Skip to content

compulim/message-port-rpc

Repository files navigation

message-port-rpc

Turns a MessagePort into an remote procedure call (RPC) stub.

Background

Modern web apps often need to deal with multiple JavaScript workers or VMs. The communication channel is often MessagePort.

By converting a dedicated MessagePort into an RPC stub, we can easily offload a Promise function to a different thread.

How to use

Make sure the pair of MessagePort used for RPC is dedicated and not started. No other RPC, listeners, or posters should be using the same pair.

It is highly recommended to create a new MessageChannel and convert them into RPC stub.

Server stub

import { messagePortRPC } from 'message-port-rpc';

messagePortRPC(port1, (x, y) => x   y);

Client stub

import { messagePortRPC } from 'message-port-rpc';

const rpc = messagePortRPC(port2);

await rpc(1, 2); // 3.

Full Web Worker example

One of the advantage of Web Worker is to offload computation-intensive functions.

On main thread (client stub)

Creates a new pair of MessagePort, pass one of the port to the worker thread, then create a RPC stub on another port.

import { messagePortRPC } from 'message-port-rpc';

// TypeScript: define the function type.
type Fn = (x: number, y: number) => Promise<number>;

// Loads a Web Worker.
const worker = new Worker('./static/worker/js/main.js');

// Creates a new pair of `MessagePort` dedicated for RPC.
const { port1, port2 } = new MessageChannel();

// Sends the dedicated port to the worker.
worker.postMessage(undefined, [port2]);

// Creates a function stub.
const callFunction = messagePortRPC<Fn>(port1);

// Calls the function stub.
const result: number = await callFunction(1, 2);

On worker thread (server stub)

Receives the MessagePort and registers an RPC function on the port.

import { messagePortRPC } from 'message-port-rpc';

// TypeScript: define the function type.
type Fn = (x: number, y: number) => Promise<number>;

// Receives the port dedicated for RPC.
addEventListener('message', ({ ports }) => {
  // Registers an RPC function on the received `MessagePort`.
  messagePortRPC<Fn>(ports[0], (x, y) => Promise.resolve(x   y));
});

If the worker takes time to start, it is okay, no invocations would be lost. Thanks to MessagePort, all messages will be queued internally until the other side signals ready to receive.

Aborting the call

Client can abort an invocation sooner by passing an AbortSignal via the withOptions function. An AbortSignal will be passed to the remote function inside this context.

In the following example, we assume the client is remotely invoking a fetch() function, which supports AbortSignal natively.

Server stub

The following code snippet will use the AbortSignal to abort the fetch() call.

messagePortRPC(ports[0], async url => {
  // During an RPC call, the `AbortSignal` is passed in the `this` context.
  const res = await fetch(url, { signal: this.signal });

  // ...
});

Client stub

The following code snippet will call the stub with additional options to pass an AbortSignal.

const abortController = new AbortController();
const remoteFetch = messagePortRPC(port);

// Calls the stub with arguments in array, and options.
const fetchPromise = remoteFetch.withOptions({ signal: abortController.signal })('https://github.com');

// Aborts the ongoing call.
abortController.abort();

// The promise will reject locally.
fetchPromise.catch(error => {});

Notes: despite the AbortSignal is passed to fetch(), when aborted, the rejection will be done locally regardless of the result of the fetch() call.

Generators and iterators

Generators and iterators are supported. This helps iterating large set of data without sending it over.

Generator and Iterator are automatically converted to AsyncGenerator and AsyncIterator respectively.

const { port1, port2 } = new MessageChannel();
const iterateValues = (): Iterator<number> => [1, 2, 3].values();

forGenerator<Fn>(port2, iterateValues);

iterateValuesRemote = forGenerator<Fn>(port1);

for await (const value of iterateValuesRemote()) {
  console.log(value); // Will print 1, 2, 3.
}

Explicit Resource Management is supported. Using await using operator on client stub will close underlying ports when the iterator/generator is being disposed. On server stub side, Symbol.asyncDispose() will be called if defined. Otherwise, Symbol.dispose() will be called if defined. This behavior is equivalent to await using on an object with both dispose functions set.

Notes: if you are calling next() to iterate instead of for-loop, and iteration stopped prematurely before reaching { done: true }, you should use using operator or withOptions({ signal: AbortSignal }) to close underlying ports properly.

API

The following is simplified version of the API. Please refer to our published typings for the full version.

function messagePortRPC<T extends (...args: unknown[]) => Promise<unknown>>(
  port: MessagePort,
  fn?: (this: { signal: AbortSignal }, ...args: Parameters<T>) => ReturnType<T>
): {
  (...args: Parameters<T>): Promise<ReturnType<T>>;

  withOptions: (init: { signal?: AbortSignal; transfer?: Transferable[] }) => (...args: Parameters<T>): Promise<ReturnType<T>>;
};

function forGenerator<
  T extends (...args: unknown[]) => Promise<Generator<TYield, TReturn, TNext>>,
  TYield = unknown,
  TReturn = any,
  TNext = unknown
>(
  port: MessagePort,
  fn?: t
): {
  (...args: Parameters<T>): AsyncGenerator<TYield, TReturn, TNext>;

  withOptions: (init: { signal?: AbortSignal; transfer?: Transferable[] }) => (...args: Parameters<T>): AsyncGenerator<TYield, TReturn, TNext>;
};

Behaviors

Why use a dedicated MessagePort?

