Design a short-video recommender for short-term interest

Scenario

You are designing a short-video recommendation system (similar to a swipe/feed product). The system must personalize the feed for each user and react quickly to the user’s short-term interests (session intent), not just long-term preferences.

Requirements

1. Primary goal: maximize user satisfaction in the current session by capturing short-term interest shifts (e.g., user starts watching cooking videos after browsing sports).
2. Scale assumptions (choose reasonable numbers and state them): tens of millions of users, millions of videos, high QPS during peak.
3. Latency: feed generation should feel instant (e.g., p95 < 200–300 ms for ranking at request time).
4. Freshness: incorporate new uploads and the user’s latest interactions quickly.
5. Safety & quality: filter harmful/low-quality/spam content; avoid repetitive or overly narrow recommendations.
6. Feedback signals: watch time, completion rate, likes, shares, comments, follows, “not interested,” skips, and dwell time.

Deliverables

• Propose an end-to-end architecture for candidate generation + ranking.
• Explain how you model and serve short-term interest (session-based signals, near-real-time features, or online learning).
• Define core offline/online metrics and experimentation approach.
• Discuss data pipelines, feature stores, model training, and serving.
• Call out key pitfalls (feedback loops, cold start, bias, exploration vs exploitation).