"Zeugswetter Andreas SB SD" <ZeugswetterA@spardat.at> writes:
> Note also, that a typical SELECT only session would not advance
> CURRENT_TIMESTAMP at all in the typical "autocommit off" mode that
> the Spec is all about.
True, but the spec also says to default to serializable transaction
mode. So in a single-transaction session like you are picturing,
the successive SELECTs would all see a frozen snapshot of the database.
Freezing CURRENT_TIMESTAMP goes right along with that, and in fact makes
a lot of sense, because it tells you exactly what time your snapshot
of the database state was taken.
This line of thought opens another can of worms: should the behavior
of CURRENT_TIMESTAMP depend on serializable vs. read-committed mode?
Maybe SetQuerySnapshot is the routine that ought to capture the
"statement-start-time" timestamp value. We could define
CURRENT_TIMESTAMP as the time of the active database snapshot.
Or at least offer a fourth parameter to that parameterized now() to
return this time.
regards, tom lane