Git Backstage, Under the Covers and on the Down Low

Don't tell my employer, but I'm an undercover git user. Officially, my company uses Clearcase UCM - but I do almost all my coding outside Clearcase. I use git behind the official version control system. I'll tell you why I risk a scolding from my boss1 and the IT department for using an unsanctioned tool. But to do that, I'll have to tell you a bit more about both Clearcase2 and git.

Clearcase Basics

Clearcase is a centrally managed, client-server type system -- a big server hosts the repository which tracks all the project's info: individual files, the directory tree, versions, branches, users, permissions, Social Security numbers, DNA sequences, etc. The client (that's you) gets a working copy of the project files and tools to interact with the server. Checking out, checking in and viewing history all require connecting to the server. You can't do a thing without being tethered to the central repository.

In the Clearcase UCM development model, programmers create new "streams" ("branches" in generic terms), each having an "activity" (which is Clearcase talk for "changeset"). When you checkin a file, it's recorded against an activity. Therefore, an activity is a set of checked-in files (and directory changes). Other developers on the project won't see your changes until you "deliver" (merge) your activity (changeset) to the parent stream (the main branch). End of vocabulary lesson.

Where Clearcase Falls Down

My biggest issue with the Clearcase model is that it inhibits both experimental and iterative programming. I assert the following:

  1. Experimental coding requires branching. Let's hope this isn't a controversial statement. The only arguments people make against branching are practical - their VCSs don't support it well. Or they don't support merging well. Clearly, if you had a VCS that could branch and merge well, you'd use it all the time.
  2. Clearcase discourages experimental branching as streams are centralized, limited and public. On my project, creating "unnecessary" experimental streams is discouraged as each stream incurs the maintenance penalty of using more resources on the server.
  3. Merging in Clearcase is also less than perfect. Instead of a patch-based approach, Clearcase restricts merges to streams that have a common baseline. And guess who has to create, maintain and recommend these baselines? You, the user. Want to merge between streams that have diverged - that don't share an exact baseline? Forget it: you have to go outside Clearcase to do it.
  4. Finally, Clearcase activities don't allow logical commits within an activity - a checkin is limited to just one file. You cannot checkin 4 files as a group representing a logical change. This makes it hard to both organize your work and revert multi-file changes if they don't work out.

Let's see if git could help. (Ok, it can - or else why would I continue writing?)

Git Comes (Quietly) to the Rescue

Git is the distributed VCS used by Linus to maintain the Linux kernel. Where Clearcase requires a centralized server, git requires none -- the repository is stored in a single hidden directory at the top of your project's file tree. You can create a git repository from an existing project very easily:

cp -R /my/clearcase/project /my/git/project
cd /my/git/project
git init
git add .
git commit

A git repository is local and under your complete control, you don't need any anybody's permission to create one. And obviously you don't need network connectivity to a centralized server. You won't need to file a single TPS report to setup a git repository. And, if needed, you can keep your repo a secret.

What was I complaining about? Oh yes, branching and logical commits.

Branching is simple and lightweight in git. Let's say you're working in your 'bug_fix' branch, and you decide that the fix could be simplified by moving some methods from class Foo to class Bar. Here's how you'd create a new branch called 'foo_refactoring', make some changes in it, and then merge those changes back to the master branch:

cd /my/git/project
git checkout bug_fix
git checkout -b foo_refactoring
// make some changes to Foo and Bar, compile, and test
git add Foo.java Bar.java
git commit -m 'Refactored Foo'
git checkout bug_fix
git merge foo_refactoring

If you decided the refactoring was unnecessary, you could have skipped the merge -- or even permanently removed the experimental branch. The branch was created quickly -- just to try out some ideas -- and it can be ignored or removed. Branching is up to you, not the sysadmin.

Finally, git lets you build logical changesets. As you can see in the foo_refactoring example, a commit can contain any number of file changes. You can build a new feature piece-by-piece, committing chunks of related work together. This is good for both you and your reviewers!

So, how are you going to use git behind your VCS?

Getting it to Git

Getting your Clearcase (or CVS, Subversion, Perforce, etc) code into git is easy: copy your working dir to some local drive space and do the "git init; git add .; git commit" sequence.

Dealing with rebases is easy too. Other users have probably made changes to the codebase (in Clearcase) and you'll need to merge your work with theirs before merging to the main stream. You can handle this by rsyncing the upstream code into a git branch, then using git rebase to merge that code into your development branch - fixing any conflicts as necessary. Git rebase basically pops your current commits off your branch, merges with the requested branch, then re-applies your commits. It helps keep the history of your changes simple. I recommending doing all development work in a sub-branch off master (master is git's default branch) and keeping master for rebases. For example:

cd /my/clearcase/project
cleartool rebase -recommended // or 'cvs up' or whatever
cd /my/git/project
git checkout master
rsync -r /my/clearcase/project /my/git/project
git commit -a -m 'rebased from clearcase'
git checkout dev_branch
git rebase master

Getting it back to Clearcase

Git has made my daily coding much nicer - but if I want my code built into my product, I still have to get it back to Clearcase.

