Using git for team packages
- git workflows
- Team workflow
- Source-full or source-less branches?
- Where do the team's git branches live?
- Transition plan
- Use cases
- More information and resources
Using git for team packages
This page is UNOFFICIAL
As of this writing (2014-08-22), the Debian/Python team still officially uses Subversion to manage all team packages. The contents of this page are PROPOSED as a team policy and recommendations for using git to manage team package. It is still under active discussion on the debian-python mailing list. Please provide your feedback there.
There are many advantages to team-based package development. The DPMT and PAPT have very excellent, knowledgeable, and experienced Python and Debian developers and our goal is to bring the very best Python experience to Debian users and developers. We're here to help each other and provide consensus (if not 100% agreement) on best practices. Collaboration has many benefits, and one of the things that aids effective collaboration is common tools and workflows surrounding packaging, while still providing individual maintainers the freedom to adapt those practices to their preferences, and the peculiarities of the packages.
One of the perks of being in a team is that all of our packages are managed in a similar way, using similar tools and workflows. This is great because any team member can immediately get the source of a package to do an update or a bug fix, and they don't have to first try to figure out how the package is version controlled. Plus, we can have nice wiki documentation which help new team members how to do things, and almost anyone on IRC can immediately give you advice. We use a common set of vcs procedures to ensure consistency among all of our packages. Both the DPMT and PAPT teams use the same version control system, layout, and workflows.
It is a team requirement that all packages be managed using the same version control system. As of this writing, that vcs is Subversion (svn).
We recognize however, that modern vcses, especially of the distributed variety, provide a lot of additional benefit for packagers. Such vcses include git, Mercurial (hg), and Bazaar (bzr), among others. git seems to be the most popular, so we are investigating using it as a possible replacement for svn for team managed projects.
We propose to relax this requirement to allow for experimentation with git-based team packages, but will eventually re-mandate one-dvcs-to-rule-them-all. I.e. an official switch to git at some point in the future.
There are two predominant git-based workflows in use within Debian for package management, git-dpm and git-buildpackage. While there is some overlap and similarities, there are some subtle differences both technically and in ease-of-use that makes it useful for the team to mandate one or the other regimes. You can gain some background on their applicability to team packages by reading these threads in the mailing list: git-dpm and git-buildpackage. dgit is an interesting possibility, since its repositories reflect the exact representation of the package in the Debian archives.
git-bp and git-dpm both impose some restrictions and/or defaults on branch names, but you have some leeway. However, as a team we are strongly recommending you use the default branch names, specifically:
master - The Debianized upstream source directory. IOW, this contains upstream source and a debian/ packaging directory.
pristine-tar - Contains the standard pristine-tar deltas.
upstream - The un-Debianized upstream source. This is what you get when you unpack the upstream tarball.
If you use other branch names, please have a good reason (not just personal preference, remember - take one for the team!), and you must document the differences in debian/README.source. The default branch names will help the repo interoperate better with other tools.
Both git-dpm and git-bp create a patch branch when you want to directly edit upstream source files for export to quilt patches, however git-dpm deletes this patch branch when you export back to the master branch. git-bp keeps its patch branch.
Q: What should we recommend/mandate about the patch branch? Some possibilities include: use a specific name, defer to the helper tool, or prohibit patch branches from being pushed to the public repos.
Q: Do we even need to pick one? Based on discussions in the mailing list, it may not be necessary for the team to mandate a particular helper. This is only possible if they interoperate seamelessly. E.g. let's say Alice and Bob both work on python-foo. Alice uses git-bp and Bob prefers git-dpm. Ideal would be if they can both work on python-foo using the tools they like, but once it's pushed and shared, it doesn't matter which tool is used. Thus branch names, tag names, and possibly other artifacts must agree, regardless of which tool they use.
Further, perhaps Carol comes along and just wants to use git and quilt without a fancy helper. She can do whatever she wants locally, but once her changes are committed and pushed to the team repository, the artifacts must be compatible with git-bp and git-dpm.
Source-full or source-less branches?
Source-less (or debian/-only) closely mirror the team conventions under svn, but it seems counter-productive to have source-less branches under git. We require team maintained packages to contain the full upstream source, and we don't believe it will be an overwhelming performance or disk-space burden to do so. It is much easier to hack on packages, including generating the quilt patches, when a checkout of the repo gives you the full source. Source-less branches defeat much of the benefit of managing packages in git.
