Re: Asymmetric partition-wise JOIN

On Tue, Oct 17, 2023 at 2:05 PM Andrei Lepikhov
<a(dot)lepikhov(at)postgrespro(dot)ru> wrote:
>
> On 16/10/2023 23:21, Ashutosh Bapat wrote:
> > On Mon, Oct 16, 2023 at 10:24 AM Andrei Lepikhov
> > Whenever I visited this idea, I hit one issue prominently - how would
> > we differentiate different scans of the non-partitioned relation.
> > Normally we do that using different Relids but in this case we
> > wouldn't be able to know the number of such relations involved in the
> > query unless we start planning such a join. It's late to add new base
> > relations and assign them new Relids. Of course I haven't thought hard
> > about it. I haven't looked at the patch to see whether this problem is
> > solved and how.
> >
> I'm curious, which type of problems do you afraid here? Why we need a
> range table entry for each scan of non-partitioned relation?
>

Not RTE but RelOptInfo.

Using the same example as Alexander Korotkov, let's say A is the
nonpartitioned table and P is partitioned table with partitions P1,
P2, ... Pn. The partitionwise join would need to compute AP1, AP2, ...
APn. Each of these joins may have different properties and thus will
require creating paths. In order to save these paths, we need
RelOptInfos which are indentified by relids. Let's assume that the
relids of these join RelOptInfos are created by union of relid of A
and relid of Px (the partition being joined). This is notionally
misleading but doable.

But the clauses of A parameterized by P will produce different
translations for each of the partitions. I think we will need
different RelOptInfos (for A) to store these translations.

The relid is also used to track the scans at executor level. Since we
have so many scans on A, each may be using different plan, we will
need different ids for those.

But if you have developed a way to use a single RelOptInfo of A to do
all this, may be we don't need all this. Will take a look at your next
version of patch.

--
Best Wishes,
Ashutosh Bapat