System Design Case Studies

Designing a News Feed

A news feed is fundamentally a fan-out problem — when someone posts, do you push it to every follower's feed immediately, or make each follower pull and merge posts on read? The answer usually depends on how famous the poster is.

Fan-out-on-write (push): when a user posts, immediately write it into every follower's precomputed feed — reads become O(1), but a celebrity with millions of followers turns one post into millions of writes
Fan-out-on-read (pull): a feed is assembled at read time by merging posts from everyone a user follows — no write amplification, but every read does real work and scales badly with follow count
Hybrid approach: push for ordinary users, pull-and-merge for high-fan-out accounts — the common real-world answer, at the cost of two code paths
Ranking, rather than plain reverse-chronological order, requires scoring candidate posts at read time — pushing at least part of the work back toward fan-out-on-read regardless of delivery strategy
Feed storage is typically a precomputed list of post IDs per user in a fast store, hydrated with full post content from a separate content store at read time
Staleness is usually acceptable — a feed a few seconds behind is fine, one of the clearer cases where eventual consistency is not just tolerable but the entire point

Fan-out-on-write path

Hybrid read path

Push vs pull vs hybrid
	Push (fan-out-on-write)	Pull (fan-out-on-read)	Hybrid
Write cost	high — one write per follower	none	high only for ordinary accounts
Read cost	low — precomputed	high — merge at read time	low, plus a small merge for celebrities
Celebrity problem	a write storm per post	not an issue	solved by routing around push

Assembling a hybrid feed

List<Post> getFeed(String userId) {
    List<String> precomputed = feedCache.get(userId);              // from push fan-out
    List<Post> celebrityPosts = celebrityFollows(userId).stream()
        .flatMap(c -> postStore.recentByAuthor(c).stream())        // merge-on-read
        .toList();
    return merge(hydrate(precomputed), celebrityPosts);
}