Resonate State Bus Pattern (TypeScript)

Overview

A long-running Resonate workflow often needs to surface progress to a live UI. If the worker is short-lived (Cloud Function, Lambda) or scales to zero, holding an SSE or WebSocket connection for the workflow's lifetime isn't an option — the connection outlives the worker process.

The state bus pattern decouples the two sides:

1. The workflow writes its current state to an external realtime bus via ctx.run.
2. The browser (or any subscriber) subscribes to that bus directly.

The worker stays stateless, progress becomes durable (ctx.run guarantees it), and the UI becomes a pure read view — the workflow can run with or without anyone watching.

Mental Model

Worker (short-lived)             Bus                    Browser (long-lived)
   │                              │                              │
   ├─ ctx.run(writeState, s1) ──▶ │                              │
   │                              ├─ notify subscribers ───────▶ render(s1)
   │  [worker returns, scales to  │                              │
   │   zero — bus persists state] │                              │
   │                              │                              │
   ├─ ctx.run(writeState, s2) ──▶ │                              │
   │                              ├─ notify subscribers ───────▶ render(s2)

The workflow never talks to a client. The client never talks to the workflow. Both sides talk to the bus.

When to use this pattern

• Your worker is serverless (Cloud Functions, Lambda, Edge Functions) and can't hold a persistent connection.
• The workflow must run whether or not anyone is subscribed (e.g. a shared game, a batch job whose output is watched by an ops dashboard).
• Multiple clients need the same live view of the same workflow.
• You want the workflow's progress to survive worker restarts and cold starts with no reconnection logic in the UI.

When NOT to use this pattern

• The worker already runs in the browser (see example-browser-worker-ts) — then the client is the worker, and in-tab state works.
• The workflow is short enough to fit inside a single request/response — just return the result.
• The consumer is another service that can long-poll the Resonate server directly.

Core pattern

Step 1 — Pick a bus

Any durable realtime system with a pub/sub shape works. The only requirement is that the worker can write and the subscriber can listen.

Common options:

The rest of this skill uses Firestore for concreteness. The pattern is identical across buses — only the client libraries change.

Step 2 — Write state from the worker

Instantiate the bus client at module scope so it's reused across warm invocations. Putting it inside the generator would reconnect on every replay and slow cold paths.

import { Firestore } from "@google-cloud/firestore";
import type { Context } from "@resonatehq/sdk";

const firestore = new Firestore();
const liveDoc = firestore.collection("myapp").doc("live");

async function writeState(_ctx: Context, state: Record<string, unknown>) {
  await liveDoc.set(state);
}

function* workflow(ctx: Context) {
  // ... compute some state ...
  yield* ctx.run(writeState, { step: "ready", t: Date.now() });
  // ... more work, more writes ...
}

Grant the worker's runtime service account roles/datastore.user on the project (Firestore Native uses Datastore IAM).

Step 3 — Subscribe from the browser

import { initializeApp } from "firebase/app";
import { getFirestore, doc, onSnapshot } from "firebase/firestore";

const app = initializeApp({ projectId: "<your-project>" /* + apiKey, etc. */ });
const db = getFirestore(app);

onSnapshot(doc(db, "myapp", "live"), (snap) => {
  const state = snap.data();
  // render state
});

Firestore rules should be read-only for anonymous users on the subscriber collection — no client should ever write through to it. Only the worker's service account does.

Design notes

One document per workflow, not per event. Overwrite the document on each state write. Subscribers want the current state, not a log. If you need history, write a log alongside the live doc and render it separately.

Idempotent writes. ctx.run replays on failure. The writeState activity must be safe to re-execute with the same input. set() (vs update()) is the simple choice — replays overwrite with identical data.

Shape the state for the UI, not the workflow. The workflow's internal state may be richer than what the UI needs. Project to a narrow DTO in writeState so the subscriber payload stays small and the schema stays stable even as the workflow evolves.

Don't stream high-frequency updates. Firestore (and most buses) bill per write. If your workflow produces 1000 updates/second, batch them — write every N events or every M milliseconds, driven from the workflow.

Contrast with "browser as worker"

example-countdown-web-ts and example-browser-worker-ts show a different model: the Resonate worker itself runs inside the browser tab. That's the right shape for a per-tab durable workflow (e.g. a client-side task that must survive a page reload). The state bus pattern is the opposite shape — the worker is server-side, possibly serverless, and the browser is a read-only subscriber. Use the bus pattern when the workflow is shared across viewers or must run when no one is watching.

Reference example

example-chess-hero-gcp-ts — perpetual chess game as a single Resonate workflow on Cloud Functions Gen 2; move-by-move state written to Firestore; any number of browser tabs subscribe with onSnapshot. Deploy flow covered in resonate-gcp-deployments-typescript (worker) and resonate-server-deployment-cloud-run (server).