Discuss
  • Are ES Modules garbage collected? If so, do they re-execute on next import?

Are ES Modules garbage collected? If so, do they re-execute on next import?

# #!/JoePea (4 years ago)

I am curious: can modules be garbage collected if the exports are not references by anything anymore? And if so, will the module be re-evaluated the next time it is imported?

I haven't tried an experiment to answer this yet. I'll be back to post findings if someone doesn't post an official answer first.

I'm thinking about code longevity. For example, if we make long-running web-based applications with many routes and features (for sake of example imagine a desktop environment, or a MMORPG game, with apps or components that are loaded within the same context). Over time, if imports are not collected, then it means we have a memory leak.

Imagine, for example, an infinite-universe MMORPG where you can land on different planets where the code for features of a planet are provided by third parties as ES Modules. I know, this might not be a safe idea to import any code into an app, but just imagine it for sake of example (imagine we have a continuous integration system to test and verify code security, or something, before that code is allowed to be consumed in the app). Imagine you play this app for many many days, and visit many places, and you leave the app running the whole time (because farming for resources is disabled if the app is not running, or something).

I would imagine that we want unused modules (when we leave a planet, for example) to be (destroyed) garbage collected so that we don't waste memory.

#!/JoePea

# Gus Caplan (4 years ago)

Modules in the spec are cached by specifier by modules that import them. Modules in major implementations are additionally cached for the entire realm by absolute URLs. I would say that for actual code (functions and classes and whatnot) leaks aren't really a problem. Even if you import a ton of levels, that's not that much memory. The main concern I've seen raised is JSON modules, where once you import them the JSON object, which can be quite large, will never be collected. Of course, there is a simple solution to that (fetch) so it isn't a world ending problem.

# Isiah Meadows (4 years ago)

Just to expand on that, if the module record itself is dereferenced (like if it's evicted from the cache somehow), then yes, it should be collected as appropriate. However, I'm not aware of any major implementation that offers that functionality.

Isiah Meadows contact at isiahmeadows.com, www.isiahmeadows.com

# Andrea Giammarchi (4 years ago)

even if dereferenced, a dynamic import could re-reference it any time, and I would expect it to still be the same module, it'd be a surprise otherwise (cached things, same namespace checks, etc).

# Isiah Meadows (4 years ago)

That's part of the caching I'm referring to. And if the cache entry for it has been evicted, I would not expect it to necessarily return the same instance, consistent with the behavior with require and require.cache in Node (and similar with most other module loaders that support cache eviction).

Isiah Meadows contact at isiahmeadows.com, www.isiahmeadows.com

# Mark S. Miller (4 years ago)

No, definitely not. The table from specifiers to module instances is indexed by specifiers. Specifiers are strings, so this table is not weak. It is not a "cache" in the sense that it is allowed to drop things. Rather it is a registry of module instances. Only a registry as a whole can be gc'ed, and which point that context is no longer around for instantiating or reinstantiating modules.

As you suggest, if it could drop things because of GC that it would then need to regenerate, that would expose the non-determinism of gc. That would be a big deal. We carefully designed WeakMaps so that gc was non-observable. WeakMaps introduce no observable non-determinism. WeakRefs alone expose the non-determinism of gc, and are kept well quarantined from the rest of the language.

# #!/JoePea (4 years ago)

How can we ensure that long-running applications (even if theoretical), that may load and unload an unlimited number of new modules over time (f.e. routes in a web page specified by 3rd parties as time progresses), not leak memory?

Even if it is theoretical, I don't like the thought of something that only ever allocates memory that will never be freed.

Is someone working on a solution for this?

#!/JoePea

# Mark S. Miller (4 years ago)

Only a module registry as a whole may be GCed. During the lifetime of any one module registry, it can only grow. No other solution is possible.

Btw, I remember being surprised ages ago when the same issue came up for the Java ClassLoader. A classLoader holds on to all the classes it ever loaded. Each class holds onto its classLoader. Each instance holds on to its class. During the lifetime of a classLoader or any of its classes, the graph of that classLoader and its classes can only grow new classes. Not until the classLoader and all of its classes are unreachable at the same time can any of them be collected. This was equally unfortunate, surprising, and inescapable.

# Guy Bedford (4 years ago)

Node.js in the CommonJS loader and dynamic loaders like SystemJS have supported module unloading for many years by permitting eviction from the loader registry. Evicting the full parent tree was the traditional reloading workflow in SystemJS, but live binding pushes are also permitted in SystemJS as well for this.

I agree there are issues and invariants of course to consider from a theoretical perspective, but those are decisions to be made in terms of what invariants are valued, and I don't feel they are necessarily absolute constraints. These decisions should be made based on what is best for the JS users and engines. Not that I feel strongly this should be a requirement but that it should still be open to consideration.

I'm not sure it was Allen's intention to ban any concept of reloading modules when defining the idempotency requirement for the host resolve function. Perhaps he could speak to that if he's around.

# Andrea Giammarchi (4 years ago)

FWIW explicit eviction is not only fine, if you own the code that does that, but the only way I can code cover 100% all the branches of my libraries. The issue here is the untrusted Web, where I'd never expect any 3rd parts library to evict a module I am using within my code ... like ... ever.

Accordingly, I think Allen, if it was Allen, made the right call for ESM.

# Andrea Giammarchi (4 years ago)

On a second thought ... couldn't import.meta.cache, or something similar, be implemented in NodeJS only?

# Herby Vojčík (4 years ago)

On 16. 7. 2020 15:23, Guy Bedford wrote:

Node.js in the CommonJS loader and dynamic loaders like SystemJS have supported module unloading for many years by permitting eviction from the loader registry. Evicting the full parent tree was the traditional reloading workflow in SystemJS, but live binding pushes are also permitted in SystemJS as well for this.

I agree there are issues and invariants of course to consider from a theoretical perspective, but those are decisions to be made in terms of what invariants are valued, and I don't feel they are necessarily

I understand there may be problems with static imports that are statically analyzed, and there "only GC module loader as whole" makes sense, but maybe, it would be possible to treat dynamically loaded modules different way? Just thinking aloud.

# Herby Vojčík (4 years ago)

On 16. 7. 2020 15:23, Guy Bedford wrote:

Node.js in the CommonJS loader and dynamic loaders like SystemJS have supported module unloading for many years by permitting eviction from the loader registry. Evicting the full parent tree was the traditional reloading workflow in SystemJS, but live binding pushes are also permitted in SystemJS as well for this.

I agree there are issues and invariants of course to consider from a theoretical perspective, but those are decisions to be made in terms of what invariants are valued, and I don't feel they are necessarily

I understand there may be problems with static imports that are statically analyzed, and there "only GC module loader as whole" makes sense, but maybe, it would be possible to treat dynamically loaded modules different way? Just thinking aloud.

# Isiah Meadows (4 years ago)

For the web, one could lock it down with CORS-like restrictions, too, so a script from one domain isn't allowed to modify cached entries from another domain. It's not insurmountable.

This is part of why I feel the invariant should be removed and the spec should just internally resolve static imports to dedupe them and then resolve them accordingly when executing them, rather than consulting external caches every time. (It's also easier for the engine to handle, assuming they don't do this already.)

Isiah Meadows contact at isiahmeadows.com, www.isiahmeadows.com

# Guy Bedford (4 years ago)

The spec and implementation are exactly defined like this - v8 directly provides HostResolveImportedModule exactly to the spec, returning a module object for a given specifier and module parent. By default in v8 modules are GC'able objects... they don't implement registries currently. There's just some work needed there to work out why, for example, in Node.js they aren't possible to GC.