Version control systems allow you to record the history of files and share files among several computers and collaborators in a professional way. Especially when working with software development or technical writing, it is essential to have file changes documented and to ensure that every computer and person involved in the project have the most recent updates of the files.

Projects that you want to share among several computers or project workers are most conveniently stored at some web site "in the cloud", here called project hosting services. For efficient access to the files, and the possibility to work offline, you interact with local copies of the files on your computers. I strongly recommend you to use such sites and version control for all serious programming and scientific writing work.

Motivation

Greg Wilson's excellent Script for Introduction to Version Control provides a detailed motivation why you will benefit greatly from using version control systems. Here follows a shorter motivation and a quick overview of the basic concepts.

Why not Dropbox or Google Drive?

The simplest services for hosting project files are Dropbox and Google Drive. It is very easy to get started with these systems, and they allow you to share files among laptops and mobile units with as many users as you want. The systems offer a kind of version control in that the files are stored frequently (several times per minute), and you can go back to previous versions for the last 30 days. However, it is challenging to find the right version from the past when there are so many of them and when the different versions are not annotated with sensible comments. Another deficiency of Dropbox and Google Drive is that they sync all your files in a folder, a feature you clearly do not want if there are many large files (simulation data, visualizations, movies, binaries from compilations, temporary scratch files, automatically generated copies) that can easily be regenerated.

However, the most serious problem with Dropbox and Google Drive arises when several people edit files simultaneously: it can be difficult detect who did what when, roll back to previous versions, and to manually merge the edits when these are incompatible. Then one needs more sophisticated tools, which means a true version control system. The following text aims at providing you with the minimum information to started with Git, the leading version control system, combined with project hosting services for file storage.

Repositories and local copies

The mentioned services host all your files in a specific project in what is known as a repository, or repo for short. When a copy of the files are wanted on a certain computer, one clones the repository on that computer. This creates a local copy of the files. Now files can be edited, new ones can be added, and files can be deleted. These changes are then brought back to the repository. If users at different computers synchronize their files frequently with the repository, most modern version control systems will be able to merge changes in files that have been edited simultaneously on different computers. This is perhaps one of the most useful features of project hosting services. However, the merge functionality clearly works best for pure text files and less well for binary files, such as PDF files, MS Word or Excel documents, and OpenOffice documents.

Installing Git

The installation of Git on various systems is described on the Git website under the Download section. Git involves compiled code so it is most convenient to download a precompiled binary version of the software on Windows, Mac and other Linux computers. On Ubuntu or any Debian-based system the relevant installation command is

Terminal> sudo apt-get install git gitk git-doc

This tutorial explains Git interaction through command-line applications in a terminal window. There are numerous graphical user interfaces to Git. Three examples are git-cola, TortoiseGit, and SourceTree.

Configuring Git

Make a file .gitconfig in your home directory with information on your full name, email address, your favorite text editor, and the name of an "excludes file" which defines the file types that Git should omit when bringing new directories under version control. Here is a simplified version of the author's .gitconfig file:

[user]
name = Hans Petter Langtangen
email = hpl@simula.no
editor = emacs

[core]
excludesfile = ~/.gitignore

The excludesfile variable is important: it points to a file called .gitignore, which must list, using the Unix Shell Wildcard notation, the type of files that you do not need to have under version control, because they represent garbage or temporary information, or they can easily be regenerated from some other source files. A suggested .gitignore file looks like

# compiled files:
*.o
*.so
*.a
# temporary files:
*.bak
*.swp
*~
.*~
*.old
tmp*
.tmp*
temp*
.#*
\#*
# tex files:
*.log
*.dvi
*.aux
*.blg
*.idx
*.nav
*.out
*.toc
*.snm
*.vrb
*.bbl
*.ilg
*.ind
*.loe
# eclipse files:
*.cproject
*.project
# misc:
.DS_Store

GitHub

Go to github.com and create an account. Then go to your account settings (icon in the upper left corner of the page), choose SSH Keys, and provide your SSH key unless you have already registered this key with another GitHub account (see Appendix: Working with multiple GitHub accounts). Often, it is just a matter of pasting the contents of id_rsa.pub or id_dsa.pub files, located in the .ssh subdirectory of your home directory, into the Key box in the web page. Make sure to just cut and paste the text from, e.g., id_rsa.pub without any extra whitespaces or other text. How to generate these files is described in the link generating SSH keys above the SSH Keys box.

