git-subset

This is a tool to filter a Git repository for a whitelist of files and folders. It has a more narrow scope than git filter-branch but is significantly faster.

Why?

This tool was created to limit access to a very large proprietary codebase (for security reasons). That is, the master branch in repo A is filtered into repo B where those with access to B only see a small subset of the real repo A.

Usage

$ git-subset --help
USAGE:
    git-subset [FLAGS] [OPTIONS] --branch <branch> [--] [revspec]

FLAGS:
    -f, --force      Overwrites the branch name if it exists.
    -h, --help       Prints help information
        --nomap      Does not use the saved map. Useful for profiling purposes.
    -q, --quiet      Don't print as much progress.
    -V, --version    Prints version information

OPTIONS:
    -b, --branch <branch>              Name of the branch to create on the rewritten commits.
        --filter-file <filter-file>    Path to the file containing paths to keep.
    -p, --path <path>...               Path to include. Can be specified multiple times.
    -r, --repo <repo>                  Path to the repository. Defaults to the current directory. [default: .]

ARGS:
    <revspec>    The ref to filter from. [default: HEAD]

Contrived Example

Suppose we want to create a subset of the Linux source tree and we have a list of the files and folders we want to keep:

$ cat linux.filter
README
COPYING
Makefile
include/
fs/btrfs/

Now, clone the Linux kernel (or another repository that isn't so YUGE):

$ git clone https://github.com/torvalds/linux.git

10 hours later...

$ cd linux

Now, to filter out everything that isn't listed in linux.filter. (Brace yourself for awesome speed!)

$ git-subtree --filter-file ../linux.filter --branch new-master
Getting list of commits...
Rewriting 4327da054142f4dbf74615918b71441d95025bad (678123/678123) - 100%
Branch 'new-master' created.

On my test machine with an SSD, this churned through 678,123 commits in about 3 minutes. This is far faster than git filter-branch. Let's just say that it probably took less time to write this tool than it would have for git filter-branch to finish running.

Running it again after pulling down the latest changes...

$ git pull
$ git-subtree --filter-file ../linux.filter --branch new-master --force

...took about 20 seconds because the mapping of old commit hashes to new commit hashes has been cached from the previous run (use --nomap to disable this).

Now, the new commit history is in the new-master branch and it contains only the history for the list of files and folders we specified:

$ git ls-tree new-master
100644 blob ca442d313d86dc67e0a2e5d584b465bd382cbf5c    COPYING
100644 blob 470bd4d9513ac42eb164cb4513300966a726fa37    Makefile
100644 blob b2ba4aaa3a71046653599aa0b3798b211a2c0d30    README
040000 tree 248cb042ad04f4b6d90a876b7ca35d1617de1e46    fs
040000 tree d3ba01442799c0b5169cc3daeb6ab7da150f47dd    include

new-master can then be pushed to a new repository that contains only the history of the files and folders we want.

Related Tools

• The BFG Repo Cleaner.

  This is similar, but doesn't have the same functionality. The BFG is more useful for filtering out specific files than it is for whitelisting file paths.

• GitRocketFilter.

  GitRocketFilter can do everything this tool can do (plus more!), but more slowly. While GitRocketFilter is more generic, git-subset is designed for one very specific use-case: creating a subset of a repository. As a result, git-subset can be very aggressive about avoiding work.

License

MIT License

Thanks

This tool was developed for internal use at Environmental Systems Research Institute (Esri) who have graciously allowed me to retain the copyright and publish it as open source software.