Wednesday, July 16, 2014

A new test runner approaches

The problem
How acceptance tests are packaged and run has morphed over time. When autopilot was originally conceived the largest user was the unity project and debian packaging was the norm. Now that autopilot has moved well beyond that simple view to support many types of applications running across different form factors, it was time to address the issue of how to run and package these high-level tests.

While helping develop testsuites for the core apps targeting ubuntu touch, it became increasingly difficult for developers to run their application's testsuites. This gave rise to further integration points inside qtcreator, enhancements to click and its manifest files, and tools like the phablet-tools suite and click-buddy. All of these tools operate well within the confines they are intended, but none truly meets the needs for test provisioning and execution.

A solution?
With these thoughts in mind I opened the floor for discussion a couple months ago detailing the need for a proper tool that could meet all of my needs, as well as those of the application developer, test author and CI folks. In a nutshell, a workflow to setup a device as well as properly manage dependencies and resolve them was needed.

Autopkg tests all the things
I'm happy to report that as of a couple weeks ago such a tool now exists in autopkgtest. If the name sounds familar, that's because it is. Autopkgtest already runs all of our automated testing at the archive level. New package uploads are tested utilizing its toolset.

So what does this mean? Utilizing the format laid out by autopkgtest, you can now run your autopilot testsuite on a phablet device in a sane manner. If you have test dependencies, they can be defined and added to the click manifest as specified. If you don't have any test dependencies, then you can run your testsuite today without any modifications to the click manifest.

Yes, but what does this really mean?
This means you can now run a testsuite with adt-run in a similar manner to how debian packages are tested. The runner will setup the device, copy the tests, resolve any dependencies, run them, and report the results back to you.

Some disclaimers
Support for running tests this way is still new. If you do find a bug, please file it!

To use the tool first install autopkgtest. If you are running trusty, the version in the archive is old. For now download the utopic deb file and install it manually. A proper backport still needs to be done.

Also as of this writing, I must caution you that you may run into this bug. If the application fails to download dependencies (you see 404 errors during setup), update your device to the latest image and try again. Note, the latest devel image might be too old if a new image hasn't been promoted in a long time.

I want to see it!
Go ahead, give it a whirl with the calendar application (or your favorite core app). Plug in a device, then run the following on your pc.

bzr branch lp:ubuntu-calendar-app
adt-run ubuntu-calendar-app --click=com.ubuntu.calendar --- ssh -s /usr/share/autopkgtest/ssh-setup/adb

Autopkgtest will give you some output along the way about what is happening. The tests will be copied, and since --click= was specified, the runner will use the click from the device, install the click in our temporary environment, and read the click manifest file for dependencies and install those too. Finally, the tests will be executed with the results returned to you.

Feedback please!
Please try running your autopilot testsuites this way and give feedback! Feel free to contact myself, the upstream authors (thanks Martin Pitt for adding support for this!), or simply file a bug. If you run into trouble, utilize the -d and the --shell switches to get more insight into what is happening while running.

1 comment:

  1. Thanks for sharing that! An important feature here is to reset the phone to the "factory state" before running the test, by appending "-- -r". This provides the "-r" (--reset) option to ssh-setup/adb, and will avoid test failures due to locally installed clicks, changed locale, etc.