Design a recommendation system
Overview
This question evaluates competence in designing large-scale recommendation systems and ML engineering skills such as candidate generation, feature store design (offline/online and point-in-time correctness), real-time signal ingestion, ranking and re-ranking, online exploration, cold-start strategies, feedback-bias mitigation, experimentation, and operational reliability. It is commonly asked to assess the ability to make architecture-level trade-offs for scalability, latency and throughput targets, multi-objective optimization, data governance and outage fallback behavior; it falls under the System Design and Machine Learning domain and requires practical, architecture-level application with conceptual trade-off reasoning rather than low-level implementation detail.
System Design: Large-Scale Home-Feed Recommendation System
Problem
Design a large-scale recommendation system for a consumer app's home feed. Describe the end-to-end architecture and address the following topics:
-
Core architecture
- Candidate generation
- Feature stores (offline/online) and point-in-time correctness
- Real-time signals ingestion
- Ranking and re-ranking
- Online exploration (e.g., multi-armed bandits)
-
Strategies
- Cold start for new users and new items
- Feedback loops and bias mitigation
-
Experimentation and operations
- A/B testing approach
- Latency and throughput targets
- Data privacy and governance
- Fallback behavior during outages
Assumptions (to scope the design)
- Consumer social app with hundreds of millions of MAU, peak 100–200k QPS feed requests.
- Each request returns ~20–40 items, drawn from a few thousand candidates.
- Multi-objective goals: short-term engagement (CTR, dwell, watch time) and long-term value (retention, creator health, diversity, safety).
Solution
Login required