Design the Recommendation System

Problem

Design Netflix's personalized recommendation system powering the home page rows.

Requirements

Functional:

• Per-user ranked rows ('Because you watched...')
• Update after each viewing session
• Cold-start for new users and new titles
• A/B test new ranking models

Non-functional:

• 200M+ users, ~17K titles
• Serve home page in < 100ms
• Models retrained continuously

Discussion points

1. Candidate generation -> ranking -> filtering
2. Offline training vs online feature store
3. Cold-start (popularity priors, onboarding taste)
4. Feedback loop and A/B testing framework
5. Fault isolation — fall back to non-personalized rows