Testing and CI strategy
We evaluated a few different CI solutions and currently CircleCI seems to be the best option for what we want to do.
- opensource project can get up to 4 concurrent builds, for free
- it is easy to setup and fully automated
- it allows us to build, test and store the build results easily
The only drawback is that it is based upon Github. We have setup an extra repository to do more tests.
Currently we can :
- have automatic builds upon each commit in each branch of the repository
- run automated routing tests using dbus, and capture a screenshot and gpx / geojson output of the result, and have that result archived in the CI ressources and/or github
- build binary packages ready for use. It's currently working for Android, for example.
What is left to do ?
- expand our set of test cases
- build packages for more different platforms
- figure out how to get svn sync'ed back automatically
Ressources available during a build
- 32x Intel(R) Xeon(R) CPU E5-2670 v2 @ 2.50GHz
- 4GB of RAM
- ~ 550GB of disk space
- a single task cannot exceed 2 hours. This is causing issue for maptool. As a workaround, we have split the world in different zones, and had to sub-divide Europe in even smaller parts (each map is a branch so that it can be built individually). The list of maps and the latest generated map for this area is automatically updated and available here
Proposed SVN + GIT + CI workflow
This example should explain a few possible use cases:
- A blue arrow represents manual action
- A small dashed arrow ( ... ) represents an automatic action triggered by a SVN hook
- A long dashed arrow ( _ _ . _ _ . ) represents an automatic action manager by the CI server
- (1) A dev commits to SVN. The code is automatically merged into git, in the trunk branch, which triggers a build on the CI server. In this example, the tests (or build) fails, the workflow ends here. The git-master branch hasn't been updated, the code in this branch should still be functional.
- (2) A dev commits a fix for the previous problem in SVN. The code is automatically merged into git, in the trunk branch, which triggers a build on the CI server. The tests succeed, the code is automatically merged into the 'master' branch, packages are built and translations templates are uploaded to launchpad. No action is required from the dev.
- (3) A dev wants to work on an experimental feature. This could be a brand new functionality, a rewrite of a core component ( e.g. project "high five" ) or even a branch related to a specific ticket in trac. He forks the git trunk in a branch for his feature. He can benefit from having the CI test, and if configured, build packages for his feature.
git clone https://github.com/navit-gps/navit.git git checkout trunk # The branch we will create needs to be based upon 'trunk' git checkout -b [name_of_your_new_branch] # We now create our own branch
- (4) The feature is now mature and is ready to be merged into the trunk. As we currently keep the SVN trunk as the main repository, we can't just merge the code into git trunk, which should only get updates from svn trunk. The correct way to merge the feature is:
- pull the latest code from git trunk to make sure that commits in svn trunk did not break the code in the branch, and merge that pull in the branch. This will trigger a CI test that will ensure the status of the branch
git merge trunk git push
- export that feature as a patch ( a diff between the branch, and git trunk).
git diff trunk
- patch svn trunk to merge the code. This will trigger another merge to git-trunk and a call to the CI workflow
Why do we want this?
The whole point of this workflow is :
- to ensure that we have a branch, somewhere, where the code is tested, and which we should never break ( this is git-master here )
- to benefit from automated steps. Each commit in SVN should trigger a regression test, coverity code analysis and package build ( instead of a daily build as we have currently )
- to rebuild a documented, publicly available workflow. Some of the scripts that are currently used for the SVN workflow have an unclear license and we can't publish them
- to leverage 3-rd party services and ressources ( CircleCI provides 4 containers for free for FLOSS projects ). These services typically require the use of Github
I already have my own awesome fork of Navit in github!
That's not a problem. Just add a remote pointing to navit-gps/trunk :
git remote add navit-gps https://github.com/navit-gps/navit.git git remote update
Now, checkout the remote trunk :
git checkout navit-gps/trunk
You will probably end up in a "detached head" state. It's ok, just create a local branch:
git checkout -b trunk
Make sure that this branch is setup to track the remote :
git branch --set-upstream-to=remotes/navit-gps/trunk trunk
Now, you have your own local copy of the trunk against which you can branch, merge and update at will.
Don't forget to push your new branch upstream :
git push origin trunk
Now, if you want to get the latest updates from svn :
git checkout trunk git pull
And you can merge/fork on up-to-date code.