Divergent specifier hazard

[GeoffreyBooth]

@jkrems and I created a repo to demonstrate the hazard mentioned by @MylesBorins in #273 (comment) and #273 (comment) where a single specifier resolves to separate things in ESM and CommonJS environments within the same runtime. This hazard is currently blocking dual packages, and logically would also block --es-module-specifier-resolution=node from becoming the default behavior.

Hopefully the repo illustrates the issue clearly enough that everyone can get a solid grasp on it. It seems to me that we have three options:

1. Accept that the hazard is a serious concern and that it therefore rules out supporting dual packages and automatic extension resolution. The current modules implementation is what ships with regard to those two features, and presumably we would remove the --es-module-specifier-resolution flag (since arguably the flag’s existence invites users to stumble into the hazard).

2. Acknowledge the hazard but decide that user education can overcome it. @MylesBorins and others concerned about the hazard would need to be persuaded as to why it isn’t a blocker, especially considering that it’s already come up in the real world. If the persuasion effort succeeds, we would then need to decide how far to lean into enabling features that invite the hazard (such as automatic extension resolution and dual packages).

3. Find a technical solution to prevent the hazard, such as making require of ESM work in a way that everyone is comfortable with (cc @weswigham). As with the previous option, we would then need to decide whether or not to support dual packages and/or automatic extension resolution.

Deciding on any of these options would lead to a resolution of the dual packages item in Phase 3 of our roadmap.

[weswigham]

I, for one, would really like to just resolve the same specifier to the same module in both resolvers - that would go a long way to making esm feel like a gradually adoptable improvement, rather than a hard break - I really hate the idea of leaving cjs users in the cold with respect to new esm libraries (and no, dynamic import isn't practical in the real world because the promise wrapper forces an unrealistic code structure compared to how code is written and consumed today).

It's worth noting that a similar mechanism to what I've looked at to support require of esm could in theory be used to support top level await in cjs without creating excessive event loop turns (just like how the TLA spec specifies esm TLA should behave, with the same caveats about possible deadlocks). I started looking at that, but am suuuuper unfamiliar with how the cjs wrapper is formed and, from what I can tell, there's currently no equivalent of the v8 create function call that supports marking the function as async, which would mean degrading to the older style wrapper (as is done when the wrapper is patched).

[weswigham]

It's worth noting that I also think the same-specifier-resolves-to-the-same-thing-in-both-resolvers also isn't actually predicated on require in esm - it's just a matter of the cjs loader prioritizing esm files in the same way as the esm loader and throwing if an esm file is found first (barring actually being able to import it). In effect, the cjs loader should be a superset of the esm loader in this scenario, except it throws when the module that is resolved to is esm.

In the pkg example, that means if pkg has a main of ./x-core in and both ./x-core.js and ./x-core.mjs exist, both loaders must always pick the same resolution. So if .js is higher priority (as would be backwards compatible and wouldn't risk picking up accidental new mains), then in both resolvers it resolves to the commonjs file. If .mjs were higher priority, then the esm entrypoint would be all both the esm and cjs loaders saw (and would, therefore, allow a silent fallback to the cjs entrypoint on older node versions which don't recognize the mjs extension).

It just comes down to both loaders need to resolve a specifier to the same thing, whatever tweaks that takes.

[AlexanderOMara]

I've mentioned this in another issue, but I think it should be mentioned here.

If two separate packages require/import a third package, they should not assume they will both get the same singleton. This is a dubious assumption, which can already lead to issues even without getting into ESM vs CJS.

With NPM, version conflicts are resolved by giving each package a version it requires, even if that means installing two different versions. With a current version of NPM, a different copy will be given to one of the packages to import, and I believe older versions would simply have installed two copies regardless of version.

That means that if package-a has dependency third-package@2.0.0 and does:

const {SomeClass} = require('third-package');

And package-b has dependency third-package@1.0.0 and does:

const {SomeClass} = require('third-package');

Both are going to get a different SomeClass object, and attempts to use instanceof for objects created outside the package will not work out. Likewise, attempting to modify the class like a pseudo-global object also won't work out.