Where do the team's git branches live?
Alioth, under the python-modules directory. To set up a new repository here, you should follow these instructions. They require ssh access to Alioth, but if you have push permissions, you should have ssh permissions. Ask a team member for help if you are unable to do this. These instructions set up the hooks for commit notifications and such:
$ ssh git.debian.org $ cd /git/python-modules $ ./setup-repository <srcpkgname> "<srcpkgname> packaging" $ exit
Of course, use your own source package name for <srcpkgname>.
You might want to make one small change after you create the bare repo on g.d.o. By default, the setup-repository script sets the HEAD symbolic ref to refs/heads/debian/sid and if your branch does not have a debian/sid ref, as by default git-dpm repos will not, you will get a warning when subsequently cloning the repo. If you get a warning such as "remote HEAD refers to nonexistent ref", the warning isn't fatal, but to prevent it, go back to g.d.o and cd into /git/python-modules/packages/your-package.git and do: git symbolic-ref HEAD refs/heads/master -- or whatever your master branch name is.
We will still require all team packages to be available on git.debian.org. Please do not host your team maintained packaging branches anywhere else; e.g. github is not appropriate for the packaging branches because it is not free software. Q: should we honor pull requests and allow for mirrors on github?
Use these Vcs-* headers in your debian/control file:
Vcs-Git: git://anonscm.debian.org/python-modules/packages/<srcpkgname>.git Vcs-Browser: http://anonscm.debian.org/cgit/python-modules/packages/<srcpkgname>.git
It seems unfeasible to mass-migrate all of the teams packages from svn to git in one fell swoop. The plan is then to migrate packages opportunistically, with a hard deadline sometime way in the future (6 months? 1 or 2 years?) after which all remaining packages will be mass migrated. The rough steps to migrate the team would then be:
- Pick a git regime (maybe), e.g. git-dpm or git-buildpackage.
- Relax the constraint that all packages must be under one vcs. All new packages should be under git. Allow maintainers to migrate packages from svn to git whenever is convenient for them (e.g. when preparing a new upstream).
- After the grace period, migrate remaining svn packages en mass.
Here are some common scenarios and use cases. There may be other ways to accomplish these tasks so if you find a more efficient way of doing something, please update these pages and discuss with the team via the debian-python mailing list. Feel free to add more use cases here too! Your experience is valuable.
Creating a new package
First, Initialize the package directory on git to set up the bare repo (with hooks installed) that you will push your branch.
- Download your package's upstream original tarball
tar -xvf srcpkgname_1.0.orig.tar.gz
git add .
git commit -m "import srcpkgname_1.0.orig.tar.gz"
git checkout -b upstream
pristine-tar commit ../srcpkgname_1.0.orig.tar.gz upstream
git-dpm init ../srcpkgname_1.0.orig.tar.gz
You'll now be left on the master branch, i.e. the packaging branch. You'll also have upstream (raw tarball contents) and pristine-tar branch.
Fill in the rest of your debian/ directory, then do this:
git add debian/*
debcommit # assuming you have a d/changelog entry, otherwise git commit
You can now build your package using whatever tools you like, including debuild, dpkg-buildpackage, and git-buildpackage.
Push your repo to git.debian.org:
git push --set-upstream ssh://git.debian.org/git/python-modules/packages/<srcpkgname>.git --all
At this point, I like to cd to a different directory and do a git clone ssh://git.debian.org/git/python-modules/packages/<srcpkgname>.git and then continue working from the <srcpkgname> directory. To prove that all the branches got pushed correctly, in this fresh clone, checkout the master, pristine-tar, and upstream branches.
Building and tagging
Once you've built and uploaded your package, you should tag the release.
git push --tags
Q: should we mandate git-dpm style tag names or git-bp style tag names? Let's work with those projects to make them more consistent, or at least provide options.
New upstream release
Sponsoring, mentoring, reviewing
- pull requests
More information and resources
Recommended/required configuration settings
Both git itself and git-bp provide ways to configure defaults. Settings that are personal to you should live only in $HOME/.gitconfig or $HOME/.gbp.conf. Settings which affect how the team interoperates should live in <repo>/debian/gbp.conf so that anyone who checks out the branch will see the same settings.