159

This question is based on Detach subdirectory into separate Git repository

Instead of detaching a single subdirectory, I want to detach a couple. For example, my current directory tree looks like this:

/apps
  /AAA
  /BBB
  /CCC
/libs
  /XXX
  /YYY
  /ZZZ

And I would like this instead:

/apps
  /AAA
/libs
  /XXX

The --subdirectory-filter argument to git filter-branch won't work because it gets rid of everything except for the given directory the first time it's run. I thought using the --index-filter argument for all unwanted files would work (albeit tedious), but if I try running it more than once, I get the following message:

Cannot create a new backup.
A previous backup already exists in refs/original/
Force overwriting the backup with -f

Any ideas? TIA

Community
  • 1
  • 1
prisonerjohn
  • 2,250
  • 2
  • 16
  • 19

11 Answers11

178

Instead of having to deal with a subshell and using ext glob (as kynan suggested), try this much simpler approach:

git filter-branch --index-filter 'git rm --cached -qr --ignore-unmatch -- . && git reset -q $GIT_COMMIT -- apps/AAA libs/XXX' --prune-empty -- --all

As mentioned by void.pointer's comment, this will remove everything except apps/AAA and libs/XXX from current repository.

Prune empty merge commits

This leaves behind lots of empty merges. These can be removed by another pass as described by raphinesse in his answer:

git filter-branch --prune-empty --parent-filter \
'sed "s/-p //g" | xargs -r git show-branch --independent | sed "s/\</-p /g"'

⚠️ Warning: The above must use GNU version of sed and xargs otherwise it would remove all commits as xargs fails. brew install gnu-sed findutils and then use gsed and gxargs:

git filter-branch --prune-empty --parent-filter \
'gsed "s/-p //g" | gxargs git show-branch --independent | gsed "s/\</-p /g"'
Josh Correia
  • 3,807
  • 3
  • 33
  • 50
David Smiley
  • 4,102
  • 2
  • 19
  • 18
  • thanks David this solution works with me unlike git stitch which duplicated each commit more than once – Mahmoud Adam Nov 03 '13 at 00:14
  • 4
    additionally, the --ignore-unmatch flag should be passed to git rm, it failed for the very first commit for me otherwise (the repository was created with git svn clone in my case) – Pontomedon Jul 24 '14 at 07:06
  • 1
    All I get from this command is lots of "duplicate parent" errors. – aaa90210 Apr 17 '15 at 04:04
  • 8
    Assuming you have tags in the mix, you should probably add `--tag-name-filter cat` to your parameters – Yonatan May 27 '15 at 19:03
  • 18
    Could you add some more information explaining what this lengthy command is doing? – Burhan Ali Sep 18 '18 at 10:26
  • 4
    I'm pleasantly surprised that this works perfectly on Windows using git bash, phew! – Dai Oct 20 '18 at 00:03
  • 6
    @BurhanAli For every commit in history, it is deleting all files except the ones you want to keep. When everything is done, you are left with only the portion of the tree you specified, along with only that history. – void.pointer Feb 15 '19 at 15:42
  • 1
    FYI: If you use `;` instead of `&&` in the index filter command, you do not need the `--ignore-unmatch` option. – void.pointer Feb 18 '19 at 15:24
  • 1
    The folder name is case sensitive even on Windows, if you specify a path name that have a different case than actual case, unexpected result will occurr (that folder will be deleted instead). – aruku7230 Apr 02 '19 at 07:59
  • 2
    If one of the paths does not exist in a given commit, I get `fatal: bad revision ` and the command aborts. To be clear, I am specifying files rather than directories. – Quolonel Questions May 24 '19 at 13:59
  • what is the purpose of the `--cached` attribute? – Martin Delille Jun 19 '19 at 17:25
  • 2
    As the name suggests, `index-filter` operates on the index, so you need `--cached` for `git rm` to also operate on the index (see `git help rm`). – kynan Jul 27 '19 at 19:29
  • I modified this and used it for 1 folder and 2 files where it had `apps/AAA libs/XXX`. Worked like a charm. No idea how, but that's what makes Stack Overflow great. – AlanSE Aug 30 '19 at 16:56
  • 2
    You should mention in your answer that the environment variable $GIT_COMMIT is to get the folders from a specific commit and if unset/empty come from HEAD – colin Nov 29 '19 at 02:04
  • If one _must_ use `git-filter-branch`—and cannot use `--subdirectory-filter` because they want to operate against multiple folders, as asked here—then this almost certainly the best approach of the options listed both here and elsewhere. `--index-filter` is much faster than operating against the working tree. And, perhaps more importantly, the nested script operates exclusively off `git` commands, and thus should behave similarly across different operating environments. – Jeremy Caney Dec 28 '19 at 02:24
  • That said, it’s worth noting that this will rewrite your history, of course, but won’t update any commit references in your commit messages. So if your team uses #refs in your commit messages, and it’s important to you that those continue to work, you’ll want to be aware of that. And, of course, as with _any_ history rewrite, you’ll probably want to push to a new repository, so you don’t end up with potential problems when other collaborators push commits from now-orphaned histories. (Alternatively, ensure that everyone on your team reclones their copy of the repository on each device.) – Jeremy Caney Dec 28 '19 at 02:31
  • `Could not delete refs/tags/v0.2.3` I receive this error. – Sohail Si Feb 03 '20 at 15:29
  • 1
    I got this error message: `Cannot create a new backup. A previous backup already exists in refs/original/ Force overwriting the backup with -f`. So this worked: `git filter-branch --prune-empty -f ... ` – jherb Jun 04 '20 at 11:08
  • 1
    `git filter-branch` is terribly slow and deprecated, I recommend the `git filter-repo` alternative https://stackoverflow.com/a/61410689/1507124 for anything non trivial – CervEd Mar 12 '21 at 07:23
55

