Suggested Workflow Question

I work for a small company with about 15 developers who work concurrently on about 10+ projects both in new development and support. We do custom software for manufacturing and production systems. Part of our contracts with our customers is a perpetual single use license of the source code at each facility. So we have a copy of the source on our office server, and another copy at each customer site. When we had only 5 developers it was easier to handle. Now that we are growing we need a source control system and I have been looking heavily into Git. Our old workflow does not seem that it will fit well with Git however, but I feel that I need a distributed system to keep track of the office version and the on-site versions of our source since development is taking place on both. (Some customers also have separate development, and testing versions on their servers as well.)

I have created git repositories on a couple of our project source directories as
test beds.  Right now (second day) I am the only one who is actually using git.
 Everyone else is simply accessing the files on the server as they have always
done, and I am making the commits when I see signifigant changes. 

My question is really a request for modified workflow ideas. My plan was to have a master repository in our office server with clones at each customer site, and multiple branches for test, QA, and production versions of the source. Since most of these customers have closed networks, we would rely on people traveling onsite, or emailing patches to get any updates back into our office repository.

Thank you for any assistance for this revision control newb.

Roger Garvin

Roger Garvin wrote:

The workflow elements you mention so far sound sane to me:

* a repository in your office
* repositories at customer sites
* emailing patches when direct connections are not possible
* test and dev branches for each customer

I presume there will also be a production branch in customer repositories on which you arrange production releases... or maybe you just tag your test branch when it passes?

When you send patches via email, keep in mind that you'll have to convey which branch they belong on "out of band", for example after the commit message and above the diffstat.

There was an interesting discussion about maintaining per-client code on different branches recently, but I can't find it.

I've found Git to be very flexible. Maybe you could start by emulating your current workflow and refine it as you go? If you find you don't know how code travels currently... well, maybe you can just let your team pass patches around until patterns emerge.

Marcel

Marcel Cary wrote:

Here's the broader thread: http://thread.gmane.org/gmane.comp.version-control.git/113278

This is an abridged summary: http://article.gmane.org/gmane.comp.version-control.git/110697

And this a nice explanation referenced by the summary: http://gitster.livejournal.com/26540.html

Marcel

On Wed, Mar 18, 2009 at 2:27 PM, Marcel M. Cary <marcel@earth.care2.com> wrote:

We've done a ton of that. Even better than emailing is that you can have a repo on a usb stick/disk.

 - At the office, update it from the 'office' repository.
 - At the client site, pull updates from the usb stick, and push any
code developed at teh client side to the stick
 - Back at the office, push to the office repo

cheers,

m

Martin Langhoff <martin.langhoff <at> gmail.com> writes:

That would be great except that most of our customers do not technically allow us to plug USB drives into any of their computers. It's a rather silly rule actually but the drives could be confiscated.

Thank you for the links to the other thread. I will read through those today. I am hoping that this is simpler in practice than it looks on paper, cause what I see on paper is a fully connected graph of all our employees with all our customers multiple source locations and with our office server. It seems like it could get out of control for tracking where changes might be (who is holding them)

Roger Garvin

onsdag 18 mars 2009 14:00:51 skrev Roger Garvin <yoyodyn@gmail.com>:

Don't forget the option of mailing bundles (man git-bundle). Those will give the effect of push/fetch via e-mail. I.e. the exact same commit SHA's get replicated, That may be relevant for the branches that contain your released/official test code.

-- robin

Robin Rosenberg <robin.rosenberg.lists <at> dewire.com> writes:

the effect of push/fetch via

for the branches that

After trying this out for a little while I and projecting some into the future I think I am going to have a problem that I am not sure if it really is a problem, or if Git will "take care of it". Let me give you a simplistic commit history to illustrate my point. In its most basic form you have three branches; dev, test, prod. In a perfect world all the development happens on dev, then updates are pushed to test, and finally after tests pass are pushed out to prod. So at a point in time you have something like this:

(P)<--(T)<--(D)

Since I would think (and maybe this is my problem) that ideally Prod is a direct descendant of Test and Test is a direct descendant of Dev. However we don't live in a perfect world, so changes get made on Test that aren't in production yet, and also in an emergency directly on the Prod branch. Eventually those changes get migrated back either through a merge or a rebase (rebase would be bad since there would be multiple developers using these repos.) Obviously making a change to Prod and then merging back to Test, and then merging back to Dev creates a lot of extra commits.

 (P)<---(T)<----(D)
  ^      ^       ^
   \      \      \
  (P')<--(T')<--(D')

Things get even more complicated when the customer decides to only push certain features into Prod from test, it looks like cherry pick would be ideal for this and make it much easier for us.

Is trying to maintain a direct ancestral relationship between Prod, Test, and Dev worth it? Am I crazy for even trying to? I am still a pretty green git user so it might not be worth it. But it does seem to me that merges might be easier and faster if the common ancestor between two commits is closer rather than farther away in the commit history.

Thanks again for your help.