Re: pg_dump is broken for partition tablespaces

On 2019-Mar-06, Andres Freund wrote:


> > I also find it far from clear that:
> >     <listitem>
> >      <para>
> >       The <replaceable class="parameter">tablespace_name</replaceable> is the name
> >       of the tablespace in which the new table is to be created.
> >       If not specified,
> >       <xref linkend="guc-default-tablespace"/> is consulted, or
> >       <xref linkend="guc-temp-tablespaces"/> if the table is temporary.  For
> >       partitioned tables, since no storage is required for the table itself,
> >       the tablespace specified here only serves to mark the default tablespace
> >       for any newly created partitions when no other tablespace is explicitly
> >       specified.
> >      </para>
> >     </listitem>
> > is handled correctly. The above says that the *specified* tablespaces -
> > which seems to exclude the default tablespace - is what's going to
> > determine what partitions use as their default tablespace. But in fact
> > that's not true, the partitioned table's pg_class.retablespace is set to
> > what default_tablespaces was at the time of the creation.
> 
> I still think the feature as is doesn't seem to have very well defined
> behaviour.

I have just started looking into this issue.  I'm not sure yet what's
the best fix; maybe for the specific case of partitioned tables and
indexes we should deviate from the ages-old behavior of storing zero
tablespace, if the tablespace is specified as the default one.  But I
haven't written the code yet.

In the meantime, here's David's patch, rebased to current master and
with the pg_upgrade and pg_dump tests fixed to match the new partition
creation behavior.

> > If we instead did:
> > 
> > CREATE TABLE public.listp1 (a integer
> > );
> > 
> > ALTER TABLE public.list1 ATTACH PARTITION public.listp FOR VALUES IN (1);
> 
> Isn't that a bit more expensive, because now the table needs to be
> scanned for maching the value? That's probably neglegible though, given
> it'd probably always empty.

I think it's always empty.  In the standard case, there are two
transactions rather than one, so yeah it's a little bit more expensive.
Maybe we should make this conditional on there actually being an
important tablespace distinction to preserve.

-- 
Álvaro Herrera                https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services