Working with Git and CVS in a team.

Hello,

Our team works on an existing project that's in CVS, but we don't have the permissions to commit directly to it. We still need to make changes and want to stay as up to date as possible, so we are trying to use git to track the CVS repository and to collaborate inside the team.

Now, I'm quite new to git, so I'm not sure if the solution I created is "correct", not way too complicated, or if there are problems waiting for us.

git-cvsimport seems to be the solution, so I've established the following structure:

+-----------------+
| dev.eclipse.org |
+-----------------+
        ^
        | (1)
        |         (2)                      (3)
+---------------+      +---------------+        +------------+
| ifs:/from-cvs | ---> | ifs:/git-proj | <----- | Merge-Dude |
+---------------+      +---------------+        +------------+
                          /           \
                         /             \
                        /               \
                    +------+         +------+
                    | User |   (4)   | User |
                    +------+         +------+

1) I use git-cvsimport to update the repository every night. /git-proj
   is a "clone --bare" of the /from-cvs repository.

2) After the import, I push the changes from CVS to /git-proj using
   'git push /git-proj origin'.

This first two steps can be done with a cronjob, there shouldn't be any conflicts.

3) Then we have the role of a "Merge-Dude" who pulls the changes from
   /git-proj and uses 'git merge origin/origin' to update his master.
   If there are any conflicts, he resolves them and pushes everything
   back to the master at /git-proj.

4) The users can just push and pull from /git-proj.

with all the changes we made and have to send upstream.

What do you guys think, is this approach feasible?

What I don't like is how we have to make the upstream patch(es). Is there an easy way we can get multiple patches, lets say for each commit we made? Or is it easier to make a lots of branches and to then create a patch from the diff between the branch and origin/origin?

Wow, this got longer than I expected. Thanks in advance!

Mirko Stocker

On Fri, Jun 13, 2008 at 04:33:34PM +0200, Mirko Stocker <mirko.stocker@hsr.ch> wrote:

Don't you mean origin/master?

In general, if you don't merge your changes but you rebase on top of the new upstream head, then you can use 'git format-patch origin/master..' to create patches for each commit.

It'll include the commit message and each commit will be written out as a separate file.

On Friday 13 June 2008 16:50:10 Miklos Vajna wrote:

<mirko.stocker@hsr.ch> wrote:

Hm, I'm not sure.. if I work on an a feature in a branch, and now I want to create a patch that applies to the CVS head, then I have to make the diff to origin/origin, right?

Ah, I didn't know about git-rebase, thanks! That's exactly what I wanted :) I think I've already fallen in love with git.

On Fri, Jun 13, 2008 at 10:43:36PM +0200, Mirko Stocker <m1stocke@hsr.ch> wrote:

'origin/origin' means the 'origin' branch of the 'origin' remote. Given that you said you created the 'origin' remote's repo using git-cvsimport, I assumed that the 'origin' remote has only one branch (named 'master').

On Friday 13 June 2008 22:47:40 Miklos Vajna wrote:

Hm, git-cvsimport created a master and an origin branch. And then I push the origin branch to the bare repository to share it with my colleagues. So for the users, origin/origin contains the original CVS content, so we need to make the diff against that, don't we? That's how I think this whole thing works, I might be wrong :-)

On Sun, Jun 15, 2008 at 10:20:05PM +0200, Mirko Stocker <me@misto.ch> wrote:

Oh, you are right. Sorry for the noise. I forgot that git-cvsimport creates an origin branch without -o master.

On Fri, 2008-06-13 at 14:33 +0200, Mirko Stocker wrote:

I'm in a team that is all CVS.  I'm the only one using git.

I can commit to CVS, unlike your situation.

I don't think there is a 'wrong' way. 'Whatever works' is the motto. The configuration should match the workflow. So, I don't have any comment on yours. I thought I'd share mine, in case there is value. I think I will learn something

My setup is:

+----------+
| cvs repo |
+----------+
     |
     | (1)
     |
+----------+
| /cvs_git |
+----------+
     |
     | (2)
    \|/
+---------------------+   (5)   +--------------------+
| john's private repo |-------->| john's public repo |
+---------------------+         +--------------------+
     |
     | (3)
    \|/
+--------------------+
| john's cvs sandbox |
+--------------------+
     |
    \|/ (4)
+----------+
| cvs repo |
+----------+

(1) git-cvsimport -i ...