If the account is a project account and not a personal account, I do not recommend to provide an SSH key although it can be done (see Appendix: Working with multiple GitHub accounts). It is easier to log in and add collaborators using their personal GitHub usernames.

Creating a new project

Click on New repository on the main page and fill out a project name, here My Project, click the check button Initialize this repository with a README, and click on Create repository. Unless you pay, all repos are public, but students and teachers can request free, private repos.

The next step is to clone the project on your personal computer. Click on the SSH button to see the address of the project, and paste this address into a terminal window, after git clone:

Terminal> git clone git://github.com:user/My-Project.git

Make sure you substitute user by your own username on GitHub.

The result of the git clone command is a new directory My-Project. It contains the file .git, which shows that it is a Git repository. It also contains a default README.md file with the project name and description. The extension .md signifies a file written in the Markdown format. You may use the reStructuredText format as an alternative (README.rst), or simply write a plain text file (README), but the git mv command must be used to change the filename.

You can now add files and directories into the My-Project directory. When your initial file collection has the desired form, you must run

Terminal> git add .
Terminal> git commit -am 'First set of files.'
Terminal> git push -u origin master

The daily file operations are explained in the section Using Git.

Collaborating

To give others permissions to push their edits of files to the repository, you click on the Settings link in the right sidebar, then click on Collaborators on the left, and fill in the name of a collaborator (her or his username on GitHub). Many find it convenient to be notified in email when others have pushed a new version of the files to the repo. Click on Service Hooks in the project's Settings menu, choose Email, fill in at most two whitespace-separated email addresses, mark the Send from Author and Active boxes, and click on Update Settings. More addresses must be dealt with through a mailing list and filling in the name of that list.

Anyone who participates in a project (has write access) or watches a project (having clicked the watch button) can monitor the development of the activity on their GitHub main page. Go to Account Settings and choose Notification Center. There you see two sections, Participating and Watching, for those participating in the project (granted write access) and those watching the project (having clicked the watch button), respectively.

Wiki pages

With every GitHub project there is an option to create wiki pages. Click on the Wiki button in the right sidebar on the main page of the project. Click on New Page to create a new page. The wiki pages can be written in different markup languages. Markdown is the default choice, but you can alternatively use MediaWiki and reStructuredText. Unfortunately, GitHub wiki pages do not allow LaTeX mathematics through MathJax, even though MediaWiki has support for LaTeX (the reason is security issues).

The wiki pages can be written and maintained through the web browser interface, but it is usually more convenient to clone them on your computer as this makes it easy to add figures and other documents you may want to link to. It also makes it straightforward to edit the wiki text in your favorite text editor. The wiki pages are stored in a separate repo and can be cloned by

Terminal> git clone git://github.com/user/My-Project.wiki.git

This command makes a local copy of the pages in the directory My-Project.wiki, which you may prefer to have at the same level as the project directory itself in your directory tree.

Each wiki page has its own file, where the extension reflects the markup language used, e.g., .md for Markdown, .rest for reStructuredText, .mediawiki for MediaWiki, and .creole for Creole wiki. The wiki files are handled as other files in a GitHub project, i.e., you need to pull before editing and then perform commit and push. After the push you can reload the page in the web browser to monitor the effect.

You may consider having the original text in doconce format and generate the wiki in the reStructuredText or MediaWiki format.

Do changes, commit the usual way, and push by

Terminal> git push git@github.com:user/My-project.wiki.git

The address can be stored as url in .git/config in the root directory of the wiki project so that just a standard git push works.

Project web pages

HTML pages stored in your repo cannot be linked to and properly rendered as web pages. Say you have some HTML file doc/file.html in the repo. The normal link to the file is

https://github.com/user/my-project/blob/master/doc/file.html

which shows up as a nicely typeset, colorful HTML code. The raw text file,

https://raw.githubusercontent.com/user/my-project/master/doc/file.html

shows up as pure text in a browser. If one wants to see the file rendered as HTML code, one can view it through htmlpreview.github.io. This means that one can use the link

http://htmlpreview.github.io/?https://raw.githubusercontent.com/user/my-project/master/doc/file.html

to produce the HTML document in a browser.

However, there is another technique available where all HTML files in a special branch gh-pages of the repository are automatically rendered correctly as HTML documents in a browser. This is the recommended technique for publishing a collection of HTML files related to the project in a simple and convenient way. The recipe is described in detail below.

