-
Notifications
You must be signed in to change notification settings - Fork 2.6k
Releasing Git for Windows
The release process of Git for Windows is ever-evolving. For now, it consists of these steps:
- Making sure that there are no unaddressed issues
- Rebasing Git for Windows' patches
- Opening a PR to kick off a PR build (and waiting for it to succeed)
- Kicking off the
/git-artifacts
slash command - Verifying that the resulting installer works
- Kicking off the
/release
slash command that publishes the release - Pushing directly to
main
to close the PR and set the stage for the Azure Pipeline (at some stage during the-rc<N>
cycle)
Note: the hardest part is traditonally preparing -rc0
of every major release.
There are essentially three-and-a-half venues where Git for Windows bugs are reported:
- The Git for Windows bug tracker at https://github.com/git-for-windows/git/issues
- The Git for Windows mailing list at https://groups.google.com/forum/#!forum/git-for-windows
- The Git mailing list at https://public-inbox.org/git/
- Twitter. Yep: ugh. Git for Windows is actually that popular that a couple of its users seriously think that Twitter would make for a perfectly fine way to report bugs. As if 280 characters were enough to make for a good bug report. Therefore, sometimes, when the Git for Windows maintainer feels particularly masochistic, they go to https://twitter.com/search?q=%23git-for-windows and see whether there is any valid bug report (one really has to pick out the needles between all that hay).
What if, say, a new Git Credential Manager Core version needs to be snuck in at the last minute (where "last minute" is anything within the past 24h)?
To integrate such an updated component, depending on the component type (MINGW or MSYS?) either the Build and publish MINGW Pacman package
or the Build and publish MSYS Pacman package
Pipeline need to be triggered (specifying the component name via the package.to.build
variable at queue time).
Once the component has been built and uploaded to Git for Windows' Pacman repositories (i.e. once that triggered Pipeline finished), the sync
workflows (32-bit and 64-bit) must be triggered, manually. These workflows are responsible for installing the updated Pacman packages into the git-sdk-64
and git-sdk-32
repositories.
Note: Historically, we used the artifacts of the git-sdk-64-extra-artifacts
and the git-sdk-32-extra-artifacts
Pipeline Pipelines that were triggered automatically by pushes to git-sdk-64
and git-sdk-32
, respectively, and we had to wait for them to complete before proceeding. As of June 17th, 2022, this is no longer necessary.
Note: this section assumes that the reader is very familiar with interactive rebases, in particular with the --rebase-merges
variant thereof. Readers without much experience in this are highly advised to read up on https://git-scm.com/docs/git-rebase#_rebasing_merges before going any further.
The trick is not just to run /usr/src/build-extra/shears.sh --merging --onto <version> merging-rebase
, it is even more important to verify that the result makes sense. To make sense, the result:
-
Should not have any
fixup!
/squash!
commits -
Should have the already-submitted and easily-submittable patches contained within the
ready-for-upstream
"sub thicket" (i.e. the patches/patch series that are merged at the top, via theMerge branch "ready-for-upstream"
commit. -
In general, the commit topology should reflect the structure
- Ready for upstream
- Still in flux/not yet cleaned up enough
- Never to be contributed to Git (e.g. Git for Windows' very own adjustments to README.md)
-
Should be verified via
git range-diff
to the previous release, to make sure that no important patches were lost.The Git for Windows maintainer likes to use a custom alias for that:
[alias] gfw-range-diff = "!sh -c 'a=; while case \"$1\" in -*) a=\"${a:+$a }$1\";; *) break;; esac; do shift; done; git range-diff --creation-factor=95 $a \"$1^{/^Start.the.merging-rebase}..$1\" \"$2^{/^Start.the.merging-rebase}..$2\"' -"
With this alias, one can conveniently generate that range-diff. For example, after rebasing Git for Windows to v2.28.0,
git gfw-range-diff v2.27.0.windows.1 HEAD | clip.exe
would generate the range-diff and put it into the clipboard, ready to paste into the PR. -
To make extra certain that nothing important has been lost, call
git diff <git-version>..<git-for-windows-version> >prev
for the previous Git version, thengit diff <new-git-version>.. >cur
and compare withgit diff --no--index prev cur
, looking for lines starting with two diff markers (/^[-+][-+]
inless.exe
).
Note: internally, the shears.sh
script runs git rebase --rebase-merges --interactive
with one quirk: a first line is injected into the todo list that creates a "dummy" merge commit: it is a merge that pulls in the pre-rebase commits, but it uses -s ours
so that the tree remains unchanged relative to the onto commit. The effect is that the patches got rebased, and the end result still fast-forwards from the previous tip of the main branch.
This is easy: just push a branch to your fork and then go to https://github.com/git-for-windows/git/pull/new/main and select that branch from your fork.
The harder part is to include the range-diff (with --creation-factor=95
; see the gfw-range-diff
alias in the previous section) and discuss it.
