Problem Framing and the Fan-Out Architecture

Framing the Problem

A social feed ranking problem differs from content recommendation in three important ways:

1. Network constraint: candidates come primarily from accounts the user follows, not the entire content catalog
2. Time-sensitivity: posts about live events, news, and conversations are highly time-sensitive; a 48-hour-old post about a basketball game in progress is worthless
3. User-generated content: quality varies enormously (spam to high-value expert posts) and there's no editorial curation layer

Scale assumptions (state these in the interview):

• 300M active users
• Average 500 follows per user
• Average 5 posts/day per active user (300M × 5 = 1.5B posts/day)
• 50K feed requests/second at peak
• Latency SLA: render first screen within 100ms

The Fan-Out Problem

When user A (with 10 million followers) posts, 10 million users need to potentially see this post in their feed. This creates a scalability challenge:

Option A — Fan-out on write (push model): When a post is created, immediately write a reference to it in each follower's feed cache.

User posts → Fan-out service → Write to Redis feed cache for each follower
10M followers × 1 post = 10M writes immediately
Feed read: just read user's pre-assembled cache (fast)
Problem: a celebrity posting generates 10M simultaneous Redis writes

Option B — Fan-out on read (pull model): At feed request time, fetch posts from all accounts the user follows.

User requests feed → Fetch timeline for each of 500 follows → Merge → Rank
Problem: merging 500 timelines at query time is slow (500 × index lookups)

Production solution: Hybrid fan-out:

• Regular users (< 100K followers): fan-out on write → pre-built feed cache
• Celebrity accounts (> 100K followers): fan-out on read at query time
• Threshold: determined by infrastructure cost; typically 10K-1M followers

Post created by regular user (1K followers):
  → Write to 1K follower caches immediately (manageable)

Post created by celebrity (10M followers):
  → Just store the post in a global index
  → At feed request time: check if user follows any celeb accounts → pull their posts
  → Merge with pre-assembled cache → rank all together

Candidate sources beyond follows:

Many platforms have expanded beyond 'only show posts from follows':

• Trending posts in topics the user engages with (two-tower retrieval)
• Posts liked/shared by people the user follows (second-degree signal)
• Recommended accounts ('you might also like') content

Mixing network + algorithmic sources: collect candidates from all sources, deduplicate, then rank them together with a unified scoring model.