I believe it is related to scroll snapping in cc. +jaydasika, +ajuma
(Just my hunch. Feel free to re-label once we have more evidence.)

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/b3bb16c205d2e2cb91f7afa6aeada4047ff6542d

commit b3bb16c205d2e2cb91f7afa6aeada4047ff6542d
Author: Tien-Ren Chen <trchen@chromium.org>
Date: Fri Jul 21 00:03:37 2017

[Blink] Add layout test for composited scroller clip snapping

This test verifies composited scrollers have consistent clip rect
regardless of current scroll offset. The four boxes in the test have
different fractional translations. The test body and reference are
identical except for scroll offset.

BUG=736052

Change-Id: I97257b29e5abf579aa028d0c04463f4c427f8a58
Reviewed-on: https://chromium-review.googlesource.com/578741
Commit-Queue: Tien-Ren Chen <trchen@chromium.org>
Reviewed-by: Peter Mayo <petermayo@chromium.org>
Cr-Commit-Position: refs/heads/master@{#488500}
[modify] https://crrev.com/b3bb16c205d2e2cb91f7afa6aeada4047ff6542d/third_party/WebKit/LayoutTests/TestExpectations
[add] https://crrev.com/b3bb16c205d2e2cb91f7afa6aeada4047ff6542d/third_party/WebKit/LayoutTests/compositing/overflow/composited-scroll-with-fractional-translation-expected.html
[add] https://crrev.com/b3bb16c205d2e2cb91f7afa6aeada4047ff6542d/third_party/WebKit/LayoutTests/compositing/overflow/composited-scroll-with-fractional-translation.html

Re #2 I strongly suspect scroll snapping as well. We can probably verify just by disabling snapping in a local build. My hypothesis is that we either only snap some of the layers of the scroller (i.e. not the scrollbars) and/or we don't apply consistent snapping to all of the layers because we snap based on each layer's position which may lead some to move down/left and others to move up/right.

This test now passes, suggesting that the issue is resolved. Is that true, or is the reference now broken?

Ping

Nope, the above CL only added a test and marked it as [ Failure ]. Just checked and it is still failing as of today.

I'll spend some time to investigate it next week, though I'm pretty preoccupied with clip-path recently.

I think my PaintLayerClipper CL today might have made it start passing.

I'm leaving the team, thus re-assigning.

It still reproducible as of today. I also tried with BGPT it is affected too. I think it is somewhere in cc/viz quad snapping. It feels like the clipped quad_rect is computed in the local space that's prior to snapping, then the draw transform of the quad getting snapped later.

Just to confirm, we're talking about physical "pixel snapping" of scrolls, rather than "scroll snap" which is a new CSS feature unused by the above tests.

I can leave this in my queue for now but since it's low-severity and long-standing I'm unlikely to get to it in the near future.