News Feed Interview Questions

The hard part is fan-out. Fan-out on write: when I post, I push the post id into every follower's feed cache. Reads are fast, but a celebrity with 50 million followers will kill the write path. Fan-out on read: at read time I pull recent posts from people I follow. Better for huge followings, slower reads. The hybrid I would propose: fan-out on write for normal users, fan-out on read for celebrities. Feed cache stores post ids, not full posts. Ranking can be recency first, then an async ML ranker. I paginate with cursors, not SQL OFFSET. Likes and view counts should not live on the same hot row as the post body.

Fan-out on write: when I post, I push the post id into each follower's timeline cache. Reads are fast. It hurts celebrities with millions of followers. Fan-out on read: I pull recent posts from people I follow at read time. Writes are cheap, reads are heavier. Hybrids push to active users and pull for the rest. I would pick hybrid for a Twitter clone. I would not only draw a queue and stop.

I would start chronological because it is explainable and debuggable. Then I add a ranker: recency, affinity, and predicted click, computed offline and applied at read. I would keep a kill switch back to chronological. I would not claim I will train a huge model in the interview. I would say I need metrics — dwell time, hides, reports — and that ranking without safety will amplify junk.

I would not write 50 million timeline entries inline with the post API. I would enqueue fan-out, shard workers by follower id, and for inactive users skip write and let them pull. The celebrity's post is stored once. I would cache the post body heavily. I would rate-limit notifications. This is why hybrid fan-out exists. I mention backpressure so the post API stays fast.

The feed stores post ids, not full comment threads. Counts can be cached and updated async. Opening a post loads comments from a separate service, paginated. I would not embed all comments in every follower's timeline copy. Likes are a counter plus a user-like relation for 'have I liked this'. Eventual consistency on counts is usually fine. Exact counts on a viral post are expensive and I would say so.