Runs in cron every 30 minutes.

The output git repo is on a shared server.  git users (me) can clone/fetch/pull.

The git repo for cvs is purely one way: no commits are to be pushed to this repo. All content comes from the cvs repo.

This is importing ALL branches from CVS because it git users (me) require access to all cvs branches.

(2) git clone -o cvs 

Using -o, my private repo uses 'cvs' as the name of the git remote that tracks cvs. This lets me do 'git fetch cvs', which feels right.

And, to create a branch from the cvs head branch:

'git checkout -b br cvs/origin'

(3) git cvsexportcommit (4) cvs commit

These close the loop: my changes are committed to cvs repo. Alternatively, the loop can be closed by emailing patch files to a cvs committer.

(5) git push public

'public' is the name of my public repo. The commits I push to public are for developers that are using git (currently zero) who want to see my work before it gets to cvs' head (which is under high contention).

I expect my team members to push to their one public repo. Then I can do: 'git fetch alice' to get Alice's work.

At the end of all this, when I do a 'git branch -r' I will see something like:

cvs/br1 cvs/br2 alice/br4 alice/br5 bob/br6 bob/br7

Thus making obvious which branch originates from which developer or from cvs.

In this configuration, the 'merge-dude' is simply another user with a private and public repo.

Issue with being the git user in a team of cvs users:

git makes merges easier and more robust.  I don't want to use cvs anymore.

My main problem is sharing my changes to cvs users. cvs head is under contention and protection, so I can't simply commit all my changes to cvs head. Our cvs strategy is to create a branch for each developer or topic. If I create a branch in cvs repo for my git commits, I'd also want to keep that cvs branch upto date wrt cvs head. I haven't sorted out how to do that easily with the git tools.

I don't know enough. I'd try it both ways to see which works for you. Doesn't git easily create a patch file in either case? (This is the part I'm not sure of.)

Regards, John

<opublikowany i wysłany>

John J. Franey wrote:

What about git-cvsserver instead?

On Fri, Jun 13, 2008 at 04:33:34PM +0200, Mirko Stocker wrote:

You could simplify this by just cvsimporting to an 'upstream' branch in git-proj.

And then it works the same: he just pulls 'upstream' occasionally, merges it into your work, and then pushes the result.

-Peff

On Friday 13 June 2008 18:09:11 Jeff King wrote:

Hm, I've tried that, but it doesnt seem to work.. git-proj is a bare repository, and git-cvsimport always creates a new .git in it. I'm using the -C option to specify the target. Is anything else needed?

Regards

Mirko

On Fri, Jun 13, 2008 at 10:47:03PM +0200, Mirko Stocker wrote:

Ah, I hadn't thought of that. Apparently git-cvsimport doesn't understand bare repos. There is even a Debian bug reported:

  http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=472873

You might be able to hack around it with:

  mkdir bare.git && (cd bare.git && git init)
  mkdir cvsimport-hack && ln -s ../bare.git cvsimport-hack/.git
  git cvsimport -C cvsimport-hack

-Peff

On Friday 13 June 2008 22:55:26 Jeff King wrote:

Uh, ok.. :) Thanks. 

I just wondered, do I even need the additional bare repository? If I use git-cvsimport with -i, then it creates only the .git without doing a checkout, then we could just clone this one from the clients and pull/push to it?

Thanks

Mirko

On Sun, Jun 15, 2008 at 10:22:49PM +0200, Mirko Stocker wrote:

Sure. From the client perspective, there's not really a difference between a bare repo and one with a working directory. Using just one repo with a working directory will mean you just have one extra checkout on the server, which wastes a little space.

-Peff

On Sunday 15 June 2008 22:48:41 Jeff King wrote:

Ok, but in my case I have the checkout on the server in any case, whether I push it to the bare repository or use it directly.

Thanks again

Mirko