An easy solution: git-filter-repo

I had a similar issue and, after reviewing the various approaches listed here, I discovered git-filter-repo. It is recommended as an alternative to git-filter-branch in the official git documentation here.

To create a new repository from a subset of directories in an existing repository, you can use the command:

git filter-repo --path <file_to_keep>

Filter multiple files/folders by chaining them:

git filter-repo --path keepthisfile --path keepthisfolder/

So, to answer the original question, with git-filter-repo you would just need the following command:

git filter-repo --path apps/AAA/ --path libs/XXX/
Adrian W
  • 4,563
  • 11
  • 38
  • 52
elmo
  • 561
  • 4
  • 5
  • 1
    This is definitely a great answer. The problem with all other solutions is that I couldn't manage to extract the contents of ALL branches of a directory. However, git filter-repo retrieved the folder from all branches and rewrote history perfectly, like cleaning the whole tree of everything I didn't need. – Teodoro Jun 18 '20 at 23:05
  • 1
    I used your answer. It is a great solution. Then somehow it was lost within many open tabs. I had to literally search from my browser history to find you and say thanks. – Ahmad Ismail Aug 26 '20 at 12:29
42

Manual steps with simple git commands

The plan is to split individual directories into its own repos, then merge them together. The following manual steps did not employ geek-to-use scripts but easy-to-understand commands and could help merge extra N sub-folders into another single repository.

Divide

Let's assume your original repo is: original_repo

1 - Split apps:

git clone original_repo apps-repo
cd apps-repo
git filter-branch --prune-empty --subdirectory-filter apps master

2 - Split libs

git clone original_repo libs-repo
cd libs-repo
git filter-branch --prune-empty --subdirectory-filter libs master

Continue if you have more than 2 folders. Now you shall have two new and temporary git repository.

Conquer by Merging apps and libs

3 - Prepare the brand new repo:

mkdir my-desired-repo
cd my-desired-repo
git init

And you will need to make at least one commit. If the following three lines should be skipped, your first repo will appear immediate under your repo's root:

touch a_file_and_make_a_commit # see user's feedback
git add a_file_and_make_a_commit
git commit -am "at least one commit is needed for it to work"

With the temp file commited, merge command in later section will stop as expected.

Taking from user's feedback, instead of adding a random file like a_file_and_make_a_commit, you can choose to add a .gitignore, or README.md etc.

4 - Merge apps repo first:

git remote add apps-repo ../apps-repo
git fetch apps-repo
git merge -s ours --no-commit apps-repo/master # see below note.
git read-tree --prefix=apps -u apps-repo/master
git commit -m "import apps"

Now you should see apps directory inside your new repository. git log should show all relevant historical commit messages.

