> On Fri, Oct 13, 2023 at 03:35:19PM +0200, Dmitry Dolgov wrote:
> > On Fri, Oct 13, 2023 at 05:07:00PM +0900, Michael Paquier wrote:
> > Now, it doesn't mean that this approach with the "powers" will never
> > happen, but based on the set of opinions I am gathering on this thread
> > I would suggest to rework the patch as follows:
> > - First implement an on/off switch that reduces the lists in IN and/or
> > ANY to one parameter. Simply.
> > - Second, refactor the powers routine.
> > - Third, extend the on/off switch, or just implement a threshold with
> > a second switch.
>
> Well, if it will help move this patch forward, why not. To clarify, I'm
> going to split the current implementation into three patches, one for
> each point you've mentioned.
Here is what I had mind. The first patch implements the basic notion of
merging, and I guess everyone agrees on its usefulness. The second and
third implement merging into groups power of 10, which I still find
useful as well. The last one adds a lower threshold for merging on top
of that. My intentions are to get the first one in, ideally I would love
to see the second and third applied as well.
> > When it comes to my opinion, I am not seeing any objections to the
> > feature as a whole, and I'm OK with the first step. I'm also OK to
> > keep the door open for more improvements in controlling how these
> > IN/ANY lists show up, but there could be more than just the number of
> > items as parameter (say the query size, different behaviors depending
> > on the number of clauses in queries, subquery context or CTEs/WITH,
> > etc. just to name a few things coming in mind).
>
> Interesting point, but now it's my turn to have troubles imagining the
> case, where list representation could be controlled depending on
> something else than the number of elements in it. Do you have any
> specific example in mind?
In the current patch version I didn't add anything yet to address the
question of having more parameters to tune constants merging. The main
obstacle as I see it is that the information for that has to be
collected when jumbling various query nodes. Anything except information
about the ArrayExpr itself would have to be acquired when jumbling some
other parts of the query, not directly related to the ArrayExpr. It
seems to me this interdependency between otherwise unrelated nodes
outweigh the value it brings, and would require some more elaborate (and
more invasive for the purpose of this patch) mechanism to implement.