Sequential recommendations at scale

Design real-time sequential recommendations where the next-item suggestion depends on the order and timing of what the user did in the last few minutes (watched A then B then skipped C to predict D), not just a static long-term profile. You must update the recommendation within ~1-2s of each user action for 50M concurrent sessions, feeding a sequence model that consumes the recent action sequence. How do you capture, store, and serve the live action sequence into the model without a slow per-request database read of the user's whole history?