[Webkit-unassigned] [Bug 191002] At least two dozen IntersectionObserver tests are flaky

bugzilla-daemon at webkit.org bugzilla-daemon at webkit.org
Mon Oct 29 06:47:01 PDT 2018


--- Comment #4 from Ali Juma <ajuma at chromium.org> ---
The root cause of this flakiness is that the test suite assumes HTMLEventLoop timing, and WebKit doesn't implement the HTMLEventLoop spec (and, when I brought this up at the Contributors' Meeting, there wasn't consensus that we should implement that spec). More specifically, the tests assume that if changes to style/layout are made in a rAF callback, then intersection observations will be updated before the next rAF callback, and events will fire within one setTimeout of that second callback. In HTMLEventLoop-based engines, this is straightforward to implement -- rAF is immediately followed by style/layout, and immediately followed by an update of intersection observations. After spending a while considering different approaches for WebKit, the most predictable (and closest to the spec) approach was to update intersection observations at the same time as we flush layers.

On iOS, flushing and rAF are both driven by didUpdate messages received from the UIProcess. None of the tests are flaky on iOS.

Mac is slightly different, but rAF still gets throttled when we fall behind on flushes. There's still a bit of flakiness, but still, the vast majority of tests are not flaky. And those that are flaky, are only flaky on Debug bots (where we're more likely to fall behind on flushes), not on Release bots.

So I suspect that on GTK, there's perhaps weaker coupling between rAF and flushes, and so on some runs we get into a state where flushing/drawing is slower for some reason, so the tests fail because the observations they're expecting haven't been computed yet.

In any case, given the above, I think that marking the entire suite as flaky in the global expectations file is too broad. So I'm going to move that over to the GTK expectations file.

You are receiving this mail because:
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/webkit-unassigned/attachments/20181029/73ed1bd7/attachment.html>

More information about the webkit-unassigned mailing list