Autotest Dynamic Test SuitesThe next iteration of CrOS testing in Autotest is more dynamic and requires fewer manual changes with respect to adding and disabling tests. The first pass at this required suites to be manually defined by hardcoding a list of tests, and added infrastructure to find a set of machines on which to run those tests. The former required manual management, and the latter can already be done by appropriately using autotest. Dynamic suites, in contrast, are defined by looking over all existing control files at runtime, and rely on the autotest infrastructure to shard the tests across all available, appropriate machines.
As far as someone scheduling a BVT run is concerned, sending an RPC to the autotest frontend telling it to schedule ‘control.bvt’ for a certain build on a certain board is sufficient; all other logic exists inside of the Autotest frontend (AFE).
Specifying job dependencies
This is how a single test specifies the kind of machine on which it can be run. DEPENDENCIES is just a list of labels. A machine satisfying all of them must be found in order for the job to be successfully scheduled.
- DEPENDENCIES=‘CDMA’, ‘POWER’, ‘GOBI3K’, ‘X86’
All current tests need to be re-evaluated and updated to encapsulate required dependencies. e.g. GSM, CDMA, POWER, GOBI3K, BLUETOOTH etc etc. For architecture-dependent tests, we can also include labels like ARM, TEGRA2, TEGRA3, X86, etc.
Board compatibilityWe ultimately decided against specifically expressing which boards a test should and should not run on. We couldn’t come up with a reason we’d need to avoid a specific board or class of boards that could not be expressed with DEPENDENCIES. Bluetooth tests don’t apply to Alex? Well, no Alex device will have BLUETOOTH. ACPI tests don’t apply to ARM? Well, they’ll DEPEND on x86.
Instead, we will rely on dynamic_suite code in the AFE to distinguish between a test has unsatisfiable dependencies at the time of scheduling, and one that can be scheduled, but does not run in the time allotted. In the latter case, the job will be ABORTed. In the former case, we will fail the test with TEST_NA. In the Alex case above, for example, all tests with a BLUETOOTH dep will be TEST_NA. As a more complicated example, consider a GOBI3K test on Mario, and assume we have only one Mario with that modem in it. If all is well, the test can be scheduled in the time allowed and things are good. If the device is down, the AFE still knows that it should have a machine that meets the test’s DEPENDENCIES, but discovers that the device is not currently ready. That’s fine, and it schedules the test. When the time allotted for the suite is up, the test would be ABORTed and included as a failure in the test results.
Job reportingFor TPM’s, one can provide a view of test results per-build-per-board that shows all jobs that could be scheduled, and the results of those tests. Specifically, jobs that are TEST_NA can be ignored if the customers of the dashboard so-desire.
For the lab team, we can show all jobs that we tried to schedule, so that we can see which jobs are getting starved and detect new, unexpected starvation.
Defining suitesWe've added a custom variable to control files called SUITE. SUITE can contain a comma-delineated list of arbitrary strings. Only suite names that have an associated control file will actually be usable in any meaningful way. By convention, we put all suite control files in a top-level directory called 'test_suites', and name the files appropriately. For example,
test_suites/control.bvt is the control file for the BVT suite.
This does require parser changes, and does make including upstream tests in our suites a bit complex. The former is not that difficult (and probably upstreamable). We could likely address the latter by adding a level of control-file-indirection, and having a site-specific file that uses SUITE and references the upstream test.
We use the already-existing “EXPERIMENTAL” flag to mark tests as candidates for inclusion in a suite, and will run them along with the non-experimental tests -- but tag them separately so that results can be separated out in the final report.
We considered making all test suites a query of sorts over the labels defined in control files (the BVT would be all SHORT, functional tests that are not experimental, for example), which would avoid these issues, but decided that this smacked too much of magical side effects for use in our primary test suites. However, we will provide a library of code that teams can use to create their own suite control files. For example, ‘kernel.soak’ could simply be a one-liner that found and scheduled jobs for all SHORT, MEDIUM, and LONG functional and stress tests in the “kernel” category.
Scheduling a suite for a given board and imageWe defined a new autotest RPC for this, with the following interface:
def create_suite_job(suite_name, board, build, pool, check_hosts=True,
Create a job to run a test suite on the given device with the given image.
When the timeout specified in the control file is reached, the
job is guaranteed to have completed and results will be available.
@param suite_name: the test suite to run, e.g. 'bvt'.
@param board: the kind of device to run the tests on.
@param build: unique name by which to refer to the image from now on.
@param pool: Specify the pool of machines to use for scheduling
@param check_hosts: require appropriate live hosts to exist in the lab.
@param num: Specify the number of machines to schedule across.
@param file_bugs: File a bug on each test failure in this suite.
@raises ControlFileNotFound: if a unique suite control file doesn't exist.
@raises NoControlFileList: if we can't list the control files at all.
@raises StageBuildFailure: if the dev server throws 500 while staging build.
@raises ControlFileEmpty: if the control file exists on the server, but
can't be read.
@return: the job ID of the suite; -1 on error.
Running the suitecreate_suite_job() starts by telling the dev server to fetch the image to be tested, and its associated autotest bundle, from the URL provided. Clients of autotest will no longer need to be aware of the dev server. The details of this may change going forward, but all this work will be handled by this RPC. Next, create_suite_job() will verify that the image meets a minimal bar for testability by reimaging a small number of machines and ensuring they boot back up successfully.
If the image is deemed NSFW, then the suite job will FAIL with an appropriate message. Once we know the image is safe enough that we aren’t going to brick our lab, we go through all the control files associated with the build being tested, and schedule the proper jobs. At that point, we let the autotest scheduler take responsibility for all the async scheduling/dependency scheduling. It is important that all jobs are scheduled via rpc calls with a name that associates them all together so that when we want to look at results we can aggregate them all in a query. This will allow callers (like the buildbot-autotest connection) to poll for the status of all jobs kicked off by the BVT suite control file, and eventually find and gather up results for reporting purposes