Design Instagram Feed

Problem statement

Design a photo-centric social feed: users upload images, follow other users, and consume a continuously refreshed home feed of posts from accounts they follow, enriched with likes, comments, and profile context. The system must support image upload with thumbnail and full-size renditions, personalized or chronological feed updates, social interactions (like, comment, follow/unfollow), and user profiles with bios, highlights, and stories.

The defining characteristic is asymmetry. Public Instagram-scale figures suggest roughly a billion daily users viewing hundreds of billions of post impressions, while only on the order of a hundred million new posts are created per day. Reads outnumber writes by thousands to one on the consumption path, yet every single write must fan out to potentially millions of readers. A food-ordering backend balances request and response; a feed system balances a tiny write stream against an almost unbounded read stream.

Why this problem is distinctive

Three structural forces shape every decision:

1. Fan-out explosion. A post by an account with N followers is logically N feed updates. With average followers in the hundreds and celebrities in the hundreds of millions, pure push and pure pull both fail. The architecture needs a hybrid fan-out strategy.
2. Media is bytes, not rows. Photos and stories are large immutable binary objects with multiple renditions. They need object storage tuned for small-file aggregation, an async transcode pipeline, and an edge CDN that absorbs nearly all delivery traffic.
3. Freshness versus cost. Users expect new posts from close friends within seconds, but serving a globally consistent, freshly ranked feed to hundreds of millions of concurrent users on every open would be economically impossible. Caching tiers and bounded staleness are architecture, not afterthoughts.

The four architectural planes

1. Media plane: resumable upload, validation, transcode into renditions, durable object storage, CDN delivery.
2. Social plane: the follow graph (followers/following adjacency), counts, and profile metadata.
3. Feed plane: fan-out on write for ordinary users, pull-merge for high-follower accounts, ranking, hydration, and pagination.
4. Engagement plane: likes, comments, counters, notifications, and seen-state deduplication.

A strong interview answer keeps these planes decoupled: the media pipeline can lag without blocking feed reads, counters can be approximate without corrupting custody of posts, and the ranking model can be rolled back without touching the social graph.

Public operating baseline versus design assumptions

Public evidence establishes that this category operates at extreme scale: Instagram announced more than 2 billion monthly active users and previously reported hundreds of millions of daily story users; the Facebook Haystack paper reported 20 billion photos stored with roughly 850 TB added per week; the TAO paper reported over 1 billion reads per second on the social graph. These are cited company and paper figures used for context. Every capacity number used for sizing in this answer is an explicit design assumption, stated where introduced.