TikTok's For You feed: learning from every swipe within minutes

How a short-video For You feed can learn from every swipe: more than a billion people a month (TikTok, 2021) and, on stated assumptions, about a million impressions a second at peak, each one a label; watch time judged against videos of the same length, with quick skips as strong negatives; the exact features served logged and joined with labels in a streaming joiner, with late labels waiting on disk and a log odds correction for sampled negatives, as in ByteDance's Monolith paper; collisionless embedding tables with admission and expiry; touched rows synced to serving every minute and dense weights daily; new videos given staged exposure on evidence, not followers; and a re-rank that keeps TikTok's published diversity rules. An illustrative design: TikTok hasn't published its architecture, and ByteDance's papers don't name TikTok.