Introduction

Checking large binary files into a source repository (Git or otherwise) is a bad idea because repository size quickly becomes unreasonable. Even if the instantaneous working tree stays manageable, preserving repository integrity requires all binary files in the entire project history, which given the typically poor compression of binary diffs, implies that the repository size will become impractically large. Some people recommend checking binaries into different repositories or even not versioning them at all, but these are not satisfying solutions for most workflows.

Features of git-fat

• clones of the source repository are small and fast because no binaries are transferred, yet fully functional (unlike git clone --depth)
• git-fat supports the same workflow for large binaries and traditionally versioned files, but internally manages the "fat" files separately
• git-bisect works properly even when versions of the binary files change over time
• selective control of which large files to pull into the local store
• local fat object stores can be shared between multiple clones, even by different users
• can easily support fat object stores distributed across multiple hosts
• depends only on stock Python and rsync

Related projects

• git-annex is a far more comprehensive solution, but with less transparent workflow and with more dependencies.
• git-media adopts a similar approach to git-fat, but with a different synchronization philosophy and with many Ruby dependencies.

Installation and configuration

Place git-fat in your PATH.

Edit .gitattributes to regard any desired extensions as fat files.

$ cat >> .gitattributes
*.png filter=fat -crlf
*.jpg filter=fat -crlf
*.gz  filter=fat -crlf
^D

Run git fat init to active the extension. Now add and commit as usual. Matched files will be transparently stored externally, but will appear complete in the working tree.

Set a remote store for the fat objects by editing .gitfat.

[rsync]
remote = your.remote-host.org:/share/fat-store

This file should typically be committed to the repository so that others will automatically have their remote set. This remote address can use any protocol supported by rsync. Most users will configure it to use remote ssh in a directory with shared access.

A worked example

Before we start, let's turn on verbose reporting so we can see what's happening. Without this environment variable, all the output lines starting with git-fat will not be shown.

$ export GIT_FAT_VERBOSE=1

First, we create a repository and configure it for use with git-fat.

$ git init repo
Initialized empty Git repository in /tmp/repo/.git/
$ cd repo
$ git fat init
$ cat > .gitfat
[rsync]
remote = localhost:/tmp/fat-store
$ mkdir -p /tmp/fat-store               # make sure the remote directory exists
$ echo '*.gz filter=fat -crlf' > .gitattributes
$ git add .gitfat .gitattributes
$ git commit -m'Initial repository'
[master (root-commit) eb7facb] Initial repository
 2 files changed, 3 insertions(+)
 create mode 100644 .gitattributes
 create mode 100644 .gitfat

Now we add a binary file whose name matches the pattern we set in .gitattributes.

$ curl https://nodeload.github.com/jedbrown/git-fat/tar.gz/master -o master.tar.gz
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100  6449  100  6449    0     0   7741      0 --:--:-- --:--:-- --:--:--  9786
$ git add master.tar.gz
git-fat filter-clean: caching to /tmp/repo/.git/fat/objects/b3489819f81603b4c04e8ed134b80bace0810324
$ git commit -m'Added master.tar.gz'
[master b85a96f] Added master.tar.gz
git-fat filter-clean: caching to /tmp/repo/.git/fat/objects/b3489819f81603b4c04e8ed134b80bace0810324
 1 file changed, 1 insertion(+)
 create mode 100644 master.tar.gz

The patch itself is very simple and does not include the binary.

$ git show --pretty=oneline HEAD
918063043a6156172c2ad66478c6edd5c7df0217 Add master.tar.gz
diff --git a/master.tar.gz b/master.tar.gz
new file mode 100644
index 0000000..12f7d52
--- /dev/null
+++ b/master.tar.gz
@@ -0,0 +1 @@
+#$# git-fat 1f218834a137f7b185b498924e7a030008aee2ae

Pushing fat files

Now let's push our fat files using the rsync configuration that we set up earlier.

$ git fat push
Pushing to localhost:/tmp/fat-store
building file list ...
1 file to consider

sent 61 bytes  received 12 bytes  48.67 bytes/sec
total size is 6449  speedup is 88.34

We we might normally set a remote now and push the git repository.

Cloning and pulling

Now let's look at what happens when we clone.

