Design a Video Recommendation Engine (Short Form)

To design a TikTok-style recommendation engine at Google scale, I would prioritize a two-stage funnel architecture to handle the massive corpus while maintaining sub-200ms latency. First, in the Candidate Generation phase, we cannot scan billions of videos. Instead, I'd use an ANN index, likely HNSW or ScaNN, built on user embeddings derived from historical interactions. This reduces the pool from billions to a few hundred relevant videos per request. Second, for real-time relevance, we need immediate feedback loops. As a user swipes, we extract features like dwell time and interaction type, updating their session vector in a low-latency store like Redis. Finally, the Ranking stage uses a Deep Neural Network to score these hundreds of candidates, incorporating rich features like video metadata, creator popularity, and the user's current context. To prevent echo chambers, I'd inject diversity by ensuring a percentage of recommendations come from new creators or different topics. For scalability, the system must be stateless where possible, relying on distributed computing for training and sharded storage for features, ensuring consistency across regions.