Personalized video feed ranking

Design the recommendation/serving system that builds a personalized home feed of videos for 1B users, choosing from a corpus of billions of items, with a strict <100ms p99 to assemble the ranked feed. Fresh content must surface quickly, watch behavior must feed back into recommendations within minutes, and the system can't re-score billions of items per request. Design the candidate-generation + ranking funnel, the feature/serving infrastructure, and the freshness path.