Git tricks

June 26, 2015 — April 12, 2024

computers are awful
information provenance
workflow
Figure 1

My own git notes, not intended to be tutorial; there are better learning resources than this. Some are noted here, in fact. See also the more universally acclaimed classic git tips.

1 Learning git

See the fastai masterclass for many more helpful tips/links/scripts/recommendations. Learn Git Branching explains the mechanics in a friendly fashion. Steve Bennett’s 10 things I hate about Git is also useful.

Mark Dominus’ master class:

2 Handy git config

2.1 git editor

I often do editing in VS code it is convenient to set it as git editor:

git config --global core.editor "code-insiders --wait"  # insiders

2.2 Ignore macOS .DS_Store files

echo .DS_Store >> .gitignore_global
git config --global core.excludesfile $HOME/.gitignore_global

3 Handy git commands

3.1 Merging ours/theirs

During a merge, git checkout --theirs filename (or --ours) will checkout respectively their (or our) version. The following sweet hack will resolve all files accordingly:

git diff --name-only --diff-filter=U | xargs git checkout --theirs --

I do this a lot and will never remember the details, so here are some aliases for fish which I can use to make this easier:

echo "alias git_theirs 'git diff --name-only --diff-filter=U | xargs git checkout --theirs --'" >  ~/.config/fish/conf.d/git_theirs.fish
echo "alias git_ours   'git diff --name-only --diff-filter=U | xargs git checkout --ours   --'" >> ~/.config/fish/conf.d/git_theirs.fish
chmod a+x ~/.config/fish/conf.d/git_theirs.fish
source ~/.config/fish/conf.d/git_theirs.fish

3.2 Searching

3.2.1 …for a matching file

git grep

3.2.2 …for a matching commit

Easy, except for the abstruse naming; It is called “pickaxe” and spelled -S.

git log -Sword

3.3 track the history of a file including renames

Kashyap Kondamudi advises Use –follow option in git log to view a file’s history.

git log --oneline --find-renames --stat --follow -- src/somefile.ts

3.4 Clone a single branch

git clone --single-branch --branch <branchname> <remote-repo>

3.5 Remove file from versioning without deleting my copy

git rm --cached blah.tmp

3.6 delete remote branch

git push <remote_name> --delete <branch_name>

3.7 Push to a non-obvious branch

git push origin HEAD:refs/heads/backdoor

This is almost obvious except the git naming of things seems… arbitrary? Why refs/heads/SOMETHING?

Read on.

4 What git calls things

By which I mean that which is formally denoted as git references. git references is the canonical description of the mechanics. tl;dr the most common names are refs/heads/SOMETHING for branch SOMETHING, refs/tags/SOMETHING and remotes/SOMEREMOTE/SOMETHING for (last known state of) a remote branch.

As alexwlchan explains, these references are friendly names for commits, and should be thought of as pointers to commits.

And yet there is something a little magical going on. How come if I pull a branch, I get the latest version of that branch, not the earliest to use that name? Other stuff is happening.

The uses are (at least partly) convention and other references can be used too. For example gerrit uses refs/for/ for code review purposes.

5 Filters

Commands applied to your files on the way in and out of the repository. Keywords, smudge, clean, .gitattr These are a long story, but not so complicated in practice. A useful one is stripping crap from jupyter notebooks.

6 Commit hooks

For doing stuff before you put it in cold storage. e.g asking DID YOU REALLY WANT TO INCLUDE THAT GIANT FILE?

Here is a commit hook that does exactly that. I made a slightly modernized version:

curl -L https://gist.github.com/danmackinlay/6e4a0e5c38a43972a0de2938e6ddadba/raw/install.sh | bash

UPDATE: I decided this was a waste of time and removed it.

After that installation you can retrofit the hook to an existing repository thusly

p -R ~/.git_template/hooks .git/

There are various frameworks for managing hooks, if you have lots. For example, pre-commit is a mini-system for managing git hooks, based on python. Husky is a node.js-based one.

I am not sure whether hook management system actually saves time overall for a solo developer, since the kind of person who remembers to install a pre-commit hook is also the kind of person who is relatively less likely to need one. Also it is remarkably labour-intensive to install the dependencies for all these systems, so if you are using heterogeneous systems this becomes tedious.

