Twitch is a social video platform with 15 million+ daily viewers. This talk explores difficulties that live content introduces to recommendations, and the recommender built to personalize many products at Twitch, including architecture decisions and learnings around offline metrics.