This improves peer dependency resolution yet again. We did not handle
scenarios like the following:
```
// a -> b -> c -> d -> c -> b (peer)
```
...which would maybe work ok the first time its run in some cases, but
then lead to a lockfile that would error on load.
This now keeps track of circular dependencies and updates nodes
accordingly. That said, there is still a lurking bug in this code
somewhere that I've added a comment for (there is a mitigation on the
tail end that seems to work well). The current state is much better than
before and I can look into it later. I think it's something small that's
incorrect.
This changes npm specifiers to be handled by deno_graph and resolved to
an npm package name and version when the specifier is encountered. It
also slightly changes how npm specifier resolution occurs—previously it
would collect all the npm specifiers and resolve them all at once, but
now it resolves them on the fly as they are encountered in the module
graph.
https://github.com/denoland/deno_graph/pull/232
---------
Co-authored-by: Bartek Iwańczuk <biwanczuk@gmail.com>
This PR fixes peer dependency resolution to only resolve peers based on
the current graph traversal path. Previously, it would resolve a peers
by looking at a graph node's ancestors, which is not correct because
graph nodes are shared by different resolutions.
It also stores more information about peer dependency resolution in the
lockfile.
The `NpmPackageId` struct is being renamed to `NpmPackageNodeId`. In a
future PR it will be moved down into only npm dependency resolution and
a `NpmPackageId` struct will be introduced in `deno_graph` that only has
the name and version of the package (no peer dependency identifier
information). So a `NpmPackageReq` will map to an `NpmPackageId`, which
will map to an `NpmPackageNodeId` in the npm resolution.
- Generalizes the npm version code (ex. `NpmVersion` -> `Version`,
`NpmVersionReq` -> `VersionReq`). This is a slow refactor towards
extracting out this code for deno specifiers and better usage in
deno_graph.
- Removes `SpecifierVersionReq`. Consolidates `NpmVersionReq` and
`SpecifierVersionReq` to just `VersionReq`
- Removes `NpmVersionMatcher`. This now just looks at `VersionReq`.
- Paves the way to allow us to create `NpmPackageReference`'s from a
package.json's dependencies/dev dependencies
(`VersionReq::parse_from_npm`).
I'm not sure this properly handles scenarios where an npm package uses
an alias that resolves to itself, we can fix that if we find a package
that actually depends on that behavior.
Closes #17420
If an optional peer dependency entry previously wasn't resolved and it's
now being resolved, then it will add it as if it were a dependency of
the previously resolved package instead of creating a new "copy package"
(seems to be what npm and pnpm does).
Closes #17240
Peer dependency resolution wasn't handling a peer dependency being
resolved without a dep higher in the tree and then with one being found
higher in the tree.
1. There was a lot of cloning going on with `NpmPackageInfo`. This is
now stored in an `Arc<NpmPackageInfo>` and cloning only happens on the
individual version.
2. The package cache is now cleared from memory after resolution.
3. This surfaced a bug in `deno cache` and I noticed it can be more
efficient if we have multiple root specifiers if we provide all the
specifiers as roots.
This adds support for peer dependencies in npm packages.
1. If not found higher in the tree (ancestor and ancestor siblings),
peer dependencies are resolved like a dependency similar to npm 7.
2. Optional peer dependencies are only resolved if found higher in the
tree.
3. This creates "copy packages" or duplicates of a package when a
package has different resolution due to peer dependency resolution—see
https://pnpm.io/how-peers-are-resolved. Unlike pnpm though, duplicates
of packages will have `_1`, `_2`, etc. added to the end of the package
version in the directory in order to minimize the chance of hitting the
max file path limit on Windows. This is done for both the local
"node_modules" directory and also the global npm cache. The files are
hard linked in this case to reduce hard drive space.
This is a first pass and the code is definitely more inefficient than it
could be.
Closes #15823