Note: as Chris noted below in the comments, for newer version(>=2.9) of git, you need to specify --allow-unrelated-histories with git merge

5 - Merge libs repo next in the same way:

git remote add libs-repo ../libs-repo
git fetch libs-repo
git merge -s ours --no-commit libs-repo/master # see above note.
git read-tree --prefix=libs -u libs-repo/master
git commit -m "import libs"

Continue if you have more than 2 repos to merge.

Reference: Merge a subdirectory of another repository with git

chfw
  • 4,502
  • 2
  • 29
  • 32
  • 4
    Since git 2.9 you need to use --allow-unrelated-histories on the merge commands. Otherwise this appears to have work well for me. – Chris Sep 11 '17 at 20:12
  • 1
    Genius! Thank you so much for this. The initial answers I'd looked at, using a tree filter on a very large repository, had git predicting taking over 26hrs to complete the git rewrites. Much happier with this simple, but repeatable approach and have successfully moved 4 sub folders into a new repo with all expected commit history. – shuttsy Jun 27 '18 at 12:12
  • Are there any downsides of doing it in two steps as described above? – Mariusz Jamro Jul 26 '18 at 06:26
  • One downside is: there is an extra merge commit, because you will need to make at least one commit otherwise this approach won't work. – chfw Jul 26 '18 at 11:40
  • 1
    You can use the first commit for a "Initial commit" which adds `.gitignore` and `README.md` files. – Jack Miller Aug 21 '18 at 12:36
  • @JackMiller, yours too. – chfw Aug 23 '18 at 20:16
  • 3
    Unfortunately this approach seems to break tracking-history for the files added in the `git merge .. git read-tree` step, as it records them as newly-added files and all of my git guis don't make the connection to their earlier commits. – Dai Oct 19 '18 at 23:47
  • This works well. However, is there a way I can merge all branches (simply create new ones in the new repo?) – jschober Dec 24 '18 at 22:52
  • @chfw do you have any idea about how to fix newly-added files issue? It seems like in merging command somehow we lost track for relative path of the directories – ksadjad Aug 07 '19 at 06:47
  • 1
    @ksadjad, No idea, to be honest. The central point of the manual merge is to select the directories to form the new repo and keep their commit histories. I am not sure how to handle such situation where a commit put files into dirA, dirB, dirDrop and only dirA and dirB are chosen for the new repo, how should the commit history relate to the original one. – chfw Aug 07 '19 at 07:36
28

Why would you want to run filter-branch more than once? You can do it all in one sweep, so no need to force it (note that you need extglob enabled in your shell for this to work):

git filter-branch --index-filter "git rm -r -f --cached --ignore-unmatch $(ls -xd apps/!(AAA) libs/!(XXX))" --prune-empty -- --all

This should get rid of all the changes in the unwanted subdirectories and keep all your branches and commits (unless they only affect files in the pruned subdirectories, by virtue of --prune-empty) - no issue with duplicate commits etc.

After this operation the unwanted directories will be listed as untracked by git status.

The $(ls ...) is necessary s.t. the extglob is evaluated by your shell instead of the index filter, which uses the sh builtin eval (where extglob is not available). See How do I enable shell options in git? for further details on that.

