Describe staging area

[git-wiki.git] / pep-git.txt

PEP: XXX
Title: Collecting information about git
Version: $Revision$
Last-Modified: $Date$
Author: Oleg Broytman <phd@phdru.name>
Status: Draft
Type: Informational
Content-Type: text/x-rst
Created: 01-Jun-2015
Post-History: 

Abstract
========

This Informational PEP collects information about git. There is, of
course, a lot of documentation for git, so the PEP concentrates on
more complex issues, scenarios and topics.

The plan is to extend the PEP in the future collecting information
about equivalence of Mercurial and git scenarios to help migrating
Python development from Mercurial to git.

The author of the PEP doesn't currently plan to write a Process PEP on
migration from Mercurial to git.


Documentation
=============

Git is accompanied with a lot of documentation, both online and
offline.

Documentation for starters
--------------------------

Git Tutorial: `part 1
<https://www.kernel.org/pub/software/scm/git/docs/gittutorial.html>`_,
`part 2
<https://www.kernel.org/pub/software/scm/git/docs/gittutorial-2.html>`_.

`Git User's manual
<https://www.kernel.org/pub/software/scm/git/docs/user-manual.html>`_.
`Everyday GIT With 20 Commands Or So
<https://www.kernel.org/pub/software/scm/git/docs/everyday.html>`_.
`Git workflows
<https://www.kernel.org/pub/software/scm/git/docs/gitworkflows.html>`_.

Advanced documentation
----------------------

`Git Magic
<http://www-cs-students.stanford.edu/~blynn/gitmagic/index.html>`_,
also with a number of translations.

`Pro Git <https://git-scm.com/book>`_. The Book about git. Buy it at
Amazon or download in PDF, mobi, or ePub form. Has translations to
many different languages. Download Russian translation from `GArik
<https://github.com/GArik/progit/wiki>`_.

`Git Wiki <https://git.wiki.kernel.org/index.php/Main_Page>`_.

Offline documentation
---------------------

Git has builtin help: run ``git help $TOPIC``. For example, run
``git help git`` or ``git help help``.


Quick start
===========

Download and installation
-------------------------

Unix users: download and install using your package manager.

Microsoft Windows: download `git-for-windows
<https://github.com/git-for-windows/git/releases>`_ or `msysGit
<https://github.com/msysgit/msysgit/releases>`_.

MacOS X: use git installed with `XCode
<https://developer.apple.com/xcode/downloads/>`_ or download from
`MacPorts <https://www.macports.org/ports.php?by=name&substr=git>`_ or
`git-osx-installer
<http://sourceforge.net/projects/git-osx-installer/files/>`_ or
install git with `Homebrew <http://brew.sh/>`_: ``brew install git``.

`git-cola <https://git-cola.github.io/index.html>`_ is a Git GUI
written in Python and GPL licensed. Linux, Windows, MacOS X.

`TortoiseGit <https://tortoisegit.org/>`_ is a Windows Shell Interface
to Git based on TortoiseSVN; open source.

Initial configuration
---------------------

This simple code is often appears in documentation, but it is
important so let repeat it here. Git stores author and committer
names/emails in every commit, so configure your real name and
preferred email::

    $ git config --global user.name "User Name"
    $ git config --global user.email user.name@example.org


Examples in this PEP
====================

Examples of git commands in this PEP use the following approach. It is
supposed that you, the user, works with a local repository named
``python`` that has an upstream remote repo named ``origin``. Your
local repo has two branches ``v1`` and ``master``. For most examples
the currently checked out branch is ``master``. That is, it's assumed
you have done something like that::

    $ git clone http://git.python.org/python.git
    $ cd python
    $ git branch v1 origin/v1

The first command clones remote repository into local directory
`python``, creates a new local branch master, sets
remotes/origin/master as its upstream remote-tracking branch and
checks it out into the working directory.

The last command creates a new local branch v1 and sets
remotes/origin/v1 as its upstream remote-tracking branch.

The same result can be achieved with commands::

    $ git clone -b v1 http://git.python.org/python.git
    $ cd python
    $ git checkout --track origin/master

The last command creates a new local branch master, sets
remotes/origin/master as its upstream remote-tracking branch and
checks it out into the working directory.


