I cannot reproduce the flaky test locally, so let's try things and see what sticks.
I think this was it: The 100 nanosecond timer was probably long enough for context cancellation to be picked up before the timer expired in our beefy machines, but not in the handicapped CI runners.
I cannot reproduce the flaky test locally, so let's try things and see what sticks.
I think this was it: The 100 nanosecond timer was probably long enough for context cancellation to be picked up before the timer expired in our beefy machines, but not in the handicapped CI runners.