Shared Reducer

Shared state management via websockets.

Designed to work with json-immutability-helper.

Install dependency

npm install --save shared-reducer json-immutability-helper

(if you want to use an alternative reducer, see the instructions below).

Usage (Backend)

This project is compatible with websocket-express, but can also be used in isolation.

With websocket-express

import {
  Broadcaster,
  WebsocketHandlerFactory,
  InMemoryModel,
  ReadWrite,
} from 'shared-reducer/backend';
import context from 'json-immutability-helper';
import { WebSocketExpress } from 'websocket-express';

const model = new InMemoryModel();
const broadcaster = new Broadcaster(model, context);
model.set('a', { foo: 'v1' });

const app = new WebSocketExpress();
const server = app.listen(0, 'localhost');

const handlerFactory = new WebsocketHandlerFactory(broadcaster);
app.ws('/:id', handlerFactory.handler((req) => req.params.id, () => ReadWrite));

const server = app.listen();

// later, to shutdown gracefully:
// send a close signal to all clients and wait up to 1 second for acknowledgement:
await handlerFactory.close(1000);
server.close();

For real use-cases, you will probably want to add authentication middleware to the expressjs chain, and you may want to give some users read-only and others read-write access, which can be achieved in the second lambda.

Alone

import { Broadcaster, InMemoryModel } from 'shared-reducer/backend';
import context from 'json-immutability-helper';

const model = new InMemoryModel();
const broadcaster = new Broadcaster(model, context);
model.set('a', { foo: 'v1' });

// ...

const subscription = await broadcaster.subscribe('a');

const begin = subscription.getInitialData();
subscription.listen((change, meta) => { /*...*/ });
await subscription.send(['=', { foo: 'v2' }]);
// callback provided earlier is invoked

await subscription.close();

Persisting data

A convenience wrapper is provided for use with collection-storage, or you can write your own implementation of the Model interface to link any backend.

import { Broadcaster, CollectionStorageModel } from 'shared-reducer/backend';
import context from 'json-immutability-helper';
import CollectionStorage from 'collection-storage';

const db = await CollectionStorage.connect('memory://something');
const model = new CollectionStorageModel(
  db.getCollection('foo'),
  'id',
  // a function which takes in an object and returns it if valid,
  // or throws if invalid (protects stored data from malicious changes)
  MY_VALIDATOR,
);
const broadcaster = new Broadcaster(model, context);

Note that the provided validator MUST verify structural integrity (e.g. ensuring no unexpected fields are added or types are changed).

Usage (Frontend)

import { SharedReducer } from 'shared-reducer/frontend';
import context from 'json-immutability-helper';

const reducer = new SharedReducer(context, () => ({
  url: 'ws://destination',
  token: 'my-token',
}));

reducer.addStateListener((state) => {
  console.log('latest state is', state);
});

reducer.addEventListener('connected', () => {
  console.log('connected / reconnected');
});

reducer.addEventListener('disconnected', (e) => {
  console.log('connection lost', e.detail.code, e.detail.reason);
});

reducer.addEventListener('warning', (e) => {
  console.log('latest change failed', e.detail);
});

const dispatch = reducer.dispatch;

dispatch([
  { a: ['=', 8] },
]);

dispatch([
  (state) => {
    return [{ a: ['=', Math.pow(2, state.a)] }];
  },
]);

dispatch([
  (state) => {
    console.log('state after handling is', state);
    return [];
  },
]);

dispatch(
  [{ a: ['add', 1] }],
  (state) => console.log('state after syncing is', state),
  (message) => console.warn('failed to sync', message),
);

dispatch([
  { a: ['add', 1] },
  { a: ['add', 1] },
]);

Specs

The specs need to match whichever reducer you are using. In the examples above, that is json-immutability-helper.

WebSocket protocol

The websocket protocol is minimal:

Client-to-server

<token>: The authentication token is sent as the first message when the connection is established. This is plaintext. The server should respond by either terminating the connection (if the token is deemed invalid), or with an init event which defines the latest state in its entirety. If no token is specified using withToken, no message will be sent (when not using authentication, it is assumed the server will send the init event unprompted).
P (ping): Can be sent periodically to keep the connection alive. The server sends a "Pong" message in response immediately.
{"change": <spec>, "id": <id>}: Defines a delta. This may contain the aggregate result of many operations performed on the client. The ID is an opaque identifier which is reflected back to the same client in the confirmation message. Other clients will not receive the ID.
x (close ack): Sent by the client in response to X (closing). Indicates that the client will not send any more messages on this connection (but may still be expecting some responses to existing messages).

Server-to-client

p (pong): Reponse to a ping. May also be sent unsolicited.
{"init": <state>}: The first message sent by the server, in response to a successful connection.
{"change": <spec>}: Sent whenever another client has changed the server state.
{"change": <spec>, "id": <id>}: Sent whenever the current client has changed the server state. Note that the spec and ID will match the client-sent values.

The IDs sent by different clients can coincide, so the ID is only reflected to the client which sent the spec.
{"error": <message>, "id": <id>}: Sent if the server rejects a client-initiated change.

If this is returned, the server state will not have changed (i.e. the entire spec failed).
X (closing): Sent when the server is about to shut down. The client should respond with x and not send any more messages on the current connection. Any currently in-flight messages will be acknowledged on a best-effort basis by the server. The server might not wait for the acknowledging x message before closing the connection.

