A news feed shows users the latest posts from people they follow. Think Twitter, Instagram, Facebook.

Functional:

• User can create a post (text, image, video)
• User sees a feed of posts from people they follow
• Feed is sorted by recency/relevance
• User can like, comment, share posts

Non-functional:

• Feed loads in <500ms
• Support 100M DAU
• Handle write spikes (celebrity posts)

POST /post
{ "text": "Hello world!", "media_ids": ["img_123"] }
→ { "post_id": "post_456", "created_at": "..." }

GET /feed?limit=20
→ { "posts": [ { "post_id": "...", "author": "...", "text": "..." }, ... ] }

flowchart LR
    Client["📱 Client"] --> LB["Load Balancer"]
    LB --> App["Web Server"]
    App --> PostDB[("Post DB<br/>PostgreSQL")]
    App --> FeedCache[("Feed Cache<br/>Redis")]
    App --> Fanout["Fan-Out Service"]
    Fanout --> Queue["Message Queue"]
    Queue --> Workers["Feed Workers"]
    Workers --> FeedCache

    style Client fill:#7c3aed,color:#fff
    style LB fill:#4f46e5,color:#fff
    style App fill:#6366f1,color:#fff
    style FeedCache fill:#8b5cf6,color:#fff
    style Fanout fill:#059669,color:#fff
    style Workers fill:#059669,color:#fff

Two approaches for building feeds:

Our approach: Hybrid

• Regular users (fewer than 5K followers): Push — pre-compute feed on write.
• Celebrities (>5K followers): Pull — merge celebrity posts into feed on read.

function fanOut(post, authorId) {
  const followerCount = getFollowerCount(authorId);

  if (followerCount < 5000) {
    // Push: add post to all followers' feed caches
    const followers = getFollowers(authorId);
    for (const followerId of followers) {
      redis.lpush(`feed:${followerId}`, post.id);
      redis.ltrim(`feed:${followerId}`, 0, 500); // keep latest 500
    }
  } else {
    // Celebrity: store post, merge on read
    redis.lpush(`celebrity_posts:${authorId}`, post.id);
  }
}

After fetching candidate posts, rank them:

score = recency_score × 0.5 + affinity_score × 0.3 + engagement_score × 0.2

• Recency: time since post (newer = higher)
• Affinity: how often user interacts with this author
• Engagement: likes + comments + shares on the post

Q: What happens to a regular user who crosses the 5,000-follower threshold overnight? The fan-out strategy needs to flip from push to pull for that author without a service interruption — on crossing the threshold, stop pushing new posts to follower feed caches and start writing to celebrity_posts, while leaving already-pushed posts in followers’ caches to age out naturally (they’re bounded by the ltrim to 500 anyway).

Q: How do you backfill a brand-new user’s feed when they haven’t followed anyone yet? There’s no push history to pull from, so cold-start with a pull-based query against a “trending” or “popular in your region/interests” set (onboarding-suggested accounts, globally popular posts) until they build a follow graph, then transition to the normal push/pull hybrid once they have follows.

Q: The ranking formula weighs recency, affinity, and engagement — how do you avoid engagement-optimization turning the feed into all viral content and no recent updates from close friends? Cap the engagement term’s influence (e.g., diminishing returns via log(likes) instead of raw counts) and consider serving distinct candidate pools — one recency-heavy for close friends/affinity, one engagement-heavy for discovery — then interleave them rather than letting one global score dominate.

Q: If a celebrity’s post is deleted after millions of pull-merge reads have already served it, how do you avoid stale cached copies? Because celebrity posts are merged on read rather than pushed, a delete just needs to remove/tombstone the post_id from celebrity_posts and post DB — no fan-out cache invalidation needed. For regular users’ push-fanned posts already sitting in follower feed caches, you need an async cleanup job or a soft-delete flag checked at render time.

Q: How do you keep a user’s feed scroll position consistent if they check the feed on phone, then desktop, then phone again? Paginate with a stable cursor (e.g., timestamp or post_id boundary) rather than offset-based pagination, and persist last-seen cursor per user server-side so switching devices resumes from the same point instead of an offset that shifts as new posts arrive.

• News feed = blend of posts from people you follow, sorted by relevance.
• Push fan-out is fast for reads but expensive for celebrity posts.
• Hybrid approach: push for regular users, pull + merge for celebrities.