1. Go to the project page on github.com and click Settings.
2. Click on Automatic Page Generator under the GitHub Pages.
3. Proceed clicking Continue to Layouts, choose a design of the index.html page that GitHub will create for you, and click Publish.
4. Go to the root directory of the project, My-Project and run git fetch origin.
5. Run git checkout gh-pages.

The index.html page is invoked by the web address

http://user.github.io/My-Project/index.html

where user is the GitHub username and My-Project is the project name.

The web pages and project files are now in two different branches. To see the branches, type git branch, and the one you are in will be marked with * in the output. Switching to the master branch is done by git checkout master. Similarly, git checkout gh-pages switches to the gh-pages branch.

My personal preference is to have the master and gh-pages synchronized, at least in projects where I want to link to various source code files or other files from the web documentation. Sometimes I also update files in the gh-pages branch without remembering to switch to the master branch. To this end, one needs to merge the branches, i.e., automatically edit files in the current branch such that they are up-to-date and identical to files in another branch.

To merge the current branch with some branch named otherbranch, run

Terminal> git merge otherbranch

Git applies smart algorithms that very often manage to merge the files without human interaction. However, occasionally these algorithms are not able to resolve conflicts between two files. A message about the failure of the merge is seen in the terminal window, and the corresponding files have markers in them showing which sections that needs manual editing to resolve the conflicts. Run git diff to show the problems (you can tailor this command to your needs as explained in the section Replacing pull by fetch and merge). After a manual edit, do git commit -a. More details on merging appears in the section Merging files with Git.

If you want to keep the master branch and the gh-pages branch synchronized, start with merging the gh-pages branch with the master branch and push the complete file collection to the gh-pages branch. Then switch to the master branch and merge with gh-pages so you get the autogenerated index.html file and associated files and directories for web pages in the root directory of the master branch as well:

Terminal> git merge master
Terminal> touch .nojekyll
Terminal> git push origin gh-pages
Terminal> git checkout master
Terminal> git merge gh-pages

You must add an empty file .nojekyll in the top directory of the project pages if you want to use Sphinx-generated HTML pages (or other pages using javascripts, style sheets, and images in subdirectories whose names start with an underscore).

You can now add the documentation to the project files and maintain them in the master branch. Before publishing documents online, make sure to update the gh-pages branch by

Terminal> git commit -am 'Ensure commit of master branch'
Terminal> git push origin master
Terminal> git checkout gh-pages
Terminal> git pull origin gh-pages
Terminal> git merge master
Terminal> git push origin gh-pages
Terminal> git checkout master

Personally, I like to move the generated index.html file and all associated scripts, stylesheets, and images from the root directory to some more isolated place, say doc/web:

Terminal> git mv index.html params.json stylesheets/ images/ \
                 javascripts/ doc/web/

The URL of the index.html file is

http://user.github.io/My-Project/doc/web/index.html

Linking to source code files or other files in the project is easy: just find the file in GitHub's web interface, choose which version of the file you want to link to (nicely HTML formatted version or the raw file), right-click on the link, choose Copy Link, and paste the link into the document you want. You can test that the link works by the Unix command curl -O <link>. Note that the link to a file is different from the source file's intuitive path in the repository. Typically, a source file dir/f.py in project prj is reached through

https://github.com/user/prj/blob/master/dir/f.py?raw=true

Sometimes you want to link to another HTML file, PDF file, movie file, or a file that is to be interpreted as a web resource by the browser. Do not use the path to the file in the repo as explained above as it will just bring the reader to the repo page. Instead, make sure the file is in the gh-pages branch and use a local link, like ../doc.pdf, or the complete gh-pages URL to the file, say

http://user.github.com/My-Project/doc/misc/doc.pdf

User web pages

GitHub also allows you to create user pages and organization pages not tied to any specific project. Your personal site has address http://user.github.com. Go to your home page on github.com and click New repository, and give it the project name user.github.com. Then follow the instructions that come up:

Terminal> mkdir user.github.com
Terminal> cd user.github.com
Terminal> git init
Terminal> # make an index.html file with some test text
Terminal> git add index.html
Terminal> git commit -m 'First commit'
Terminal> git remote add origin \
          git@github.com:user/user.github.com.git
Terminal> git push -u origin master

Go to http://user.github.com and see how the index.html is rendered. You can now add various contents as in any ordinary Git repository. If you want to use Sphinx generated HTML pages, recall to add an empty file .nojekyll.

Using Git

Most Mac and Linux users prefer to work with Git via commands in a terminal window. Windows users may prefer a graphical user interface (GUI), and there are many options in this respect. There are also GUIs for Mac users. Here we concentrate on the efficient command-line interface to Git.

