I suspect changes for this issue may be creating test_io failures on my windows builders, most notably my x86 Windows 7 builder where test_io has been failing both attempts on almost all builds.  It fails with a lock failure during interpreter shutdown, and commit ef4ac967 appears the most plausible commit out of the set introduced at the first failing build on Feb 24.

See https://buildbot.python.org/all/#/builders/58/builds/1977 for the first failure.  test_io has failed both attempts on all but 3 of the subsequent 16 tests of the 3.x branch.

It might be worth noting that this builder is slow, so if there are timeouts involved or a race condition of any sort it might be triggering it more readily than other builders.  I do, however, see the same failures - albeit less frequently and not usually on both tries - on the Win8 and Win10 builders.

For what it's worth one other oddity is that while having test_multiprocessing* failures are relatively common on the Win7 builder during the first round of tests due to exceeding the timeout, they usually pass on the retry, but over this same time frame have begun failing - not due to timeout - even on the second attempt, which is unusual.  It might be coincidental but similar failures are also showing up sporadically on the Win8/Win10 builders where such failures are not common at all.