To skip the pre-commit hook,

git commit --no-verify

7 Subtrees/submodules/subprojects/subdirs/subterranean mole people

Sub-projects inside other projects? External projects? The simplest way of integrating external projects seems to be as subtrees. Once this is set up we can mostly ignore them and things work mostly as expected. Alternatively there are submodules, which have various complications. More recently, there is the subtrac system, which I have not yet used.

7.1 Submodule

Include external projects as separate repositories within a repository is possible, but I won’t document it, since it’s well documented elsewhere, and I use it less often, because it is fiddly, in that Some discipline is required to make it go; you need to remember to git submodule init etc.

7.2 Subtrac

Have not yet tried.

subtrac is a helper tool that makes it easier to keep track of your git submodule contents. It collects the entire contents of the entire history of all your submodules (recursively) into a separate git branch, which can be pushed, pulled, forked, and merged however you want.

7.3 Subtree

Subtree subsumes one git tree into another in a usually-transparent way (no separate checkout as with submodules.) It can be used for temporary merging or for splicing and dicing projects.

7.3.1 Splicing a subtree onto a project

Creatin’:

git fetch remote branch
git subtree add --prefix=subdir remote branch --squash

Updatin’:

git fetch remote branch
git subtree pull --prefix=subdir remote branch --squash
git subtree push --prefix=subdir remote branch --squash

Con: Rebasin’ with a subtree in your repo is slow and involved.

7.3.2 Taking a cutting to make a sub-project

Use subtree split to prise out one chunk. It has various wrinkles but is fast and easy.

pushd superproject
git subtree split -P project_subdir -b project_branch
popd
mkdir project
pushd project
git init
git pull ../superproject project_branch

Alternatively, to comprehensively rewrite history to exclude everything outside a subdir:

pushd superproject
cd ..
git clone superproject subproject
pushd subproject
git filter-branch \
    --subdirectory-filter project_subdir \
    --prune-empty -- \
    --all

7.4 Download a sub-directory from a git tree

This works for github at least. I think anything running git-svn?

Heinous hack

  1. replace tree/master => trunk
  2. svn co the new url
svn co https://github.com/buckyroberts/Source-Code-from-Tutorials/trunk/Python

8 Deleting all tags

git tag -l | xargs -I %% git push -v origin :refs/tags/%%
git tag -l | xargs git tag -d && git fetch -t

9 Conventions

Many possible workflows are feasible with git. Large teams often have elaborate conventions for what to name branches and who gets to merge what with whom. Here are some I have seen in the wild:

Some of these systems have associated helpers, see next.

10 Helpers

Git has various layers of abstraction, from a very basic infrastructure of plumbing, through notionally use-friendly, higher-level porcelain commands which are supposed by authors of git to be user friendly but universally regarded as “a passable first attempt at best”, through to various helper commands that ease various workflows and pain points.

10.1 git-worktree

10.2 git-branchless

arxanas/git-branchless: Branchless workflow for Git

The branchless workflow is designed for use in a repository with a single main branch that all commits are rebased onto. It improves developer velocity by encouraging fast and frequent commits, and helps developers operate on these commits fearlessly.

In the branchless workflow, the commits you’re working on are inferred based on your activity, so you no longer need branches to keep track of them. Nonetheless, branches are sometimes convenient, and git-branchless fully supports them. If you prefer, you can continue to use your normal workflow and benefit from features like git sl or git undo without going entirely branchless.

10.3 git-undo

Also : git undo: We can do better

How is it so easy to “lose” your data in a system that’s supposed to never lose your data?

Well, it’s not that it’s too easy to lose your data — but rather, that it’s too difficult to recover it. For each operation you want to recover from, there’s a different “magic” incantation to undo it. All the data is still there in principle, but it’s not accessible to many in practice.

…To address this problem, I offer git undo

10.4 gerrit

Gerrit is a code review system for git.

10.5 legit

legit simplifies feature branch workflows.

10.6 rerere

Not repeating yourself during merges/rebases? git rerere automates this:

git config --global rerere.enabled true
git config --global rerere.autoupdate true

11 Importing some files across a branch

git checkout my_branch -- my_file/

12 Garbage collecting

In brief, this will purge a lot of stuff from a constipated repo in emergencies:

git reflog expire --expire=now --all && git gc --prune=now

In-depth explanation here.