Branches and branches
=====================

Git terminology can be a bit misleading. Take, for example, the term
"branch". In git it has two meanings. A branch is a directed line of
commits (possibly with merges). And a branch is a label or a pointer
assigned to a line of commits. It is important to distinguish when you
talk about commits and when about their labels. Lines of commits are
by itself unnamed and are usually only lengthening and merging.
Labels, on the other hand, can be created, moved, renamed and deleted
freely.


Remote repositories and remote branches
=======================================

Remote-tracking branches are branches (pointers to commits) in your
local repository. They are there for you to remember what branches and
commits have been pulled from and pushed to what remote repos (you can
pull from and push to many remotes). Remote-tracking branches live
under ``remotes/$REMOTE`` namespaces, e.g. ``remotes/origin/master``.

To see the status of remote-tracking branches run::

    $ git branch -rv

To see local and remote-tracking branches (and tags) pointing to
commits::

    $ git log --decorate

You never do your own development on remote-tracking branches. You
create a local branch that has a remote branch as upstream and do
development on that local branch. On push git pushes commits to the
remote repo and updates remote-tracking branches, on pull git fetches
commits from the remote repo, updates remote-tracking branches and
fast-forwards, merges or rebases local branches.

When you do an initial clone like this::

    $ git clone -b v1 http://git.python.org/python.git

git clones remote repository ``http://git.python.org/python.git`` to
directory ``python``, creates remote-tracking branches, creates a
local branch ``v1``, configure it to track upstream remotes/origin/v1
branch and checks out ``v1`` into the working directory.

Updating local and remote-tracking branches
-------------------------------------------

There is a major difference between

::

    $ git fetch $REMOTE $BRANCH

and

::

    $ git fetch $REMOTE $BRANCH:$BRANCH

The first command fetches commits from the named $BRANCH in the
$REMOTE repository that are not in your repository, updates
remote-tracking branch and leaves the id (the hash) of the head commit
in file .git/FETCH_HEAD.

The second command fetches commits from the named $BRANCH in the
$REMOTE repository that are not in your repository and updates both
the local branch $BRANCH and its upstream remote-tracking branch. But
it refuses to update branches in case of non-fast-forward. And it
refuses to update the current branch.

The first command is used internally by ``git pull``.

::

    $ git pull $REMOTE $BRANCH

is equivalent to

::

    $ git fetch $REMOTE $BRANCH
    $ git merge FETCH_HEAD

Certainly, $BRANCH in that case should be your current branch. If you
want to merge a different branch into your current branch first update
that non-current branch and then merge::

    $ git fetch origin v1:v1  # Update v1
    $ git pull --rebase origin master  # Update the current branch master
                                       # using rebase instead of merge
    $ git merge v1

If you have not yet pushed commits on ``v1``, though, the scenario has
to become a bit more complex. Git refuses to update
non-fast-forwardable branch, and you don't want to do force-pull
because that would remove your non-pushed commits and you would need
to recover. So you want to rebase ``v1`` but you cannot rebase
non-current branch. Hence, checkout ``v1`` and rebase it before
merging::

    $ git checkout v1
    $ git pull --rebase origin v1
    $ git checkout master
    $ git pull --rebase origin master
    $ git merge v1

It is possible to configure git to make it fetch/pull a few branches
or all branches at once, so you can simply run

::

    $ git pull origin

or even

::

    $ git pull

Default remote repository for fetching/pulling is origin. Default set
of references to fetch is calculated using matching algorithm: git
fetches all branches having the same name on both ends.

Push
''''

Pushing is a bit simpler. There is only one command ``push``. When you
run

::

    $ git push origin v1 master

git pushes local v1 to remote v1 and local master to remote master.
The same as::

    $ git push origin v1:v1 master:master

Git pushes commits to the remote repo and updates remote-tracking
branches. Git refuses to push commits that aren't fast-forwardable.
You can force-push anyway, but please remember - you can force-push to
your own repositories but don't force-push to public or shared repos.
If you find git refuses to push commits that aren't fast-forwardable,
better fetch and merge commits from the remote repo (or rebase your
commits on top of the fetched commits), then push. Only force-push if
you know what you do and why you do it. See the section `Commit
editing and caveats`_ below.

It is possible to configure git to make it push a few branches or all
branches at once, so you can simply run

::

    $ git push origin

or even

::

    $ git push

Default remote repository for pushing is origin. Default set
of references to push in git before 2.0 is calculated using matching
algorithm: git pushes all branches having the same name on both ends.
Default set of references to push in git 2.0+ is calculated using
simple algorithm: git pushes the current branch back to its
@{upstream}.

