Online learning for feed ranking

Design an online-learning system for a news/feed product where user preferences shift within hours (breaking news, trends). The ranking model must incorporate fresh click/engagement feedback on a minutes-scale, not the usual daily batch retrain, while serving ~60k ranking requests/sec at 30ms p99. You need the freshness of near-real-time updates but must guard against the model overreacting to a transient spike or to bot/abuse traffic, and you must be able to roll back a bad online update quickly.