Re: Assert failure on 'list_member_ptr(rel->joininfo, restrictinfo)'

On Mon, Dec 25, 2023 at 2:56 AM Ashutosh Bapat
<ashutosh.bapat.oss@gmail.com> wrote:
> On Sun, Dec 24, 2023 at 5:32 PM Alexander Korotkov <aekorotkov@gmail.com> wrote:
> >
> >
> > Thank you Ashutosh for your work on this matter.  With a large number of partitions, it definitely makes sense to reduce both Bitmapset's size as well as the number of Bitmapsets.
> >
> > I've checked the patchset [1] with your test suite to check the memory consumption.  The results are in the table below.
> >
> > query                             | no patch   | patch      | no self-join removal
> > ----------------------------------------------------------------------------------
> > 2-way join, non partitioned       | 14792      | 15208      | 29152
> > 2-way join, no partitionwise join | 19519576   | 19519576   | 19519576
> > 2-way join, partitionwise join    | 40851968   | 40851968   | 40851968
> > 3-way join, non partitioned       | 20632      | 21784      | 79376
> > 3-way join, no partitionwise join | 45227224   | 45227224   | 45227224
> > 3-way join, partitionwise join    | 151655144  | 151655144  | 151655144
> > 4-way join, non partitioned       | 25816      | 27736      | 209128
> > 4-way join, no partitionwise join | 83540712   | 83540712   | 83540712
> > 4-way join, partitionwise join    | 463960088  | 463960088  | 463960088
> > 5-way join, non partitioned       | 31000      | 33720      | 562552
> > 5-way join, no partitionwise join | 149284376  | 149284376  | 149284376
> > 5-way join, partitionwise join    | 1663896608 | 1663896608 | 1663896608
> >
> >
> > The most noticeable thing for me is that self-join removal doesn't work with partitioned tables.  I think this is the direction for future work on this subject.  In non-partitioned cases, patchset gives a small memory overhead.  However, the memory consumption is still much less than it is without the self-join removal.  So, removing the join still lowers memory consumption even if it copies some Bitmapsets.  Given that patchset [1] is required for the correctness of memory manipulations in Bitmapsets during join removals, I'm going to push it if there are no objections.
>
> I am missing the link between this work and the self join work. Can
> you please provide me relevant pointers?

This thread was started from the bug in self-join removal [1].  The
fix under consideration [2] makes replace_relid() leave the argument
unmodified.  I've used your test set [3] to check the memory overhead
of this solution.

Links.
1. https://www.postgresql.org/message-id/CAMbWs4_wJthNtYBL%2BSsebpgF-5L2r5zFFk6xYbS0A78GKOTFHw%40mail.gmail.com
2. https://www.postgresql.org/message-id/CAPpHfdtLgCryACcrmLv%3DKoq9rAB3%3Dtr5y9D84dGgvUhSCvjzjg%40mail.gmail.com
3. https://www.postgresql.org/message-id/CAExHW5stmOUobE55pMt83r8UxvfCph%2BPvo5dNpdrVCsBgXEzDQ%40mail.gmail.com

------
Regards,
Alexander Korotkov