Design Instagram

Instagram is a photo-sharing social platform where users publish images, follow other accounts, and browse a home feed composed of recent photos from the accounts they follow. On the surface it appears as a straightforward photo grid, but underneath the system must ingest and deliver images while constructing feeds from the same follow graph and durable post store.

Requirements

Instagram presents a feed problem with a notable twist. Each photo post initiates an asynchronous media processing lifecycle that must complete before a feed can display the image. The design must keep home-feed reads inexpensive, wait for images to become renderable, and avoid pushing a celebrity post into millions of follower feeds synchronously.

Functional requirements

• Users can follow and unfollow other accounts; relationship changes affect which authors appear in the reader's home feed.
• An author can upload a photo with a caption. The system durably accepts the post after persisting metadata and kicking off the media processing pipeline.
• A reader can fetch the newest page of their home feed and paginate older items using an opaque cursor.
• The home feed displays recent photo posts from followed accounts with eventual consistency, including a scalable path for celebrity authors whose follower count makes full synchronous fanout too expensive.
• Uploaded images are processed into multiple resolutions (thumbnail, standard, full) and served through a CDN for low-latency delivery.

Out of scope (below the line)

• Video, Stories, Reels, and live streaming
• Full ML ranking, recommendation engine, and explore/discover algorithms
• Comments, likes, saves, reposts, and direct messages
• Hashtag and location-based discovery (mentioned in Other Considerations)
• Ads auctioning and sponsored post insertion
• Private-account approval workflows and rich privacy matrices
• Real-time chat and push notification delivery
• Multi-region active-active conflict resolution

Non-functional requirements

• Target 99.9% availability for photo upload and home-feed reads in a healthy region.
• Target p95 < 300 ms for upload acknowledgement (ACK) after durable metadata persist, before full processing completes.
• Target p95 < 300 ms for the first home-feed page from a warm regional stack.
• Assume 50M daily active users (DAU), 300k/s peak home-feed reads, 5k/s peak photo uploads, 15B+ follow edges, and celebrity accounts with 10M+ followers.
• Home-feed reads may lag the latest post by a few seconds, but the author's own profile and authoritative post record stay readable immediately after upload ACK.
• Accepted post metadata, follow changes, and original uploaded media must survive process or node failure. Materialized feed rows and processed image variants are rebuildable.
• The hot post path avoids synchronous O(follower_count) work for large authors.

Back-of-the-envelope estimation

At 5k uploads per second, the media pipeline processes roughly 15k renditions per second. Each original photo is 2-5 MB, so raw upload ingress alone is 10-25 GB/s into object storage. These numbers justify why image bytes must never flow through the application server and why presigned URLs are the standard upload pattern.