Designing a Social Network Feed System

What you'll learn

• Problem Statement

  Design a social media platform where users can post short messages, follow others, and view a personalized home timeline aggregating posts from all followed accounts. The system handles billions of daily post views with sub-200ms timelin…

• Requirements

• Functional Requirements

  1. Post short messages (280 characters + optional media) 2. Follow/Unfollow other users 3. Home Timeline — aggregated feed of posts from followed users 4. User Timeline — all posts by a specific user 5. Search posts by keyword/hashtag 6.…

• Non-Functional Requirements

  1. High availability — service should never go down 2. Low latency — timeline generation What breaks next — the timeline JOIN itself does not scale by adding replicas. Each replica still has to do the same fan-in sort per request, and th…

• Stage 3 — Bottleneck #1: Timeline generation at read time (the fan-in problem)

  At 325K timeline reads/sec, doing a JOIN across follows and posts on every request is architecturally impossible — no amount of read replicas fixes it. A user following 200 people needs the top 20 posts merged from 200 possibly-scattered…

• Stage 4 — Bottleneck #2: Hot-shard write amplification from popular accounts

  Even with hybrid fan-out, the celebrity carve-out only handles the top ~0.01% of users. The tier just below — "moderately famous" accounts with 10K–100K followers — still fan-out on write, and their posts create write storms into Redis.…

• Stage 5 — Bottleneck #3: Media, search, and cross-region reads

  Two smaller-but-real problems appear once fan-out is under control: 1. Media — post attachments (images/video) are 40× the size of post text. Serving 24 TB/day of media through the API tier is wasteful; that traffic never touches busines…

• Cost vs Complexity — When to Pick What

  | Concern | MVP choice | Growth choice | Enterprise choice | |---------|------------|---------------|-------------------| | Compute | Single VM, monolith | K8s deployment, N API pods | Multi-region K8s, autoscaling | | Database | Managed…

• Migration Path (MVP → Enterprise)

  The ordering matters. Each step should be shippable behind a feature flag, measurable, and reversible. 1. Introduce read replicas — no code change to writers; readers gradually shift via a config flag. Measure replication lag before maki…

• Fault-Tolerant Key Generation

  - Each server independently generates sequences - No coordination/lock needed between them - If one dies: other continues (IDs have gaps, uniqueness preserved) - Both out of sync? Fine — uniqueness comes from even/odd split, not from syn…

• Timeline Generation (The Core Challenge)

• The Problem

  User X follows 200 people. Loading their timeline requires: 1. Get list of 200 followed accounts 2. For each, find their latest posts (scattered across shards) 3. Merge and rank 4,000+ posts 4. Return top 20 At 325K timeline reads/sec, t…

• Feed Storage Data Structure

  The timeline cache for each user can be modeled as: A LinkedHashMap is chosen because it provides both random access (jump directly to any feed item by ID for cursor-based pagination) and sequential iteration (traverse items in insertion…

• Solution: Pre-Generation (Fan-out)

• Fan-out Approaches (Critical Trade-off)

  | Approach | Mechanism | Best For | Problem | |----------|-----------|----------|---------| | Fan-out on Write | Push to all followers on post creation | Users with 50% failure rate, stop routing to it ---

• Extended Features

• Retweets

• Trending Topics

• Who to Follow Suggestions

  ---

• Monitoring and Metrics

  | Metric | Alert Threshold | Why | |--------|----------------|-----| | Timeline latency p99 | > 200ms | SLA violation | | Fan-out lag | > 5 seconds | Followers see stale timelines | | Replication lag | > 5 seconds | Reads may be too stal…

• Summary: Architecture Decisions

  | Decision | What it Solves | |----------|---------------| | PostID = epoch + sequence | Time-ordered primary key, no secondary index, even sharding | | Even/odd key-gen servers | Fault-tolerant ID generation without coordination | | Pos…