Om's Brain

3. FB News Feed

4 min read

1️⃣ High Level Design

News-Feed.excalidraw

Client
   |
API Gateway
   |
+-------------------+
| Post Service      |
+-------------------+
   |
 Post Table

Post Flow

User creates post
    ↓
Post Service
    ↓
Post Table

Follow Flow

User follows another user
      ↓
Follow Service
      ↓
Follow Table

Feed Flow (Naive Approach)

User requests feed
        ↓
Feed Service
        ↓
Get followed users
        ↓
Fetch posts of each user
        ↓
Merge + Sort
        ↓
Return Feed

6️⃣ Pagination Design

Use cursor-based pagination. Cursor = timestamp of oldest item seen.

Example:

Feed Page 1
 
Post A (10:00)
Post B (09:59)
Post C (09:58)
 
Cursor = 09:58

Next request:

GET /feed?cursor=09:58

Returns older posts.

Benefits:

  • No duplicate records
  • Infinite scrolling
  • Better than offset pagination

7️⃣ Problem With Naive Feed Generation

Suppose:

User follows 10,000 people

Feed Service must:

Read 10,000 timelines
Fetch millions of posts
Merge all results
Sort them

Problems:

  • Huge network calls
  • High latency
  • Expensive sorting
  • Doesn’t scale

8️⃣ Fanout on Write (Precomputed Feed)

Idea

Instead of computing feed during reads:

Compute feed when post is created

Create a new table:

PrecomputedFeed

Schema

UserIdPostIds
UserA[P1,P2,P3…]

Store latest ~200 posts per user.

Post Creation Flow

User creates post
      ↓
Post Service
      ↓
Post Table
      ↓
Find Followers
      ↓
Insert PostId into follower feeds

Feed Read Flow

Feed Service
      ↓
PrecomputedFeed Table
      ↓
Return Feed

Benefits:

✅ Extremely fast reads

✅ Simple feed retrieval

9️⃣ Problem: Celebrity Accounts

Suppose:

Justin Bieber
100M followers

One post requires:

100M feed writes

This causes:

  • Massive write amplification
  • Huge spikes
  • Long post creation delays

🔟 Async Fanout

Never perform fanout synchronously.

Use Queue + Workers.

Post Service
      ↓
Queue
      ↓
Feed Workers
      ↓
PrecomputedFeed

Workflow

Post Created
      ↓
Publish Event
      ↓
Worker Consumes Event
      ↓
Fetch Followers
      ↓
Write Feed Entries

Large Fanout

Split work.

100M followers

becomes

1000 jobs
   ×
100K followers each

Distributed across worker fleet.

1️⃣1️⃣ Hybrid Fanout Strategy

Observation:

Most users → few followers
Few users → millions of followers

Use:

Fanout on Write

For normal users.

Write feed during post creation

Fanout on Read

For celebrities.

Generate feed dynamically

Follow Table Addition

isPrecomputed
FollowingFollowedisPrecomputed
ABtrue
ACelebrityfalse

Feed Generation

Feed Request
      ↓
Read Precomputed Feed
      ↓
Read Celebrity Posts
      ↓
Merge Results
      ↓
Return Feed

Best of both worlds.

1️⃣2️⃣ Hot Key Problem

Scenario

A celebrity creates a post.

Millions of users request:

Post #123

All requests hit same partition.

DynamoDB Partition X

Result:

Hot Partition
Throttling
Failures

1️⃣3️⃣ Solution: Distributed Cache

Add cache before database.

Feed Service
      ↓
Redis Cache
      ↓
DynamoDB

Benefits:

  • Most reads served from memory
  • Reduces DB traffic drastically

1️⃣4️⃣ Cache Hot Key Problem

Even cache can become hot.

All requests
      ↓
Same Cache Key

Single Redis node overloaded.

1️⃣5️⃣ Solution: Cache Replication

Instead of sharding:

Redis-1
Redis-2
Redis-3
Redis-4

Feed Service randomly picks one.

Request
   ↓
Random Cache Instance

Benefits:

  • Load distributed
  • No single hot cache node
  • Much higher throughput

1️⃣6️⃣ Key Interview Concepts

Fanout on Read

Generate feed during read time

Pros:

  • Cheap writes

Cons:

  • Expensive reads

Fanout on Write

Generate feed during write time

Pros:

  • Fast reads

Cons:

  • Expensive writes

Hybrid Model

Normal Users  → Fanout on Write
Celebrities   → Fanout on Read

Industry standard approach.

1️⃣7️⃣ Final Architecture

                +----------------+
                | API Gateway    |
                +----------------+
                         |
       +-----------------+-----------------+
       |                                   |
+-------------+                    +-------------+
| Post Service|                    |Feed Service |
+-------------+                    +-------------+
       |                                   |
       v                                   |
+-------------+                            |
| Post Table  |                            |
+-------------+                            |
       |                                   |
       v                                   |
     Queue ------------------> Feed Workers
                                       |
                                       v
                              +------------------+
                              |Precomputed Feed  |
                              +------------------+
                                       |
                                       v
                                Redis Cache
                                       |
                                       v
                                  Feed User

Interview Sound-Bite 🎯

News Feed is fundamentally a fanout problem. The key tradeoff is between fanout-on-read and fanout-on-write. A scalable design typically uses a hybrid approach, precomputing feeds for normal users while generating celebrity content dynamically, combined with asynchronous workers, queues, and caching to handle scale.

Mindmap

Graph View