Specs

The specs need to match whichever reducer you are using. In the examples above, that is json-immutability-helper.

Alternative reducer

To enable different features of json-immutability-helper, you can customise it before passing it to the constructor. For example, to enable list commands such as updateWhere and mathematical commands such as Reverse Polish Notation (rpn):

// Backend
import { Broadcaster, InMemoryModel } from 'shared-reducer/backend';
import listCommands from 'json-immutability-helper/commands/list';
import mathCommands from 'json-immutability-helper/commands/math';
import context from 'json-immutability-helper';

const broadcaster = new Broadcaster(
  new InMemoryModel(),
  context.with(listCommands, mathCommands),
);

// Frontend
import { SharedReducer } from 'shared-reducer/frontend';
import listCommands from 'json-immutability-helper/commands/list';
import mathCommands from 'json-immutability-helper/commands/math';
import context from 'json-immutability-helper';

const reducer = new SharedReducer(
  context.with(listCommands, mathCommands),
  () => ({ url: 'ws://destination' }),
);

If you want to use an entirely different reducer, create a wrapper:

import context from 'json-immutability-helper';

const myReducer = {
  update: (value, spec) => {
    // return a new value which is the result of applying
    // the given spec to the given value (or throw an error)
  },
  combine: (specs) => {
    // return a new spec which is equivalent to applying
    // all the given specs in order
  },
};

// Backend
const broadcaster = new Broadcaster(new InMemoryModel(), myReducer);

// Frontend
const reducer = new SharedReducer(myReducer, () => ({ url: 'ws://destination' }));

Be careful when using your own reducer to avoid introducing security vulnerabilities; the functions will be called with untrusted input, so should be careful to avoid attacks such as code injection or prototype pollution.

Other customisations (Backend)

The Broadcaster constructor can also take some optional arguments:

new Broadcaster(model, reducer[, options]);

options.subscribers: specify a custom keyed broadcaster, used for communicating changes to all consumers. Required interface:
```
{
  add(key, listener) {
    // add the listener function to key
  },
  remove(key, listener) {
    // remove the listener function from key
  },
  broadcast(key, message) {
    // call all current listener functions for key with
    // the parameter message
  },
}
```
All functions can be asynchronous or synchronous.

The main use-case for overriding this would be to share messages between multiple servers for load balancing, but note that in most cases you probably want to load balance documents rather than users for better scalability.
options.taskQueues: specify a custom task queue, used to ensure operations happen in the correct order. Required interface:
```
{
  push(key, task) {
    // add the (possibly asynchronous) task to the queue
    // for the given key
  },
}
```
The default implementation will execute the task if it is the first task in a particular queue. If there is already a task in the queue, it will be stored and executed once the existing tasks have finished. Once all tasks for a particular key have finished, it will remove the queue.

As with subscribers, the main reason to override this is to provide consistency if multiple servers are able to modify the same document simultaneously.
options.idProvider: specify a custom unique ID provider. Must be a function which returns a unique string ID when called. Can be asynchronous.

The returned ID is used internally and passed through the configured taskQueues to identify the source of a change. It is not revealed to users. The default implementation uses a fixed random prefix followed by an incrementing number, which should be sufficient for most use cases.

Other customisations (Frontend)

If the connection is lost, the frontend will attempt to reconnect automatically. By default this uses an exponential backoff with a small amount of randomness, as well as attempting to connect if the page regains focus or the computer rejoins a network. You can fully customise this behaviour:

import { SharedReducer, OnlineScheduler, exponentialDelay } from 'shared-reducer/frontend';

const reducer = new SharedReducer(context, () => ({ url: 'ws://destination' }), {
  scheduler: new OnlineScheduler(
    exponentialDelay({
      base: 2,
      initialDelay: 200,
      maxDelay: 10 * 60 * 1000,
      randomness: 0.3,
    }),
    20 * 1000, // timeout for each connection attempt
  ),
});

The exponentialDelay helper returns:

min(initialDelay * (base ^ attempt), maxDelay) * (1 - random(randomness))

All delay values are in milliseconds.

You can also provide a custom function instead of exponentialDelay; it will be given the current attempt number (0-based), and should return the number of milliseconds to wait before triggering the attempt.

Finally, by default when reconnecting SharedReducer will replay all messages which have not been confirmed (AT_LEAST_ONCE delivery). You can change this to AT_MOST_ONCE or a custom mechanism:

import { SharedReducer, AT_MOST_ONCE } from 'shared-reducer/frontend';

const reducer = new SharedReducer(context, () => ({ url: 'ws://destination' }), {
  deliveryStrategy: AT_MOST_ONCE,
});

Custom strategies can be defined as functions:

function myCustomDeliveryStrategy(serverState, spec, hasSent) {
  return true; // re-send all (equivalent to AT_LEAST_ONCE)
}

serverState is the new state from the server after reconnecting.
spec is a spec that has not been confirmed as delivered to the server.
hasSent is true if the spec has already been sent to the server (but no delivery confirmation
was received). It is false if the message was never sent to the server.

Note that the function will be invoked multiple times (once for each change that is pending). It should return true for messages to resend, and false for messages to drop.