You already know: multi-task ranking (the video feed case), graph-based sources (the friend suggestions case), long-term holdouts, and from the ad case that position and exposure bias logged data. This case adds a feed built from people's own connections: posts must be gathered from friends, groups and pages efficiently, ranking decides which one in ten posts people see, comment-heavy content can amplify conflict, and policy decisions (removal, demotion) must be applied consistently and visibly. The design questions: how to build each person's inventory, how to define value, where integrity lives, and how to give people control.
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: "pull posts from friends and rank them by predicted engagement". A staff candidate first asks where posts come from (friends, groups, pages, very large accounts), what value means (engagement, or what people say was worth their time), which integrity rules apply (misinformation, engagement bait, borderline content), and what control people have. Those answers produce hybrid fan-out, a multi-task value model checked by surveys, a separate integrity layer, and a chronological option with explanations.
Assumptions you state: