Clarify first: user scale, refresh style (pull-to-refresh vs. autoplay), primary objective (watch time, engagement, or retention). Different objectives change everything downstream.
Multi-channel retrieval: Collaborative filtering (Swing / item-to-item), two-tower embedding retrieval, hot/followed/geo channels. Narrow millions down to a few thousand.
Ranking: Light pre-ranking cuts to a few hundred; heavy ranking uses multi-objective models (completion, likes, follows, watch time fused together). Cover the feature system (user profile, behavior sequence, item stats, cross features) and how samples are built.
Cold start: New users get interest tags + hot pool + exploration traffic. New videos get small test traffic, then scale by real-time feedback.
Feedback loop: Impressions without play and quick swipes are signals too. Near-real-time features and frequent model updates keep the hot pool fresh. Prevent filter bubbles with diversity re-ranking and shuffling.
Get asked follow-ups live, then receive a scored report — like a real MLE interview loop.
Start AI mock interview