Meta ML System Design: Real-Time Personalized Feed Ranking

Overview

You are asked to design a real-time machine learning ranking system that personalizes the order of articles and comments in Meta's news feed. The system must serve personalized ranked lists per user with sub-100ms latency, incorporate real-time interactions (clicks, likes, dwell time), handle millions of users and content items, and support A/B testing, monitoring, and explainability.

Core requirements and scope

Explain how you would perform candidate generation, feature engineering, and final ranking. Describe offline model training and an online serving stack that ingests streaming signals, updates user state, and recomputes scores or re-ranks candidates quickly. Address filtering (inappropriate or already-seen items), cold-start for new users/content, diversity and fairness constraints, and a feedback loop that logs interactions for continual improvement.

High-level flow / stages

1. Candidate generation (recall) — fast, high-recall retrieval (e.g., embeddings, inverted indices, graph-based or heuristic filters).
2. Feature enrichment — join user profile, session/context signals, content metadata, and recent events from a feature store or stream layer.
3. Scoring & ranking — low-latency model (tree-based, neural pointwise/pairwise, or learned-to-rank) with diversity/fairness post-processing.
4. Logging & feedback — collect impressions, clicks, dwell time for training and online evaluation.
5. Experimentation — support A/B tests, shadow runs, and canary deployments.

Skill signals interviewers look for

You should demonstrate understanding of scalable retrieval, latency/throughput trade-offs, feature stores and real-time feature pipelines, online vs offline learning (including bandits), evaluation metrics (CTR, dwell, novelty, fairness), monitoring/SLAs, and practical deployment strategies (caching, sharding, graceful degradation). Discussability on explainability, cost optimizations, and mitigation for filter bubbles is a plus.