perf: bundle runtime dependencies

[pi0]

🔗 Linked issue

❓ Type of change

• 📖 Documentation (updates to the documentation, readme, or JSdoc annotations)
• 🐞 Bug fix (a non-breaking change that fixes an issue)
• 👌 Enhancement (improving an existing functionality like performance)
• ✨ New feature (a non-breaking change that adds functionality)
• 🧹 Chore (updates to the build process or auxiliary tools and libraries)
• ⚠️ Breaking change (fix or feature that would cause existing functionality to change)

📚 Description

Nitro output uses file tracing for all node modules by default. It has many benefits such as supporting native modules and the best compatibility with minimized rollup involvement.

However, there are downsides with externals too at least for built-in and standard unjs packages inlining is better.

• Issues with package hosting and multi-version handling
• we can optimize tree-shaking output up to file level instead of function exports
• Extra data such as package.json files are copied over the output directory
• Extra startup overhead because of node_modules resolution
• Possible wrong export condition resolutions
• Inability to tree-shake with predictable platform injections (such as AsyncLocalStorage)

This PR uses a list of nitro built-in runtime dependencies and explicitly inlines them in the final output bundle.

TODO: This method however has a downside in that inlined sourcemaps add to the bundle overhead. Looking for a way to disable inline-source maps only for node_modules.

Benchmark Results

Measurement / Mode	external deps	inlined deps
number of files	120	13
disk size (without sourcemaps)	840 KB	476 KB
start	31.823 ms	20ms
fetch	14.18 ms	14ms
start-to-fetch	47.464ms	35ms

Config:

export default defineNitroConfig({
  sourceMap: false,
  analyze: {
    filename: "./analyze.html",
  },
});

Timing bench:

console.time("start");
console.time("start-to-fetch");

await import("./b/server/index.mjs").then(async () => {
  console.timeEnd("start");
  console.time("fetch");
  await fetch("http://localhost:3000/").then((r) => {
    console.timeEnd("fetch");
    console.timeEnd("start-to-fetch");
  });
});

Node.js version: v18.16.0 on Macbook Air M2

📝 Checklist

• I have linked an issue or discussion.
• I have updated the documentation accordingly.

[codecov]

Codecov Report

Merging #1554 (4aeda33) into main (6430ded) will increase coverage by 0.19%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main    #1554      +/-   ##
==========================================
+ Coverage   76.00%   76.20%   +0.19%     
==========================================
  Files          74       76       +2     
  Lines        7711     7778      +67     
  Branches      760      773      +13     
==========================================
+ Hits         5861     5927      +66     
- Misses       1848     1849       +1     
  Partials        2        2              

Files Changed	Coverage Δ
src/deps.ts	100.00% <100.00%> (ø)
src/index.ts	100.00% <100.00%> (ø)
src/rollup/config.ts	91.48% <100.00%> (+0.26%)	⬆️
src/rollup/plugins/sourcemap-ignore.ts	100.00% <100.00%> (ø)
src/types/nitro.ts	100.00% <100.00%> (ø)

... and 1 file with indirect coverage changes

[danielroe]

In future we might need an opt out for major versions somehow? (If user, or other library is using dependency.)

Or equally if one of these is a dependency of an external we might have an issue too.

[pi0]

In future we might need an opt out for major versions somehow? (If user, or other library is using dependency.) Or equally if one of these is a dependency of an external we might have an issue too.

Yes, i was thinking to support externals.exclude (and filter-out default inlines when user/framework specifically requires opt-out for a feature) + a flag to completely disable this behavior.

Regarding major versions (or any custom version by user), we still use rollup's search path, resolution, and aliases so overriding is possible by all means, however even totday, we require them (defined list) to be in Nitro-Expected semver-major range as nitro internally uses all these deps. (if an external dependency in .output/server/node_modules/x requires defu@v1 for example and nitro itself rely on v2, we still just like today also have .output/server/node_modules/defu@1 and bundle the v2 for nitro runtime itself as external resolution is handled separately from rollup)

I want to make sure mainly it is good to you too for becoming the default behavior .

[danielroe]

It is fine to me as default behaviour. I'm observing the potential issue that if a library in externals uses ufo/pathe then this will be included twice (once in node_modules and once bundled). If other major versions of libraries like ufo are included in externals that entirely addresses my first point.

[pi0]

if a library in externals uses ufo/pathe then this will be included twice (once in node_modules and once bundled).

That's a valid point if both inline and external use same semver/major of ufo. And was initial reason that we made all externals by default. I guess we can think of some ways to automatically "deoptimize" inlining however it involves traced file dependency "pre-analysis" and sub-dependencny deptimization.

I guess for default behavior, it should benefit most of users both Nitro and Nuxt with both bundle size but more importantly the stability of resolutions with rollup comparing to current externals.

[danielroe]

Yes, I see this as a good change, too.

[pi0]

Updated: Sourcemap size issue is (kinda) resolved by using a sourcemap minify plugin that scans emited sourcemap chunks and dynamically disables the ones that include node modules (reducing basic build from 2.1 MB to 454 kB.

I think there is way more area for improving this plugin. We can tree-shake content only for example. Considering we use lazy chunks for user code, this should be relatively ok. This can be disabled with experimental.sourcemapMinify: false in case of any issues.

experimental.bundleRuntimeDependencies: false flag can also be used to disable the whole experiment.

Merging to main to test against edge about effect of this change and fine-tune until release.