To configure git before 2.0 to the new behaviour run::

$ git config push.default simple

To configure git 2.0+ to the old behaviour run::

$ git config push.default matching

Git refuses to push a branch if it's the current branch in the remote
non-bare repository: git refuses to update remote working directory.
You really should push only to bare repositories. For non-bare
repositories git prefers pull-based workflow.

When you want to deploy code on a remote host and can only use push
(because your workstation is behind a firewall and you cannot pull
from it) you do that in two steps using two repositories: you push
from the workstation to a bare repo on the remote host, ssh to the
remote host and pull from the bare repo to a non-bare deployment repo.

That changed in git 2.3, but see `the blog post
<https://github.com/blog/1957-git-2-3-has-been-released#push-to-deploy>`_
for caveats; in 2.4 the push-to-deploy feature was `further improved
<https://github.com/blog/1994-git-2-4-atomic-pushes-push-to-deploy-and-more#push-to-deploy-improvements>`_.

Tags
''''

Git automatically fetches tags that point to commits being fetched
during fetch/pull. To fetch all tags (and commits they point to) run
``git fetch --tags origin``. To fetch some specific tags fetch them
explicitly::

    $ git fetch origin tag $TAG1 tag $TAG2...

For example::

    $ git fetch origin tag 1.4.2
    $ git fetch origin v1:v1 tag 2.1.7

Git doesn't automatically pushes tags. That allows you to have private
tags. To push tags list them explicitly::

    $ git push origin tag 1.4.2
    $ git push origin v1 master tag 2.1.7

Don't move tags with ``git tag -f`` or remove tags with ``git tag -d``
after they have been published.

Private information
'''''''''''''''''''

When cloning/fetching/pulling/pushing git copies only database objects
(commits, trees, files and tags) and symbolic references (branches and
lightweight tags). Everything else is private to the repository and
never cloned, updated or pushed. It's your config, your hooks, your
private exclude file.

If you want to distribute hooks, copy them to the working tree, add,
commit, push and instruct the team to update ind install the hook
manually.


Commit editing and caveats
==========================

A warning not to edit published (pushed) commits also appears in
documentation but it's repeated here anyway as it's very important.

It is possible to recover from forced push but it's PITA for the
entire team. Please avoid it.

To see what commits have not been published yet compare the head of the
branch with its upstream remote-tracking branch::

    $ git log origin/master..
    $ git log origin/v1..v1

For every branch that has an upstream remote-tracking branch git
maintains an alias @{upstream} (short version @{u}), so the commands
above can be given as::

    $ git log @{u}..
    $ git log v1@{u}..v1

To see the status of all branches::

    $ git branch -avv

To compare the status of local branches with a remote repo::

    $ git remote show origin

Read `how to recover from upstream rebase
<https://git-scm.com/docs/git-rebase#_recovering_from_upstream_rebase>`_.
It is in ``git help rebase``.

On the other hand don't be too afraid about commit editing. You can
safely edit, remove, reorder, combine and split commits that haven't
been pushed yet. You can even push commits to your own (backup) repo,
edit them later and force-push edited commits to replace what have
already been pushed. Not a problem until commits are in a public
or shared repository.


Undo
====

Whatever you do, don't panic. Almost anything in git can be undone.

git checkout: restore file's content
------------------------------------

``git checkout``, for example, can be used to restore the content of
file(s) to that one of a commit. Like this::

    git checkout HEAD~ README

The commands restores the contents of README file to the last but one
commit in the current branch. By default the commit ID is simply HEAD;
i.e. ``git checkout README`` restores README to the latest commit.

(Do not use ``git checkout`` to view a content of a file in a commit,
use ``git cat-file -p``; e.g. ``git cat-file -p HEAD~:path/to/README``).

git reset: remove (non-pushed) commits
--------------------------------------

``git reset`` moves the head of the current branch. The head can be
moved to point to any commit but it's often used to remove a commit or
a few (preferably, non-pushed ones) from the top of the branch - that
is, to move the branch backward in order to undo a few (non-pushed)
commits.

``git reset`` has three modes of operation - soft, hard and mixed.
Default is mixed. ProGit `explains
<https://git-scm.com/book/en/Git-Tools-Reset-Demystified>`_ the
difference very clearly. Bare repositories don't have indices or
working trees so in a bare repo only soft reset is possible.

