On Fri, Feb 04, 2022 at 04:29:19PM -0500, Tom Lane wrote:
> > + A previous bug fix disabled building of extended statistics for
> > + old-style inheritance trees, but any existing statistics data was
> > + not removed, and that data would become more and more out-of-date
> > + over time. Adjust the planner to ignore such data. Extended
> > + statistics for the individual child tables are still built and used,
> > + however.
>
> > The issue here isn't that old stats were never updated. For inheritance, they
> > *were* updated with non-inherited stats (for SELECT FROM ONLY). But then
> > "SELECT FROM tbl*" used the stats anyway...
>
> I'm confused about this bit. Are we still building bogus stats for
> inheritance parents, or has that stopped?
To make a long story long:
- before 859b3003de, an ERROR occurred when a stats object was created on an
inheritance parent.
- To avoid the error, 859b3003de changed to no longer build "whole tree" stats
on the table heirarchy. Non-inheried stats were still collected.
- However, the stats were *also* applied to inherited queries (FROM tbl*).
36c4bc6 stops applying stats that shouldn't be applied (and doesn't change
their collection during ANALYZE).
20b9fa3 then changes to collect inherited stats on partitioned tables, since
they have no non-inherited stats, and since extended stats on partitioned
tables were intended to work since v10, and did work until 859b3003de stopped
collecting them.
In back branches, pg_statistic has inherited stats for partitioned tables, and
non-inherited stats otherwise.
--
Justin
