From: Tom Lane <tgl@sss.pgh.pa.us>
To: Steve Crawford <scrawford@pinpointresearch.com>
cc: PG-General Mailing List <pgsql-general@postgresql.org>
Subject: Re: Unexpected date conversion results
In-reply-to: <CAEfWYyzvM4XcfKunhvT1_xs_9rGnbXbRvnn_znQD4-Wg-aA5Vg@mail.gmail.com>
References: <CAEfWYyzvM4XcfKunhvT1_xs_9rGnbXbRvnn_znQD4-Wg-aA5Vg@mail.gmail.com>
Comments: In-reply-to Steve Crawford <scrawford@pinpointresearch.com>
	message dated "Fri, 21 Nov 2025 16:09:01 -0800"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <2554464.1763772684.1@sss.pgh.pa.us>
Content-Transfer-Encoding: quoted-printable
Date: Fri, 21 Nov 2025 19:51:24 -0500
Message-ID: <2554465.1763772684@sss.pgh.pa.us>
Archived-At: <https://www.postgresql.org/message-id/2554465.1763772684%40sss.pgh.pa.us>
Precedence: bulk

Steve Crawford <scrawford@pinpointresearch.com> writes:
> However, extracting the epoch from current_date returns 4pm the prior da=
y
> (i.e. 2025-11-21 00:00:00-00), in other words midnight 2025-11-21 UTC wh=
ich
> seems to be inconsistent behavior:

> steve=3D> select to_timestamp(extract(epoch from current_date));
>       to_timestamp
> ------------------------
>  2025-11-20 16:00:00-08

The reason this is misbehaving is that there are two versions of
extract(), one for timestamp-with-timezone input and one for
timestamp-without-timezone input.  The latter applies no
timezone correction, so it won't give true Unix-epoch results
unless you are in UTC zone to start with.

By default, a date will be promoted to timestamp-without-timezone not
timestamp-with-timezone, so the above doesn't give what you want.
It'd work better with a cast to force the right interpretation:

regression=3D# set timezone =3D 'America/Los_Angeles';
SET
regression=3D# select to_timestamp(extract(epoch from current_date));
      to_timestamp      =

------------------------
 2025-11-20 16:00:00-08
(1 row)

regression=3D# select to_timestamp(extract(epoch from current_date::timest=
amptz)); =

      to_timestamp      =

------------------------
 2025-11-21 00:00:00-08
(1 row)

> There was a time, like version 9-dot-something, when the above queries
> performed as expected returning midnight in the current time zone but I
> haven't been able to find a change document indicating this as an expect=
ed
> change.

A bit of experimenting says the current behavior dates to 9.2.
I've not checked the release notes to see if it was documented,
but in any case it's stood for long enough now that I doubt
we'd change it.

			regards, tom lane