Basic Git commands

Cloning

You get started with your project on a new machine, or another user can get started with the project, by running

Terminal> git clone git@github.com:user/My-Project.git
Terminal> cd My-Project
ls

Recall to replace user by your real username and My-Project by the actual project name.

The pull-change-push cycle

The typical work flow with the "My Project" project starts with updating the local repository by going to the My-Project directory and writing

Terminal> git pull origin master

You may want to do git fetch and git merge instead of git pull as explained in the section Replacing pull by fetch and merge, especially if you work with branches.

You can now edit files, make new files, and make new directories. New files and directories must be added with git add. There are also Git commands for deleting, renaming, and moving files. Typical examples on these Git commands are

Terminal> git add file2.* dir1 dir2  # add files and directories
Terminal> git rm file3
Terminal> git rm -r dir2
Terminal> git mv oldname newname
Terminal> git mv oldname ../newdir

When your chunk of work is ready, it is time to commit your changes (note the -am option):

Terminal> git commit -am 'Description of changes.'

If typos or errors enter the message, the git commit --amend command can be used to reformulate the message. Running git diff prior to git commit makes it easier to formulate descriptive commit messages since this command gives a listing of all the changes you have made to the files since the last commit or pull command.

You may perform many commits (to keep track of small changes), before you push your changes to the global repository:

Terminal> git push origin master

Terminal> git add .

Viewing the history of files

A nice graphical tool allows you to view all changes, or just the latest ones:

Terminal> gitk --all
Terminal> gitk --since="2 weeks ago"

You can also view changes to all files, some selected ones, or a subdirectory:

Terminal> git log -p                 # all changes to all files
Terminal> git log -p filename        # changes to a specific file
Terminal> git log --stat --summary   # compact summary
Terminal> git log --stat --summary subdir

Adding --follow will print the history of file versions before the file got its present name.

To show the author who is responsible for the last modification of each line in the file, use git blame:

Terminal> git blame filename
Terminal> git blame --since="1 week" filename

A useful command to see the history of who did what, where individual edits of words are highlighted (--word-diff), is

git log -p --stat --word-diff filename

Removed words appear in brackets and added words in curly braces.

Looking for when a particular piece of text entered or left the file, say the text def myfunc, one can run

Terminal> git log -p --word-diff --stat -S'def myfunc' filename

This is useful to track down particular changes in the files to see when they occurred and who introduced them. One can also search for regular expressions instead of exact text: just replace -S by -G.

Retrieving old files

Occasionally you need to go back to an earlier version of a file, e.g., a file called f.py. Start with viewing the history:

Terminal> git log f.py

Find a commit candidate from the list that you will compare the present version to, copy the commit hash (string like c7673487...), and run

Terminal> git diff c7673487763ec2bb374758fb8e7efefa12f16dea f.py

where the long string is the relevant commit hash. You can now view the differences between the most recent version and the one in the commit you picked (see the section Replacing pull by fetch and merge for how to configure the tools used by the git diff command). If you want to restore the old file, write

Terminal> git checkout c7673487763ec2bb374758fb8e7efefa12f16dea f.py

To go back to another version (the most recent one, for instance), find the commit hash with git log f.py, and do get checkout <commit hash> f.py.

If f.py changed name from e.py at some point and you want e.py back, run git log --follow f.py to find the commit when e.py existed, and do a git checkout <commit hash> e.py.

In case f.py no longer exists, run git log -- f.py to see its history before deletion. The last commit shown does not contain the file, so you need to check out the next last to retrieve the latest version of a deleted file.

Often you just need to view the old file, not replace the current one by the old one, and then git show is handy. Unfortunately, it requires the full path from the root git directory:

Terminal> git show \
          c7673487763ec2bb374758fb8e7efefa12f16dea:dir1/dir2/f.py

Reset the entire repo to an old version

Run git log on some file and find the commit hash of the date or message when want to go back to. Run git checkout <commit hash> to change all files to this state. The problem of going back to the most recent state is that git log has no newer commits than the one you checked out. The trick is to say git checkout master to set all files to most recent version again.

If you want to reset all files to an old version and commit this state as the valid present state, you do

Terminal> git checkout c7673487763ec2bb374758fb8e7efefa12f16dea .
Terminal> git commit -am 'Resetting to ...'

Note the period at the end of the first command (without it, you only get the possibility to look at old files, but the next commit is not affected).

