Use case: GIT to manage transactions in a CMS?

Hello,

I am working on a project (a content management system) where the data is stored as files and folders.

Currently, for persistance and transactions we use the ZODB [1] object database. But our goal is to move away from the ZODB and use directly the file system, as it will allow us to use all the good unix tools.

We are using git to manage the source code. And now we are exploring git to see if it can do the job of transactions, so that each transaction in the system will be a git commit.

One problem we have found is that we can not commit empty directories (what we need to do). Any idea how to solve or work-around this constraint?

Any suggestions and input on this exotic use case for git will be very welcomed.

[1] http://www.zope.org/Wikis/ZODB/FrontPage

J. David Ibáñez wrote:

$ touch empty/dir/.placeholder

Sounds cool. I'll have to give it a whirl when you've got something to show.

Andreas Ericsson wrote:

I would like to have something better, no so much of a work-around.

Maybe I could hack something, are there some docs explaining the internal format of git objects? or, where to look in the source to find this info?

Another point, the application is written in Python, right now I have to open a shell to run the git commands.

It would be nice if the core functionality was implemented in a C library, then with a Python wrapper I could use git without going through the shell.

Is this something to be expected? There will be a "libgit" sometime in the future?

Probably I will have to hack something else to have ACID transactions, in the short-term. But "git" is at least interesting, since it could open a world of possibilities. If I ever have something to show, I will tell you.

"J. David Ibáñez" <jdavid@itaapy.com> writes:

The index file is mostly dealt with by read-cache.c (despite its name, it has both read and write routines). The easiest way to understand the tree object format is by reading write-tree.c.

But the format is the easiest part. You have to realize that this is a nontrivial task.

The merge algorithm assumes that it can detect a presense of directory by finding a blob under that. You need to teach 'read-tree -m' that sometimes there is an empty directory recorded in the index file. This is very important to avoid directory-file conflicts during the merge.

Tree object side is probably easier because we _can_ write an empty tree object, and presumably three tree-walker implementations (one used in read-tree and object layer, one in diff-tree and another in merge-tree) should all handle empty trees gracefully, but still its callers need to be verified.

I always have a shell open so it does not sound a big deal to me ;-).

Yes, read the list archive.

But not yet. The core level has been fluctuating quite a bit. The latest round being the object layer bookkeeping changes.