Online feed ranking updates

Design an online (continual) learning system for a feed-ranking model that must incorporate fresh engagement signal within minutes, a video going viral, a breaking news event, rather than waiting for a daily batch retrain. The risk: continually updating a live model on streaming feedback can let a feedback loop, a data spike, or a bad batch corrupt the model in production. Walk through the streaming-update architecture, how you push updated weights to the serving fleet safely, and how you contain or roll back a bad update.