Unstaging
'''''''''

Mixed mode reset with a path or paths can be used to unstage changes -
that is, to remove from index changes added with ``git add`` for
committing. See `The Book
<https://git-scm.com/book/en/Git-Basics-Undoing-Things>`_ for details
about unstaging and other undo tricks.

git reflog: reference log
-------------------------

Removing commits with ``git reset`` or moving the head of a branch
sounds dangerous and it is. But there is a way to undo: another
reset back to the original commit. Git doesn't remove commits
immediately; unreferenced commits (in git terminology they are called
"dangling commits") stay in the database for some time (default is two
weeks) so you can reset back to it or create a new branch pointing to
the original commit.

For every move of a branch's head - with ``git commit``, ``git
checkout``, ``git fetch``, ``git pull``, ``git rebase``, ``git reset``
and so on - git stores a reference log (reflog for short). For every
move git stores where the head was. Command ``git reflog`` can be used
to view (and manipulate) the log.

In addition to the moves of the head of every branch git stores the
moves of the HEAD - a symbolic reference that (usually) names the
current branch. HEAD is changed with ``git checkout $BRANCH``.

By default ``git reflog`` shows the moves of the HEAD, i.e. the
command is equivalent to ``git reflog HEAD``. To show the moves of the
head of a branch use the command ``git reflog $BRANCH``.

So to undo a ``git reset`` lookup the original commit in ``git
reflog``, verify it with ``git show`` or ``git log`` and run ``git
reset $COMMIT_ID``. Git stores the move of the branch's head in
reflog, so you can undo that undo later again.

In a more complex situation you'd want to move some commits along with
resetting the head of the branch. Cherry-pick them to the new branch.
For example, if you want to reset the branch ``master`` back to the
original commit but preserve two commits created in the current branch
do something like::

    $ git branch save-master # create a new branch saving master
    $ git reflog # find the original place of master
    $ git reset $COMMIT_ID
    $ git cherry-pick save-master~ save-master
    $ git branch -D save-master # remove temporary branch

git revert: revert a commit
---------------------------

``git revert`` reverts a commit or commits, that is, it creates a new
commit or commits that revert(s) the effects of the given commits.
It's the only way to undo published commits (``git commit --amend``,
``git rebase`` and ``git reset`` change the branch in
non-fast-forwardable ways so they should only be used for non-pushed
commits.)

There is a problem with reverting a merge commit. ``git revert`` can
undo the code created by the merge commit but it cannot undo the fact
of merge. See the discussion `How to revert a faulty merge
<https://www.kernel.org/pub/software/scm/git/docs/howto/revert-a-faulty-merge.html>`_.

One thing that cannot be undone
-------------------------------

Whatever you undo, there is one thing that cannot be undone -
overwritten uncommitted changes. Uncommitted changes don't belong to
git so git cannot help preserving them.

Most of the time git warns you when you're going to execute a command
that overwrites uncommitted changes. Git warns you when you try to
switch branches with ``git checkout``. It warns you when you're going
to rebase with non-clean working tree. It refuses to pull new commits
over non-committed files.

But there are commands that do exactly that - overwrite files in the
working tree. Commands like ``git checkout $PATHs`` or ``git reset
--hard`` silently overwrite files including your uncommitted changes.

With that in mind you can understand the stance "commit early, commit
often". Commit as often as possible. Commit on every save in your
editor or IDE. You can edit your commits before pushing - change,
reorder, combine, remove. But save your changes in git database,
either commit changes or at least stash them with ``git stash``.


Merge or rebase?
================

Internet is full of heated discussions on the topic: "merge or
rebase?" Most of them are meaningless. When a DVCS is being used in a
big team with a big and complex project with many branches there is
simply no way to avoid merges. So the question's diminished to
"whether to use rebase, and if yes - when to use rebase?" Considering
that it is very much recommended not to rebase published commits the
question's diminished even further: "whether to use rebase on
non-pushed commits?"

That small question is for the team to decide. The author of the PEP
recommends to use rebase when pulling, i.e. always do ``git pull
--rebase`` or even configure automatic setup of rebase for every new
branch::

    $ git config branch.autosetuprebase always

and configure rebase for existing branches::

    $ git config branch.$NAME.rebase true

For example::

    $ git config branch.v1.rebase true
    $ git config branch.master.rebase true

After that ``git pull origin master`` becomes equivalent to ``git pull
--rebase origin master``.

