On 2019-Aug-14, Etsuro Fujita wrote:
> On Tue, Aug 13, 2019 at 11:01 PM Alvaro Herrera
> <alvherre@2ndquadrant.com> wrote:
> > On the subject of FDW support: I did look into supporting that before
> > submitting this. I think it's not academically difficult: just have the
> > FDW's acquire_sample_rows callback invoke the update_param functions
> > once in a while. Sadly, in practical terms it looks like postgres_fdw
> > is quite stupid about ANALYZE (it scans the whole table??) so doing
> > something that's actually useful may not be so easy. At least, we know
> > the total relation size and maybe we can add the ctid column to the
> > cursor in postgresAcquireSampleRowsFunc so that we have a current block
> > number to report (becing careful about synchronized seqscans).
>
> I don't follow this thread fully, so I might miss something, but I
> don't think that's fully applicable, because foreign tables managed by
> postgres_fdw can be eg, views on the remote side.
Oh, hmm, well I guess that covers the tables and matviews then ... I'm
not sure there's a good way to cover foreign tables that are views or
other stuff. Maybe that'll have to do. But at least it covers more
cases than none.
> > I do wonder why doesn't postgres_fdw use TABLESAMPLE.
>
> Yeah, that's really what I'm thinking for PG13; but I think we would
> still need to scan the whole table in some cases (eg, when the foreign
> table is a view on the remote side), because the TABLESAMLE clause can
> only be applied to regular tables and materialized views.
Sure.
> > I did not look at other FDWs at all, mind.
>
> IIUC, oracle_fdw already uses the SAMPLE BLOCK clause for that. Right?
Yeah, it does that, I checked precisely oracle_fdw this morning.
--
Álvaro Herrera https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services