On Sun, 18 Sept 2022 at 07:00, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:
> > After a bit of trawling through the archives, I found it here:
> > https://www.postgresql.org/message-id/20180124162006.pmapfiznhgngwtjf%40alvherre.pgsql
> > I think there was insufficient discussion and you're probably right that
> > it wasn't the best fix. I don't object to finding another fix.
>
> FWIW, I don't see any big problem with what you did. We'd need to
> do something more like what David suggests if the planner ever has
> a reason to consider partitioned indexes. But as long as it does
> not, why expend the time to build data structures representing them?
Did you miss the report about left join removals not working with
partitioned tables due to lack of unique proofs? That seems like a
good enough reason to me.
> And we'd have to add code in quite a few places to ignore them,
> once they're in indexlist.
I think the same is true for "hypothetical" indexes. Maybe that would
be a good field to grep on to find the places that need to be
addressed.
David
