[ trivia warning ]
I wrote:
> We don't make any great effort to expose that though. It looks like
> the closest value that timestamptzin makes different from zero is
> regression=# select extract(epoch from '1999-12-31 19:00:00.00000000001-05' - '1999-12-31 19:00:00-05'::timestamptz)
;
> date_part
> ----------------------
> 1.45519152283669e-11
> (1 row)
Actually, it looks like the precision is being limited by the rotation
from EST zone. In GMT zone I can do this:
regression=# select extract(epoch from '2000-01-01
00:00:00.0000000000000000000000000000000000000000000000000000000000000000001'::timestamptz- '2000-01-01 00:00:00');
date_part
-----------
1e-67
(1 row)
and it could go a lot smaller except there's an arbitrary limit on
the length of input string that timestamptzin will take.
If float timestamps weren't deprecated it might be worth trying to make
this behave less surprisingly.
regards, tom lane