Community
  • 1
  • 1
kynan
  • 13,235
  • 6
  • 79
  • 81
  • 1
    Interesting idea. I have a similar problem but could not get it to work, see http://stackoverflow.com/questions/8050687/how-do-i-enable-shell-options-in-git – manol Nov 10 '11 at 10:43
  • This is pretty much what I needed, though I had sprinkling of both files and folders across my repo... Thanks :) – notlesh Dec 05 '11 at 00:58
  • 1
    hm. even with extglob turned on I'm getting an error near my parenthesis: syntax error near unexpected token `(' my command looks like: git filter-branch -f --index-filter "git rm -r -f --cached --ignore-unmatch src/css/themes/!(some_theme*)" --prune-empty -- --all an ls with src/css/themes/!(some_theme*) returns all the other themes so extglob does appear to be working... – robdodson Dec 02 '12 at 19:01
  • You can get your shell to evaluate the glob as described in http://stackoverflow.com/a/8079852/396967 – kynan Dec 03 '12 at 19:56
  • all i had to do was properly escape the characters. (eg "\!", "\(", "\|", "\)" . – Mike Graf Jan 30 '13 at 19:24
  • 2
    @MikeGraf I don't think that will give the desired result: escaping would match a literal "!" etc. in your path. – kynan Jan 31 '13 at 01:05
  • Kynan you're right, it didnt actually work. It ran, but didnt give the result desired.. ("I dont understand why it doesnt work, it compiled!!" :P ) – Mike Graf Jan 31 '13 at 17:22
  • On Ubuntu, I'm finding that ls -x doesn't put this on one line. I had to use $(ls -md apps/!(AAA) libs/!(XXX) | sed -e 's/, /,/g' | sed -e 's/ /\\ /g' | sed -e 's/&/\\&/g' | sed -e 's/,/ /g' | tr -d '\r\n') The -m separated the output by commas instead of the column based output I was getting. Ands seds to deal with spaces, ampersands in file names before replacing the commas with spaces and getting rid of newlines. – Bae Mar 20 '14 at 00:25
  • Kynan pointed out (stackoverflow.com/a/8079852/396967) that ls -w would get the ls back on one line. – Bae Mar 23 '14 at 22:55
  • I was getting errors like: "/usr/lib/git-core/git-filter-branch: line 415: foo.txt: command not found" It turned out that this was caused by multi-line output from ls, and adding -w 1000 solved the problem. – retroj Jul 23 '19 at 14:49
  • Another issue that I ran into was that I wanted to preserve files which matched the patterns: foo* and quux/foo*. The pattern !(foo*) was matching 'quux' so all of its contents were being deleted despite the second pattern. The solution was: !(foo*|quux) quux/!(foo*) – retroj Jul 23 '19 at 14:52
  • 1
    @david-smiley’s (more recent) answer uses a very similar approach, but has the advantage of relying exclusively on `git` commands, and thus isn’t as susceptible to differences in how the `ls` is interpreted across operating systems, as @Bae discovered. – Jeremy Caney Dec 28 '19 at 02:38
21

Answering my own question here... after a lot of trial and error.

I managed to do this using a combination of git subtree and git-stitch-repo. These instructions are based on:

First, I pulled out the directories I wanted to keep into their own separate repository: 

cd origRepo
git subtree split -P apps/AAA -b aaa
git subtree split -P libs/XXX -b xxx

cd ..
mkdir aaaRepo
cd aaaRepo
git init
git fetch ../origRepo aaa
git checkout -b master FETCH_HEAD

cd ..
mkdir xxxRepo
cd xxxRepo
git init
git fetch ../origRepo xxx
git checkout -b master FETCH_HEAD

I then created a new empty repository, and imported/stitched the last two into it:

cd ..
mkdir newRepo
cd newRepo
git init
git-stitch-repo ../aaaRepo:apps/AAA ../xxxRepo:libs/XXX | git fast-import

This creates two branches, master-A and master-B, each holding the content of one of the stitched repos. To combine them and clean up:

git checkout master-A
git pull . master-B
git checkout master
git branch -d master-A 
git branch -d master-B

Now I'm not quite sure how/when this happens, but after the first checkout and the pull, the code magically merges into the master branch (any insight on what's going on here is appreciated!)

Everything seems to have worked as expected, except that if I look through the newRepo commit history, there are duplicates when the changeset affected both apps/AAA and libs/XXX. If there is a way to remove duplicates, then it would be perfect.

Josh Lee
  • 171,072
  • 38
  • 269
  • 275
prisonerjohn
  • 2,250
  • 2
  • 16
  • 19
  • Neat tools you found here. Insight on "checkout": "git pull" is the same as "git fetch && git merge". The "fetch" part is innocuous since you are "fetching locally". So I think this checkout command is the same as "git merge master-B", which is a bit more self-evident. See http://www.kernel.org/pub/software/scm/git/docs/git-pull.html – Phil Hord Jul 27 '10 at 22:37
  • 1
    Unfortunately the git-stitch-repo tool is broken due to bad dependencies nowadays. – Henrik Jan 28 '13 at 13:18
  • @Henrik What problem were you experiencing exactly? It works for me, although I had to add `export PERL5LIB="$PERL5LIB:/usr/local/git/lib/perl5/site_perl/"` to my bash config so that it could find Git.pm. Then I installed it with cpan. –  Mar 27 '13 at 03:54
  • It's possible to use `git subtree add` to perform this task. See https://stackoverflow.com/a/58253979/1894803 – laconbass Oct 06 '19 at 04:11
7

I have writen a git filter to solve exactly this problem. It has the fantastic name of git_filter and is located at github here:

https://github.com/slobobaby/git_filter

It is based on the excellent libgit2.

