I am speculating that the rerun logic issue where we saw the hang may be more of a serial test runner issue. Would you mind if I re-enabled the arm/aarch64 inclusion in the rerun logic now that I made a change based on this speculation? It would be a relatively quick way to check if the serial test runner is the issue, since now the rerun logic will not use the serial test runner but rather the normal parallel runner with a single worker (so, the same intent but expressed another way, using the test runners we use all the time). If we still hit the issue, it is unrelated to the serial test runner strategy. If we don’t see the issue, then: (1) great, we have a solution, and (2) I know I need to look into the serial test runner strategy which may need some updates for recent changes.
How does that sound? If I enable it and it times out, I’ll just revert the change and we’ll go back to normal. (And I’ll know more about the issue, albeit with more investigation necessary). If it works just fine, we’ll leave it this way (and I’ll know I need to look into the serial test runner).
(And, as an aside, I may just nuke the serial test runner anyway, since we can do it with a multi-worker runner with a single worker just fine, and reduce the code size — I really don’t see a good reason to keep the serial test runner strategy anymore except for a purely theoretical sense).
Yes, you could use android builder to run that experiment.
Please watch test 7 of this builder after your change goes in(Another test for aarch64 which was previously timed out has been disabled for offline debugging of other unrelated problems).
Excellent. I’ll try this in the afternoon. I need to run out now but I’ll check in what we discussed later on when I get back.
I just put this change in that reverted the aarch64 and arm removal from test-rerun eligibility:
I’ll watch this builder now and see what happens. If it hangs on reruns, I’ll revert r255935.
Great, thanks! Yes I was just checking after dinner and saw that it looks to be rerunning just fine.
Okay I am likely to just strip out the serial test runner strategy at this point. One less thing to keep running.