YouTube's watch page shows a small rail of recommended videos next to whatever you are watching. That short list decides whether you keep watching or jump away. We are designing the system that produces it.
The product goal is not "predict the next click." It is to increase high-quality watch time across a session so creators reach their audience and the platform earns trust. Pure click-through rewards clickbait, raw watch time rewards bloated content, and pure recency collapses the rail into a trending feed that starves long-tail creators. We therefore define a composite engagement-and-satisfaction signal before choosing a model.
The primary surface is the up-next rail, which renders five ranked videos and feeds autoplay. Other surfaces (home feed, search, subscriptions) are out of scope.
The ML decision boundary
Given a user, currently playing video, and session context, we predict a personalized engagement-and-satisfaction score for each candidate video, then return a ranked slate of five. The decision consumer is the watch-page client, which uses position one as the autoplay target. The constraints are concrete: 99th-percentile latency around 250 ms, a catalog of about one billion videos, and traffic from about one billion daily active users.
The shape of the system follows from those numbers. On the read path, a watch-page request fans out to a retrieval stage that produces a few thousand candidates, a ranker that scores them, and a slate-level re-ranker that returns five with diversity and safety applied. Behind that, an offline and nearline loop logs engagement events, retrains models, and refreshes the retrieval index. The following sections develop each part of this architecture.