Hi,
there's a bit of a discussion within Debian on collaborating using Git.
One of the long-standing issues is that there are multiple ways Debian
packaging can be represented in a git tree, and none of them are optimal.
The problem at hand is that the packaging workflow consists of
1. importing an upstream release
2. optionally stripping out undistributable parts
3. adding packaging metadata
4. optionally adding a patch stack
The workflow for upgrading a package is
1. import a new upstream release
2. apply and possibly modify the exclusion list
3. apply the packaging metadata, updating it in the process
4. rebase the patch stack
Right now, git is used mainly as a network file system, and only tagged
releases are expected to be consistent enough to compile, because often
going from one consistent state to another as an atomic operation would
require multiple changes to be applied in the same commit.
The imported archive is represented either directly as a tree (which may
be imported from the upstream project if no files are undistributable
for Debian), or via a mechanism that can reproduce a compressed archive
that is bitwise identical to the upstream release, from a tree and some
additional patch data.
The patch stack is stored as a set of patches inside a directory, and
rebased using quilt.
An alternate representation stores the patch stack as a branch that is
rebased using git, and then exported to single files.
The Debian changelog is stored as a file inside Git, but some automation
exists to update this from Git commit messages.
Debian changelog entries refer to bugs in the Debian Bug Tracking
system. There is a desire to also incorporate forges (currently, GitLab)
and refer to the forges' issue tracker from commit messages (where the
issue tracker is used for team collaboration, while the Debian BTS is
used for user-visible bugs).
All of this is very silly, because we're essentially storing metadata as
data because we cannot express in Git what we're actually doing, and the
conflicting priorities people have have led to conflicting solutions.
I'd like to xkcd 927 this now, and find a common mapping.
From a requirements perspective, I'd like to be able to
- express patches as commits:
- allow cherry-picking upstream commits as Debian patches
- allow cherry-picking Debian patches for upstream submission
- generate the Debian changelog from changes committed to Git
- express filter steps for generating the upstream archive(s) from a
tree‑ish and some metadata
- store upstream signatures inside Git
- keep a history of patches, including patches applied to previously
released packages
A possible implementation would be a type of Git "user extension" object
that contains
- an extension name
- an object type (interpreted by the extension)
- type-tagged references to other objects
- other type-tagged data
Validity of the object would be determined by the extension, and git
would treat this object as mostly opaque (i.e. whenever one is
encountered, the extension needs to be called). The only exception would
be references, because we need to be able to transfer these objects and
all their dependencies efficiently (so the extension would generate a
list of references that should be recursively packed or omitted).
On top of that, we could represent a Debian package through special
objects, such as
- debian::debian-dir (a tree-like object referenced from the root
tree, contains a tree for plain files plus links to special objects for
generated items, such as patch stacks)
- debian::upstream-archive (a tree-like object that marks the boundary
between objects imported from upstream, and objects that are part of
packaging, and gives instructions for regenerating the upstream archives
without storing them as blobs)
- debian::update-upstream (a commit-like object to move to a new
upstream-archive object, this contains the upstream version number that
the following upload object must use)
- debian::changelog-entry (a commit-like object that adds an item to
the Debian changelog)
- debian::upload (a commit-like object that adds a version to the
Debian changelog)
- debian::rebase-patches (a commit-like object that links the patch
stacks before and after a rebase)
- ...
Changes to packaging would still be represented as commit objects
containing a tree, but that tree would contain a special entry for the
"debian" subdirectory that points to the last packaging change.
This is very high-level so far, because I'd like to get some feedback
first on whether it makes sense to pursue this further. This would use
up the last unused three-bit object type in Git, so it will have to be
very generic on this side to not block future development -- and it
would require a lot of design effort on the Debian side as well to
hammer out the details.
Any feelings/objections/missed requirements?
Simon
