Branch/tag resolution for git dependencies

[ouranos]

Do you want to request a feature or report a bug?

a bug

What is the current behavior?

We have a repository where we have a branch for the minor version (eg 1.1) and then tag releases for patch versions in this branch (v1.1.0, v1.1.1, ...)

I want to point a project the branch so I can have changes that are not released yet, if I put:

{
  "mno-enterprise-angular": "git+https://git@github.com/<org>/<project>.git#1.1"
}

It actually resolves to the last patch release (eg: v1.1.1) instead of the tip of the branch.

If the current behavior is a bug, please provide the steps to reproduce.

See above, an example is:

{
  "package": "git+https://git@github.com/maestrano/mno-enterprise-angular.git#1.1"
}

The resulting yarn.lock is:

"mno-enterprise-angular@git+https://git@github.com/maestrano/mno-enterprise-angular.git#1.1":
  version "1.1.0"
  resolved "git+https://git@github.com/maestrano/mno-enterprise-angular.git#4fd4c179f5ca315bcf32e25c2cdbfe85102a0ae1"

4fd4c17 is the v1.1.0 tag

What is the expected behavior?

It is expected to resolve to the tip of the branch.
At the moment, I specify the commit in the package.json as a workaround but I can't use yarn upgrade.

Please mention your node.js, yarn and operating system version.

• yarn v0.24.6
• node v6.11.0
• ubuntu 14.04

[bestander]

@sheerun wanted to address a similar issue #3624 (comment)

[bestander]

The feature of yarn.lock is that it guarantees repeatable builds.
I think yarn.lock should "freeze" the branch/tag to a specific commit when the lockfile is generated.

[ouranos]

Thanks @bestander.

In this case, we'd still freeze the branch to its current commit.
When doing yarn upgrade, it'd update to the new commit (if it has changed).
The problem is that it currently resolves #1.1 as a tag/version instead of a branch.

[bestander]

I see, there is still confusion, I think we need to have more fields tracking git dependencies: branch, tag, commit hash, tarball hash (if using offline mirror).

Regarding this bug, help from community members is welcome

[Volune]

Is there any documentation or specification about the algorithms in src/util/git.js?
The algorithm that resolves the commit hash from the git url is there, and seems to be confusing. And it was already there in the "first commit", so the git history is not helping.

If I understand well, the git-url-hash, if not a commit-hash, is first used as "SemVer" version constraints, then fallback as a branch/tag name.
I don't think that's an intuitive behavior, because in this case 1.1 is an exact match for the branch name, but v1.1.1 is preferred.

I suggest that we first try to match the git-url-hash against the tags and branches.
If someone needs to force the use of "SemVer" version constraints, he can add a leading space, eg: # 1.1. This should not match any branch/tags, as spaces are not valid in branch/tag names.

[bestander]

@Volune, there is no documentation, the code evolved trying to satisfy the use cases that Yarn faced, we call this "self documenting code" :)
The git resolution logic is covered with some tests here https://github.com/yarnpkg/yarn/blob/master/__tests__/package-resolver.js.

Feel free to explore this code, we can afford making a breaking change here if it is justified.

[bestander]

So what priority of these references would you set up:

(none) master => #tag => #branch => #git-hash?
That is at resolution time.

What hash should go to the lockfile?

[ouranos]

I think we should still keep the git sha1 in the lockfile. This way we keep the deterministic property when using branches.
If matching on tags, we could use the full tag as it's supposed to be an immutable reference.

For the resolution priority, I'd suggest as @Volune:
(none) master => commit-hash => tag/branch (exact match) => SemVer

[Volune]

I tested a little on NPM5, it doesn't seem to use the "SemVer" algorithm on tags.

I think we should always keep the git sha1 in the lockfile (or the tarball hash, I saw another discussion about that). The tag is only supposed to be immutable, it can still be messed with. In my opinion, using sometimes the sha1 sometimes the full tag would be both more complex and less secure.

Also I've seen some repositories not using master as their default branch. Not having a version in the git url should use that default branch. I think we can get the default branch with git ls-remote --symref <remote> HEAD.

---

No version git+proto://host/path/repo.git

• Try to use default branch
• Try to use master
• Error

With version git+proto://host/path/repo.git#version

• Try to use version as a git sha1
• Try to use version as a tag name
• Try to use version as a branch name
• Try to use version as SemVer against existing tags
• Try to use version as SemVer against existing branches -- (tests branches and tags at same time?)
• Error

[bestander]

Sounds good, @Volune.
When the PR is coming? :)
Any idea how to cover it with tests so that we don't regress?