Node: Worker eval as ES Module

[targos]

Unless I missed it, there doesn't seem to be a way of creating a Worker from an ESM source string (with eval: true).

Example that should be possible somehow:

import { Worker } from 'worker_threads';

new Worker(
  `
  import { join } from 'path';
  console.log(join('hello', 'world'));
`,
  { eval: true },
);

Related: https://github.com/nodejs/node/issues/21502

[targos]

Adding execArgv: ['--input-type=module'] to the options doesn't help (and wouldn't be the best UX).

[mko-io]

Can I use Es6 module in Worker by using new Worker("./foo.js", { type: "module" }) Chrome 80 will enable this feature by default https://bugs.chromium.org/p/chromium/issues/detail?id=680046

[addaleax]

@nodejs/modules-active-members

[targos]

@mko-io if your file would be interpreted as a module when run with node foo.js (in a "type": "module" package, or if the extension is .mjs), yes.

[jkrems]

I hope we'll support type: module for workers. I assume in that case we could treat it as an implicit "also treat the source text as .mjs" (as opposed to supporting different kinds of modules in eval: true):

import { Worker } from 'worker_threads';

new Worker(
  `
  import { join } from 'path';
  console.log(join('hello', 'world'));
`,
  { eval: true, type: 'module' },
);

(Not working code, just as a possible example of what it would look like in node.)

[addaleax]

If we do add a type: 'module' option to the Worker constructor, should that also apply to non-eval code, i.e. be treated like a package.json type: option?

[jkrems]

I assume it should only import instead of require (in terms of loading semantics). Generally speaking we've been hesitant to allow the importer to change the static semantics of the imported resource. import-semantics may also imply that the "filename" is a URL (e.g. support for data: URLs).

Since we don't support bare specifiers for workers and we don't support conditional mappings of effectively "absolute URLs", I assume that type: module is mostly opting into an async worker bootstrap but not much else outside of eval.

[guybedford]

Do workers already support the package.json "type": "module" ok? If so then perhaps we don't need type: 'module' as an option at all actually. The major argument for that would be web compatibility but we already don't have web compatibility for this API as a goal right?

Perhaps eval can be supported via either eval: 'module' or using a data URI with a MIME type corresponding to a module like we already support in the module loader as well?

[devsnek]

I think we should try to match the web worker API, which uses type: classic and type: module. lacking URL.createObjectURL, combining type:module with eval:true makes the most sense to me. The collision with the "type" name in package.json is unfortunate but not a huge deal imo.

[guybedford]

The question is - why though?

[devsnek]

why work towards compat with other runtimes? it's annoying and confusing when there's a difference. we also do try to work in compat where possible, like supporting on{event} properties.

[guybedford]

As mentioned Worker is already not web compatible though as it uses paths not URLs and also has API differences. So since a compatibility layer is needed regardless any attempt towards web compat is just "going through the motions", while also invalidating the work we've done that already ensures module types are well-defined.

[GeoffreyBooth]

I guess one question is, _could_ Worker someday be Web compatible? Like we may be lacking URL.createObjectURL right now, but it seems like something we could conceivably add someday, isn’t it? Since we have data URIs already. Likewise we’ve been batting around the idea of import from http URLs, which isn’t possible now but my understanding is that that was mostly punted for difficulty (working out the mechanics of what allowing that would mean) rather than some philosophical opposition.

Anyway my point is, if theoretically Worker might be web-compatible someday once we work out all the other incompatible pieces around it, then we might as well stay compatible whenever possible now, to avoid breaking changes later when the only parts that are incompatible are the APIs we deliberately designed differently. And I agree with @devsnek, it’s a better UX to have the same (or as similar as possible) API for Worker in both environments, if nothing else to help users remember it.

[devsnek]

Also this compat isn't theoretical, i've been able to run wasm threading demos in node just by creating globals of MessagePort and Worker from worker_threads.

[guybedford]

Here is my best shot at a single worker.mjs file that works in both browsers and Node.js:

const isNode = typeof process === 'object' && process.versions && process.versions.node;

// Give me Top-Level Await!
Promise.resolve()
.then(async () => {
  const workerThreads = isNode && await import('worker_threads');
  const fileURLToPath = isNode && (await import('url')).fileURLToPath;

  const isMainThread = isNode ? workerThreads.isMainThread : typeof WorkerGlobalScope === 'undefined' || self instanceof WorkerGlobalScope === false;

  if (isMainThread) {
    const { Worker } = isNode ? workerThreads : window;

    const worker = new Worker(isNode ? fileURLToPath(import.meta.url) : import.meta.url, { type: 'module' });

    worker.postMessage('hello');

    if (isNode)
      worker.on('message', onMessage);
    else
      worker.addEventListener('message', onMessage);

    function onMessage (e) {
      console.log(isNode ? e : e.data);
    }
  }

  // Worker
  else {
    if (isNode)
      workerThreads.parentPort.on('message', onMessage);
    else
      addEventListener('message', onMessage);

    function onMessage(e) {
      (isNode ? workerThreads.parentPort : self).postMessage((isNode ? e : e.data) + ' world');
    }
  }
});

The story works, but it took me at least half an hour to get all the details right. All in all it needs 9 branches to handle platform checks. Supporting "type": "module" was not even an afterthought in that process.

This is what I mean when I say that arguing for "type": "module" for workers as if it somehow is a compatibility issue at this point, is to ignore those 9 other conditional branches which truly do affect this workflow, while that one does not.

[addaleax]

@guybedford I wouldn’t call it a compatibility issue so much as a familiarity issue – the Worker API was written to match Node’s needs and differences from the Web platform, but it was also intended to be familiar to those who use Workers on the Web.

[guybedford]

@addaleax I am only arguing against the points made by @devsnek and @GeoffreyBooth that we have to use new Worker(..., { "type": "module" }) here in the name of browser compatibility, by clearly trying to show the API differences. I am fully behind the approach of worker_threads in that the browserify implementation of worker_threads would effectively be acting as an adapter for the browser though, exactly based on web workers API being an overall guide.

The reason I'm making this argument is because Node.js modules have a way to know the module type of every single module, which we've carefully worked out. We specially don't allow inline type information for the loader, so I don't think we should make workers an exception here. If we were to support "type": "module" for worker instantiation this introduces a new out-of-band mechanism that further complicates module interpretation, whereas the current techniques of module detection currently work perfectly fine for worker threads.

This is why my suggestion is to rely on the content-type for data URIs like we do in the module loader, or to have something like "eval": "module" specifically not to create the assumption for users that "type": "module" for workers is supported.

[devsnek]

using onmessage and a separate file for the worker (which is the common case on the web) you end up code that i would feel comfortable describing as portable. obviously one can write code that is not compatible (using isMainThread, addEventListener, etc) but it isn't required.

[aduh95]

Related https://github.com/nodejs/node/pull/34584