[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Feedback on 3.0 source format problems

Russ Allbery writes ("Re: Feedback on 3.0 source format problems"):
> Even if we never used tarballs, and instead our unit of operation was the
> upstream Git repository plus Debian branches, I would maintain a rebased
> branch of Debian changes to upstream

I think this is definitely the right thing for many packages.

I don't think this depends on the interchange format representing the
changes as patches rather than commits.  At a technical level, they
represent roughly the same information and it is possible to
interconvert.  However:

  * There are many changes that patches cannot represent.

  * git (working with a rebasing branch) is a much more powerful tool
    for editing a patch series than quilt or diff/patch or emacs
    (working with a series of patches as files).  Ie, patches are a
    poor editing format.  Consequently most people actually gateway to
    git at the first opportunity, and out again for export.

  * Few people outside Debian work with quilt stacks any more.

All of these things mean that patches are a really poor interchange

The only difficulty is this one:

> This is actually, in a way, *harder* if we were using pure Git, since if I
> have a rebased branch of Debian changes on top of upstream, and I need a
> place to integrate that with Debian packaging, what does that
> debian/master branch look like?  I don't really want it to be a constantly
> rebased branch; I want it to be a conventional branch.  But I want to keep
> merging the changes against upstream into it (but not maintain them on
> that branch, only maintain the Debian packaging files on that branch).

My preferred answer is that it is a constantly rebasing branch topped
with a series of pseudomerges to make it fast-forwarding.

git-dpm sort of does this.  I have been experimenting with and
blundering towards another approach, which is closer to raw git.

Something like this:

       	 ------/--A-----/---B3---/--> interchange view
       	      /	       /       	/
       	     / 	      /	       3
       	    /	     /	      /
       	   2	    2	     2
       	  /	   /   	    /
       	 1     	  1        1
       	/  	 /     	  /
    ---p-----p--A----p---B---> `packaging-only' branch
      /     /  	    /
   --#-----#-------#-----> upstream

    Key:       1,2,3   commits touching upstream files
	       A,B     commits touching debian/ only
               B3      mixed commit (eg made by an NMUer)
	       #       upstream releases

       	      -p-      special merge, takes contents of debian/
              /          from the previous `packaging only'
	                 commit and rest from upstream

       	      -/-      pseudomerge; contents are identical to
       	      /         parent lower on diagram.

Looking from the tip of the interchange view, it is I think always
possible to classify these commits appropriately: pseudomerges are
fairly obvious (if all three trees are identical, we descend to the
parent with the most recent commit date), and the `p' special merge is
the only non-psuedo merge and has a special form.

So it would be possible to write a `git-debian-rebase' tool which
would take (for example) B3, above, and be able to perform functions

 * Strip pseudomerges: Rewrite the current branch so it contains no
   pseudomerges, turning ...B3 into ...p-A-1-2-B3.  (This should make
   a note, in your .git/ somewhere, of the latest pseudomerge to be

 * Cleanup branch: Reorganise the current branch so that the debian/
   changes come first, turning -p-A-1-2-B3 into ...p-A-B-1-2.

 * New upstream rebase: Start rebasing onto a new upstream version,
   turning ...#...p-A-B-1-2 into ...#'...p'-1-2.  This would be a
   wrapper around git-rebase, which prepares p' and then tries to
   rebase 1 2 onto p'.  So if you ask for an interactive rebase p'
   doesn't appear in your commit list.

   Note that the rebasing of p into p' cannot fail because p' simply
   copies debian/ from B and and everything else from #'.  (Rebasing A
   and B is undesirable.  We want the debian/ files to be non-rebasing
   so we can `git log' and get the packaging history.)

 * Record pseudomerge.  This is like "committing" your patch queue.
   The LH parent is taken from the previous strip pseudomerge.  (We
   should probably check that this is consistent with what we see in
   debian/changelog, but that is not a sufficient check.)

Maybe some of these operations should automatically edit

The only thing that this can't easily do is permit nonlinear (ie,
merging) history on the `packaging-only' branch, because upstream
might contain debian/ files too, so it is not possible to distinguish
a merge of two `packaging-only' branches from the special merge `p'.
(Indeed I since upstream might copy debian/ from us, I think it is not
easy to reliably distinguish the two parents of a `p'.  In the most
exciting edge case, upstream might `git merge' a previous instance of
our interchange view, but I think even then everything still works.)


Ian Jackson <ijackson@chiark.greenend.org.uk>   These opinions are my own.

If I emailed you from an address @fyvzl.net or @evade.org.uk, that is
a private address which bypasses my fierce spamfilter.

Reply to: