On Wed, Sep 26, 2018 at 02:38:25PM +0200, Peter Eisentraut wrote:
> On 22/09/2018 00:35, Bruce Momjian wrote:
> > I have always thought of clock/statement/transation as decreasing levels
> > of time precision, and it might be odd to change that. I don't think we
> > want to change the behavior of statement_timestamp() in procedures, so
> > that kind of requires us not to change transaction_timestamp() inside of
> > procedures.
> >
> > However, no change in behavior causes the problem that if you have a
> > transaction block using transaction_timestamp() or CURRENT_TIMESTAMP,
> > and you move it into a procedure, the behavior of those functions will
> > change, but this was always true of moving statement_timestamp() into a
> > function, and I have never heard a complaint about that.
>
> Right. Statement timestamp actually means the timestamp of the
> top-level statement, because it's set where the protocol interaction
> happens. The transaction timestamp is implemented as the statement
> timestamp when the transaction starts, but for intra-procedural
> transactions, the statement timestamp does not advance, so the
> transaction timestamp doesn't change.
>
> Note that this also affects the xact_start column in pg_stat_activity.
>
> We could certainly address this by adding three or four or five new
> timestamps that cover all these varieties. But perhaps it's worth
> asking what these timestamps are useful for and which ones we really need.
Frankly, we might be fine with just documenting it and see if anyone
complains. I think the top-level statement part is obvious, but I am
not sure the top-level transaction part is. This is because it is clear
the top-level statement causes all the lower statements underneath it,
but the top-level transaction doesn't control all the transactions under
it in the same way.
--
Bruce Momjian <bruce@momjian.us> http://momjian.us
EnterpriseDB http://enterprisedb.com
+ As you are, so once was I. As I am, so you will be. +
+ Ancient Roman grave inscription +