Re: POC, WIP: OR-clause support for indexes

On Tue, Mar 19, 2024 at 7:17 AM Andrei Lepikhov
<a(dot)lepikhov(at)postgrespro(dot)ru> wrote:
> On 14/3/2024 16:31, Alexander Korotkov wrote:
> > On Wed, Mar 13, 2024 at 2:16 PM Andrei Lepikhov
> > As you can see this case is not related to partial indexes. Just no
> > index selective for the whole query. However, splitting scan by the OR
> > qual lets use a combination of two selective indexes.
> I have rewritten the 0002-* patch according to your concern. A candidate
> and some thoughts are attached.
> As I see, we have a problem here: expanding each array and trying to
> apply an element to each index can result in a lengthy planning stage.
> Also, an index scan with the SAOP may potentially be more effective than
> with the list of OR clauses.
> Originally, the transformation's purpose was to reduce a query's
> complexity and the number of optimization ways to speed up planning and
> (sometimes) execution. Here, we reduce planning complexity only in the
> case of an array size larger than MAX_SAOP_ARRAY_SIZE.
> Maybe we can fall back to the previous version of the second patch,
> keeping in mind that someone who wants to get maximum profit from the
> BitmapOr scan of multiple indexes can just disable this optimization,
> enabling deep search of the most optimal scanning way?
> As a compromise solution, I propose adding one more option to the
> previous version: if an element doesn't fit any partial index, try to
> cover it with a plain index.
> In this case, we still do not guarantee the most optimal fit of elements
> to the set of indexes, but we speed up planning. Does that make sense?

Thank you for your research Andrei. Now things get more clear on the
advantages and disadvantages of this transformation.

The current patch has a boolean guc enable_or_transformation.
However, when we have just a few ORs to be transformated, then we
should get less performance gain from the transformation and higher
chances to lose a good bitmap scan plan from that. When there is a
huge list of ORs to be transformed, then the performance gain is
greater and it is less likely we could lose a good bitmap scan plan.

What do you think about introducing a GUC threshold value: the minimum
size of list to do OR-to-ANY transformation?
min_list_or_transformation or something.

------
Regards,
Alexander Korotkov