Re: git-svn: handling merge-base failures

[Eric Wong <normalperson@yhbt.net>]

Andrew Myrick <amyrick@apple.com> wrote:
> On Dec 23, 2009, at 11:54 AM, Andrew Myrick wrote:
> > One of my projects is failing to clone because merge-base is failing
> > on one of the revisions; the branch is a partial branch, so
> > merge-base can't find a common ancestor with trunk.  I'd like to
> > catch the exception that command_oneline should throw when
> > merge-base fails,
> 
> Instead of using the Error.pm module, I was able to handle the
> exception with the more basic eval block.  However, there are some
> details here I would like to discuss with the community.
> 
> When git-svn fetches a partial branch, it appears to refetch all of
> the history of the subdirectory from which the branch was created.
> This creates new commits for the old revisions, and these new commits
> exist as a separate history for the partial branch.   When git-svn
> fetches the revision at which this partial branch is merged back to
> trunk, git-svn attempts to run merge-base to find a common ancestor.
> However, because the partial branch has its own history, the
> merge-base fails, and git-svn dies.
> 
> Naively handling the exception with an eval block and skipping the
> merge ticket works fine in that it brings us back to parity with git
> 1.6.5.7, but it means that the merge parent relationship between trunk
> and the partial branch is lost.  Is there any way to preserve this
> information, or does the separate commit history of the partial branch
> make it fundamentally impossible?

Hi Andrew,

A method of preserving the $SVN_PATH <=> $PARENT@$REV mapping for
reusing follow_parent-created branches is definitely desired.

I've just been lacking time and motivation these days with other
projects taking priority.  Help (even if it's just to refactor/cleanup
existing code) would definitely be appreciated here.

> I've created a small svn repository that demonstrates this failure
> with git v1.6.6.  Its dump is attached.

Thanks.  This (and a similar dump a few weeks back) will definitely
come in handy for writing test cases and fixing this long-standing
issue.

-- 
Eric Wong