Going back to a previous commit

Sometimes accidents with many files happen and you want to go back to the last commit. Find the hash of the last commit and do

Terminal> git reset --hard c867c487763ec2

This command destroys everything you have done since the last commit. To push it as the new state of the repo, do

Terminal> git push origin HEAD --force

Merging files with Git

The git pull command fetches new files from the repository and tries to perform an automatic merge if there are conflicts between the local files and the files in the repository. Alternatively, you may run git fetch and git merge to do the same thing as described in the section Replacing pull by fetch and merge. We shall now address what do to if the merge goes wrong, which occasionally happens.

Git will write a message in the terminal window if the merge is unsuccessful for one or more files. These files will have to be edited manually. Merge markers of the type >>>>>, ======, and <<<<< have been inserted by Git to mark sections of a file where the version in the repository differ from the local version. You must decide which lines that are to appear in the final, merged version. When done, perform git commit and the conflicts are resolved.

Graphical merge tools may ease the process of merging text files. You can run git mergetool --tool=meld to open the merge tool meld for every file that needs to be merged (or specify the name of a particular file). Other popular merge tools supported by Git are araxis, bc3, diffuse, ecmerge, emerge, gvimdiff, kdiff3, opendiff, p4merge, tkdiff, tortoisemerge, vimdiff, and xxdiff.

Below is a Unix shell script illustrating how to make a global repository in Git, and how two users clone this repository and perform edits in parallel. There is one file myfile in the repository.

#!/bin/sh
# Demo script for exemplifying git and merge

rm -rf tmp1 tmp2 tmp_repo   # Clean up previous runs

mkdir tmp_repo   # Global repository for testing
cd tmp_repo
git --bare init --shared
cd ..

# Make a repo that can be pushed to tmp_repo
mkdir _tmp
cd _tmp
cat > myfile <<EOF
This is a little
test file for
exemplifying merge
of files in different
git directories.
EOF
git init
git add .   # Add all files not mentioned in ~/.gitignore
git commit -am 'first commit'
git push ../tmp_repo master
cd ..
rm -rf _tmp

# Make a new hg repositories tmp1 and tmp2 (two users)
git clone tmp_repo tmp1
git clone tmp_repo tmp2
# Change myfile in the directory tmp1
cd tmp1
# Edit myfile: insert a new second line
perl -pi -e 's/a little\n/a little\ntmp1-add1\n/g' myfile
# Register change in local repository
git commit -am 'Inserted a new second line in myfile.'
# Look at changes in this clone
git log -p
# or a more compact summary
git log --stat --summary
# or graphically
#gitk
# Register change in global repository tmp_repo
git push origin master
cd ..

# Change myfile in the directory tmp2 "in parallel"
cd tmp2
# Edit myfile: add a line at the end
cat >> myfile <<EOF
tmp2-add1
EOF
# Register change locally
git commit -am 'Added a new line at the end'
# Register change globally
git push origin master
# Error message: global repository has changed,
# we need to pull those changes to local repository first
# and see if all files are compatible before we can update
# our own changes to the global repository.
# git writes
#To /home/hpl/vc/scripting/manu/py/bitgit/src-bitgit/tmp_repo
# ! [rejected]        master -> master (non-fast-forward)
#error: failed to push some refs to ...

git pull origin master
# git writes:
#Auto-merging myfile
#Merge made by recursive.
# myfile |    1 +
# 1 files changed, 1 insertions(+), 0 deletions(-)
cat myfile  # successful merge!
git commit -am merge
git push origin master
cd ..

# Perform new changes in parallel in tmp1 and tmp2,
# this time causing hg merge to fail

# Change myfile in the directory tmp1
cd tmp1
# Do it all right by pulling and updating first
git pull origin master
# Edit myfile: insert "just" in first line.
perl -pi -e 's/a little/tmp1-add2 a little/g' myfile
# Register change in local repository
git commit -am 'Inserted "just" in first line.'
# Register change in global repository tmp_repo
git push origin master
cd ..

