Design Instagram

Instagram lets people upload photos and videos, follow others, and scroll a feed of what they post. The design combines three things: a media pipeline for uploads, object storage plus a CDN for serving images, and a feed built from the follow graph.

Requirements#

Functional

• Upload a photo with a caption.
• Follow and unfollow users.
• View a feed of posts from followed accounts, and any user's profile grid.
• Like and comment on posts.

Non-functional

• Images load fast worldwide; the feed renders in under a second.
• Uploaded media is never lost.
• 500 million daily users; reads far outnumber uploads.

Estimates#

Upload flow#

Sending large files through API servers wastes their bandwidth. Instead, clients upload straight to object storage with a pre-signed URL:

1. The app asks the upload API for a pre-signed URL and uploads the photo directly.
2. Storage emits an event; workers create thumbnail, feed and full-size versions in modern formats like WebP or AVIF.
3. The post is published only once its versions exist.

Serving images#

All media is served through a CDN. URLs include the size and a content hash, such as /p/9f2c.../640.webp, so they can be cached forever. Clients request the size that matches the screen.

Data model#

users   (id, username, bio, avatar_url, created_at)
posts   (id, user_id, caption, media_key, created_at)
follows (follower_id, followee_id, created_at)   -- PK (follower_id, followee_id)
likes   (post_id, user_id, created_at)           -- PK (post_id, user_id)

• Posts are sharded by user_id, so a profile grid is one shard query sorted by time.
• Follows are stored twice, indexed by follower and by followee, to answer both "who do I follow" and "who follows me".
• Like counts are kept as a counter on the post (updated asynchronously) instead of COUNT(*) on every view.

The feed#

Like any social feed, use fan-out on write for most accounts (push post IDs into followers' feed caches) and fan-out on read for celebrities, merged at read time. See the News Feed design for the full trade-off.

Scaling and reliability#

• Durability: object storage replicates across zones; keep originals in a cheaper storage class after 30 days.
• Counters: likes on viral posts can mean thousands of writes per second to one row. Buffer increments in Redis and flush them periodically.
• Read replicas: profile and post metadata are read-heavy and tolerate small delays.
• Search and explore: a separate pipeline indexes posts and hashtags into a search cluster.