dog

Durable Object Groups

Usage no npm install needed!

<script type="module">
  import dog from 'https://cdn.skypack.dev/dog';
</script>

README

DOG CI

Durable Object Groups

Features

  • Supports Replica workloads using the HTTP and/or WS protocols
  • Creates or reuses a Replica based on configured connection limit
  • Includes Replica-to-Replica (peer-to-peer) communication
  • Ready for strongly-typed, strict TypeScript usage
  • Allows an active connection to:
    • broadcast messages to the entire cluster
    • emit messages to Replica-owned connections
    • send a whisper a single connection within the cluster

Overview

With DOG, it's easy to setup named clusters of related Durable Objects. Each cluster is controlled by a Group, which directs an incoming Request to a specific Replica instance. A Group adheres to the user-defined limit of active connections per Replica and, in doing so, will reuse existing or create new Replica instances as necessary.

DOG includes convenience methods that allow a Replica to directly communicate with another Replica belonging to the same Group – effectively a peer-to-peer/gossip network. Additionally, when dealing with active client connections, a Replica class allows you to:

  • broadcast a message to all active connections within the entire cluster
  • emit a message only to active connections owned by the Replica itself
  • whisper a message to a single, targeted connection (via your own identification system); even if it's owned by another Replica instance!

Group and Replica are both abstract classes, which means that you're allowed — and required — to extend them with your own application needs. You may define your own class methods, add your own state properties, or use Durable Storage to fulfill your needs.

Please see Usage, the API docs, and the example application for further information!

Install

$ npm install dog

Usage

Refer to the /example for a complete Chat Room application.

import { identify, Group, Replica } from 'dog';

// deployed as `POOL` binding
export class Pool extends Group {
  limit = 50; // each Replica handles 50 connections max

  link(env: Bindings) {
    return {
      child: env.TASK, // receiving Replica
      self: env.POOL, // self-identifier
    };
  }
}

// deployed as `TASK` binding
export class Task extends Replica {
  link(env) {
    return {
      parent: env.POOL, // parent Group
      self: env.TASK, // self-identifier
    };
  }

  async onmessage(socket, data) {
    let message = JSON.parse(data);
    console.log('[task] onmessage', message);

    if (message.type === 'crawl:url') {
      let { url } = message;
      // ...
      let output = { url, done: true };
      // alert everyone that the task is complete
      return socket.broadcast(JSON.stringify(output), true);
    }

    // other events
  }

  receive(req) {
    // Receive & handle the request
    // NOTE: This is the original, forwarded request
    let { pathname } = new URL(req.url);

    // Rely on internal util for WebSocket upgrade
    if (pathname === '/ws') return this.connect(req);

    // Any other custom routing behavior(s)
    if (pathname === '/') return new Response('OK');

    return toError('Unknown path', 404);
  }
}

function toError(msg, status) {
  return new Response(msg, { status });
}

// Module Worker
export default {
  fetch(req, env, ctx) {
    // Accept: /tasks/<taskname>
    let match = /[/]tasks[/]([^/]+)[/]?/.exec(req.url);
    if (match == null) return toError('Missing task name', 404);

    let taskname = match[1].trim();
    if (taskname.length < 1) return toError('Invalid task name', 400);

    // Generate Durable Object ID from taskname
    let group = env.POOL.idFromName(taskname);

    // Custom request identifier logic
    let reqid = req.headers.get('x-request-id');

    // Identify the `Replica` stub to use
    let replica = await identify(group, reqid, {
      parent: env.POOL,
      child: env.TASK,
    });

    // (Optional) Save reqid -> replica.id
    // await KV.put(`req::${reqid}`, replica.id.toString());

    // Send request to the Replica instance
    return replica.fetch(req);
  }
}

API

identify

Note: Refer to the TypeScript definitions for more information.

The utility function to identify a Replica to be used and, if necessary, will create a new Replica if none are available. Returns the Replica stub directly.

Group

Note: Refer to the TypeScript definitions for more information.

Required:

  • limit: number – the maximum number of active connections a Replica can handle
  • link(env: Bindings): { self, child } – define the relationships between this Group and its Replica child class

A Group is initial coordinator for the cluster. It receives a user-supplied request identifier, ReqID, and replies with the Durable Object ID for the Replica instance to be used. If the ReqID has been seen before, the Group will attempt to target the same Replica that the ReqID was previously connected to. If the ReqID is unknown, the Group will send the request to the least-utilized Replica instance or generate a new Replica ID to be used.

When targeting an existing Replica instance, the Group verifies that the Replica actually has availability for the request, as determined by the user-supplied limit value. If a new Replica instance needs to be created, the Group's clusterize() method is called to generate a new Replica instance identifier. You may override this method with your own logic – for example, including a jurisdiction – but by default, the Group calls newUniqueId() for a system-guaranteed identifier.

The number of active connections within each Replica instance is automatically tracked and shared between the Replica and its Group parent. The Replica's count is decremented when the connection is closed. This means that when a Replica works with WebSockets, open connections continue to reserve Replica quota until closed. Non-upgraded HTTP connections close and decrement the Replica count as soon as a Response is returned.

Important: Do not define your own fetch() method!
Doing so requires that super.fetch() be called appropriately, otherwise the entire cluster's inter-communication will fail.

You may attach any additional state and/or methods to your Group class extension.

Replica

Note: Refer to the TypeScript definitions for more information.