# Change myfile in the directory tmp2 "in parallel"
cd tmp2
# Edit myfile: replace little by modest
perl -pi -e 's/a little/a tmp2-replace1\ntmp2-add2\n/g' myfile
# Register change locally
git commit -am 'Replaced "little" by "modest"'
# Register change globally
git push origin master
# Not possible: need to pull changes in the global repository
git pull origin master
# git writes
#CONFLICT (content): Merge conflict in myfile
#Automatic merge failed; fix conflicts and then commit the result.
# we have to do a manual merge
cat myfile
echo 'Now you must edit myfile manually'

You may run this file git_merge.sh named by sh -x git_merge.sh. At the end, the versions of myfile in the repository and the tmp2 directory are in conflict. Git tried to merge the two versions, but failed. Merge markers are left in tmp2/myfile:

<<<<<<< HEAD
This is a tmp2-replace1
tmp2-add2

=======
This is tmp1-add2 a little
>>>>>>> ad9b9f631c4cc586ea951390d9415ac83bcc9c01
tmp1-add1
test file for
exemplifying merge
of files in different
git directories.
tmp2-add1

Launch a text editor and edit the file, or use git mergetool, so that the file becomes correct. Then run git commit -am merge to finalize the merge.

Git working style with branching and stashing

Branching and stashing are nice features of Git that allow you to try out new things without affecting the stable version of your files. Usually, you extend and modify files quite often and perform a git commit every time you want to record the changes in your local repository. Imagine that you want to correct a set of errors in some files and push these corrections immediately. The problem is that such a push will also include the latest, yet unfinished files that you have committed.

Branching

A better organization of your work would be to keep the latest, ongoing developments separate from the more official and stable version of the files. This is easily achieved by creating a separate branch where new developments takes place:

Terminal> git branch newstuff      # create new branch
Terminal> git checkout newstuff
Terminal> # extend and modify files...
Terminal> git commit -am 'Modified ... Added a file on ...'
Terminal> git checkout master      # swith back to master
Terminal> # correct errors
Terminal> git push origin master
Terminal> git checkout newstuff    # switch to other branch
Terminal> git merge master         # keep branch up-to-date w/master
Terminal> # continue development work...
Terminal> git commit -am 'More modifications of ...'

At some point, your developments in newstuff are mature enough to be incorporated in the master branch:

Terminal> git checkout newstuff
Terminal> git merge master        # synchronize newstuff w/master
Terminal> git checkout master
Terminal> git merge newstuff      # synchronize master w/newstuff

You no longer need the newstuff branch and can delete it:

Terminal> git branch -d newstuff

This command deletes the branch locally. To also delete the branch in the remote repo, run

Terminal> git push origin --delete newstuff

You can learn more in an excellent introduction and demonstration of Git branching.

Stashing

It is not possible to switch branches unless you have committed the files in the current branch. If your work on some files is in a mess and you want to change to another branch or fix other files in the current branch, a "global" commit affecting all files might be immature. Then the git stash command is handy. It records the state of your files and sets you back to the state of the last commit in the current branch. With git stash apply you will update the files in this branch to the state when you did the last git stash.

Let us explain a typical case. Suppose you have performed some extensive edits in some files and then you are suddenly interrupted. You need to fix some typos in some other files, commit the changes, and push. The problem is that many files are in an unfinished state - in hindsight you realize that those files should have been modified in a separate branch. It is not too late to create that branch! First run git stash to get the files back to the state they were at the last commit. Then run git stash branch newstuff to create a new branch newstuff containing the state of the files when you did the (last) git stash command. Stashing used this way is a convenient technique to move some immature edits after the last commit out in a new branch for further experimental work.

Replacing pull by fetch and merge

The git pull command actually performs two steps that are sometimes advantageous to run separately. First, a get fetch is run to fetch new files from the repository, and thereafter a git merge command is run to merge the new files with your local version of the files. While git pull tries to do a lot and be smart in the merge, very often with success, the merge step may occasionally lead to trouble. That is why it is recommended to run a git merge separately, especially if you work with branches.

To fetch files from your repository at GitHub, which usually has the nickname origin, you write

Terminal> git fetch origin

You now have the possibility to check out in detail what the differences are between the new files and local ones:

Terminal> git diff origin/master

This command produces comparisons of the files in the current local branch and the master branch at origin (the GitHub repo). In this way you can exactly see the differences between branches. It also gives you an overview of what others have done with the files. When you are ready to merge in the new files from the master branch of origin with the files in the current local branch, you say

Terminal> git merge origin/master

Especially when you work with multiple branches, as outlined in the section Git working style with branching and stashing, it is wise to first do a get fetch origin and then update each branch separately. The git fetch origin command will list the branches, e.g.,

* master
  gh-pages
  next

After updating master as described, you can continue with another branch:

Terminal> git checkout next
Terminal> git diff origin/next
Terminal> git merge origin/next
Terminal> git checkout master

