Terry: " The Travis retest (for 3.7 backport) just failed with the same errors.  Why can't we disable this tests.  It now takes hours to do a merge with two backports."

Usually, I prefer to try to understand a bug before disabling, even temporarely, a test. Most of the time, a test failure means a regression. If we disable a test, we reduce the code coverage and more generally reduces the quality of the code.

I would prefer to only disable the test if nobody understand the issue and the issue blocks the workflow longer than a few days. This week, there was a CPython sprint during Pycon US, so yeah, it was annoying.