System Design

Recommendation System

Design a personalized recommendation system like Netflix's.

Functional requirements

• Serve a personalized, ranked list of ~50 titles for a user's homepage on request.
• Incorporate user behavior signals (watches, completions, ratings, browse/click events) into future recommendations.
• Reflect recent user activity in recommendations within minutes (e.g., a just-finished show influences the next homepage load).
• Support new titles being recommendable within hours of catalog ingestion, and provide reasonable recommendations for brand-new users.
• Allow filtering by business rules at serve time (regional licensing, maturity ratings, already-watched exclusions).

Non-functional requirements

• Recommendation serving must return within a p99 of 200ms end-to-end (it blocks homepage render).
• Serving path prioritizes availability over freshness: always return some recommendation list, degrading to less-personalized results rather than erroring.
• Behavioral event ingestion must be durable: losing interaction data degrades model quality; at-least-once delivery is acceptable.
• Model/feature updates need only eventual consistency; users may briefly see slightly stale personalization, but never a broken or empty homepage.
• User viewing history is sensitive: access-controlled, encrypted at rest, and deletable on user request (compliance).

Scaling & constraints

• 300M MAU, 120M DAU; average user loads the homepage 4 times per day, with a 3x global peak-to-average factor.
• Catalog of ~20K titles, growing ~5% per year; each title has ~2KB of metadata plus learned embeddings.
• Each active user generates ~200 behavioral events per day; each event is ~500 bytes.
• Read:write ratio at the serving layer is roughly 1:50 in favor of writes (events vastly outnumber homepage requests, but reads are latency-critical).
• Raw event data retained for 2 years for model training; user-facing history retained indefinitely until deletion request.
• Per-user feature vectors are ~5KB; per-title feature vectors are ~10KB.

Out of scope

• Video encoding, storage, and playback/CDN delivery.
• The details of ML model architectures and training algorithms (treat models as black boxes with defined inputs/outputs).
• Search and explicit browse/catalog navigation.
• Billing, subscriptions, and account management.