You already know: two-tower retrieval and approximate nearest-neighbour search (D4), multi-task learning (D5), the recommendation and search case (staged funnels), and from the ad case that position and feedback loops bias logs. This case adds a feed where every swipe is a signal: a catalogue that grows by millions of uploads a day, signals that arrive in seconds, an objective that is easy to over-optimise, and responsibilities to creators and younger users. The design questions: how to retrieve fresh content, what to optimise, how to build a good slate, and how to know you are not trading next month for today.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud.
1. The prompt, and the questions that change the design
A mid-level answer: "two-tower retrieval and a ranker that predicts watch time". A staff candidate first asks what success means (minutes today, or people coming back next month), how big and how fresh the catalogue is, which signals exist, and what responsibilities apply to creators, integrity and younger users. Those answers produce a multi-task ranker with long-term guardrails, retrieval that indexes new uploads within minutes, real-time session features, and policy-aware re-ranking.
Assumptions you state: