Fixing layout test flakiness

We'd like to stamp out all the tests that have ordering dependencies. This helps make the tests more reliable and, eventually, will make it so we can run tests in a random order and avoid new ordering dependencies being introduced. To get there, we need to weed out and fix all the existing ordering dependencies.

Diagnosing test ordering flakiness

These are steps for diagnosing ordering flakiness once you have a test that you believe depends on an earlier test running.

Bisect test ordering

  1. Run the tests such that the test in question fails.
  2. Run ./Tools/Scripts/print-test-ordering and save the output to a file. This outputs the tests run in the order they were run on each content_shell instance.
  3. Create a file that contains only the tests run on that worker in the same order as in your saved output file. The last line in the file should be the failing test.
  4. Run ./Tools/Scripts/bisect-test-ordering --test-list=path/to/file/from/step/3

The bisect-test-ordering script should spit out a list of tests at the end that causes the test to fail.

At the moment bisect-test-ordering only allows you to find tests that fail due to a previous test running. It‘s a small change to the script to make it work for tests that pass due to a previous test running (i.e. to figure out which test it depends on running before it). Contact ojan@chromium if you’re interested in adding that feature to the script.

Manual bisect

Instead of running bisect-test-ordering, you can manually do the work of step 4 above.

  1. run-webkit-tests --child-processes=1 --order=none --test-list=path/to/file/from/step/3
  2. If the test doesn‘t fail here, then the test itself is probably just flaky. If it does, remove some lines from the file and repeat step 1. Continue repeating until you’ve found the dependency. If the test fails when run by itself, but passes on the bots, that means that it depends on another test to pass. In this case, you need to generate the list of tests run by run-webkit-tests --order=natural and repeat this process to find which test causes the test in question to pass (e.g. crbug.com/262793).
  3. File a bug and give it the LayoutTestOrdering label, e.g. crbug.com/262787 or crbug.com/262791.

Finding test ordering flakiness

Run tests in a random order and diagnose failures

  1. Run run-webkit-tests --order=random --no-retry
  2. Run ./Tools/Scripts/print-test-ordering and save the output to a file. This outputs the tests run in the order they were run on each content_shell instance.
  3. Run the diagnosing steps from above to figure out which tests

Run run-webkit-tests --run-singly --no-retry. This starts up a new content_shell instance for each test. Tests that fail when run in isolation but pass when run as part of the full test suite represent some state that we‘re not properly resetting between test runs or some state that we’re not properly setting when starting up content_shell. You might want to run with --time-out-ms=60000 to weed out tests that timeout due to waiting on content_shell startup time.

Diagnose especially flaky tests

  1. Load https://test-results.appspot.com/dashboards/overview.html#group=%40ToT%20Blink&flipCount=12
  2. Tweak the flakiness threshold to the desired level of flakiness.
  3. Click on webkit_tests to get that list of flaky tests.
  4. Diagnose the source of flakiness for that test.