Design an ML-powered system to rank comments under posts in a news feed product. ### Requirements - For each feed item (post/story), users can open the comment section and should see an ordered list of comments. - Ranking should be **personalized** (depends on the viewer) and should balance: - relevance/engagement (e.g., likes, replies, dwell time), - quality (avoid spam/toxicity/low-effort), - freshness (new comments can surface), - author/viewer relationships. ### Constraints / considerations - High QPS and strict latency (assume p95 end-to-end budget ~100–200 ms for the comment list). - Support new comments arriving continuously. - Handle abuse (spam, brigading, adversarial behavior), deleted comments, and block lists. - You should describe: data, labels, modeling approach, online serving architecture, evaluation (offline + A/B), and monitoring/guardrails.