In a recent discussion with ctalbert, we discussed what is the right course to take for getting automation running on windows mobile and how we could avoid this 6 month nightmare for the next platform. The big question we want to answer is how can we implement our automation for talos and unittests on Windows Mobile and not have to reinvent the wheel for the next platform?
Our current approach to date has been to take what we have for Firefox (which works really good) and adapt it to work on a mobile device. This approach has not been very successful in the development of Fennec (functionality, UI, performance) nor the automation of unittests and talos.
After further thought on this subject, I find there are 4 area to focus on:
- Infrastructure and Tools
- Porting Harnesses and Tests
- Managing Automation and Failures
- Mobile Specific Automation
Each of these 4 areas are tightly coupled yet require a much different solution than the others. Let me outline each area in a bit more detail describing the problem, our solution, and a longer term solution.
Infrastructure and Tools:
This is the area of all the assumptions. The OS you run on, network connectivity, available disk space for test and log files, tools to manage your environment and processes and a method for doing this all automatically. I am going to leave out the build and reporting infrastructure from this topic as those take place separately and don’t run on the device.
Our first port of this to maemo was not as difficult because we could use python to manage files, network, and processes just as we do on a desktop OS. Minor adjustments had to be made for running on storage cards, using different python libraries (we have a limited set available on maemo) and system tools, as well as changing requirements for process names and directory structures. Also maemo has ssh functionality, a cli terminal, and a full set of command line tools to manage stuff.
Moving onto Windows Mobile, we don’t have the tools and infrastructure like we do on Maemo. This means we need to spend a lot of time hacking our communications required for automation and scripting tools like python. Almost all process interaction (create, query, kill), need custom code to take care of it. This has presented us with a problem where we don’t have the luxury of our OS support and tool support. Our approach to date has been to write (or in the case of python, port) our tools to make them work on the device. Unfortunately after 4 months we don’t have a working set of automation that people are happy with.
How can we create infrastructure that is scalable to all platforms? From what I have seen, we need to move away from our reliance on all tools on the device. This means no python or web server on the device, no test data on the device, and assume we won’t be able to store log files or use common system tools. I would like to see a custom communication layer for each OS. So for Windows Mobile, we would create a server that lives on the device which: ensures we have ip connectivity, provides file system tools, process management tools, and allows for the OS to reboot and come back connected. The other half of this is a job server which sends commands to the device and serves/collects test data via a web interface. I know this is a big step from where we are now, but in the future it seems like the easiest approach.
Porting Harnesses and Tests:
This focus area is more about making sure the environment is setup correctly, the right tests are run and useful logs are created. This can be developed without a full infrastructure in place, but it really requires some knowledge about the infrastructure.
For Maemo, a lot of work was done to extract the unittests from the source tree and retrofit the tools and harnesses to manage tests in “chunks” rather than running them all at once. We have done a lot of work to clean up tests that assume preferences or ones that look for features.
The challenge on Windows Mobile is without an infrastructure the tests rely on we need to do things differently. Very few bugs were found that prevented tests from running while porting tests to Maemo. For WinMo, that is a different story. We cannot run the web server locally, cannot load our mochitests in an iframe, and have trouble creating log files. All of these issues force us to morph our testing even further away from where it was and realize that we need to do this differently.
What I see as the ultimate solution here is to setup a “test server” which serves all our test data. Each test would remove the dependencies it has on localhost and work with a test server given that it has an IP address. We would then have an extension for Firefox/Fennec which would serve as the test driver to setup, run, and report the test results. This would allow for any mobile device (that supports addons), desktop, or community member to install the addon and start testing.
Managing Automation and Failures:
This is a much smaller topic than the previous two, but once we do have data generated, how do we keep generating it reliably and use it to our advantage?
Right now our toolset of Tinderbox and Buildbot do a great job for Firefox. I believe that we can still utilize them for mobile. There might be specific issues with them, but for the purposes of running Talos and Unittests, we need something that will take a given build, send it to a machine, run tests, and figure out when it is done. We even have great tools to notify us via IRC when a test suite fails.
The danger here is when testing on a new device or product, we find hundreds if not thousands of failures. The time required to track those down and file bugs is a big job by itself. When you have a large number of bugs waiting to be fixed it won’t happen in the same week (or quarter). This brings up a problem where nobody pays attention to the reported data because there are always so many failures.
The other problem that occurs is with so many crashes and running tests in smaller chunks or 1 by 1, we end up with smaller log files and lose the ability to get the pass/fail status that our test harnesses for Firefox rely upon. I know simply looking for a TEST-UNEXPECTED string in *.log is a reasonable solution, but as we have learned there are a lot of corner cases and that doesn’t tell you which tests were not run.
How can we make this better? Our first step to solving this problem is LogCompare. This is a log parser that uploads data to a database and lets us compare build by build. This solves the problem of finding new failures and ignoring known failures if we want to. A final solution to this would be to expand on this idea and have test runners (via the addon) upload test result blobs to a database. Adding more tools to query status of a test run, missed tests, etc… can be done giving us more detailed reports than just pass/fail. In the long term a tool like this can be used to look at random orange results as well as getting status of many community members pitching in CPU time.
Mobile Specific Automation:
The last piece of the puzzle is to create specific test cases to exercise mobile features. This is fairly trivial and great work has already been done for bookmarks and tabs using browser-chrome. This is important as the more developers we have working on mobile and the more branches we have, the greater the frequency of regressions.
Here is the problem, the mobile UI is changing so rapidly that it would take a full time job to keep up with the automation. This is assuming you have comprehensive tests written. It is actually faster to install a build and poke around the UI for 10 minutes than it is to keep the tests maintained. I know in reality there will be moving pieces on a regular basis, but right now we are ripping big pieces out and rewriting everything else. As a reference point the tab functionality has changed 4 times in the last year.
Looking at the regressions found in the last couple weeks we would not have found those with automation. There is a great list of stuff we want to automate in the Fennec UI and almost none of those would have failed as the result of a recent regression. This means we need many more tests than a few hundred core functionality tests. It also points out that we are not going to catch everything even if we all agree that our tests are comprehensive.
What is the best way to utilize automation? Until a 1.0 release, we have to expect that things will be very volatile. We should fix the automation we currently have and require developers to run it on their desktop or device before checking in. This should be a 1.0 requirement. If a developer is changing functionality, fix the tests. Why this works is we don’t have a lot of tests right now. This will serve more of a purpose to fix the process than finding bugs. Post 1.0, lets build up the automation to have decent coverage of all the major areas (pan, zoom, tabs, controls, awesomebar, bookmarks, downloads, addons, navigation), and keep the requirement that these tests need to run for each patch we checkin. The time to run a test will be fast on a desktop <5 minutes.
While we seem to be flopping around like a fish out of water, we just need some clear focus and agreement from all parties about the direction and we can have a great solution. My goal is to be forward looking and not dwell on the existing techniques that work, yet are still being improved. After looking at this from a future point of view I see that developing a great solution to meet our needs now can also allow for greater community involvement leading to greater test coverage for Fennec and Firefox. The amount of work required to generalize this is equivalent to the same work for a specialized solution for Windows Mobile.
I encourage you to think about ways we can reduce our test requirements while allowing for greater flexibility.