relying upon the ordering of a dictionary is wrong, even with py3.6. Lots of people test for that by using an explicitly random dictionary ordering per test run via PYTHONHASHSEED (way more feasible than "shuffling test data" since not all dictionary use is that simplistic. sys.modules is a dict, for example, how do you "shuffle" that?) That apparently accidental feature is being removed.
it means code will suddenly break when you move it to an interpreter that does not have this implementation detail. This kind of breakage is really easy in the area class instrumentation libraries where order of class attributes affects something. Every non-cPython will implement this anyway, unwitting reliance upon it will be widespread, and there really won't be much of a "problem", other than they really should make this behavior official someday.
