Design Facebook Live Comments

#Problem Statement

Design a real-time comment system for live video streams. Millions of viewers can comment simultaneously; all viewers should see comments in near-real-time.

📖

Real-world: This is a massive fan-out problem: one comment must reach millions of concurrent viewers in near-real-time, which is the inverse of chat (few recipients, guaranteed delivery) — here you have enormous recipients but can tolerate dropping/coalescing some comments (nobody reads all 50,000 comments/second on a viral stream anyway). So the design leans on pub/sub broadcast over WebSocket/SSE, reading from a cache rather than the database, and often sampling or rate-limiting what's displayed. The hard, underappreciated layer for a security audience is real-time moderation at scale: live comments are a firehose of potential spam, harassment, hate speech, and scams, and unlike a stored post you can't review before publishing — so you need inline automated moderation (ML classifiers, slur filters, rate limits per user) running in the fan-out path itself. "How do you moderate a million comments a minute in real time?" is the question that separates a complete answer from a naive one.

#Requirements

#Functional

• Post a comment on a live stream
• All viewers see new comments in near-real-time (< 2s)
• Comment feed is paginated (most recent N comments visible)
• Comment count displayed live
• Moderation: hide/remove comments

#Non-Functional

• Viral stream: 1M concurrent viewers, 1k comments/s
• Comment delivery latency: < 2s
• Not necessary for every viewer to see every comment (sampling acceptable for very high rate)

#Core Design

#Fan-Out at 1M Viewers

The core problem: 1 comment posted → deliver to 1M viewers. Classic fan-out challenge.

Can't do direct WebSocket fan-outpushing to 1M WebSocket connections from a single app server is not feasible.

Solution: Pub/Sub at scale

Comment Posted → Kafka (topic: comments:{stream_id})
                      ↓
            Comment Server cluster (each server handles N viewers)
                      ↓
                  Each server subscribes to Kafka topic
                  → pushes to its connected WebSocket clients

Each comment server handles ~50k WebSocket connections. For 1M viewers: ~20 comment servers. Each subscribes to the same Kafka topic. Kafka delivers each message to every server (broadcast). Each server fans out to its 50k clients.

#Comment Sampling for High Rate

At 1k comments/s with 1M viewers, no viewer can read every comment. Sample:

• Show at most 5-10 comments per second in the UI
• Server samples 1/10 of comments for broadcast; store all in DB
• UI smoothly streams the sampled comments

#Comment Storage

Cassandra:
  comments(
    stream_id   UUID,
    comment_id  TIMEUUID,  -- auto-sorted by time
    user_id     UUID,
    content     TEXT,
    status      ENUM('visible','hidden','deleted')
    PRIMARY KEY (stream_id, comment_id DESC)
  )

#Architecture

Viewer (WebSocket) ←→ Comment Server (one of 20) ← Kafka ← Comment API
                                                      ↓
                                             Comment DB (Cassandra)
                                                      ↓
                                             Moderation Service (ML + human queue)

#Security Considerations

#Interview Tips