Required:

  • link(env: Bindings): { self, child } – define the relationships between this Replica and its Group parent class
  • receive(req: Request): Promise<Response> | Response – a user-supplied method to handle an incoming Request

A Replica is the cluster's terminating node. In other words, it's your workhorse and is where the bulk of your application logic will reside. By default, a Replica actually does nothing and requires your user-supplied code to become useful. It does, however, provide you with utilities, lifecycle hooks, and event listeners to organize and structure your logic.

A Replica can only receive a Request from its parent Group or from its Replica siblings/peers. Because of this, you cannot define a fetch() method in your Replica class extension, otherwise all internal routing and inter-communication will break.

However, this does not mean that you cannot deploy your own external-facing routing solution!

If an incoming request to a Replica is not an internal DOG event, the request is passed to your receive method, which receives the original Request without any modifications. This means that the execution order for a client request looks like this:

client request
└──> dog.identify(...)
      │   ├──> Group#fetch (internal)
      │   └──> Group#clusterize (optional)
      └──> Replica
          └──> Replica.fetch (user)
              └──> Replica#receive

Your receive method is the final handler and decides what the Replica actually does.

If you'd like to remain in the HTTP protocol, then you can treat receive() as if it were the underyling fetch() method. Otherwise, to upgrade the HTTP connection into a WebSocket connection, then you may reach for the Replica.connect() method, which handles the upgrade and unlocks the rest of the Replica abstractions.

Internally, a Socket interface is instantiated and passed to WebSocket event listeners that you chose to define. For example, to handle incoming messages or to react to a new connection, your Replica class may including the following:

import { Replica } from 'dog';

export class Counter extends Replica {
  #counts = new Map<string, number>;

  onopen(socket) {
    // via dog.identify
    // ~> your own ReqID
    let reqid = socket.uid;
    this.#counts.set(reqid, 0);

    // notify others ONLY in Replica
    socket.emit(`"${reqid}" has joined`);
  }

  onmessage(socket, data) {
    let reqid = socket.uid;
    let current = this.#counts.get(reqid);

    // data is always a string
    let msg = JSON.parse(data);

    if (msg.type === '+1') count++;
    else if (msg.type === '-1') count--;
    else return; // unknown msg type

    this.#counts.set(reqid, count);

    // tell EVERYONE in cluster about new count
    socket.broadcast(`"${reqid}" now has ${count}`);
  }

  receive(req) {
    // Only accept "/ws" pathname
    let isWS = /^[/]ws[/]?/.test(req.url);
    // Handle upgrade, return 101 status
    if (isWS) return this.connect(req);
    // Otherwise return a 404
    return new Response('Invalid', { status: 404 });
  }
}

The Replica class allows you to optionally define event listeners for the underlying WebSocket events. Whether or not you define onclose and/or onerror listeners, the Replica will always notify the Group parent when the WebSocket connection is closed. The event listeners may be asynchronous and their names follow the browser's WebSocket event names:

  • onopen – the Replica established a WebSocket connection
  • onmessage – the Replica received a message from the WebSocket connection
  • onerror – the WebSocket connection terminated due to an error
  • onclose – the WebSocket connection was closed

Note: If defined, the onclose listener will be called in the absence of an onerror listener.

Finally, a Replica may communicate directly with its Replica peers in the cluster. This does not rely on WebSockets nor does it require you to use them! It can, however, be leveraged at any point during your HTTP and/or WebSocket handlers.

In DOG, this peer-to-peer communication is called gossip – because Replicas are typically talking about their connections but without involving the connections; AKA, behind their backs!

In order for a Replica to hear gossip, it must define an ongossip method handler. It will receive a decoded JSON object and must return a new JSON object so that DOG can serialize it and deliver it to sender. In practice, this internal communication is happening over HTTP which means that each Gossip.Message must represent point-in-time information.

Returning to the Counter example, suppose the Counter objects needs to coordinate with one another to determine a leaderboard. Refreshing this leaderboard could be done through a new refresh:leaderboard message, for example:

import { Replica } from 'dog';

export class Counter extends Replica {
  #counts = new Map<string, number>;
  #lastupdate = 0; // timestamp
  #leaders = []; // Array<string, number>[]

  // NOTE: now `async` method
  async onmessage(socket, data) {
    let reqid = socket.uid;
    let current = this.#counts.get(reqid);

    // data is always a string
    let msg = JSON.parse(data);

    // ...

    if (msg.type === 'refresh:leaderboard') {
      // Only gossip if cache is older than 60s
      if (Date.now() - this.#lastupdate > 60e3) {
        // `ongossip` returns Array<[string,number][]>
        let results = await this.gossip({ type: 'ask:scores' });
        let leaders = results.flat(1); // [ [reqid,count], [reqid,count], ... ]

        // sort by highest scores, keep top 10 only
        this.#scores = leaders.sort((a, b) => b[1] - a[1]).slice(0, 10);
        this.#lastupdate = Date.now();
      }

      // Tell EVERYONE in cluster
      return socket.broadcast({
        leaders: this.#scores,
        timestamp: this.#lastupdate,
      });
    }
  }

  ongossip(msg) {
    // Return array of tuples: Array<[string, number]>
    if (msg.type === 'ask:scores') return [...this.#counts];
    throw new Error(`Missing "${msg.type}" handler in ongossip`);
  }

  // ...
}

License

MIT © Cloudflare