Following my previous article, I had some interesting comments introducing me to git-filter-branch (which is a new function coming from cogito’s cg-admin-rewritehist). This command is really designed to rewrite the history and you can do much more changes… it enabled me to fix the dates/authors/committers/logs of all the commits that were created with git_load_dirs. It can also be used to add one or more “parent commits” to any commit.

In parallel I discovered some problems with the git repository that I created: the tags were no more pointing to my master branch. This is because git rebase won’t convert them while rewriting history.

This lead me to redo everything from scratch. This time I used git-filter-branch instead. The man page even gives an example of how to link two branches together as if one was the predecessor of the other. Here’s how you can do it: let’s bind together “old” and “new”… the resulting branch will be “new-rewritten”.

$ git rev-parse old
0975870bb1631379f2da798fa78736a4fe32960a
$ git checkout new
$ git-filter-branch --tag-name-filter=cat --parent-filter \
"sed -e 's/^$/-p 0975870bb1631379f2da798fa78736a4fe32960a/'" \
new-rewritten
[...]
Rewritten history saved to the new-rewritten branch

Short explanation: the only commit without a parent commit (thus matching the empty regex “^$”) is the root commit and this one is changed to have a parent (-p) which is the last commit of the branch “old”.

At the end, you remove all the temporary branches, keep only what’s needed and repack everything to save space:


$ git branch -D old new
$ git prune
$ git repack -a -d


The dpkg team has a nice history of changing VCS over time. At the beginning, Ian Jackson simply uploaded new tarballs, then CVS was used during a few years, then Arch got used and up to now Subversion was used. When the subversion repository got created, the arch history has not been integrated as somehow the conversion tools didn’t work.

Now we’re likely to move over git for various reasons and we wanted to get back the various bits of history stored in the different VCS. Unfortunately we lost the arch repository. So we have disjoints bits of history and we want to put them all in a single nice git branch… git comes with git-cvsimport, git-archimport and git-svnimport, so converting CVS/SVN/Arch repositories is relatively easy. But you end up with several repositories and several branches.

Git comes with a nice feature called “git rebase” which is able to replay history over another branch, but for this to work you need to have a common ancestor in the branch used for the rebase. That’s not the case… so let’s try to create that common ancestor! Extracting the first tree from the newest branch and committing it on top on the oldest branch will give that common ancestor because two identical trees will have the same identifier. Using git_load_dirs you can easily load a tree in your git repository, and “git archive” will let you extract the first tree too.

In short, let’s see how I attach the “master” branch of my “git-svn” repository to the “master” branch of my “git-cvs” repository:

$ cd git-svn
$ git-rev-list --all | tail -1
0d6ec86c5d05f7e60a484c68d37fb5fc31146c40
$ git-archive --prefix=dpkg-1.13.11/ 0d6ec86c5d05f7e60a484c68d37fb5fc31146c40 | (cd /tmp && tar xf -)
$ cd ../git-cvs
$ git checkout master
$ git_load_dirs -L"Fake commit to link SVN to older CVS history" /tmp/dpkg-1.13.11
[...]
$ git fetch ../git-svn master:svn
$ git checkout svn
$ git rebase master

That’s it, your svn branch now contains the old cvs history. Repeat as many times as necessary…