John Naylor <john.nay...@enterprisedb.com> writes: > On Wed, Sep 1, 2021 at 5:32 AM Aleksander Alekseev <aleksan...@timescale.com> > wrote: >> By looking at timestamptz_bin() implementation I don't see why it >> should be STABLE. Its return value depends only on the input values. >> It doesn't look at the session parameters. timestamptz_in() and >> timestamptz_out() are STABLE, that's true, but this is no concern of >> timestamptz_bin().
> I'm not quite willing to bet the answer couldn't change if the timezone > changes, but it's possible I'm the one missing something. After playing with it for awhile, it seems like the behavior is indeed not TZ-dependent, but the real question is should it be? As an example, regression=# set timezone to 'America/New_York'; SET regression=# select date_bin('1 day', '2021-11-01 00:00 +00'::timestamptz, '2021-09-01 00:00 -04'::timestamptz); date_bin ------------------------ 2021-10-31 00:00:00-04 (1 row) regression=# select date_bin('1 day', '2021-11-10 00:00 +00'::timestamptz, '2021-09-01 00:00 -04'::timestamptz); date_bin ------------------------ 2021-11-08 23:00:00-05 (1 row) I see that these two answers are both exactly multiples of 24 hours away from the given origin. But if I'm binning on the basis of "days" or larger units, I would sort of expect to get local midnight, and I'm not getting that once I cross a DST boundary. If this is indeed the behavior we want, I concur with Aleksander that date_bin isn't TZ-sensitive and needn't be marked STABLE. regards, tom lane