Skip to content

[Feature Request]: Support resolve the modules with hard-links #410

Open
@colinaaa

Description

@colinaaa

What problem does this feature solve?

pnpm creates hard links from the global store to the project's node_modules folders.

For example, imagine you have the following directory structure: packages/a and packages/b have the same version of lodash as a dependency. But they are different symlink that points to different files.

.
├── node_modules
│   ├── a -> ../packages/a
├── package.json
├── packages
│   ├── a
│   │   ├── index.ts
│   │   ├── node_modules
│   │   │   ├── b -> ../../b
│   │   │   └── lodash -> ../../../node_modules/.pnpm/[email protected]/node_modules/lodash
│   │   └── package.json
│   └── b
│       ├── index.ts
│       ├── node_modules
│       │   └── lodash -> .pnpm/[email protected]/node_modules/lodash
│       ├── package.json
│       ├── pnpm-lock.yaml
│       └── pnpm-workspace.yaml
├── pnpm-lock.yaml
└── pnpm-workspace.yaml

Since two lodash/lodash.js points to different files, both rspack(or oxc internally) and webpack(or enhanced-resolve internally) treat the two files as unrelated. So at build time, two copies of lodash code will be packaged in the output bundle, even if their content is the same.

However, it can be found that in the file system, the two lodash.js point exactly to the same inode node. Which means they are the same file in physical storage.

$ ls -i packages/a/node_modules/lodash/lodash.js
2249211 packages/a/node_modules/lodash/lodash.js

$ ls -i packages/b/node_modules/lodash/lodash.js
2249211 packages/b/node_modules/lodash/lodash.js

To solve the problem, we created an enhanced-resolve plugin to cache the request:

import fs from 'node:fs'

import type { ResolveRequest, Resolver } from 'enhanced-resolve'

class InodeWebpackPlugin {
  static #source = 'resolved'

  apply(resolver: Resolver) {
    resolver
      .getHook(InodeWebpackPlugin.#source)
      .tapAsync('INodeCachePlugin', (request, _, callback) => {
        if (!request.path) {
          return callback()
        }

        try {
          const { ino } = fs.statSync(request.path)

          const cachedRequest = this.#cache.get(ino)
          if (cachedRequest) {
            // Note that the query may change for the same path
            // with a different query.
            Object.assign(request, { ...cachedRequest, query: request.query })
            return callback()
          }

          this.#cache.set(ino, request)
        } catch {
          // explicitly ignore error
        }

        return callback()
      })
  }

  #cache = new Map<number, ResolveRequest>()
}

And it works as expected in webpack, the duplicated modules have been eliminated.

So I wonder if this could be added to enhanced-resolve(just like the symlinks options).

Related issue: web-infra-dev/rspack#5912

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions