I mentioned this on general a while ago.
I had the problem when I dumped my 7.0.3 db to upgrade to 7.1. I had to
modify the dump because there were some 60 seconds in there. It was
obvious in the code in backend/utils/adt/datetime that it was using
sprintf to do the formatting, and sprintf was taking the the float the
represented the seconds and rounding it.
select '2001-07-10 15:39:59.999'::timestamp;
?column?
---------------------------
2001-07-10 15:39:60.00-04
(1 row)
Thomas Lockhart wrote:
>
> > > Have you observed otherwise?
> > Yes. Specifically timestamps are dumped in a way that (1) they lose
> > percision (2) sometimes have 60 in the seconds field which prevents the
> > dump from being restored.
>
> The loss of precision for timestamp data stems from conservative
> attempts to get consistant behavior from the data type. It is certainly
> not entirely successful, but changes would have to solve some of these
> problems without introducing more.
>
> I've only seen the "60 seconds problem" with earlier Mandrake distros
> which combined normal compiler optimizations with a "fast math"
> optimization, against the apparent advice of the gcc developers. What
> kind of system are you on, and how did you build PostgreSQL?
>
> Regards.
>
> - Thomas
--
Joseph Shraibman
jks@selectacast.net
Increase signal to noise ratio. http://www.targabot.com
