Re: Problems around compute_query_id

On Mon, Apr 12, 2021 at 03:12:40PM +0900, Michael Paquier wrote:
> Hi all,
> 
> Starting a new thread as the one that has introduced compute_query_id
> is already long enough.
> 
> Fujii-san has reported on Twitter that enabling the computation of
> query IDs does not work properly with log_statement as the query ID is
> calculated at parse analyze time and the query is logged before that.
> As far as I can see, that's really a problem as any queries logged are
> combined with a query ID of 0, and log parsers would not really be
> able to use that, even if the information provided by for example
> log_duration gives the computed query ID prevent in pg_stat_activity.

I don't see any way around that.  The goal of log_statements is to log all
syntactically valid queries, including otherwise invalid queries.  For
instance, it would log

SELECT notacolumn;

and there's no way to compute a queryid in that case.  I think that
log_statements already causes some issues with log parsers.  At least pgbadger
recommends to avoid using that:

"Do not enable log_statement as its log format will not be parsed by pgBadger."

I think we should simply document that %Q is not compatible with
log_statements.

> While making the feature run on some test server, I have noticed that
> %Q would log some garbage query ID for autovacuum workers that's kept
> around.  That looks wrong.

I've not been able to reproduce it, do you have some hint on how to do it?

Maybe setting a zero queryid at the beginning of AutoVacWorkerMain() could fix
the problem?