Re: git-svn set-tree bug

[Junio C Hamano <junkio@cox.net>]

Eric Wong <normalperson@yhbt.net> writes:

> Joakim Tjernlund <joakim.tjernlund@transmode.se> wrote:
>> > -----Original Message-----
>> > From: Steven Grimm [mailto:koreth@midwinter.com] 
>> > Sent: den 11 juni 2007 01:37
>> > To: Joakim Tjernlund
>> > Cc: 'Eric Wong'; 'git'
>> > Subject: Re: git-svn set-tree bug
>> > 
>> > Joakim Tjernlund wrote:
>> > > Is there a way to tell set-tree to commit the whole "merge" branch
>> > > as one svn commit?
>> > > If I merge the latest kernel into my tree there will
>> > > be a lot of commits that I don't want in svn.
>> > >   
>> > 
>> > You want a "squash" merge. Something like this:
>> > 
>> > git checkout -b tempbranch origin/svn-branch-to-commit-merge-to
>> > git merge --squash branch-with-commits-you-want-to-merge
>> > git commit
>> > git svn dcommit
>> > 
>> > The "merge" command will merge in the changes but will not commit 
>> > anything; when you do the explicit "commit" command 
>> > afterwards, you get 
>> > the contents of the merge but from git's point of view it's just a 
>> > regular commit so git-svn doesn't get confused.
>> > 
>> > After you do git svn dcommit, you may want to edit 
>> > .git/info/grafts to 
>> > tell git after the fact that this commit was a merge. It won't hurt 
>> > git-svn at that point and it will mean you can do another merge later 
>> > without git getting confused about what has already been merged.
>> > 
>> > Take a look at the script I posted a while back, which does something 
>> > similar:
>> > 
>> > http://www.spinics.net/lists/git/msg29119.html
>
> I must have missed this message the first time around.
>
>> Hi Steven
>> 
>> That looks promising, especially Junos comment about making git-svn
>> able to deal with merges. Eric, do you feel this is doable?
>
> Doable?  Yes.  However, I think using grafts is quite hackish and
> unreliable[1].  I'd rather just have users using set-tree if
> they want to deal with non-linear history in the first place.
>
> I'd personally avoid any sort of non-linear history when interacting
> with SVN repositories, however.

I've been wondering if you can do a moral equilvalent of the
graft trick but without using graft inside dcommit.  Perform a
merge --squash of the other branch (call the tip commit $B),
then dcommit on the git side as usual, and call it commit $C.
Steven's procedure would do a graft trick here, but instead of
doing that, rewrite $C to have the two parents.  Using the tree
object of $C, create a new git commit $D that is a merge between
the parent of $C (i.e. $C^) and the squashed branch tip $B.
Replace the tip of the current branch (which is $C) with $D.
Finally, replace the mapping between svn commit and git side
recorded in the revdb (which currently says $C on the git side
corresponds to the HEAD of SVN side) with this new commit $D.

Wouldn't that let the git side know what was merged into the
branch, so that later merges on the git side would go smoothly?

Or am I grossly misunderstanding how dcommit, tracking of svn vs
git commit mappings and the graft trick work?