$ cd ..
$ git clone repo repo2
Cloning into 'repo2'...
done.
$ cd repo2
$ git fat init                          # don't forget
$ ls -l                                 # file is just a placeholder
total 4
-rw-r--r--  1 jed  users  53 Nov 25 22:42 master.tar.gz
$ cat master.tar.gz                     # holds the SHA1 of the file
#$# git-fat 1f218834a137f7b185b498924e7a030008aee2ae

We can always get a summary of what fat objects are missing in our local cache.

Orphan objects:
1f218834a137f7b185b498924e7a030008aee2ae

Now get any objects referenced by our current HEAD. This command also accepts the --all option to pull full history, or a revision to pull selected history.

$ git fat pull
receiving file list ...
1 file to consider
1f218834a137f7b185b498924e7a030008aee2ae
        6449 100%    6.15MB/s    0:00:00 (xfer#1, to-check=0/1)

sent 30 bytes  received 6558 bytes  4392.00 bytes/sec
total size is 6449  speedup is 0.98
Restoring 1f218834a137f7b185b498924e7a030008aee2ae -> master.tar.gz
git-fat filter-smudge: restoring from /tmp/repo2/.git/fat/objects/1f218834a137f7b185b498924e7a030008aee2ae

Everything is in place

$ git status
git-fat filter-clean: caching to /tmp/repo2/.git/fat/objects/1f218834a137f7b185b498924e7a030008aee2ae
# On branch master
nothing to commit, working directory clean
$ ls -l                                 # recovered the full file
total 8
-rw-r--r-- 1 jed users 6449 Nov 25 17:10 master.tar.gz

Summary

• Set the "fat" file types in .gitattributes.
• Use normal git commands to interact with the repository without thinking about what files are fat and non-fat. The fat files will be treated specially.
• Synchronize fat files with git fat push and git fat pull.

Retroactive import using git filter-branch [Experimental]

Sometimes large objects were added to a repository by accident or for lack of a better place to put them. If you are willing to rewrite history, forcing everyone to reclone, you can retroactively manage those files with git fat. Be sure that you understand the consequences of git filter-branch before attempting this. This feature is experimental and irreversible, so be doubly careful with backups.

Step 1: Locate the fat files

Run git fat find THRESH_BYTES > fat-files and inspect fat-files in an editor. Lines will be sorted by the maximum object size that has been at each path, and look like

something.big           filter=fat -text #    8154677 1

where the first number after the # is the number of bytes and the second number is the number of modifications that path has seen. You will normally filter out some of these paths using grep and/or an editor. When satisfied, remove the ends of the lines (including the #) and append to .gitattributes. It's best to git add .gitattributes and commit at this time (likely enrolling some extant files into git fat).

Step 2: filter-branch

Copy .gitattributes to /tmp/fat-filter-files and edit to remove everything after the file name (e.g., sed s/ \+filter=fat.*$//). Currently, this may only contain exact paths relative to the root of the repository. Finally, run

git filter-branch --index-filter                 \
    'git fat index-filter /tmp/fat-filter-files --manage-gitattributes' \
    --tag-name-filter cat -- --all

(You can remove the --manage-gitattributes option if you don't want to append all the files being enrolled in git fat to .gitattributes, however, future users would need to use .git/info/attributes to have the git fat fileters run.) When this finishes, inspect to see if everything is in order and follow the Checklist for Shrinking a Repository in the git filter-branch man page, typically git clone file:///path/to/repo. Be sure to git fat push from the original repository.

See the script test-retroactive.sh for an example of cleaning.

Implementation notes

The actual binary files are stored in .git/fat/objects, leaving .git/objects nice and small.

$ du -bs .git/objects
2212    .git/objects/
$ ls -l .git/fat/objects                # This is where the file actually goes, but that's not important
total 8
-rw------- 1 jed users 6449 Nov 25 17:01 1f218834a137f7b185b498924e7a030008aee2ae

If you have multiple clones that access the same filesystem, you can make .git/fat/objects a symlink to a common location, in which case all content will be available in all repositories without extra copies. You still need to git fat push to make it available to others.

Some refinements

• Allow pulling and pushing only select files
• Relate orphan objects to file system
• Put some more useful message in smudged (working tree) version of missing files.
• More friendly configuration for multiple fat remotes
• Make commands safer in presence of a dirty tree.
• Private setting of a different remote.
• Gracefully handle unmanaged files when the filter is called (either legacy files or files matching the pattern that should some reason not be treated as fat).