> Victor Stinner wrote:
>> Is it a virtual machine or a physical machine? Was your Windows busy?
>> Did you run tests in parallel?
>
> Physical, not really other than the tests, and I ran with -j0 (on a machine with 2 single core CPUs).

Ah yes, I missed the -j0. I didn't try with -j0. Some tests should be
fixed to support shorter sleep, or default sleep should be increased.

> I'm not sure what I think of the TEST_SLEEP/TEST_SHORT_SLEEP scheme, but I do like the idea behind support.check_time_delta.

If we cannot agree on the whole patch, I will split it in two parts:
check_time_delta() and TEST_SLEEP/TEST_SHORT_SLEEP.

> 1) define a support.sleep function that multiplies the value given by some definable constant (default of 1) before passing the value to time.sleep.

Oh, I like this idea. It's closer to what is done currently.

> 2) define a support.sleep_until function, which would sleep for a given interval repeatedly until some condition is satisfied or a timeout is reached.

I'm not sure that such helper will provide useful debug info if the test fails.