Real-time personalization shelf

Design a real-time personalization service for a streaming-media app that re-ranks the home shelves based on what the user did in the last few minutes (played, skipped, searched), updating within 2 seconds of an action for 30M concurrent users. The personalized response must return in under 100ms p95. Cover how you capture and process in-session signals, the serving path, and how you reconcile fast session signals with the slower batch user profile.