Testing and CI strategy
We evaluated a few different CI solutions (Travis, Cloudbees, drone.io) 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
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 resources and/or github
- build binary packages ready for use. It's currently working for Android and Windows, for example.
- push automatic updates to the Google Play Store
What is left to do ?
- expand our set of test cases
- build packages for more different platforms (iphone / wince / ubuntu ?)
- push updates automatically to f-droid
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. These maps are usable for routing tests, but most of them are not available for real navigation because of the "flood" bug.
Proposed 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
Example (from the svn + git era, needs update) :
- (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 master here )
- to benefit from automated steps. Each commit in trunk triggers 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 were used for the SVN workflow have an unclear license and we can't publish them
- to leverage 3-rd party services and resources ( CircleCI provides 4 containers for free for FLOSS projects ).
OK, I've pushed my commits, where is my CI output?
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.
It won't work, I get weird merge conflicts for files I've never touched!
The folder layout on github differs from that on SVN in one way: The root folder on github corresponds to the navit/ folder in SVN. Thus, if you have used git-svn to create your local repo, you won't be able to just add navit-gps/navit as another remote because folder layouts won't match and merges will create a mess.
To work around this issue, do the following:
Clone navit-gps/navit into a new folder. We'll refer to this as the navit-gps repo:
git clone https://github.com/navit-gps/navit.git FOLDER_NAME
Check out trunk and pull:
git checkout trunk git pull
In your git-svn repo, get the latest revision from SVN and merge it into your feature branch, so your feature branch will be based on the latest revision:
git checkout master git svn rebase git checkout your-feature-branch git merge master
Now go back to the navit-gps repo. If you don't have a branch for your feature yet, create one and check it out:
git branch your-feature-branch git checkout your-feature-branch
If your feature branch exists already, do the following instead:
git checkout your-feature-branch git merge trunk
Now you have two options. Read them carefully and understand their limitations and side effects, then choose the one that works best for your sitation:
- Create a patch in your git-svn repo and apply it to the navit-gps repo. This works well for a freshly forked branch, but if the branch has had other commits since it was forked from trunk, the patch may fail to apply.
- Copy the contents of the navit/ folder in your git-svn repo into the root folder of the navit-gps repo, overwriting any existing files. This will work even for branches that had other commits since they were forked. However, any changes that did not originate either from SVN trunk or your feature branch would be overwritten by this. Also, this method will not catch any file deletions, and will create duplicates for moved files. If you decide to go this way, be sure to examine the changes in the git-svn repo before you stage them.
Now stage all changed files and commit.
Finally push your changes:
If you are updating an existing feature branch, you have pulled from trunk earlier, thus pushing the changes will trigger a CI build. When CircleCI detects that trunk has been merged into a branch, it will build that branch. The build will include everything that got pushed together with the merge, thus your changes will be included.
(to do: figure out how to trigger CI manually if I can't merge trunk because it hasn't changed since the last push to the feature branch...)