Choose your backbone

A single super-line server uses an in-memory adapter — rooms and topics fan out within that one process. To run more than one process (behind a load balancer), give every server a shared adapter so fan-out crosses nodes.

The adapter is the server↔server axis. (The client↔server wire is the separate, independent transport axis — you pick each on its own.)

The adapter seam

Rooms, topics, and the cluster event bus all compile down to channel pub/sub behind the Adapter interface. Swap the implementation; the rest of your code is unchanged — only the adapter: line differs:

import { createRedisAdapter } from '@super-line/adapter-redis'
import { webSocketServerTransport } from '@super-line/transport-websocket'

const srv = createSuperLineServer(api, {
  transports: [webSocketServerTransport({ server })],
  authenticate,
  adapter: createRedisAdapter('redis://localhost:6379'), // ← the only line that changes
})

Point every server process at the same backbone, and room.broadcast, srv.publish / forRole(r).publish, and the cluster bus all reach clients (and peers) on any node. At-most-once delivery is preserved.

No adapter for a single node

You don't need an adapter for one process — the default in-memory adapter handles it. Add a backbone only when you scale out.

Which adapter?

Adapter	Reach for it when…	Presence	Shape
Redis	the pragmatic default for a backend cluster	strong, central	one central broker, broker-enforced expiry
RabbitMQ	you already run it, or want selective per-channel routing	gossip (eventual)	broker delivers only subscribed channels
ZeroMQ	you want the lightest broker-less option	gossip (eventual)	direct sockets; mesh O(N²) or proxy mode
libp2p	decentralized self-hosted / edge	gossip (eventual)	NAT traversal, encrypted, discovery; ESM-only

All four implement the same Adapter seam, so switching is a one-line change. Redis is the default; the three broker-less options trade central presence for one fewer service to run.

Running it

The scaling example boots a real cluster with Docker Compose — Redis, a Caddy load balancer, three server nodes, and six client containers — so you can watch a publish, a room broadcast, and a shared stats topic gossiped over the bus fan out across separate processes:

cd examples/scaling && docker compose up

Each broker-less adapter ships the same cluster with the broker swapped out: scaling-libp2p, scaling-rabbitmq, and the react-chat-cluster-* variants.

Next: pick a backbone — Redis · libp2p · RabbitMQ · ZeroMQ.