In case when merge is preferred it is recommended to create new
commits in a separate feature or topic branch while using rebase to
update the mainline branch. When the topic branch is ready merge it
into mainline. To avoid a tedious task of resolving large number of
conflicts at once you can merge the topic branch to the mainline from
time to time and switch back to the topic branch to continue working
on it. The entire workflow would be something like::

    $ git checkout -b issue-42  # create a new issue branch and switch to it
        ...edit/test/commit...
    $ git checkout master
    $ git pull --rebase origin master  # update master from the upstream
    $ git merge issue-42
    $ git branch -d issue-42  # delete the topic branch
    $ git push origin master

When the topic branch is deleted only the label is removed, commits
are stayed in the database, they are now merged into master::

    o--o--o--o--o--M--< master - the mainline branch
        \         /
         --*--*--*             - the topic branch, now unnamed

The topic branch is deleted to avoid cluttering branch namespace with
small topic branches. Information on what issue was fixed or what
feature was implemented should be in the commit messages.


Null-merges
===========

Git has a builtin merge strategy for what Python core developers call
"null-merge"::

    $ git merge -s ours v1  # null-merge v1 into master


Advanced configuration
======================

Line endings
------------

Git has builtin mechanisms to handle line endings between platforms
with different EOL styles. To allow git to do CRLF conversion assign
``text`` attribute to files using `.gitattributes
<https://www.kernel.org/pub/software/scm/git/docs/gitattributes.html>`_.
For files that have to have specific line ending assign ``eol``
attribute. For binary files the attribute is, naturally, ``binary``.

For example::

    $ cat .gitattributes
    *.py text
    *.txt text
    *.png binary
    /readme.txt eol=CRLF

To check what attributes git uses for files use ``git check-attr``
command. For example::

$ git check-attr -a -- \*.py


Advanced topics
===============

Staging area
------------

Staging area aka index aka cache is a distinguishing feature of git.
Staging area is where git collects patches before committing them.
Separation between collecting patches and commit phases provides a
very useful feature of git: one can review collected patches before
commit and even edit them - remove some hunks, add new hunks and
review again.

To add files to the index use ``git add``. Collecting patches before
committing means you need to do that for every change, not only to add
new (untracked) files. To simplify committing in case you just want to
commit everything without reviewing run ``git commit --all`` (or just
``-a``) - the command adds every changed tracked file to the index and
then commit.

To add hunks of patches to the index use ``git add --patch`` (or just
``-p``). To remove collected files from the index use ``git reset HEAD
-- $FILE...`` To add/inspect/remove collected hunks use ``git add
--interactive`` (``-i``).

To see the diff between the index and the last commit (i.e., collected
patches) use ``git diff --cached``. To see the diff between the
working tree and the index (i.e., uncollected patches) use just ``git
diff``.

See `WhatIsTheIndex
<https://git.wiki.kernel.org/index.php/WhatIsTheIndex>`_ and
`IndexCommandQuickref
<https://git.wiki.kernel.org/index.php/IndexCommandQuickref>`_ in Git
Wiki.


ReReRe
======

https://git-scm.com/book/en/Git-Tools-Rerere


Database maintenance
====================

TODO: dangling objects, git gc, git repack.

https://gcc.gnu.org/ml/gcc/2007-12/msg00165.html

http://vcscompare.blogspot.ru/2008/06/git-repack-parameters.html


Tips and tricks
===============

TODO: sticky options; example: git grep -O.

TODO: tricky options; example: git log -p3.

TODO: bash/zsh completion, bash/zsh prompt.
https://git.kernel.org/cgit/git/git.git/tree/contrib/completion


git on server
=============

TODO: anonymous access; git over ssh; gitolite; gitweb; cgit; gitlab.

http://gitolite.com/gitolite/index.html

https://git.kernel.org/cgit/git/git.git/tree/gitweb

http://git.zx2c4.com/cgit/

From Mercurial to git
=====================

Mercurial for Git users https://mercurial.selenic.com/wiki/GitConcepts

https://github.com/felipec/git-remote-hg

https://hg-git.github.io/


References
==========

.. [] 


Copyright
=========

This document has been placed in the public domain.


\f
..
   Local Variables:
   mode: indented-text
   indent-tabs-mode: nil
   sentence-end-double-space: t
   fill-column: 70
   coding: utf-8
   End:
   vim: set fenc=us-ascii tw=70 :