MLE System Design: Design a Real-Time Personalized Recommendation System

Question

## Round 1 - System Design

Design a real-time personalized recommendation system for a video streaming platform. The system should serve the right video recommendations to each user within 100ms.

**Clarifying questions you should ask:**
- How many daily active users? (assume 50M)
- Is latency measured at p50 or p99?
- Cold start: how do we handle new users with no history?
- Can we pre-compute recommendations or must they be online?

## Architecture Components

**Candidate Generation:**
- Two-tower model: user embedding + item embedding, trained on watch history.
- ANN index (e.g., FAISS or ScaNN) for fast top-K retrieval at query time.

**Ranking:**
- Lightweight pointwise ranker (logistic regression or shallow NN) with user context features.
- Feature store (Redis) for low-latency feature lookup.

**Serving:**
- Pre-compute candidate sets every 30 min; refresh user features in real time.
- Fallback to popularity-based recommendations for cold-start users.

**Feedback Loop:**
- Stream click/watch events via Kafka; retrain candidate model daily, ranker hourly.

## Follow-ups

1. How do you detect and mitigate popularity bias in the candidate model?
2. What metrics would you use to evaluate the system offline vs. online (A/B test)?
3. How do you handle a sudden spike in new content (breaking news, viral video)?
4. How does your architecture change if recommendations must be explainable to users?

MLE System Design: Design a Real-Time Personalized Recommendation System

Question Details

Full Details

Topics

More from Airbnb