I wrote:
> My feeling is we ought to preserve the old behavior here, which would
> involve making JULIAN_MAXYEAR_FOR_TIMESTAMPS format-dependent and
> adjusting the float values for the two derived constants; not much of a
> problem code-wise. I think though that it would break quite a number of
> the proposed new regression tests for the float case. TBH, I thought
> the number of added test cases was rather excessive anyway, so I wouldn't
> have a problem with just leaving out whichever ones don't pass with both
> build options.
Actually, it seems a lot of the provided test cases fail for float
timestamps anyway; there's an assumption that294276-12-31 23:59:59.999999294277-01-01 00:00:00.000000
are distinct timestamps, which they are not in float mode.
regards, tom lane