13 Editing history

13.1 Cleaning out all big files

Figure 2: Every time I find a good picture of an octopus on the internet I put in on my git blog pages

bfg does that:

git clone --mirror git://example.com/some-big-repo.git
cd some-big-repo.git
git repack
bfg --strip-blobs-bigger-than 10M .
git reflog expire --expire=now --all && git gc --prune=now --aggressive
git push -f

13.2 Deleting specific things

I think bfg also does this. There is also native support:

git filter-branch -f \
    --index-filter
    'git rm -r --cached --ignore-unmatch unwanted_files'

13.3 Exporting a minimal history from some repository

i.e. Exporting a branch for a client/collaborator, which they should still operate on in git, but which does not contain all the potentially proprietary stuff in the main repo. Ideally they should see one commit with no history.

If the merge history is clean, there is no need to be fancy; if I have a branch which ahs never merged in any secret information then I can just push it to a new repository and it won’t bring along any of the secret stuff.

OTOH, research code is often unhygienic and chaotic, so we might need to be more careful.

Option 0: Export a tarball and then forget about git:

git archive HEAD --format=zip > archive.zip

13.3.1 Option 1: squash the whole thing onto a single commit.

I don’t know a faster way of doing this than the classic:

git checkout -b temp_branch
git rebase --root -i

Addendum: possibly this bash hack is superior?

git reset $(git commit-tree HEAD^{tree} -m "A new start")

13.3.2 Option 2: create an orphan branch and copy the files over

Here is an example.

git checkout --orphan temp_branch
git add -A
git commit -am "Initial commit"
git branch -D master
git branch -m master
git push -f origin master

Merging with branches created this way can be tedious, however.

13.3.3 Option 3: Serendipitous orphan

create an orphaned commit that exactly matches an existing commit:

TREE=`git cat-file -p master |sed '1,/^$/s/^tree //p;d;'`
COMMIT=`echo Truncated tree | git commit-tree $TREE`
git branch truncated-master $COMMIT
git branch backup-master-just-in-case-i-regret-it-later master
git push -f origin truncated-master:master

I think this possibly allows us to easily cherry-pick commits against the new tree and return to the original.

14 Making git work with a broken-permission FS

e.g. you are editing a git repo on NTFS via Linux and things are silly.

git config core.filemode false

15 Detecting if there are changes to commit

Thomas Nyman:

if output=$(git status --porcelain) && [ -z "$output" ]; then
  # Working directory clean
else
  # Uncommitted changes
fi

16 Emergency commit

Oh crap I’m leaving the office in a hurry and I just need to get my work into git ASAP for continuing on another computer. I don’t care about sensible commit messages because I am on my own private branch and no-one else will see them when I squash the pull request.

I put this little script in a file called gitbang to automate the this case.

#!/usr/bin/env bash
# I’m leaving the office. Capture all changes in my private branch and push to server.
if output=$(git status --porcelain) && [ -z "$output" ]; then
  echo "nothing to commit"
else
  git add --all && git commit -m bang
fi
git pull && git submodule update --init --recursive  && git push

Or: this could be a fish function why not

function gitbang --argument-names remoteref --description "push to remote"
  set pushrel ""
  set -q remoteref; or set pushrel HEAD:remoteref
  set gitstatus (git status --porcelain)
  if (test -z "$gitstatus" )
    echo "nothing to commit"
  else
    git add --all ; and git commit -m bang
  end
  git pull ; and git submodule update --init --recursive ; and git push
end

Pro tip: if you use VS code there is a feature called Cloud Changes that synchronises your changes to the cloud, so you can pick up where you left off on another computer without arsing about with git.

17 Git hosting

Or maybe avoid the need for hosting by using…

18 git email workflows

Learn to use email with git.

19 Content-specific diffing

Tools such as git-latexdiff provide custom diffing for, in this case, LaTeX code. These need to be found on a case-by-case basis.

20 SSH credentials

Managing SSH credentials in git is non-obvious. See SSH.

21 Jupyter

For sanity in git+jupyter, see jupyter.

22 Decent GUIs

See Git GUIs.

23 Which repo am I in?

For fish and bash shell, see bash-git-prompt.

24 Data versioning

See data versioning.

25 Incoming