?

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

Liked this part.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the feature's goal and the engagement metrics being tracked, then propose guardrail metrics that ensure the feature doesn't harm user experience, system performance, or business goals. Structure your answer around categories like user well-being, technical performance, and long-term ecosystem health, and explain how you'd monitor and act on them.

Pro tip: Emphasize that guardrails should be leading indicators of potential harm, not just lagging metrics, and mention the importance of setting thresholds for automatic rollback to prevent negative user impact.

1. Clarify feature goals and engagement metrics

Understand what the auto-scroll feature aims to achieve (e.g., increased time spent, content discovery) and which engagement metrics are primary (e.g., scroll depth, session time). This sets the context for identifying potential side effects.

2. Identify potential negative impacts

Brainstorm ways the feature could harm users, the platform, or the business—such as accidental scrolling, reduced user control, increased bounce rates, or technical issues like jank.

3. Select guardrail metrics across categories

Choose specific metrics that capture those risks: user well-being (e.g., scroll-back rate, session abandonment), technical performance (e.g., frame drops, load time), and business health (e.g., ad viewability, retention).

4. Define thresholds and monitoring plan

Set acceptable bounds for each guardrail (e.g., no more than 2% increase in bounce rate) and outline how you'll monitor them during the experiment, including automated alerts and rollback criteria.

5. Plan for analysis and iteration

Explain how you'll analyze guardrail metrics alongside engagement metrics, looking for trade-offs, and how you'll iterate on the feature if guardrails are breached.

Key Points to Mention

  • User well-being metrics: scroll-back rate, session abandonment, frustration signals (e.g., rapid scrolling), and time to first interaction.
  • Technical performance metrics: frame rate, latency, crash rate, and battery consumption.
  • Business metrics: ad viewability, click-through rate, retention, and long-term engagement.
  • Statistical power and multiple comparisons: ensure guardrails are adequately powered and adjust for multiple testing.
  • Automated rollback: set thresholds for automatic experiment termination to prevent harm.
  • Long-term holdout: consider a holdout group to measure long-term effects beyond the experiment window.

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