We formulate a set of rules for quick reference, while the prose tries to motivate each of them. Do not always take them literally; you should value good reasons for your actions higher than manpages such as this one.
To achieve this, try to split your work into small steps from the very beginning. It is always easier to squash a few commits together than to split one big commit into several. Don’t be afraid of making too small or imperfect steps along the way. You can always go back later and edit the commits with git rebase --interactive before you publish them. You can use git stash push --keep-index to run the test suite independent of other uncommitted changes; see the EXAMPLES section of git-stash(1).
Merges have many advantages, so we try to solve as many problems as possible with merges alone. Cherry-picking is still occasionally useful; see "Merging upwards" below for an example.
Most importantly, merging works at the branch level, while cherry-picking works at the commit level. This means that a merge can carry over the changes from 1, 10, or 1000 commits with equal ease, which in turn means the workflow scales much better to a large number of contributors (and contributions). Merges are also easier to understand because a merge commit is a "promise" that all changes from all its parents are now included.
There is a tradeoff of course: merges require a more careful branch management. The following subsections discuss the important points.
There is a fourth official branch that is used slightly differently:
Each of the four branches is usually a direct descendant of the one above it.
Conceptually, the feature enters at an unstable branch (usually next or pu), and "graduates" to master for the next release once it is considered stable enough.
Example 1. Merge upwards
Always commit your fixes to the oldest supported branch that requires them. Then (periodically) merge the integration branches upwards into each other.
This gives a very controlled flow of fixes. If you notice that you have applied a fix to e.g. master that is also required in maint, you will need to cherry-pick it (using git-cherry-pick(1)) downwards. This will happen a few times and is nothing to worry about unless you do it very frequently.
Committing everything directly on the integration branches leads to many problems: Bad commits cannot be undone, so they must be reverted one by one, which creates confusing histories and further error potential when you forget to revert part of a group of changes. Working in parallel mixes up the changes, creating further confusion.
Use of "topic branches" solves these problems. The name is pretty self explanatory, with a caveat that comes from the "merge upwards" rule above:
Example 2. Topic branches
Make a side branch for every topic (feature, bugfix, ...). Fork it off at the oldest integration branch that you will eventually want to merge it into.
Many things can then be done very naturally:
Note that the last point clashes with the other two: a topic that has been merged elsewhere should not be rebased. See the section on RECOVERING FROM UPSTREAM REBASE in git-rebase(1).
We should point out that "habitually" (regularly for no real reason) merging an integration branch into your topics — and by extension, merging anything upstream into anything downstream on a regular basis — is frowned upon:
Example 3. Merge to downstream only at well-defined points
Do not merge to downstream except with a good reason: upstream API changes affect your branch; your branch no longer merges to upstream cleanly; etc.
Otherwise, the topic that was merged to suddenly contains more than a single (well-separated) change. The many resulting small merges will greatly clutter up history. Anyone who later investigates the history of a file will have to find out whether that merge affected the topic in development. An upstream might even inadvertently be merged into a "more stable" branch. And so on.
The solution, of course, is to make a merge that we can undo: merge into a throw-away branch.
Example 4. Throw-away integration branches
To test the interaction of several topics, merge them into a throw-away branch. You must never base any work on such a branch!
If you make it (very) clear that this branch is going to be deleted right after the testing, you can even publish this branch, for example to give the testers a chance to work with it, or other developers a chance to see if their in-progress work will be compatible. git.git has such an official throw-away integration branch called pu.
A feature release is created from the master branch, since master tracks the commits that should go into the next feature release.
The master branch is supposed to be a superset of maint. If this condition does not hold, then maint contains some commits that are not included on master. The fixes represented by those commits will therefore not be included in your feature release.
To verify that master is indeed a superset of maint, use git log:
Example 5. Verify master is a superset of maint
git log master..maint
This command should not list any commits. Otherwise, check out master and merge maint into it.
