- Initial design
When do we delete merged Git branches?
After a new Tails release has been shipped, we should review which Git branches has been merged in the release and remove them accordingly. This will make sure that we keep a clean-ish Git repository for a new release cycle.
Deleting obsolete Git branches
What we call an obsolete Git branch here is one that has been merged into
It's not the same as an old, stalled branch: an old branch may not have been merged, it could've been around for longer then 6 months or even a year, the work might never be done but we will consider it a WIP branch and will thus not be deleted.
After a Git branch has been merged and we discover that it introduces a regression, the cleanest way would be to work on a resolution in a new Git branch, and not reuse the old Git branch. Hence, we consider it safe to delete branches that were already merged.
Exception: branches created just to get an APT suite
Update: tails#8654 (closed) has been implemented in a way that
makes all this moot (any branch that's created to get an APT suite
will have one commit to add that suite to
the only exception being branches that would be created to create an
APT suite that they don't plan to use, but for that use case, so far
we've instead hard-coded the additional suite names we need in the
scripts that generate the reprepro configuration).
Given how our custom APT repository
currently works, we sometimes push
Git branches merely to have an APT suite created, and be able to upload Debian
packages to it. Such a branch will look like it has been merged (it has no
commit on top of current
master), but still it should not be deleted:
otherwise, our reprepro will remove the corresponding APT suite from its
configuration; the APT suite indices and packages will stay around, but when one
resumes work on this branch and pushes it again, then reprepro will most
probably either bail out (and cause additional work for our sysadmins), or just
create an empty APT suite (and then the packages previously uploaded to this APT
suite won't be part of it anymore, causing additional work for the developer).
So, on the short-term, the person who deletes Git branches should manually scan the list of candidates for deletion, and avoid deleting branches that 1. were created merely to get an APT branch; and 2. were actually not merged yet. Of course that's a bit painful and error-prone.
Longer term, if/once we have each branch contain a config file with the list of APT suites that shall be used (as mentioned on tails#8654 (closed)), then for each branch that is a candidate for deletion we can:
- Check if that APT suite list file, in that branch, contains anything but a base branch's APT suite. If it doesn't then it can be deleted. Otherwise, go on:
- Check each listed non-base APT suite for packages that either are not in some base branch (XXX: which one?) at all, or are newer than the version in some base branch (XXX: which one?). If such packages are found, then drop this branch from the list of candidates for deletion, on the grounds that its APT suite has work that hasn't been merged.
Mid-term, we could do the check from step 2 above on the APT suite that corresponds to each branch we're going to delete. It's going to be sub-optimal, as most branches don't introduce Debian packages, but it would probably be good enough, and the code will be reusable for the long-term solution anyway.
Another mid-term solution would be to piggy-back on tails#8689 (closed): a branch that 1. has no commit on top of master; and 2. has no such change file; probably was created only to get an APT suite, and should not be deleted.
Who does delete these Git branches?
Ideally, the one who deletes them has access to push to the official Git repository. If we would chose for having the release manager do this, it can be done post-release and it wouldn't require people from the sysadmin team to step in and issue a command.
What kind of privileges are required to do that?
Push access to the official main Git repository.
So far, only humans have write access there and we would like to not replace them with robots yet. That is for the future (e.g. the system that hosts our APT repository already has privileges akin to Git commit rights).