Design Content Recommendation System

Netflix's "Design Content Recommendation System" is a popular system design interview question, often tagged with System Design, ML Systems, Personalization, and Machine Learning. It typically asks candidates to architect a scalable, personalized recommendation engine like Netflix's, handling millions of users and vast content libraries. No single canonical "full problem statement" exists across sources, but it's commonly framed around building a two-stage ML pipeline (candidate generation and ranking) to maximize user engagement via metrics like CTR and watch time.[1][4][6]

Problem Statement

Design a content recommendation system for a streaming platform like Netflix that delivers personalized video suggestions (e.g., movies, TV shows) to users on the homepage, search, and rows. The system must process user interactions in real-time, handle cold starts for new users/items, mitigate popularity bias, and scale to millions of daily active users with billions of content interactions. Key goals: predict ranked lists of relevant items to boost metrics like click-through rate (CTR), watch time, retention, and mean average precision/recall (MAP/MAR).[4][6][1]

Functional Requirements

• Generate top-K (e.g., 20-1000) candidate items per user using collaborative filtering, content-based methods, or hybrid ML models (e.g., matrix factorization, embeddings via VAEs).
• Rank candidates with deep learning models (e.g., RNNs/LSTMs for sequences, CNNs for thumbnails) incorporating user history, metadata, context (time/day), and real-time signals.
• Support diverse canvases: homepage rows, personalized artwork, search autocomplete.
• Offline batch processing for model training; nearline/online for fresh recs.[2][6][1]

Non-Functional Requirements

• Scale: 200M+ users, 100M+ items, 1B+ daily events; low latency (<200ms per request).
• Availability: 99.99% uptime with fault tolerance (e.g., microservices, caching).
• Data: Handle sparse matrices, privacy (anonymized data), global diversity.
• Extensibility: A/B testing, continuous learning, integration with data pipelines (Kafka, Spark).[1][2]

Input/Output Examples

Input: User ID, profile (viewing history: titles watched/rated/duration; searches; demographics), context (device, time), content catalog (metadata: genres, cast, tags).

• Example: User watches sci-fi thrillers late at night → input includes history vector, session timestamp.

Output: Ranked list of 20 personalized recs (e.g., JSON: [{"id": "title123", "score": 0.95, "artwork": "action_poster.jpg"}, ...]).

• Candidate gen: Top 1000 items (e.g., similar users' likes).
• Ranking: Filtered top 20 by predicted engagement score.[6][4]

No explicit constraints like strict input formats are standardized, but common ones include QPS (10K+), storage (PB-scale), and cold-start handling (popularity baselines).<>