Video feed ranking serving

Design the online serving stack for a video recommendation feed that must return a personalized ranked list of 20 items in under 100ms p99 for 200M daily users, drawing from a catalog of 500M videos. Candidate generation must surface fresh content (uploaded in the last hour) alongside long-tail personalization. Walk through the two-stage retrieval+ranking architecture, how you keep tail latency bounded, and how a freshly uploaded video gets a chance to be served before it has any engagement signal.