Design a home feed that shows a user recent posts from people/pages they follow, ranked, with infinite scroll, at social-network scale.

• 300M DAU × 10 feed loads/day = 3B loads/day ≈ 35k/s, peak 100k/s.
• Posts: 300M × 0.5/day = 150M posts/day ≈ 1,700/s.
• Fan-out on write: 150M posts × 300 followers = 45B feed-cache inserts/day ≈ 520k/s — feasible in Redis but heavy.
• Feed cache: 300M users × 500 post ids × 8 B = 1.2 TB → sharded Redis cluster.
• Celebrity post (50M followers) → 50M inserts for one post: the reason hybrid fan-out exists.

POST /v1/posts                { text, mediaIds[] }         → { postId }
GET  /v1/feed?cursor=&limit=20 → { items:[{postId, author, text, media, likeCount, ts}], nextCursor }
POST /v1/posts/{id}/like
POST /v1/users/{id}/follow

Cursor = (rank score, post id) so pagination is stable while new posts arrive.

posts      (post_id, author_id, content, media, created_at)          -- sharded by post_id
follows    (follower_id, followee_id)                                  -- graph DB or sharded table, both directions indexed
feed_cache Redis ZSET  feed:{user_id} → {post_id: score}  capped at 500-1000 entries
post_cache Redis HASH  post:{post_id} → content + counters
counters   like_count/comment_count kept in Redis, flushed to DB periodically

1. Post service writes the post, then emits a post.created event.
2. Fan-out service (consumer): fetches the author's follower ids and pushes the post id into each follower's feed_cache ZSET (fan-out on write). For authors above a follower threshold (e.g. 100k), it skips fan-out (fan-out on read).
3. Feed service on read: pulls the user's ZSET (pre-computed), merges in posts from followed celebrities fetched directly (small list), runs the ranker, hydrates post content from post_cache/DB, returns a page.
4. Ranking: features (recency, affinity, engagement counts, content type) → a lightweight model scoring a candidate set of ~500; heavy models run offline to produce per-user affinity scores.
5. Media via CDN; the feed carries URLs only.
6. Inactive users: skip fan-out for users inactive > 30 days; build their feed lazily on next login.

Hybrid fan-out is the core decision: write-fan-out gives O(1) reads but O(followers) writes; read-fan-out the reverse. Split by follower count.

Feed cache sizing: keep only ids (not content) and cap the ZSET; older items fall off and are rebuilt from the DB if a user scrolls far.

Counters: likes are written to Redis with periodic flush; the feed shows slightly stale counts — fine.

Deletion/privacy: on post delete, remove from the post cache and let feeds filter out missing posts at hydration time rather than scanning every feed.

• Fan-out workers lag: feeds go stale; feed service can supplement with a direct query of recent posts from top-N followed accounts.
• Feed cache loss: rebuild lazily from the posts DB on next read (slow first load) — keep the DB indexed on (author_id, created_at).
• Ranker down: fall back to reverse-chronological.
• Hot post (viral): post cache replicated; counters batched to avoid write contention.