remove the warning about performance hit in --experimental-detect-module

[mcollina]

I did some experiments with --experimental-detect-module and I've found that the added cost of "typeless" package.json is almost negligible, around 1ms on my system.

I recommend we remove the warning.

[mcollina]

cc @GeoffreyBooth @joyeecheung

[GeoffreyBooth]

It’s certainly much reduced since the initial commit where that warning was added; probably next to nothing for CommonJS files, as that’s what we’ve optimized for. I think it’s probably still a noticeable impact for ESM files that are ambiguous (like a .js file using import/export where the package.json has no type field). Did you measure that case?

[mcollina]

yes... 1ms on my machine.

[GeoffreyBooth]

Well that’s great news, I wasn’t expecting that. 1ms for what, a single file? Or an entire app, like if you had a thousand ESM files in a "type": "module" scope and you removed the type field?

I guess the question then is what are the downsides of not trying to discourage detection? If I remember correctly, the primary intent with detection was to cover two cases:

• Extensionless “loose” files like /usr/local/bin/my-node-script where the user can’t reasonably put a package.json file nearby
• Beginners writing a hello world single file app

With the thinking being that the performance hit of detection was outweighed in these cases by the UX improvement, of having the my-node-script command work or having the beginner’s code work without first needing to learn about .mjs or type. But that aside from these niche cases, most apps could and should specify the type field to avoid the performance penalty.

But if there is no performance hit, or a negligible one, then why wouldn’t everyone just rely on detection all the time? Why bother setting the type field? I feel like there must be reasons that people would want to still do so, like compatibility with build or linting tools maybe; but I don’t know concrete examples. Maybe there aren’t any, or not any compelling ones, and we shouldn’t bother trying to encourage the type field or .mjs/.cjs? It feels wrong even to suggest that, but off the top of my head I can’t think of the argument why not.

[mcollina]

I would not push our luck too much, as 1ms for 1000 modules is 1s, and I've seen apps with that many.

I would likely not emit the warning for the top level/entrypoint, but still for everything inside node_modules.

[GeoffreyBooth]

I would likely not emit the warning for the top level/entrypoint, but still for everything inside node_modules.

The problem there is that warning for node_modules is really annoying, as the code in there is generally outside of the user’s control. Like what are they supposed to do to address the warning; patch the dependency? Ask its maintainer to add the type field? Choose something else?

We discussed warning on node_modules early on and we were thinking that it wouldn’t matter so much for there because library authors would hear from users really fast if they published libraries that only work on Node versions with detection enabled by default. Eventually once detection has been enabled by default for years and the older versions of Node are EOL, that might be more of a reason to warn; though who knows what assumptions we should make by then. Maybe by 2027 we flip to attempting to parse as ESM first and fall back to CommonJS, if the ESM migration is finally over the hump by then, and so the only detection performance penalty is for CommonJS files relying on it.

[mcollina]

Then not warning at all is actually ok.

[RomainLanz]

I have run a benchmark on one of my "huge" project using ESM on Windows. Here are the results.

Node 20 with "type": "module"

Benchmark 1: node bin/server.js
  Time (mean ± σ):     728.4 ms ±  10.6 ms    [User: 87.5 ms, System: 42.2 ms]
  Range (min … max):   713.6 ms … 744.4 ms    10 runs

Node 22 with "type": "module"

Benchmark 1: node bin/server.js
  Time (mean ± σ):     739.5 ms ±  35.4 ms    [User: 78.1 ms, System: 34.4 ms]
  Range (min … max):   644.4 ms … 767.0 ms    10 runs

Node 22 with --experimental-detect-module

Benchmark 1: node --experimental-detect-module .\bin\server.js
  Time (mean ± σ):     744.8 ms ±  16.2 ms    [User: 118.8 ms, System: 37.5 ms]
  Range (min … max):   728.8 ms … 785.5 ms    10 runs

[GeoffreyBooth]

I have run a benchmark on one of my “huge” project using ESM on Windows

How many ESM files is this, and how many lines of code?

I also don’t quite know what to make of these results. The Node 20 numbers are interesting but not really relevant; I’m comparing Node 22 with and without the flag. If I wanted to boil it down to “here’s how much slower the flag makes the app,” am I comparing “Time (mean)” so 739.5 ms to 744.8 ms, or 0.7% slower? Or is it more relevant to compare the “User” numbers, 78.1 to 118.8, so 34% slower?

Or is what’s really happening here that the extra CPU cost of detection is irrelevant because the async file I/O takes so much longer? As in, without the flag the CPU is more idle than with the flag, but in the end the process takes almost exactly the same amount of wallclock time because the bottleneck is the I/O and not the CPU?

[joyeecheung]

The warning is defending against the case where there is a package.json, but it doesn’t contain “type”: “module” even though the .js modules inside are using ESM syntax. To hit the case that it is defending against, you will need to recursively remove all the “type” : “modules” entry in your package.jsons (assuming you have any, and that most of your dependencies use real ESM and count on this to work), and run it with the flag; then compare it with the case where this detection flag is unnecessary (when you still have the type field everywhere). you can’t compare it with a case where you still have package.json with type fields because without the flag that module graph would not load at all currently.

[RomainLanz]

How many ESM files is this, and how many lines of code?

I am not sure how I can count the loaded file since I am basically starting one of my application and killing it once the http server start (so it is the initialization time of the application).

I could re-run it many times to see if the number differ and if the "User" number seems related to the removal of the type in the package.json.

you will need to recursively remove all the “type” : “modules” entry in your package.jsons

Do you mean I should also remove it from all my dependencies to create a fair test?

[GeoffreyBooth]

@RomainLanz I just mean count the number of JavaScript files and lines of code in your project. So delete/move away node_modules and then run commands to count the number of .js files, and sum the lines of code. (Not sure what those commands would be for Windows but I’m sure you can Google it.)

The benchmarks that we want to compare are:

• Status quo, Node 22 with your app as normal
• Node 22 with your app’s package.json edited to remove the type field, if it was there; and with all your ESM files using the .js extension, if you had any with .mjs

Assuming you probably have only one package.json for your project, that’s all you need to do. I wouldn’t modify anything in your node_modules, as that’s not realistic that typical users would do. If you’re using Hyperfine, please include the summary lines at the end (“such and such is 2x faster than such and such”).

[joyeecheung]

Do you mean I should also remove it from all my dependencies to create a fair test?

I would say yes because removing the warning effectively encourages packages to publish without adding a type field in their package.json and that’s what we don’t want. The warning here would show up in their CI to prevent packages from accidentally publish without the type field. Otherwise it could very well be interpreted as “it’s fine to publish ESM in .js without a type field now!” And the ecosystem start to get slower as it becomes the norm and Node.js commonly has to second guess the module type.

[mcollina]

Usually, the community is pretty good at correcting those things, so I don't expect it to be a problem.