I find the simplest and safest way to do this is by applying a series of patches to the Clearcase controlled working dir. You can ask git to generate a patch for a single commit using git diff:

git diff 345983 > my-change.patch

But if you give git-diff a branch name instead of a commit, it will generate patch files for each diverging commit between the two branches. Assuming your master branch represents the rebased Clearcase branch and your dev branch has been 'git rebased' to master, this is exactly what you need!

cd /my/project/git
git checkout bug_fix
git diff master > bug_fix.patch
cd /my/clearcase/project
// checkout any files if necessary
patch -p2 < /my/git/project/bug_fix.patch

Build, test and submit in Clearcase - you're done!

Final Thoughts

Having a full understanding of your VCS's data model is essential to using a it correctly. Perhaps the root cause of why I prefer git over almost any other system is its simple conceptual model (Git for Computer Scientists does a nice job explaining the data model). Clearcase is typical enterprise software -- its feature sheet is very long and highlights words that CIOs love like "reliable", "maintainable", and "support contract", but the documentation is thrifty when discussing the systems internals. Version control is too essential and too difficult to trust to a system you don't understand. So I use git - behind the scenes if necessary.

Update (2008/8/5): As several commentators have noticed, if you can use Clearcase snapshot views instead of dynamic views, then you can directly init a git repo in the view storage directory. Then you can use git directly out of that directory or clone it. By creating the git repo directly in your snapshot view, you can substitute the rsync step with a "git pull" and the patch step with a "git push". This is a big win. Alas, my employer requires dynamic views due to tool limitations, so I didn't discuss this method in the original entry.


1 Hi Boss! What I'm discussing here is not really any less safe than having un-committed work in any working dir (which everybody does). But the threat of a knuckle slapping adds some drama to this post - don't you think? ;-)

2 Almost everything discussed in this post is true of any centralized version control system (CVS, Subversion, etc), not just Clearcase.

Technorati tags for this post:

Comments on "Git Backstage, Under the Covers and on the Down Low":

Comment by Turbodad

Posted at Apr 11th 2008. Turbodad's URL
Thanks for this article, it really saved my sanity. I wrote a follow-up at http://www.turbodad.com/2008/04/07/improving-clearcase-workflow-with-git/ to describe my experience with your suggested workflow applied to Windows/Cygwin environment.

Comment by coder

Posted at Apr 12th 2008. coder's URL
Take a look at this funny post on a clearcase break-up story... I guess you'll find it familiar:

http://codicesoftware.blogspot.com/2008/04/break-up-clearcase-bad-love-affair.html

Comment by Matt

Posted at Apr 14th 2008. Matt's URL
Turbodad: thanks for the hint re using xargs to checkout the files in your clearcase view before applying the patch.

Comment by Alex

Posted at Aug 2nd 2008. Alex's URL
Hi Matt,

Inspiring article. Thanks. I've run with your ideas and added what I think are improvements. Most significantly, instead of copying and rsync'ing your snapshot view, perhaps you'd initialize a git repository within it, then clone that repository. The ClearCase snapshot thus acts as a remote (albeit local) repository, and synchronization is just a matter of standard pull and push. As long as you do not branch in the snapshot (that's what the clone is for), Git and ClearCase play nice.

http://genaud.net/2008/08/clearcase-globally-git-locally/

Thanks again,
Alex

Comment by not@this.time.net

Posted at Sep 12th 2008.
Hi,
Being quite fond of ClearCase, even though it needs a lot of setup adaption to work nice, I found this article interesting since distributed development interests me.
This concept works fine - if you do not care about history, traceability and similar concepts. I tried the suggested concept, and found out that git is not able to differ between (1) a rename and (2) deletion of a file/folder and addition of another.
This "feature" - to my regret - makes use of the suggested procedure a nightmare for any Configuration Manager in ClearCase since it will sabotage the content in the ClearCase VOB if you have done e.g. refactoring of java classes/package names in your git repository.

Comment by Charles

Posted at Oct 30th 2008. Charles's URL
Here is a Python script I put together that handles file adding, removing and renaming as well as history importing:

http://github.com/charleso/git-cc/tree/master

I hope this helps someone other than me.

Comment by antz

Posted at Jan 17th 2009.
I work for a company where I use the clearcase snapshot view. And I have made the snapshot view itself the git repository. This is working pretty smoothly for me. If you are keeping git branches, just make sure that one branch ('master' perhaps) mirrors clearcase activities (for example, do a 'cleartool update' only while on branch master). Do note that 'git rebase' can come in handy in such a setup :)
Also, I came across instances of people finding a problem in such a setup because git apparently recognises permission changes (clearcase snapshot view causes a LOT of these!). However, I have not come across such a problem. I am using git version 1.5.4.

Comment by antz

Posted at Jan 17th 2009.
I just learnt that git did notice permission changes earlier. It has been long fixed.

Comment by Steve

Posted at Mar 6th 2009. Steve's URL
I'm linking to a tool that we switched to from Clearcase because our boss wanted to make sure we didn't do exactly what you're doing ;-) They're very IT governance and security focused here. I have to say though, it's 10x better than Clearcase and SVN, but I haven't tried Git.

Good luck captain!

Comment by Pat O'Neil

Posted at Apr 28th 2009.
One ClearCase feature no-one has mentioned is clearmake, which gives a configuration record that is a very reliable record of what's actually in a build. Does any other SCM have a capability like that?

Allowed html tags: br p blockquote i b em code strong u ol ul li a
(type "HuMaN" here)