I needed to split a large repository with many commits (~100000) and the solutions based on git filter-branch took several days to run. git_filter takes a minute to do the same thing.

slobobaby
  • 696
  • 7
  • 10
7

Use 'git splits' git extension

git splits is a bash script that is a wrapper around git branch-filter that I created as a git extension, based on jkeating's solution.

It was made exactly for this situation. For your error, try using the git splits -f option to force removal of the backup. Because git splits operates on a new branch, it won't rewrite your current branch, so the backup is extraneous. See the readme for more detail and be sure to use it on a copy/clone of your repo ( just in case!).

  1. install git splits.

  2. Split the directories into a local branch #change into your repo's directory cd /path/to/repo #checkout the branch git checkout XYZ
    #split multiple directories into new branch XYZ git splits -b XYZ apps/AAA libs/ZZZ

  3. Create an empty repo somewhere. We'll assume we've created an empty repo called xyz on GitHub that has path : git@github.com:simpliwp/xyz.git

  4. Push to the new repo. #add a new remote origin for the empty repo so we can push to the empty repo on GitHub git remote add origin_xyz git@github.com:simpliwp/xyz.git #push the branch to the empty repo's master branch git push origin_xyz XYZ:master

  5. Clone the newly created remote repo into a new local directory
    #change current directory out of the old repo cd /path/to/where/you/want/the/new/local/repo #clone the remote repo you just pushed to git clone git@github.com:simpliwp/xyz.git

Community
  • 1
  • 1
AndrewD
  • 4,924
  • 3
  • 30
  • 32
  • It does not seem to be possible to add files to the split and update them later, right? – Alex May 21 '17 at 14:44
  • This seems to slow to run on my repo with tons of commits – Shinta Smith Oct 18 '17 at 15:57
  • git-split seems to use [git --index filter](https://github.com/ajdruff/git-splits/blob/master/git-splits#L78) which is extremely slow compared to --subdirectory-filter. For some repos it may still be a viable option, but for big repos (multiple gigabytes, 6-digit commits) --index-filter effectively takes weeks to run, even on dedicated cloud hardware. – Jostein Kjønigsen Mar 14 '18 at 06:21
6
git clone git@example.com:thing.git
cd thing
git fetch
for originBranch in `git branch -r | grep -v master`; do
    branch=${originBranch:7:${#originBranch}}
    git checkout $branch
done
git checkout master

git filter-branch --index-filter 'git rm --cached -qr --ignore-unmatch -- . && git reset -q $GIT_COMMIT -- dir1 dir2 .gitignore' --prune-empty -- --all

git remote set-url origin git@example.com:newthing.git
git push --all
Richard Barraclough
  • 2,625
  • 3
  • 36
  • 54
  • Reading through all the other comments got me on the right track. However, your solution just works. It imports all branches, and works with multiple directories! Great! – jschober Dec 24 '18 at 23:38
  • 1
    The `for` loop is worth acknowledging, since other similar answers don’t include it. If you don’t have a local copy of each branch in your clone, then `filter-branch` won’t account for them as part of its rewrite, which could potentially exclude files introduced in other branches, but not yet merged with your current branch. (Though it‘s also worth doing a `git fetch` on any branches you have previously checked out to ensure that they remain current.) – Jeremy Caney Dec 28 '19 at 08:44
3

Yeah. Force overwriting the backup by using the -f flag on subsequent calls to filter-branch to override that warning. :) Otherwise I think you have the solution (that is, eradicate an unwanted directory at a time with filter-branch).

Jakob Borg
  • 23,685
  • 6
  • 47
  • 47
0

I think it is easier to just export the commits that touch those paths:

git log --pretty=email --patch-with-stat --reverse --full-index --binary -- /apps/{AAA,BBB,CCC} /libs/{XXX,YYY,ZZZ} > subdir.patch

and then import those commits into a new repo:

git am < subdir.patch

If you have merge commits that cannot be rebased, you may want to try with -m --first-parent:

git log --pretty=email --patch-with-stat --reverse --full-index --binary -m --first-parent -- <your paths>
Christian Rauch
  • 86
  • 3
  • 10
-6

Delete the backup present under the .git directory in refs/original like the message suggests. The directory is hidden.

user5200576
  • 1