[webkit-dev] Handling flaky layout-test failures in EWS

[Aakash Jain]

> On Dec 3, 2019, at 1:54 PM, Ryosuke Niwa <rniwa at webkit.org> wrote:
> 
> 
> On Tue, Dec 3, 2019 at 9:29 AM Alexey Proskuryakov <ap at webkit.org <mailto:ap at webkit.org>> wrote:
> 
> Yes, I think that this makes more sense than retrying.
> 
> What is the current behavior when a patch introduces substantial flakiness? E.g. this scenario:
> 
> - First test run produces 5 failures.
> - Second test run produces 5 different failures.
> - Clean re-run produces no failures.
> 
> This looks like decent evidence to make the EWS bubble red. I don't know what exactly the threshold should be, but that should certainly be below 30.
> 
> This makes sense to me.
> 
> Another scenario where flaky failures might be a real regression is when tests which never failed before starts failing flakily.
> 
> Can we fetch data from the flakiness dashboard and see if we’ve ever observed the flakiness on a given test? That would let us get out of the guessing game.

Yes, that's a good idea, and something we have been exploring lately. Now we have new results database (https://results.webkit.org <https://results.webkit.org/>) as well which have REST API support which EWS can use. Using data from results database, we can have significant improvements in efficiency and flakiness detection in EWS. We started with some ideas in https://bugs.webkit.org/show_bug.cgi?id=204368 <https://bugs.webkit.org/show_bug.cgi?id=204368>.

> 
> I understand some flakiness might be specific to EWS but I’d imagine that’d be more of minority. If it is common, we could also make EWS bots periodically run full tests without patches and report results to flakiness dashboard so that we have a corpse of flaky teat failures on EWS bots as well.
> 
> - R. Niwa
> 
>> 3 дек. 2019 г., в 8:40 AM, Aakash Jain <aakash_jain at apple.com <mailto:aakash_jain at apple.com>> написал(а):
> 
>> 
>> Hi Everyone,
>> 
>> We have various layout-tests which are flaky (which sometimes pass and sometimes fail/crash/timeout). EWS needs to work despite these flaky tests, and need to be able to tell whether the patch being tested introduced any test failure or not.
>> 
>> In EWS, we have logic (same logic in both old and new EWS) on how to deal with flaky tests. The logic is roughly following:
>> 
>> - We apply the patch and build.
>> 
>> - Run layout-tests. During the test-run, we retry the failed tests. If those tests pass in retry, the run is considered to have no failing test (this retry part is same for EWS / build.webkit.org <http://build.webkit.org/> / manual-run and part of run-webkit-test script itself).
>> 
>> - If a test-run has some failures, EWS retry the test-run one more time (to check if those test failures were flaky). 
>> 
>> - Then we do one more test-run on clean-tree (without the patch), and compare the results of first run, second run, and clean tree run.
>> 
>> - After that we analyze the results from all three test-runs (which we call first_run, second_run and clean_tree_run). 
>> 
>> 
>> If there are different test failures in first and second runs (i.e.: flaky test failures), and the patch being tested did not introduce any new test failure, we retry the entire build (probably hoping that next time we will not see the flakiness). This retry can result in almost infinite loop, if someone commits a very flaky test (which is not rare), and would cause EWS to be almost stuck until the flakiness is fixed.
>> 
>> From https://trac.webkit.org/browser/webkit/trunk/Tools/Scripts/webkitpy/tool/bot/patchanalysistask.py#L352 <https://trac.webkit.org/browser/webkit/trunk/Tools/Scripts/webkitpy/tool/bot/patchanalysistask.py#L352>  ('Defer' means retrying the build).
>> '''
>> 350    # At this point we know that at least one test flaked, but no consistent failures
>> 351    # were introduced. This is a bit of a grey-zone.
>> 352    return False  # Defer patch
>> '''
>> 
>> 
>> Retrying the entire build, just because of few flaky tests seems excessive and inefficient. This doesn't help identify flaky tests, and only delays the EWS result. Instead, we should mark the build as SUCCESS (since the patch did not introduce any new consistent test failure).
>> 
>> In other EWS test-suites like API tests and JSC tests, we do not retry the entire build on flaky test failures, we ignore the flaky tests, and only focus on tests which failed consistently. We should do the similar thing for layout-tests as well.
>> 
>> I am going to make that change for layout tests in https://bugs.webkit.org/show_bug.cgi?id=204769 <https://bugs.webkit.org/show_bug.cgi?id=204769>. Please let me know if anyone has a different opinion.
>> 
>> Thanks
>> Aakash
> 
> 
> _______________________________________________
> webkit-dev mailing list
> webkit-dev at lists.webkit.org <mailto:webkit-dev at lists.webkit.org>
> https://lists.webkit.org/mailman/listinfo/webkit-dev <https://lists.webkit.org/mailman/listinfo/webkit-dev>
> -- 
> - R. Niwa

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/webkit-dev/attachments/20191205/09699ff3/attachment.htm>