On Piwigo we typically have two active branches:

• trunk holds the main source code with features for the next major version
• 2.7 (in instance) holds the source code of the current major version, only bug fixes and minor features are pushed to this branch

As the SVN concept of branches is very loose we used to develop either on master or 2.7 and then merge specific commits to the other branch if needed. This concept is totally applicable in Git, it's called cherry pick.

But this is not how Git is supposed to work. Git tends to track every changes, and links them to others, and because cherry pick basically forge a new commit from another one, a part of the history is lost.

The workflow we follow is based on Git flow as described by Vincent Driessen in his excellent article A successful Git branching model. But with important changes because of our way to work and the translation process.

Don't worry it's actually very simple.

Contains the latest stable version of the source code. Production tags are created on this branch. Never commit on this branch.

Contains the latest development code. It's like the trunk on Subversion. It is advised to only make small commits on this branch, prefer using a bug/ or feature/ branch.

develop is merged into master just after a new release.

Contain the code of a specific major version and all its minor versions. It is initiated by a single commit making necessary changes (version change, production config, etc.). Like develop it's advised to only make small commits on this branch.

Remember that we NEVER make database changes in minor releases, neither we add translatable strings.

Minor releases are tagged on the release branch.

The current release branch is merged into develop just after a new release.

Are used to… fix bugs. Create one bug branch by bug. It should be named after a specific issue number (eg: bug/1324).

If the bug has to be ported to the current release branch, then you must init the bug branch from the release branch, and NOT from develop, otherwise you won't be able to merge it.

Once the fix is ready merge it to develop and to the current release branch if needed.
You notice that, unlike Vincent Dressen, we don't merge bug fixes in master but in the current branch, which is then merged back into master when ready.

Don't forget to delete the branch locally and remotely when finished.

Almost the same as bug/ but for new features. It can be named after a specific issue number, or an arbitrary name for unlisted features.

Once the feature is ready, merge it too develop and in the current release branch only if it's a minor feature.

Like bugs, feature branches that will be merges to the release branch must be init from the release branch, and not from develop;

Don't forget to delete the branch locally and remotely when finished.

This branch is only used for translations committed by Lexiglot and has a special behavior.

Most importantly the translation branch tracks the develop branch. Which means you have to merge develop to translation after merging a feature which add new language strings.

Because translation is tracking develop, it cannot be merged to the current release branch. It's the only case where we use cherry-picking to apply new translations to the release branch. This is only done before a release.

These are some random advices you should really apply in order to have a clean Git history.

When merging a branch into another, the default behavior of Git is to fast-forward it, that means if no changes has been made on the target branch, the source branch will simply disappear from the tree, making it difficult to localize merging point.

By adding the –no-ff option to git merge you will force Git to create an empty technical commit when merging.

This is of course not always required, it's up to you to evaluate if a specific set of commits belongs to a separated branch, but don't forget than the branch will still be visible if someone else committed in the mean time.

Say you commit on branch A, and Bob also committed on branch A, and he pushed the branch to the central server. Then you want tu push your commit: Git will refuse to it, he will ask you to pull the remote branch. Ok so you do a git pull. If you are lucky there won't be any merge conflict and you will be able to push.

But this will create a one-shot branch, and a new commit “Merge branch 'master' on 'master'”. And this is a complete mess!

(from gitready.com)

The fact is pull is a shorthand for fetch & merge. But you don't want a merge, you want to apply your modifications after the ones of Bob.

The solution is to use the –rebase switch, which will try to reapply your commits at the end of the tree.

Say we are working on the branch 2.7, the command is: git pull –rebase origin 2.7

And that's it! Much cleaner.

Of course you may have conflicts during the rebase and have to manually solve them, like a normal merge.

A golden rule is “commit often, push once”. That mean you should make very small commits, assigning a very single issue, and push everything when you are really sure it's ready.

You can even pause your work on a branch, go to another one, commit a bit more, return to the first one, finish your work, and finally push everything.

Don't forget some very handy tools of git to help you to craft “perfect” commits:

• stash to store your current modifications and clean your working copy
• add -p to stage only some parts of a file
• squash to regroup sequential commits into one
• commit –ammend to modify your latest commit

(You must not abuse of these last two, only use them for correcting typos, and this kind of things;)

All this “rewriting history” can be done when you work locally, once you pushed everything on Github, it's too late. Hence the “push once”.