Configuring the git diff command

The git diff command launches by default the Unix diff tool in the terminal window. Many users prefer to use other diff tools, and the desired one can be specified in your ~/.gitconfig file. However, a much recommended approach is to wrap a shell script around the call to the diff program, because git diff actually calls the diff program with a series of command-line arguments that will confuse diff programs that take the names of the two files to be compared as arguments. In ~/.gitconfig you specify a script to do the diff:

[diff]
external = ~/bin/git-diff-wrapper.sh

It remains to write the git-diff-wrapper.sh script. The 2nd and 5th command-line arguments passed to this script are the name of the files to be compared in the diff. A typical script may therefore look like

#!/bin/sh

diff "$2" "$5" | less

Here we use the standard (and quite primitive) Unix diff program, but we can replace diff by, e.g., diffuse, kdiff3, xxdiff, meld, pdiff, or others. With a Python script you can easily check for the extensions of the files and use different diff tools for different types of files, e.g., latexdiff for LaTeX files and pdiff for pure text files.

Terminal> git fetch --all
Terminal> git reset --hard origin/master

Terminal> git checkout A
Terminal> git checkout --patch B f.txt

Terminal> git checkout A
Terminal> git checkout B doc/f.txt src/files/g.py

Team work with forking and pull requests

In small collaboration teams it is natural that everyone has push access to the repo. On GitHub this is known as the Shared Repository Model. As teams grow larger, there will usually be a few people in charge who should approve changes to the files. Ordinary team members will in this case not clone a repo and push changes, but instead fork the repo and send pull requests, which constitutes the Fork and Pull Model.

Say you want to fork the repo https://github.com/somebody/proj1.git. The first step is to press the Fork button on the project page for the somebody/proj1 project on GitHub. This action creates a new repo proj1, known as the forked repo, on your GitHub account. Clone the fork as you clone any repo:

Terminal> git clone https://github.com/user/proj1.git

When you do git push origin master, you update your fork. However, the original repo is usually under development too, and you need to pull from that one to stay up to date. A git pull origin master pulls from origin which is your fork. To pull from the original repo, you create a name upstream, either by

Terminal> git remote add upstream \
              https://github.com/somebody/proj1.git

if you cloned with such an https address, or by

Terminal> git remote add upstream \
              git@github.com:somebody/proj1.git

if you cloned with a git@github.com (SSH) address. Doing a git pull upstream master would seem to be the command for pulling the most recent files in the original repo. However, it is not recommended to update the forked repo's files this way because heavy development of the sombody/proj1 project may lead to serious merge problems. It is much better to replace the pull by a separate fetch and merge. The typical workflow is

Terminal> git fetch upstream           # get new version of files
Terminal> git merge upstream/master    # merge with yours
Terminal> # Your files are up to date - ready for editing
Terminal> git commit -am 'Description...'
Terminal> git push origin master       # store changes in your fork

At some point you would like to push your changes back to the original repo somebody/proj1. This is done by a pull request. Make sure you have selected the right branch on the project page of your forked project. Press the Pull Request button and fill out the form that pops up. Trusted people in the somebody/proj1 project will now review your changes and if they are approved, your files are merged into the original repo. If not, there are tools for keeping a dialog about how to proceed.

Also in small teams where everyone has push access, the fork and pull request model is beneficial for reviewing files before the repo is actually updated with new contributions.

Cloning a repo with multiple branches

An annoying feature of Git for beginners is the fact that if you clone a repo, you only get the master branch. There are seemingly no other branches:

Terminal> git branch
* master

To see which branches that exist in the repo, type

Terminal> git branch -a
* master
  remotes/origin/HEAD -> origin/master
  remotes/origin/gh-pages
  remotes/origin/master
  remotes/origin/next

If there is only one remote repo that you pull/push from/to, you can simply switch branch with git checkout the usual way:

Terminal> git checkout gh-pages
Terminal> git branch
* gh-pages
  master
Terminal> git checkout next
Terminal> git branch
  gh-pages
  master
* next

You might need to do git fetch origin to see new branches made on other machines.

When you have more than one remote, which is usually the case if you have forked a repo, see the section Team work with forking and pull requests, you must use do a checkout with specifying the remote branch you want:

Terminal> git checkout -b gh-pages --track remote/gh-pages
Terminal> git checkout -b next --track upstream/next

Files can be edited, added, or removed as soon as you have done the local checkout.

It is possible to write a little script that takes the output of git branch -a after a git clone command and automatically check out all branches via git checkout.

