Design a service where users upload photos, follow others, and browse a feed of photos with likes and comments — optimised for fast image delivery worldwide.

• 100M uploads/day × 2 MB original + ~1 MB of derived sizes = 300 TB/day → ~100 PB/year. Object storage with lifecycle tiering.
• Feed reads: 500M × 10 = 5B/day ≈ 60k/s, peak 200k/s.
• Metadata per photo ~1 KB → 100 GB/day; sharded DB by user.
• CDN egress dominates: 5B feed loads × 10 images × 200 KB = 10 PB/day.

POST /v1/media/upload/init   → presigned URL(s)
POST /v1/posts               { mediaIds[], caption, tags[] }
GET  /v1/feed?cursor=
GET  /v1/users/{id}/posts?cursor=
POST /v1/posts/{id}/like     /   POST /v1/posts/{id}/comments
GET  /v1/tags/{tag}/posts

posts     (post_id, user_id, caption, created_at)                 -- sharded by user_id
media     (media_id, post_id, original_key, variants:{thumb, medium, large}, width, height)
follows   (follower_id, followee_id)
likes     (post_id, user_id, ts)  PK(post_id, user_id)
comments  (post_id, comment_id, user_id, text, ts)
feed      Redis LIST per user (post ids)
counters  Redis: post:{id}:likes
Object storage keys: media/{media_id}/{variant}.jpg

1. Upload: client uploads the original directly to object storage via presigned URL; the media service enqueues a processing job.
2. Image pipeline: workers generate variants (thumbnail 150, feed 640, full 1080), strip EXIF (privacy), run moderation/NSFW models, store variants, update media.
3. Post service: creates the post once media is ready (or marks processing and lets the feed skip it); emits post.created.
4. Feed: same hybrid fan-out as a news feed — push post ids to followers' Redis lists, read-time merge for large accounts; hydrate metadata; image URLs point to the CDN.
5. CDN: all variants served from a global CDN with long cache TTLs; image URLs are immutable (media id + variant), so no invalidation is needed.
6. Engagement: likes/comments write to sharded tables + Redis counters; notification events published.

Immutable image URLs are the simplest CDN strategy: a new upload = new id; edits create new media. Cache TTL can be a year.

Variant generation on upload (eager) vs on first request (lazy): eager for the 3 standard sizes; lazy resize-on-the-fly (with caching) for unusual sizes.

Storage cost: originals to cold storage after 30 days; keep variants hot.

Privacy: strip location EXIF before storing variants; originals access-controlled.

• Processing worker fails: job retried; post stays processing; feed skips it until ready.
• CDN outage: multi-CDN or direct-to-origin fallback with rate limiting.
• Object storage regional outage: cross-region replication for variants; uploads fail over to another region.
• Feed cache loss: rebuild lazily from follows + posts DB.