Page History

Freaky Git shit

Nigel Stanger edited this page on 28 Mar

Clone this wiki locally

LFS

Migrate a large repository to LFS in GitBucket

  1. Create a new repository newrepo on the server by cloning the old one.
  2. Clone newrepo to local.
  3. Initialise git-lfs in newrepo as per the git-lfs tutorial:
    git lfs track "*.thing1" "*.thing2"
    git add .gitattributes
    git commit -m "Track thing1, thing2 files (LFS)"
    git rm --cached "*.thing1" "*.thing2"
    git add .
    git commit -m "Convert last commit to LFS"
  4. Migrate the history as per the git-lfs tutorial:
    git lfs migrate import --everything --include="*.thing1,*.thing2"
    If you specify multiple --include switches only the last one will take effect, so to migrate multiple file types, put the glob expressions inside the same string as shown above.
  5. We now need to force push newrepo to remote, but it may be too large for nginx’s request size limits. If you happen to have release tags handily scattered throughout history (or you can identify some well-spaced commit hashes) you can reduce the request size by only pushing up to a certain commit, e.g.:
    git push --force origin 9d42ad26508d94429ce2179358db5c2bfc1dc03d:master
    # ...etc...
    # one last push to catch everything since the last tag:
    git push --force
    Do this for each of the tags/commits in chronological order. You can find the commit hashes for the tags using git show-ref --tags.
  6. Also force push the rewritten tags to prevent push/pull errors due to the different commit hashes:
    git push --force --tags
  7. Clean up local as per the git-lfs tutorial:
    git reflog expire --expire-unreachable=now --all
    git gc --prune=now

Migrate LFS repo to a different remote

Normally you just clone the old remote, create the new remote on GitBucket, then git push -u origin master (and maybe git push --tags). However this isn’t aware of LFS files, so it will only push the pointers and all your LFS files will be missing on the remote. The correct procedure:

  1. git clone <old remote>
  2. git lfs fetch --all in the local clone to make sure the clone has all of the LFS objects (including historical versions). If you don’t do this the LFS push later on may fail due to missing objects.
  3. git remote set-url <new remote> in the local clone.
  4. git push -u origin master as usual in the local clone to push all the non-LFS objects to the remote.
  5. git push --tags in the local clone if needed.
  6. git lfs push --all origin master in the local clone to push all the LFS objects to the remote. If this is too large you can use the same tag-based partial push trick as in the previous section. It may or may not like having :master on the end — I found it worked fine with just a commit hash.

Finding repositories that haven’t been pushed

Something along these lines (from comments on https://stackoverflow.com/a/12499489):

find . -name .git -type d -print -exec git --git-dir={} --work-tree={}/.. cherry -v \;

Migrate a wiki repo to another GitBucket server

  1. Clone the original repository that contains the wiki (old-repo).
  2. Clone the actual wiki itself (old-wiki).
  3. Create empty new-repo on new server.
  4. Transfer ownership of new-repo if necessary. Do this before any of the remaining steps.
  5. In old-repo clone:
    git remote set-url origin <new-repo-url>
    git push
  6. In old-wiki clone:
    git remote set-url origin <new-wiki-url>
    git push -f
  7. Update wiki link in README.md of new-repo.
  8. Profit!

Ignoring local changes

A common problem with PDF outputs under Gradle is that they frequently end up being marked as out of date even though the original source (e.g., PlantUML or SVG) hasn’t changed. (I think this is because the PDF embeds metadata like its creation date and is therefore technically different every time.) This is annoying if the output is under version control. To avoid accidentally committing and pushing such changes, it’s useful to tell Git to ignore the changes (https://compiledsuccessfully.dev/git-skip-worktree/):

git update-index --skip-worktree <file>

This isn’t the intended use of skip-worktree; it’s really designed for sparse checkouts. However, it works quite well for the case of “I want to have this file checked out but not to worry about changes most of the time”. This is as opposed to git update-index --assume-unchanged, which assumes that the marked file shouldn’t be changed (https://stackoverflow.com/a/13631525.)

According to the Compiled Successfully link git pull should still work.

If you do need to commit the changes to the file, you will need to undo the ignorance:

git update-index --no-skip-worktree <file>

To list all skipped files:

git ls-files -v|grep '^S'

Managing a master/release branch (e.g., for assignments)

Consider the INFO 201 project, privately developed on isgb/teaching and publicly released on isgb/info201. The simplest way to manage the release process without accidentally pushing to the wrong repo is to have local master track the remote master on isgb/teaching (which should already be the case if things are correctly configured) and a local release branch tracking remote master on isgb/info201. Use git -vv to check which branches are tracking which remotes.

Create your remote repos on isgb/teaching and isgb/info201, then:

git remote add origin https://isgb.otago.ac.nz/teaching/git/INFO201/project.git
git push -u origin master
# or just git clone if the dev repo already exists on isgb/teaching
git remote add release https://isgb.otago.ac.nz/info201/git/shared/project.git
git branch release
git push -u release release:master

Development occurs on on local master and local master is pushed only to origin/master. Release only occurs when local master is merged onto local release. A plain git push therefore can’t accidentally release things. You have to explicitly merge, checkout local release and push from there:

git checkout release
git merge master
git push release HEAD:master

This could be done without the local release branch (git push origin master and git push release master), but there’s still the danger that a plain git push will accidentally release something. A local release branch adds more belts and braces against that:

  • Release doesn’t occur at all unless you merge locally.
  • Plain git push yells at you when the local and remote branch names are different.