Design Instagram - Photo Sharing Platform

Instagram includes several products. I would ask whether this is photo publishing and the following feed, or Reels, Stories, Explore, messaging or live video. I will begin with publishing and a following feed.

Feed order, audience privacy and follower skew are the most consequential questions; they matter more than naming a cache early.

I will support durable publication, follow/unfollow, paginated reads, likes and comments. Reverse chronological order is the baseline. Ranking later reorders eligible candidates; it must not change authorization rules.

A post becomes visible only when required approved renditions are ready and metadata commits. Feed entries are references, not an authority for content visibility. A stale feed cache cannot overrule deletion or privacy.

Assume 10M DAU, 20 feed pages per user/day and 20 posts/page. Assume 1M new photo posts/day and 200 eligible followers per ordinary author. I explicitly separate the celebrity tail; averages hide its operational impact.

I begin with indexed relational metadata for users, follows, posts, likes and comments, plus private object storage. The feed index contains bounded post references and can be rebuilt.

A like is a unique relationship, not an unprotected increment. Retried operations must not inflate counts; displayed counters can be derived and reconciled.

Clients upload through scoped capabilities, but a client saying “complete” does not publish a post. A server pipeline validates content and produces required renditions.

Pagination uses an opaque stable ordering cursor. Offsets drift when new posts arrive. A ranked feed needs an explicit ranking-session/snapshot policy rather than re-sorting every page independently.

A simple correct feed pulls recent posts from followed authors and merges by publication time. Publishing is cheap, but read work grows with follow count. After measuring that bottleneck I add a derived inbox for ordinary authors.

Both versions share the same authoritative published-post/visibility rules. I do not begin with a large ranking platform before establishing reliable publication and candidate delivery.

Transfer completion, processing completion and visible publication are distinct states. I make each retryable and protect publication against stale jobs. A delete racing a worker must never make content reappear.

For a hybrid feed I load ordinary-author inbox references, fetch bounded recent windows for followed popular authors, merge/deduplicate, check visibility and hydrate a page. Ranking is optional and follows candidate eligibility.

The celebrity threshold is a measured cost decision based on active followers, posting/read rates and lag. One universal follower count is not a defensible threshold.

The feed derivation is disposable and observable. I can degrade ranking or counters, but I cannot bypass a required authorization check to fill the page.

I measure publish success/age, fan-out lag by author class, first-page latency, media failures, duplicate entries, privacy denials and ranking fallback rate. Aggregate averages can hide a broken celebrity workflow.

I covered publish, follow, feed, likes and comments. Publication is authoritative; the derived feed can lag or be rebuilt. Hybrid fan-out addresses skew that average QPS hides.

Senior depth includes retries, privacy, deletion and end-to-end read/write flow. Staff depth adds graph ownership, hot-author isolation, regional behavior, rebuild strategy and measured ranking quality within latency/safety budgets.