The coolest merge EVER!

Ok, Junio had some cool octopus merges, but I just one-upped him.

I just merged the "gitk" repository into git, and I did it as a real git merge, which means that I actually retained all the original gitk repository information intact. IOW, it's not a "import the data" thing, it's literally a merge of the two trees, and the result has two roots.

Now, the advantage of this kind of merge is that Paul's original gitk repository is totally unaffected by it, yet because I now have his history (and the exact same objects), the normal kind of git merge should work fine for me to continue to import Paul's work - we have the common parent needed to resolve all differences.

Now, I don't know how often this ends up being actually used in practice, but at least in theory this is a totally generic thing, where you create a "union" of two git trees. I did the union merge manually, but in theory it should be easy to automate, with simply something like

	git fetch <project-to-union-merge>
	GIT_INDEX_FILE=.git/tmp-index git-read-tree FETCH_HEAD
	GIT_INDEX_FILE=.git/tmp-index git-checkout-cache -a -u
	git-update-cache --add -- (GIT_INDEX_FILE=.git/tmp-index git-ls-files)
	cp .git/FETCH_HEAD .git/MERGE_HEAD
	git commit

(this is not exactly how I did it, but that's just because I'd never done this before so I didn't think it through and I had some stupid extra steps in between that were unnecessary).

Of course, in order for the union merge to work, the namespaces have to be fit on top of each other with no clashes, otherwise future merges will be quite painful. In the case of gitk, Paul's repository only tracked that single file, so that wasn't a problem.

Anyway, you shouldn't notice anything new, except for the fact that "gitk" now gets automatically included with the base git distribution. And the git repository has two roots, but hey, git itself doesn't care.

			Linus

On 6/23/05, Linus Torvalds <torvalds@osdl.org> wrote:

You do realise what this means, don't you?

The Borg can now start accumulating software into one massive repository of everything ever written and it will never forget - resistance will be futile.

Of course, it may run out of disk space.

jon.

Jon Seymour wrote:

The kernel.org robots will melt before that happens :)

git has already increased the kernel.org file publish time over an hour. 
  Hopefully inotify and an improved rsync will fix some of that.

	Jeff

LT> Now, the advantage of this kind of merge is that Paul's original gitk LT> repository is totally unaffected by it, yet because I now have his history LT> (and the exact same objects), the normal kind of git merge should work LT> fine for me to continue to import Paul's work - we have the common parent LT> needed to resolve all differences.

This clearly shows that you are in the "project lead" integrator mindset. Making it easy for you, the integrator, to pull changes made in Paul's tree is what this "coolest merge ever" is all about, but I suspect there would be a massive additional support needed if you want to make it easy for Paul to pull changes made to gitk in your tree.

I am not saying I do not like it, however, to make it easy for Paul to pull changes made in your tree relevant only to gitk part, off the top of my head:

 - If a contributor to GIT wants to make a change to gitk that
   adds a new file that is used by gitk (say, some common tcl
   library thing to be included), we either need to feed them to
   Paul and get the changes propagate to you through him, or we
   need a way to mark that file somehow belonging to the
   development history rooted at gitk root, not GIT root.

 - Similarly, if a contributor to GIT wants to make a change to
   gitk file itself, feeding the change through Paul to you
   would lesson both your burden and Paul's.

 - Alternatively Paul can keep track of which files are relevant
   to gitk, slurp and merge commits while ignoring the files the
   gitk package does not care about.  This one would involve
   interesting scripting somebody may want to tackle (hint,
   hint).

Long time ago I had a discussion with somebody (I vaguely recall it was with Dan Barkalow but I am not sure) about this exact issue. He wanted to have a way to distinguish Cogito-only part and core GIT part in a Cogito source tree, and help developers feed core changes to you and Cogito changes to Petr.

Back then I dismissed that approach by saying that what is broken was Cogito's source tree structure, which overlays two projects that are theoretically separable and practically managed separately, and it is not worth supporting such a broken source tree structure. Now you are making GIT source tree such an overlayed one.

Hi, Junio C Hamano wrote:

I don't think that's possible; after all, the trees are now merged, so any pull would fetch all of Linus' tree.

Paul could do it as patches, or Linus could do it in a branch, or we could write something entirely different from git that happens to support cherrypicking. ;-)

On Fri, 24 Jun 2005, Matthias Urlichs wrote:

Linus could do:

 git-read-tree gitk-head
 git-update-cache gitk
 git-commit-tree `write-tree` -p gitk-head > gitk-patched-head
 git-read-tree HEAD
 git merge gitk-patched-head

(or, better, use a separate index file for the gitk index)

(to commit changes to the gitk script made in a git working directory)

The change I proposed earlier would be so that the system would know what was going on and users wouldn't have to. Then someone who didn't know that gitk was (also) a separate project and just committed changes to it would still generate gitk commits when appropriate.

	-Daniel
*This .sig left intentionally blank*

On Fri, 24 Jun 2005, Matthias Urlichs wrote:

No no. 

A merge is a one-way thing. I merged Paul's tree, but Paul didn't merge mine. His is still independent, and you can still pull his tree without getting the rest of git, and Paul can still continue to work on his tree as if I never merged it at all.

Now, merging back isn't as easy: if any gitk changes get done in my "union" tree, Paul can't just pull those, becasue they now end up being linked to the history of the unified thing, so when pulling, he'd now end up getting all of the regular git stuff too.

Which is probably acceptable, but Junio's point was that this is not a symmetric setup: git is like a black hole that never lets any information escape, and once you've been sucked into a git archive, you end up not being able to separate it.

Or rather, you _can_ separate out pieces of it, but now it's a matter of cherry-picking, not automatic merges. Of course, people want to be able to do that anyway, and normally that will also merge back perfectly, so there's no huge downside, except that we should make it fairly easy.

		Linus