Elegant subdirectory checkout of remote-tracking branch?

I'm trying to figure out the most efficient way to keep an up to date `todo` branch checked out in Meta [1]. I've tried a few things like:

  $ git submodule add -b refs/remotes/origin/todo --reference ./ -- ./ Meta

and:

  $ git clone --single-branch --branch refs/remotes/origin/todo ./ Meta

These fail because I can't use a remote tracking branch as a source for the clone. It should be possible to do:

  $ git clone --reference . --single-branch --branch todo git://git.kernel.org/pub/scm/git/git.git Meta

but that will require (I think) network access during a fetch. Since I'm already fetching `origin` from the superproject, I don't want to have to refetch them for the submodule (or whatever Meta ends up being). Here's what I think happens with a submodule fetch:

1. Query the remote URL to dereference its current `todo` branch.
2. Check if we have that object in our local object share (which
   we should, due to --reference and a recent superproject
   fetch).
3. Fetch any missing objects from the remote URL.

I want to replace step 1 with:

1b. Query the superproject to dereference its current
    `origin/todo` branch.

and step 3 with:

3b. Access objects from the superproject directly (as with
    --reference / --shared).

Do I need to setup something like:

  [remote "origin"]
    url = ../.git
    fetch = +refs/remotes/origin/todo:refs/remotes/origin/todo

by hand, or is there an easier way?

I can, of course, clone a local `todo` branch if I've set one up in my superproject. However, then I'd have to update-ref that branch to sync with origin/todo after every fetch (that updates origin/todo). This could be handled with a `git fetch` wrapper, but… yuck :p.

Any suggestions for an elegant solution would be appreciated :). Once we figure something out, I can write it up and stick it in howto/maintain-git.txt.

Cheers, Trevor

[1]: http://article.gmane.org/gmane.comp.version-control.git/144748

"W. Trevor King" <wking@tremily.us> writes:

My Meta checkout is not a submodule of anything. It is a totally independent repository and is not linked with git.git in any way. It started as an independent repository, and it still is.

That it has been pushed to the same repository at k.org was a mere historical accident (I only had that repository writable there when I started); it not even a recommended practice to publish unrelated histories in a single repository. You are seeing 'todo' in some of my public mirrors only due to historical inertia.

Just treat the git.git primary directory and the Meta directory as what they are: two checkouts of two totally independent projects, one of which happens to be designed to be checked out at a fixed location relative to the other. When running "pull --ff-only", do so in both independently.

That is what I do in my secondary repositories.

You are free to write and do whatever suites yourself, but a description of a workflow that is not what I use does not belong to the howto/maintain-git.txt document ;-)

On Fri, Mar 01, 2013 at 10:22:53AM -0500, W. Trevor King wrote:

Yes, it will. Junio mentioned already that for him, "Meta" is really a separate repository, and I think the simplest thing is to just treat it that way (that's how I handle my personal "meta" branch).

But if you really want to save the extra network round trip during a fetch, you can either:

  1. Just fetch from the surrounding repository using a custom refspec,
     like:

       git init Meta
       cd Meta
       git config add remote.origin.url ..
       git config add remote.origin.fetch \
         refs/remotes/origin/todo:refs/remotes/origin/todo
       git fetch
       git checkout todo

or

  2. Look into the git-new-workdir script in contrib/workdir, which lets
     you check out an alternate branch in a separate directory.

The latter would probably be the most seamless, but it's also the most likely to have bugs. :)

-Peff