Re: date bug (again)

Darcy Buskermolen <darcy(at)ok-connect(dot)com> writes:
> SELECT date_part('day', (('2000-10-1 0:00'::datetime + '1 month') + '-1
> day'))::int4 AS days_in_month;

> days_in_month
> --------------
> 30
> (1 row)

This is not a bug, or at least it's not entirely clear what the behavior
ought to be. The issue is what happens at a daylight-savings
transition. The results I get (US Eastern timezone) are

regression=# select '2000-10-1 0:00'::timestamp;
?column?
------------------------
2000-10-01 00:00:00-04
(1 row)

regression=# select '2000-10-1 0:00'::timestamp + '1 month';
?column?
------------------------
2000-10-31 23:00:00-05
(1 row)

See what's happening? You get a result that's exactly 31 days times
24 hours later, but that date_trunc()'s down to only 30 days. A
finer-grain example is

regression=# select '2000-10-29 0:00'::timestamp ;
?column?
------------------------
2000-10-29 00:00:00-04
(1 row)

regression=# select '2000-10-29 0:00'::timestamp + '1 day';
?column?
------------------------
2000-10-29 23:00:00-05
(1 row)

The real question is whether "+ 1 day" ought to mean "+ 24 hours"
or not, and if not what it *should* mean...

regards, tom lane