Jeff Davis <pgsql@j-davis.com> writes:
> Introduced in commit 4318daecc959886d001a6e79c6ea853e8b1dfb4b.
> Attached regression failure and patch. It looks like we're no longer
> running floating point timestamps on the buildfarm?
Evidently :-(. Who wants to crank up an animal doing that?
> To make the output for float/integer timestamps the same, we either need
> to remove the fractional seconds or reduce the number of hours.
> It looks like the test is for maximum output length of an interval, so
> it seems better to reduce the number of hours to one million, which is
> what my patch does. It still loses 3 characters of output length, which
> is a bit unfortunate given the purpose of the test, but it's better than
> losing 7. I don't have a better idea.
Hm. As the test stands, it requires a float-timestamps implementation
to store a value of 3600000610.000001 seconds, which is 16 decimal digits
or about one more than can reliably be extracted from an IEEE float8.
At least for IEEE machines, I think we could remove just one digit
and it would work (or if not, that would say that some additional thought
needs to be put into the rounding behavior in interval_out).
However, removing *any* digits seems like it mostly defeats the point of
the test. Maybe we should just lose the test?
A different solution is to add a variant expected-output file, though
I'm not terribly thrilled with that answer.
regards, tom lane
