Thomas Munro <thomas.munro@gmail.com> writes:
> Before release 10, it was possible for it to use floating point
> storage instead of integers; I wonder if that could be a factor here.
I wondered about that briefly as well. It doesn't seem to fit the
facts though.
(1) integer timestamps have been the default since 8.4;
(2) if the OP were using float timestamps, he'd have to be working
with dates millions of years away from the epoch to lose any
substantial fraction of a second in precision. 2^53 seconds is
~285 million years if I've not miscounted. Even dropping the last
microsecond digit won't happen for dates within a couple hundred
years of AD 2000.
regards, tom lane