At Thu, 13 May 2021 13:26:37 +0900 (JST), Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote in
> At Thu, 13 May 2021 12:11:12 +0900 (JST), Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote in
> > As the result, even if we take the DLL approach, still not need to
> > split out the query-id provider part. By the following config:
> >
> > > query_id_provider = 'pg_stat_statements'
> >
> > the core can obtain the entrypoint of, say, "_PG_calculate_query_id"
> > to call it. And it can be of another module.
> >
> > It seems like the only problem doing that is we don't have a means to
> > call per-process intializer for a preload libralies.
>
> So this is a crude PoC of that.
>
> pg_stat_statemnts defines its own query-id provider function in
> pg_stat_statements which calls in-core DefaultJumbeQuery (end emits a
> log line).
>
> If server started with query_id_provider='auto' and pg_stat_statements
> is not loaded, pg_stat_activity.query_id is null.
>
> If query_id_provider='auto' and pg_stat_statements is loaded,
> pg_stat_activity.query_id is filled in with a query id.
>
> If query_id_provider='default' or 'pg_stat_statements' and
> pg_stat_statements is not loaded, pg_stat_activity.query_id is filled
> in with a query id.
>
> If query_id_provider='none' and pg_stat_statements is loaded, it
> complains as "query id provider is not available" and refuss to start.
>
> If showing the variable, it shows the real provider name instead of
> the setting in postgresql.conf.
The change contains needless things that tries to handle per-backend
change case, so it would be simpler assuming we don't want on-the-fly
change of provider (and I believe so since that change surely causes
inconsistency)
regards.
--
Kyotaro Horiguchi
NTT Open Source Software Center
