Design a short-video recommender system

## ML System Design — Short-video recommendation Design an end-to-end recommendation system for a short-video feed (TikTok/Reels-style). Walk through the full pipeline: ### 1) Objective and constraints - Define the product goal (e.g., maximize long-term user value). - State constraints: latency budget, freshness, exploration, safety/compliance. ### 2) Labels and training data - What labels would you use (watch time, completion, likes, shares, follows, skips)? - How do you handle **delayed feedback**, **position bias**, and **negative sampling**? - How do you construct training examples (impressions, sessions, user-video pairs)? ### 3) Features - User features (history, embeddings, time-of-day, locale) - Item/video features (content, creator, freshness) - Context features (device, network, entry point) - Cross features / interactions ### 4) “Three-stage” recommendation architecture Describe a standard 3-stage system: 1. **Candidate generation (retrieval)** 2. **Ranking** 3. **Re-ranking / post-processing** (diversity, constraints, business rules) For each stage: - Model family choices (e.g., two-tower, GBDT, deep ranker) - Serving architecture and latency considerations - How you ensure freshness and handle cold start ### 5) Offline evaluation and model selection - Which offline metrics would you use and why? - How do you validate correlation with online metrics? - How do you avoid offline-to-online mismatch? ### 6) Online experimentation - A/B test design, primary/guardrail metrics, ramp strategy - Debugging when offline improves but online regresses ### 7) Safety, robustness, and monitoring - How to prevent harmful content amplification - Monitoring (data drift, performance drift, quality regressions) Be specific and justify trade-offs.