It looks like between 3.4.3 and 3.5, datetime_best_possible changed from using _PyTime_gettimeofday with a _PyTime_timeval to using _PyTime_AsTimeval with struct timeval. The difference is that _PyTime_timeval appears to have been defined with a proper time_t for secs, where struct timeval is defined (on Windows) with a long which is still 32 bits on Windows, so we reintroduced the Y2K38 problem.

The bug was introduced while trying to support #22117: "Rewrite pytime.h to work on nanoseconds" in reve93eeadef0c3: https://hg.python.org/cpython/rev/e93eeadef0c3