To make it easier to see what happened to patches that are no longer part of Git for Windows' branch thicket, it is highly advised to use a variant of the shell script used in https://github.com/git-for-windows/git/pull/2626, and to perform a similarly thorough analysis, especially for any -rc0
version:
$ x="$(git range-diff -s origin/main^{/^Start.the}..origin/main v2.27.0..v2.28.0-rc0 |
sed -n 's/^[ 0-9]*: \([0-9a-f][0-9a-f]*\) [=!] [ 0-9]*: \([0-9a-f][0-9a-f]*\).*/-e "s\/\1\/\1 (upstream: \2)\/"/p')"
$ git gfw-range-diff origin/main HEAD |
sed -e 's/^[ 0-9]*: [0-9a-f]* [=!]/ &/' \
-e 's/^[ 0-9]*: [0-9a-f]* </-&/' \
-e 's/^[ 0-9]*: [0-9a-f]* >/+&/' |
eval sed ${x:-\'\'} |
clip.exe
For convenience, this can be automated via this alias:
[alias]
pre-release-pr-range-diff = "!set -x && x=\"$(git range-diff -s origin/main^{/^Start.the}..origin/main origin/main^{/^Start.the}^..HEAD^{/^Start.the}^ | sed -n 's/^[ 0-9]*: \\([0-9a-f][0-9a-f]*\\) [=!] [ 0-9]*: \\([0-9a-f][0-9a-f]*\\).*/-e \"s\\/\\1\\/\\1 (upstream: \\2)\\/\"/p')\" && git gfw-range-diff origin/main HEAD | sed -e 's/^[ 0-9]*: [0-9a-f]* [=!]/ &/' -e 's/^[ 0-9]*: [0-9a-f]* </-&/' -e 's/^[ 0-9]*: [0-9a-f]* >/+&/' | eval sed ${x:-\\'\\'} | clip.exe"
When a new version is tagged in upstream git, the "Monitor Component Updates" pipeline in Git for Windows creates an issue for the version (example). For tracking & automation purposes, you should:
- link the issue to the pull request created in the previous step
- add the issue to the current "Next release" milestone (example)
Add a PR comment with the slash command /git-artifacts
. This will kick off several GitHub workflow runs in https://github.com/git-for-windows/git-for-windows-automation: one to create the tag and adjust the release notes, and then one each for every CPU architecture supported by Git for Windows generating:
- The installer
- The Portable Git
- The
.tar.bz2
archive - The MinGit
- The BusyBox-based MinGit
- The NuGet packages (x86_64-only)
The idea here is to download the Git-<version>-64-bit.exe
artifact from the workflow run, install it, and run through the "pre-flight check list" at https://github.com/git-for-windows/build-extra/blob/HEAD/installer/checklist.txt.
This one is really easy (as long as nothing is broken...): add a PR comment with the /release
slash command.
Note: The pacman
upload always takes this long.
Sadly, things are broken a lot. In those cases, the logs have to be analyzed, and the GitHub workflow definition needs to be edited (on the release
branch of https://github.com/git-for-windows/git-for-windows-automation), and the failing jobs of the workflow run have to be rerunExamples for failures that happened in the past:
-
Timeouts while uploading the GitHub Release. In that case, the partially-populated draft release has to be deleted manually (first delete the assets, or it won't delete the release), and then re-deploy.
-
Outage of github.com. In that case, simply re-deploying the stage that pushes the commits was sufficient.
-
Something was pushed to the
main
branch in the meantime, and the Release Pipeline could not handle that. In this instance, the Pipeline was edited togit pull && git push
if thegit push
failed. More complex scenarios might involve pulling a branch from a personal fork, e.g. if merge conflicts had to be resolved.
Pushing directly to main
to close the PR and set the stage for uploading the release as a new snapshot
This step is trivial: git push origin <branch>:main
where <branch>
is something like rebase-to-v2.37.0
.
This will trigger another "Git artifacts" run, which will figure out that there is a GitHub Release for that commit, download those artifacts, then trigger a run of the Snapshots
Release Pipeline, which will then upload the snapshot.
It is important to wait with pushing to main
until there is a GitHub Release, otherwise the "Git artifacts" Pipeline would build another set of artifacts and upload those, but we do want to use the same artifacts as were uploaded to GitHub Releases.
Note: The idea is to push to main
relatively soon after the Release Pipeline finished, to keep the Pacman repository, the snapshots and the main
branches as aligned as possible.
Outside of the -rc<N>
phase, just follow the process as described above. This will publish one of those (2)
versions, e.g. v2.24.1(2). Typically this happens very close after an official release e.g. when really serious bugs crept into Git for Windows unnoticed, such as Git GUI crashing before even showing the window.
In the hypothetical case that a quick-fix release is necessary during the -rc<N>
phase, after main
was pushed, i.e. needing to branch off of an earlier revision of main
, all the above steps can be performed with ease (as the Pipelines do not run directly on main
, but instead require use.branch
to be set). The only tricky part is pushing the result to main
. The recommended way is to use the next merging rebase to pull in the changes by inserting another fake merge before the one inserted by shears.sh
.