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

Senior

Jul 2026

Summary

Meta ML Engineer system design round focused entirely on building a real-time nearby place recommendation system. Pretty comprehensive scope, they wanted the full stack from metrics down to A/B testing, so it was a lot to cover in one session.

Questions Asked (1)

Design a real-time system that recommends nearby places to users. Walk through your objectives and success metrics, data sources, candidate generation, feature engineering, ranking model, cold-start handling, exploration vs exploitation, spam filtering, privacy constraints, latency requirements, and an A/B testing plan.

System DesignA/B Testing & ExperimentationProduct Analytics & Metrics

Author's notes

This is basically a full ML system design in one question.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the product goals and success metrics, then walk through the ML pipeline from data sources to serving, addressing each component in a logical order. Emphasize trade-offs and how you would validate the system with A/B tests, while covering edge cases like cold-start, spam, and privacy.

Pro tip: Anchor your design around the user experience and business metrics—interviewers at Meta care about how ML drives engagement and revenue, so tie every technical choice back to a measurable impact.

1. Define Objectives and Metrics

Clarify the product goal (e.g., increase user engagement with nearby places) and define success metrics like CTR, conversion rate, and user retention. Also consider counter-metrics like spam reports or latency.

2. Data Sources and Candidate Generation

Identify data sources (user location, place database, user history, social graph). Generate candidates via multiple methods: geo-based retrieval, collaborative filtering, and trending places.

3. Feature Engineering and Ranking Model

Engineer features for user, place, and context (distance, popularity, time of day, user preferences). Use a two-stage ranking: a lightweight model for filtering and a heavy model (e.g., DNN) for final ranking.

4. Handle Edge Cases and Constraints

Address cold-start (new users/places) with heuristics or content-based methods, balance exploration vs exploitation (e.g., epsilon-greedy or Thompson sampling), filter spam using rules and ML, and ensure privacy (differential privacy, on-device processing) and latency (<100ms).

5. A/B Testing and Iteration

Design an A/B test with control and treatment groups, define primary and guardrail metrics, ensure sufficient power, and plan for long-term holdouts to measure sustained impact.

Key Points to Mention

Two-stage ranking architecture (candidate generation + ranking) for scalability and latency.
Cold-start strategies: using content-based features, user demographics, and contextual signals.
Exploration vs exploitation: multi-armed bandits or reinforcement learning to balance short-term and long-term gains.
Spam filtering: combining rule-based filters, anomaly detection, and user feedback loops.
Privacy: on-device processing, differential privacy, and anonymization of location data.
A/B testing: proper randomization, metric selection, and guardrail metrics to avoid negative side effects.

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