This isn't a new issue, and I don't think that it's a compelling reason to not support dual packages. If a package wants to be a dual package, it should move to not assuming there will only be one version of a package... actually, all packages should do this anyway.

[MylesBorins]

@AlexanderOMara the issue is not only across package boundaries but within the same package if you are moving between CJS and ESM (which is likely to happen as teams transition code bases).

Packages like react which have global state that is shared will break in very mysterious and hard to debug ways if you end up with a different singleton in ESM + CJS.

[ljharb]

@MylesBorins that's true, but that's already the case with duplicates or different versions. The solution is peer deps for "across package boundaries"; there's no solution node can provide that would be any different (npm or another package manager could perhaps provide one; bower for example's solution was "pick which one you want and that's the only one you'll get", but that was untenable for many use cases).

Within the same package, I don't see why that's a hazard we're concerned with; that seems like something that package needs to test for?

[weswigham]

As you said in the other thread - without a package, the issue can be fixed with peer dependencies - within a package, it's pretty much impossible as long as two side by side "seperate but equivalent" builds of the package are there (currently). Forcing only one implementation to be active (for all usages) would be my prefered solution~

[GeoffreyBooth]

It’s worth noting that I also think the same-specifier-resolves-to-the-same-thing-in-both-resolvers also isn’t actually predicated on require in esm - it’s just a matter of the cjs loader prioritizing esm files in the same way as the esm loader and throwing if an esm file is found first (barring actually being able to import it). In effect, the cjs loader should be a superset of the esm loader in this scenario, except it throws when the module that is resolved to is esm.

I remember asking @guybedford about this when I was struggling to get dual packages to work, and trying to find a workaround for this hazard. My suggestion was basically “well once 'pkg' is loaded as one module type, when 'pkg' tries to get loaded as the other module type can’t we just throw an exception?” And the answer was basically no, not if we want Node to remain deterministic, which apparently we do. I argued that since ESM is statically analyzable, wouldn’t the CommonJS versions of any specifier always be loaded second? But he said no, because of dynamic import().

He can surely explain it better than I can, but assuming that Node remaining deterministic is a requirement of the project, then I think throwing on this hazard isn’t a solution.

It just comes down to both loaders need to resolve a specifier to the same thing, whatever tweaks that takes.

This is the current behavior of the --experimental-modules implementation, at least under the default --es-module-specifier-resolution=explicit. And if that “same thing” is a CommonJS thing, then it works great; it’s only when you want to make that “same thing” ESM that everything gets tricky, since both require and import support CommonJS sources while only import supports ESM. This is perhaps one of the strongest arguments for supporting require of ESM, so that ESM has parity with CommonJS in terms of support within Node’s module systems.

@weswigham would it be possible to produce a PR that implements require of ESM, at least for relative files with explicit extensions, e.g. require('./file.mjs')? Just to demonstrate that it actually works, and let people see what the tradeoffs are. Then we’d need to figure out how to define separate package entry points for CommonJS versus ESM, but that’s obviously a solvable problem. I don’t know if the group will ever go for optional extensions, since there were objections to that besides this hazard, but I think this hazard was the only objection raised against dual packages.

[weswigham]

I remember asking @guybedford about this when I was struggling to get dual packages to work, and trying to find a workaround for this hazard. My suggestion was basically “well once 'pkg' is loaded as one module type, when 'pkg' tries to get loaded as the other module type can’t we just throw an exception?” And the answer was basically no, not if we want Node to remain deterministic, which apparently we do.

You seem to misunderstand - basing which you get on how it's first imported is the issue there. Just dont do that. (None of the logic in the esm resolver at least is dependent on the parent module kind, and neither is the cjs core algorithm, nor should it be). The key is just to make it resolve to the same thing every time, regardless of how it is accessed.

[weswigham]

@weswigham would it be possible to produce a PR that implements require of ESM, at least for relative files with explicit extensions, e.g. require('./file.mjs')? Just to demonstrate that it actually works, and let people see what the tradeoffs are.

There was a branch with that already linked in the proposal issue. :S