Instead of multiplexing multiple calls into a single MessagePort, a dedicated MessagePort simplifies the code, easier to secure and audit the channel, and eliminates crosstalk.

Internally, for every RPC call, we create a new pair of MessagePort. The result of the call is passed through the MessagePort. After the call is resolved/rejected/aborted, the MessagePort will be shutdown.

With a new pair of MessagePort, messages are queued until the event listener call MessagePort.start(). In other words, with dedicated MessagePort, calls are less likely to get lost due to false-start.

What can be passed as arguments and return value?

All arguments and return value will be send over the MessagePort. The values must be transferable using the Structured Clone Algorithm by the underlying MessagePort.

In other words, you cannot pass function or class as an argument or return value.

Will it pass the this context?

No, because the this context is commonly a class object or globalThis. Structured Clone Algorithm probably will not work for most this.

If you need to pass this, please pass it as an argument.

Can I use it with <iframe>?

Yes, you can use it with <iframe>.

However, despite the communication channel in <iframe> is very similar to MessagePort and supports Structured Clone Algorithm, it is not MessagePort.

You will need to create a new MessageChannel and use HTMLIFrameElement.contentWindow.postMessage() to send one of the MessagePort to the <iframe> content window. Then, you can convert the MessagePort into RPC with this package.

Why hosting a single function vs. multiple functions?

We think a single function is much simpler, less responsibility, and more flexible approach.

To create a pool of stubs, you should create multiple MessagePort, one for each stub. Then, send it through an initializer stub. The receiver side receiving these ports should set up stubs for each of the port, registering their respective subroutine.

Can I call from the other side too?

Yes, our implementation supports bidirectional and asymmetrical calls over a single pair of MessagePort.

You can register different functions on both sides and call from the other side.

// On main thread:
// - a power function is hosted on the port;
// - the return value is the stub of the worker, which is a sum function.
const sum = messagePortRPC(port1, (x ** y) => x ** y);

await sum(1, 2); // 1   2 = 3
// On worker thread:
// - a sum function is hosted on the port;
// - the return value is the stub of the main thread, which is a power function.
addEventListener('message', ({ ports }) => {
  const power = messagePortRPC(ports[0], (x   y) => x   y);

  await power(3, 4); // 3 ** 4 = 81
});

Do I need to sequence the calls myself?

No, you do not need to wait for a call to return before making another call.

Internally, all calls are isolated by their own pair of MessagePort and processed asynchronously.

Can I send Error object?

Yes, thanks to the Structured Clone Algorithm, you can send objects of Error class.

However, there are slight differences in the error object received.

const obj = await stub();

obj instanceof Error; // False. Error object from SCA has a different prototype.
Object.prototype.toString.call(obj) === '[object Error]'; // True.

Alternatively, you can recreate the error object.

Can I provide my own marshalling function?

No, we do not support marshalling function.

Alternatively, you can channel MessagePort to a pair of marshal and unmarshal functions. Make sure you implement both marshal and unmarshal functions on both sides of the port.

Can I offload a Redux store or useReducer to a Web Worker?

Yes, you can offload them to a Web Worker. Some notes to take:

  • action and state must be serializable through Structured Clone Algorithm
    • no classes, functions, DOM elements, no thunk, etc.
  • middleware must not contains code that does not work in worker
    • no DOM access, etc.

You can look at sample useBindReducer and useReducerSource to see how it work.

We will eventually made these React hooks available. Stay tuned.

How could I stop the stub from listening to a port?

To stop the stub, you should close the port by calling MessagePort.close().

The port for the stub must be dedicated and not to be reused. When you want to stop the stub from listening to a port, you should simply close the port.

Why don't you create MessagePort for me?

We understood there are hassles to create MessagePort yourself.

We spent a lot of time experimenting with different options and landed on this design for several reasons:

  • you own the resources and control the lifetime of the resources, less likely to resources leak
  • you do not need to create the stub before sending the port to the other side
  • you can control which side creates the ports and do not need to pipe them yourself
  • you can build marshal/unmarshal function without too much piping
  • you can build a MessagePort-like custom channel without extra piping

There are downsides:

  • you forget to dedicate the MessagePort to a stub
  • you need to write one more line of code

At the end of the day, we think channel customization outweighted the disadvantages and made a bet on this design.

Can I use BroadcastChannel to listen to many client stubs at once?

No, you cannot use BroadcastChannel.

BroadcastChannel does not support sending MessagePort and other transferable objects, which is critical to the operation of the stub.

Why should I use this implementation of RPC?

We are professional developers. Our philosophy makes this package easy to use.

  • Standards: we use MessagePort and Structured Clone Algorithm as-is
  • Airtight: we wrap everything in their own MessagePort, no multiplexing = no crosstalks
  • Small scope: one MessagePort host one function only, more flexibility on building style
  • Simple: you almost know how to write this package
  • Maintainability: we relies heavily on tooling and automation to maintain this package

Contributions

Like us? Star us.

Want to make it better? File us an issue.

Don't like something you see? Submit a pull request.