git-svn-Cloning repository with complicate nesting

Hi, this is my first message in the list: this may be a newbie question and my English may not be very good.

I've an SVN repository structured like this:

http://<url>/path/to/repo
    |
  HEAD
    |----- root
    |
  BRANCHES
    |----- V1.0
    |         |----- root
    |
    |----- V1.1
    |         |----- root
    |
    |----- V1.2
    |         |----- root
    |
    |----- DEV
    |         |----- FEATURE1
    |         |            |----- root
    |         |
    |         |----- FEATURE2
    |         |            |----- root
    |         |
    |         |----- FEATURE3
    |                      |----- root
    |
    |----- BUILDS
              |----- BUILD1
              |            |----- root
              |
              |----- BUILD2
              |            |----- root
              |
              |----- BUILD3
                           |----- root

the same for TAGS.

I did this:

git init git svn init <url> vim .git/config

[core]
	repositoryformatversion = 0
	filemode = true
	bare = false
	logallrefupdates = true
[svn-remote "svn"]
	url = <url>
	fetch = <url>/HEAD/root:refs/remotes/trunk
	branches = <url>/BRANCHES/*/root:refs/remotes/branches/*
	branches = <url>/BRANCHES/*/*/root:refs/remotes/devel/*
	tags = <url>/TAGS/*/root:refs/remotes/tags/*

git svn fetch

It is now cloning the repo (it is a really big repo)

It is my configuration ok for the repository structure?

if from another terminal I execute "git branch -r" I get:
  tags/V1.3.0
  tags/V1.3.0@3260
  tags/V1.3.1
  tags/V1.3.1@3359
  tags/V1.3.2
  tags/V1.3.2@4256
  tags/V1.4.0-COMMUNITY-FINAL
  tags/V1.4.0-ENTERPRISE-BETA@4241
  trunk
  trunk@4475

It should have already created some branch but I don't see any...

what are those @XXXX number for some of those branches?

Is the syntax of the svn-remote configuration correct?

with this: branches = <url>/BRANCHES/*/*/root:refs/remotes/devel/* how does git choose the name of the branch? ( refs/remotes/devel/WHAT_GOES_HERE ? )

I would like it to use the tuple */* of the directory

If the syntax of my configuration is not correct, where can I found a documentation about it? I couldn't find one.

Thanks Regards, Daniele

Hi Daniele,

I think you're stuck.  Someone who knows git-svn better than I might be able to figure out a solution (I haven't played with using wildcards in the middle of branches refspecs), but as far as I can tell git-svn can't support your repository's structure.

The fundamental problem is that the BRANCHES/ hierarchy isn't consistent.  git-svn expects the directories under a 'branches' path to be branch names.  In your case, you can't specify a configuration that covers your whole repository.

This:

	branches = <url>/BRANCHES/DEV/*:refs/remotes/svn/dev/*
	branches = <url>/BRANCHES/BUILDS/*:refs/remotes/svn/builds/*

will let you track the stuff under BRANCHES/DEV and BRANCHES/BUILDS, but won't let you see the V*.* branches.

This:

	branches = <url>/BRANCHES/*:refs/remotes/svn/*

will let you see the V*.* branches, but will get confused over the DEV and BUILDS stuff (because it's expecting the layout to be DEV/root and BUILDS/root).

So I think git-svn would need to be modified to support your situation.

One possible approach is to make git-svn smart about what it considers a branch's name.  In theory, what you'd like is to just specify

	branches = <url>/BRANCHES/*:refs/remotes/svn/*

and have git-svn properly identify all the branches:

	V1.0
	V1.1
	V1.2
	DEV/FEATURE1
	DEV/FEATURE2
	DEV/FEATURE3
	BUILDS/BUILD1
	BUILDS/BUILD2
	BUILDS/BUILD3

To do this, git-svn could track the current 'trunk' directory structure (just the top-level contents of the root should suffice).  Then when it detects a new path under BRANCHES it could search through that path until it finds the same 'trunk' directory structure, and then use the path as the full branch name.

For example, say a repository's trunk has 3 directories and a file:
	trunk/
		foo/...
		bar/...
		baz/...
		readme.txt
When a commit creates a new branch:
	branches/some/path/
		foo/...
		bar/...
		baz/...
		readme.txt
git-svn sees that the commit is to a new path under branches/ and looks through branches/some/ and branches/some/path/ to find the trunk's contents, deciding in this case that 'some/path' is the branch's name.

Anyway, this is just an idea of how things might work.  There are probably some corner-cases that could make this a bit tricky to implement...

		M.

Daniele Segato wrote:

On Thu, Sep 3, 2009 at 4:31 PM, Marc Branchaud<marcnarc@xiplink.com> wrote:

Hi Marc and thank you for your reply.

yes.. and developers should be smart enought to keep consistency on their own.. unfurtunately they oftern aren't

I'm not feeling that lucky but I would be happy about it. :-)

hum... It could be an idea but I don't know how hard could it be to check it: how much deep inside the tree should Git search for branch? what if there are branches that are completely differents?

may be the user could create and pass a "branch model" to git svn to make it able to decide what is a branch and what isn't.

I think there are many options

In my case It would have worked defining a structure like: BRANCHES/*/root -> remote/svn/branch/* BRANCHES/*/*/root -> remote/svn/branch/*/*

But it would be better with something like: BRANCHES/\([^\/]+\)/root -> remote/svn/branch/$1 BRANCHES/\([^\/]+\)/\([^\/]+\)/root -> remote/svn/branch/$1/$2

like a grouped regex.

every branch not matching the regex will be skipped.. only an idea.. It couln't work in every situation but it would allow a greater degree of freedom in the configuration.

I don't know if Git developers are interested in thinking on some features like this.. it sound like an SVN-only hack to me.

Regards, Daniele

Daniele Segato wrote:

Well, as far as it takes.  Actually, the search should be fairly well constrained.  After checking that the entries under a particular path aren't a copy of the trunk, each entry can only represent either a branch that git-svn already knows about, or a new branch.

Not sure what you mean there.  Can you come up with an example of this?

Turning the branch refspecs into true regular expressions sounds like a good idea to me, and would probably be easier to implement than the trunk-detection stuff I proposed.  Though there'd have to be a way to stay compatible with or upgrade the current non-regex refspecs.

All of git-svn is already a hack for dealing with SVN.  :)  I'm sure a patch for this would be welcome.

		M.