Development Workflow for Trac
The public branches listed in TracDownload#LatestDevelopmentSourceCode should remain bug free and always usable from an end-user perspective, as we support installing directly from the branches. To achieve this, we follow the practices listed below.
Initial code review
Except for trivial fixes, it's usually a good idea to start with a patch or an experimental branch, to provide some visibility of the changes to the other contributors.
A patch is always attached to a ticket. If there is no ticket, then create one and attach the patch to it.
When there are many iterations or spin-off changes needed, it's a good idea to start a branch, either in the svn sandbox for those who have the commit permission or inside an external DVCS repository, by forking our Mercurial or Git mirrors (see TracRepositories).
Writing Commit Messages
Commits should be atomic. If three separate issues are being fixed (unless they are all fixed by one change), they need to be done as three separate commits. This also applies to whitespace and style changes, which should be done in their own commit. Whitespace commits should not include code or content changes. Accordingly, code change commits should not include whitespace changes (unless the whitespace changes are on the same lines as the code being changed).
Commit messages should describe what changed, and reference the issue number if the commit closes or is associated with a particular issue. Example:
Version: Short Description Optional Long Description Refs #zzz
This is the first line. It consists of the development version, like
1.2.2dev. This line must be 72 characters or less. There should be no full stop (period) at the end and the imperative form should be used (example: Add a command for setting wiki attributes).
There are two line breaks between the subject and the long description. The description can have any length and formatting, like lists, but it must be hard-wrapped at 80 characters.
List the issues being addressed, using either
Fixes. A comma-separated list of issues can be used:
Refs #12, #34.
Integration in release branches
Merging in this direction (porting or forward porting) makes it quite easy to merge all pending changes from one stable branch to the next, eg 0.12-stable to 1.0-stable, then to trunk. This workflow is much simpler than the opposite one, back porting, which involves cherry-picking or careful "blocking" of changes which shouldn't be merged back. The SCM tools have generally better support for merging in this forward direction, even Subversion since v1.5.
While this pattern of forward porting makes it easy to merge all pending changes, in practice we cherry-pick changes as they are committed in order to keep all branches in sync and prevent pending changes from accumulating on a branch. Multiple changesets can be cherry-picked in a single merge command using a comma-separated list of changesets:
0.12-stable$ ed trac/attachment.py # or your other favorite editor 0.12-stable$ make test ... (all is good) 0.12-stable$ svn ci -m "0.12.6dev: fixed ... (#xyz)" ... Committing transaction... Committed revision 15410.
0.12-stable$ cd ../1.0-stable 1.0-stable$ svn merge -c 15410 ^/branches/0.12-stable 1.0-stable$ make test ... xxx 1.0-stable$ # some fixes needed for API changes 1.0-stable$ svn ci -m "1.0.2dev: Merged from 0.12-stable." ... Committing transaction... Committed revision 15411.
1.0-stable$ cd ../trunk trunk$ svn merge -c 15411 ^/branches/1.0-stable trunk$ make test ... xxx trunk$ # some fixes needed for API changes trunk$ svn ci -m "1.1.2dev: Merged from 1.0-stable." ... Committing transaction... Committed revision 15412.
Among the possible porting related fixes that should be done when porting:
- Use Python idioms adapted to the minimal version on the branch, eg for Trac 0.12 the baseline is Python 2.4, for Trac 1.0 and trunk it's 2.5; this means that among other things we can use
with ...as appropriate.
- Use newer APIs and conventions, eg the DatabaseApi#Trac1.0API; see the ApiChanges subpage for the corresponding target branch.
If a changeset should not be forward ported, for example when extracting new messages to the catalog, the commit should be blocked using a record only merge. The record only merge is the same as a cherry-pick merge, with the addition of the
Pushing from a DVCS to SVN
If the changes were staged in your dev Git or Hg repository, some additional steps are needed to commit the changes to Subversion. The steps will be described for a Git repository, but will be similar for an Hg repository.
You should first interactively rebase (
git rebase -i) your changes to get them in a form that is appropriate for committing to the Subversion repository. In doing so, consider how you'd like the changes to be represented in the repository history, taking into account that we frequently interrogate the repository history to discover the cause of regressions or understand the purpose and intent of code. For example, logically related changesets may be squashed if they were staged as multiple changesets for the purpose of easing code review. However, unrelated changes and refactorings should be pushed as separate changesets so that they don't obfuscate functional changes.
Next, reword your log messages in a form that is appropriate for the central repository: prefix each log message with the target version (e.g.
1.0.10dev:) and reference the appropriate ticket(s). See the log for examples.
Once you've interactively rebased your Git branch and prepared your log messages, the process to push changes to Subversion is:
- Rebase your staging branch against the HEAD of the branch you'll be committing to
- Checkout an svn working copy of the branch you'll be committing to
- Copy the svn metadata directory (
.svn) of the working copy into the root of your git repository
- Step through checkouts of your repository changesets and commit each of them to svn
Here are the same steps described in command line form:
$ svn checkout <target-branch> trac-svn-wc $ cp -r trac-svn-wc/.svn trac-git-repos/ $ cd trac-git-repos $ git status -sb ## tXYZ $ git checkout tXYZ~3 $ git log --format="%B" > commit.txt $ svn add ...; svn del ... $ svn ci -F commit.txt $ git checkout tXYZ~2 ... $ git checkout tXYZ~1 ... $ git checkout tXYZ ...
Note that it's also perfectly workable to keep a single checkout under the control of multiple revision control systems simultaneously, for the long run.