Now you can proceed with the creation of the feature release. Apply a tag to the tip of master indicating the release version:
Example 6. Release tagging
git tag -s -m "Git X.Y.Z" vX.Y.Z master
You need to push the new tag to a public Git server (see "DISTRIBUTED WORKFLOWS" below). This makes the tag available to others tracking your project. The push could also trigger a post-update hook to perform release-related items such as building release tarballs and preformatted documentation pages.
Similarly, for a maintenance release, maint is tracking the commits to be released. Therefore, in the steps above simply tag and push maint rather than master.
First, if you wish to continue to release maintenance fixes for the feature release made before the recent one, then you must create another branch to track commits for that previous release.
To do this, the current maintenance branch is copied to another branch named with the previous release version number (e.g. maint-X.Y.(Z-1) where X.Y.Z is the current release).
Example 7. Copy maint
git branch maint-X.Y.(Z-1) maint
The maint branch should now be fast-forwarded to the newly released code so that maintenance fixes can be tracked for the current release:
Example 8. Update maint to new release
If the merge fails because it is not a fast-forward, then it is possible some fixes on maint were missed in the feature release. This will not happen if the content of the branches was verified as described in the previous section.
Example 9. Rewind and rebuild next
The advantage of doing this is that the history of next will be clean. For example, some topics merged into next may have initially looked promising, but were later found to be undesirable or premature. In such a case, the topic is reverted out of next but the fact remains in the history that it was once merged and reverted. By recreating next, you give another incarnation of such topics a clean slate to retry, and a feature release is a good point in history to do so.
If you do this, then you should make a public announcement indicating that next was rewound and rebuilt.
The same rewind and rebuild process may be followed for pu. A public announcement is not necessary since pu is a throw-away branch, as described above.
Roughly speaking, there are two important workflows: merge and patch. The important difference is that the merge workflow can propagate full history, including merges, while patches cannot. Both workflows can be used in parallel: in git.git, only subsystem maintainers use the merge workflow, while everyone else sends patches.
Note that the maintainer(s) may impose restrictions, such as "Signed-off-by" requirements, that all commits/patches submitted for inclusion must adhere to. Consult your project’s documentation for more information.
There are three main tools that can be used for this:
Note the last point. Do not use git pull unless you actually want to merge the remote branch.
Getting changes out is easy:
Example 10. Push/pull: Publishing branches/topics
git push <remote> <branch> and tell everyone where they can fetch from.
You will still have to tell people by other means, such as mail. (Git provides the git-request-pull(1) to send preformatted pull requests to upstream maintainers to simplify this task.)
If you just want to get the newest copies of the integration branches, staying up to date is easy too:
Example 11. Push/pull: Staying up to date
Use git fetch <remote> or git remote update to stay up to date.
Then simply fork your topic branches from the stable remotes as explained earlier.
If you are a maintainer and would like to merge other people’s topic branches to the integration branches, they will typically send a request to do so by mail. Such a request looks like
Please pull from <url> <branch>
In that case, git pull can do the fetch and merge in one go, as follows.
Example 12. Push/pull: Merging remote topics
git pull <url> <branch>
Occasionally, the maintainer may get merge conflicts when they try to pull changes from downstream. In this case, they can ask downstream to do the merge and resolve the conflicts themselves (perhaps they will know better how to resolve them). It is one of the rare cases where downstream should merge from upstream.
Example 13. format-patch/am: Publishing branches/topics
If the maintainer tells you that your patch no longer applies to the current upstream, you will have to rebase your topic (you cannot use a merge because you cannot format-patch merges):
Example 14. format-patch/am: Keeping topics up to date
git pull --rebase <url> <branch>
You can then fix the conflicts during the rebase. Presumably you have not published your topic other than by mail, so rebasing it is not a problem.
If you receive such a patch series (as maintainer, or perhaps as a reader of the mailing list it was sent to), save the mails to files, create a new topic branch and use git am to import the commits:
Example 15. format-patch/am: Importing patches
git am < patch
One feature worth pointing out is the three-way merge, which can help if you get conflicts: git am -3 will use index information contained in patches to figure out the merge base. See git-am(1) for other options.