Recommendation System

Source: 30-System-Design/Framework & Systems

Prompt

• Design a recommendation system for a large-scale video platform (e.g., home feed).

Requirements

• Functional: personalized recommendations, diversity/freshness, cold start handling.
• Non-functional: low latency, high throughput, explainability, privacy.
• Constraints: strict latency budget, heavy-tailed catalogs, fast content churn.

Success Metrics

• Offline: NDCG, MAP, Recall@K.
• Online: CTR, watch time, retention, satisfaction.
• Guardrails: skips, hides, complaints, diversity metrics.

Data

• Sources: user events, item metadata, embeddings, context signals.
• Labeling: implicit feedback (clicks, watch time), explicit feedback (likes).
• Negative signals: skips, short dwell, “not interested”.

Modeling

• Candidate generation: two-tower / ANN retrieval.
• Ranking: gradient boosting or deep ranker with context features.
• Re-ranking: diversity/novelty constraints.
• Cold start: content-based features, popularity priors, exploration.

Serving

• Retrieval: ANN index, caching by segment.
• Latency budget: retrieval + ranking + re-ranking.
• Personalization: real-time features, session intent.
• Fallbacks: popular/trending when features are missing.

Evaluation & Monitoring

• Offline eval: train/val splits, leakage checks.
• Online eval: A/B tests with guardrail metrics.
• Drift/abuse: data drift, feedback loops.
• Long-term health: measure satisfaction beyond short-term CTR.

Risks & Tradeoffs

• Exploration vs exploitation.
• Popularity bias and filter bubbles.
• Novelty vs relevance in short sessions.

Notes