SSE Scaling

From MVP to production: Handle 100+ concurrent users with low latency.

Scaling Patterns

1. In-Memory (Development Only)

Using EventEmitter for single-server setups:

typescript

import { EventEmitter } from 'events';

const notificationEmitter = new EventEmitter();

// Listener in route handler
notificationEmitter.on('notification', (data) => {
  controller.enqueue(encoder.encode(`data: ${JSON.stringify(data)}\n\n`));
});

// Trigger from other route
notificationEmitter.emit('notification', { userId, message });

Limitations: Single process only, loses messages on server restart.

2. Redis Pub/Sub (Recommended for Production)

Broadcast events across multiple server instances:

typescript

import Redis from 'ioredis';

const redis = new Redis();

// In route handler: Subscribe to channel
redis.subscribe(`notifications:${userId}`, (err, count) => {
  if (err) controller.error(err);
});

// Receive messages
redis.on('message', (channel, message) => {
  controller.enqueue(encoder.encode(`data: ${message}\n\n`));
});

// Trigger from elsewhere
redis.publish(`notifications:${userId}`, JSON.stringify({ type: 'like', userId: 1 }));

Benefits: Works across multiple server instances, persistent message broker.

3. Fan-Out Pattern

Send same message to multiple recipients efficiently:

typescript

// In Payload hook: Broadcast to all followers
const followers = await getFollowers(postAuthorId);

for (const follower of followers) {
  await redis.publish(
    `notifications:${follower.id}`,
    JSON.stringify({ type: 'new-post', postId, authorId })
  );
}

Concurrency Limits

Vercel Limits

Plan	Timeout	Concurrent Connections
Pro	60s	Limited by runtime
Enterprise	900s (15m)	Limited by runtime
Fluid Compute	Unlimited	Scales dynamically

Node.js Limits

Default file descriptor limit: 1024 connections per process.

Increase for production:

bash

# Linux: Check current limit
ulimit -n

# Increase to 65536
ulimit -n 65536

Memory Per Connection

Estimate ~50KB-100KB per active SSE connection. For 1000 users:

•1000 connections × 100KB = ~100MB memory
•With 4 server instances: 25MB per instance

Load Testing

Artillery.io Test

yaml

# sse-load-test.yml
config:
  target: 'http://localhost:3000'
  phases:
    - duration: 60
      arrivalRate: 10

scenarios:
  - name: 'SSE Connection Test'
    flow:
      - get:
          url: '/api/notifications'
          capture:
            - json: '$.message'

Run:

bash

artillery run sse-load-test.yml

Manual Test with curl

bash

# Spawn 100 concurrent connections
for i in {1..100}; do
  curl -N http://localhost:3000/api/notifications &
done

Performance Optimization

1. Connection Pooling

Reuse Redis connections across instances:

typescript

const redis = new Redis({
  host: process.env.REDIS_HOST,
  maxRetriesPerRequest: null, // Important for pub/sub
  enableReadyCheck: false,
});

2. Message Batching

Batch multiple notifications into single message:

typescript

const batch = await collectNotificationsForMs(100); // Wait 100ms for batch

redis.publish(
  `notifications:${userId}`,
  JSON.stringify({ batch: batch })
);

3. Selective Delivery

Only send to users who care about event type:

typescript

// In hook: Get only users interested in this notification type
const subscribers = await getSubscribers(userId, 'likes');

for (const subscriber of subscribers) {
  redis.publish(`notifications:${subscriber.id}`, JSON.stringify(notification));
}

Monitoring & Observability

Key Metrics

•Connection count: Current active SSE connections
•Message throughput: Messages/sec through Redis
•Memory usage: Per instance and total
•Latency: Message publish → client receive (p50, p95, p99)

Logging

typescript

console.log(`[SSE] Client connected: ${clientId}`);
console.log(`[SSE] Published to ${recipientCount} users`);
console.log(`[SSE] Client ${clientId} disconnected after ${connectionDuration}ms`);

Migration Path (MVP → Production)

•Week 1: Deploy MVP with EventEmitter (single instance)
•Week 2: Add basic monitoring, identify bottlenecks
•Week 3: Migrate to Redis (multi-instance ready)
•Week 4: Optimize delivery, add batching
•Ongoing: Monitor latency and connection count

AI Prompts for Scaling

Prompt: "Refactor my SSE code to use Redis pub/sub for scaling across multiple server instances."

Prompt: "Write a load test for SSE that simulates 1000 concurrent users connecting and receiving notifications."

Prompt: "How can I optimize SSE delivery to reduce latency when broadcasting to 10,000 users?"

Related Concepts

•Message queues (vs pub/sub): Use RabbitMQ or AWS SQS if message persistence needed
•WebSockets: For bidirectional communication (more complex than SSE)
•gRPC: For internal service-to-service events (not client-facing)