Design Instagram: media upload, CDN, and feed ranking
Redis Cache: Instagram's O(1) Feed Reads
The Feed Service writes pre-assembled timelines to Cache. Instagram reportedly stored 300M photos per day at peak — Redis timelines meant feed reads were entirely from memory, enabling 60M DAU without proportional DB scaling.
Request Flow
Feed Service writes to Redis
On fanout, the Feed Service pushes each new post's CDN URL to the front of every follower's timeline list in Cache. Each follower's timeline is a Redis sorted set or list capped at ~1000 entries.
LPUSH timeline:user456 '{"url":"cdn.example.com/photo.jpg","ts":1700000000}'
LTRIM timeline:user456 0 999 // keep latest 1000Viewer fetches from Redis
GET /feed reads the first 20 entries from the viewer's timeline list in Cache. P99 latency < 2ms. No database involved.
GET /feed?page=0 → LRANGE timeline:user456 0 19
Cache miss — fall back to DB
If the user's timeline isn't in Cache (cold start, eviction), the Feed Service rebuilds it from the Database and repopulates Redis.
if (!cache.exists("timeline:user456")) {
const posts = db.getFollowedPosts(user, limit=1000);
cache.lpush("timeline:user456", ...posts);
}Why We Need This
Without Redis, every feed page requires a multi-table join: follows → posts → photos → users. With 500M users this is untenable. Redis reduces the fan-in problem to a single key lookup.
Instagram reportedly stored 300M photos per day at peak. Redis timelines meant feed reads were served entirely from memory — enabling 60M daily active users at the time without proportional DB scaling.
Key Insight
This is the complete Instagram read path: Viewer → Redis (timeline list) → CDN (image bytes). Two hops, both from cache. The database and S3 origin are only touched for writes and cold starts.
Step 7 of 7