Git workflows

Although the purpose of these notes is just to get the reader started with Git, it must be mentioned that there are advanced features of Git that have led to very powerful workflows with files and people, especially for software development. There is an official Git workflow model that outlines the basic principles, but it can be quite advanced for those with modest Git knowledge. A more detailed explanation of a recommended workflow for beginners is given in the developer instructions for the software package PETSc. This is highly suggested reading. The associated "quick summary" of Git commands for their workflow is also useful.

Git tips

How can I see which files are tracked by Git?

git ls-files is the command:

Terminal> git ls-files            # list all tracked files
Terminal> git ls-files -o         # list non-tracked files
Terminal> git ls-files myfile     # prints myfile if it's tracked
Terminal> git ls-files myfile --error-unmatch

The latter command prints an error message if myfile is not tracked. See man git-ls-files for the many options this utility has.

How can I reduce the size of a repo?

The command git gc can compress a git repository and should be run regularly on large repositories. Greater effect is achieved by git gc --aggressive --prune=all. You can measure the size of a repo before and after compression by git gc using du -s repodir, where repodir is the name of the root directory of the repository.

Occasionally big or sensitive files are removed from the repo and you want to permanently remove these files from the revision history. This is achieved using git filter-branch. To remove a file or directory with path doc/src/mydoc relative to the root directory of the repo, go to this root directory, make sure all branches are checked out on your computer, and run

Terminal> git filter-branch --index-filter \
         'git rm -r --cached --ignore-unmatch doc/src/mydoc' \
         --prune-empty -- --all
Terminal> rm -rf .git/refs/original/
Terminal> git reflog expire --expire=now --all
Terminal> git gc --aggressive --prune=now
Terminal> git push origin master --force  # do this for each branch
Terminal> git checkout somebranch
Terminal> git push origin somebranch --force

You must repeat the push command for each branch as indicated. If other users have created their own branches in this repo, they need to rebase, not merge, when updating the branches!

How can I restore missing files?

Sometimes you accidentally remove files from a repo, either by git rm or a plain rm. You can get the files back as long as they are in the remote repo. In case of a plain rm command, run

Terminal> git checkout `git ls-files`

to restore all missing files in the current directory.

In case of a git rm command, use git log --diff-filter=D --summary to find the commit hash corresponding to the last commit the files were in the repo. Restoring a file is then done by

Terminal> git checkout <commit hash> filename

More documentation on Git

• Git - the simple guide
• Web course on Git
• Everyday GIT With 20 Commands Or So
• Git cheat sheet
• Atlassian Git Tutorials
• Atlassian Git Workflows
• GitHub Guides:
• Getting your project on GitHub
• Repositories, branches, commits, issues, and pull requests
• GitHub (web) pages
• Understanding the GitHub Flow
• Git branching
• Git top 10 tutorials
• Lars Vogel's Git Tutorial
• How to use Git with Dropbox
• Git Community Book (explains Git very well)
• Git for Designers (aimed a people with no previous knowledge of version control systems)
• Git Magic: Basic Tricks
• The official Git Tutorial
• Git Tutorial Video on YouTube
• Git Questions
• Git Reference (can also be used as a tutorial on Git)
• Git User Manual
• Git home page
• Quick intro to Git and GitHub (somewhat like the present guide)
• Git/GitHub GUIs on Windows and Mac
• 10 Things I hate about Git

Appendix: Working with multiple GitHub accounts

Working against different GitHub accounts is easy if each project you work with on each account adds you as a collaborator. The term "you" here means your primary username on GitHub. My strong recommendation is to always check out a project using your primary GitHub username.

Occasionally you want to create a new GitHub account, say for a project XYZ. For such a non-personal account, do not provide an SSH key of any particular user. The reason is that this user will then get two GitHub identities, and switching between these identities will require some special tweakings. Just forget about the SSH key for a project account and add collaborators to repos using each collaborators personal GitHub username.

If you really need to operate the XYZ account as a personal account, you must provide an SSH key that is different from any other key at any other GitHub account (you will get an error message if you try to register an already registered SSH key, but it is possible to get around the error message by providing an id_rsa.pub key on one account and an id_dsa.pub on another - that will cause trouble). Jeffrey Way has written a recipe for how to operate multiple GitHub accounts using multiple identities.

To debug which identity that is used when you pull and push to GitHub accounts, you can first run

Terminal> ssh -Tv git@github.com

to see your current identity and which SSH key that was used to identify you. Typing

Terminal> ssh-add -l

lists all your SSH keys. The shown strings can be compared with the string in the SSH key field of any GitHub account.