Re: Allowing NOT IN to use ANTI joins

On 24 June 2014 11:32, David Rowley <dgrowleyml(at)gmail(dot)com> wrote:

> So if anyone can point me in the right direction then that would be
> really useful.

Many things can be added simply, but most things can't. It seems we
just don't have that information. If we did, Tom would have done this
already.

> On a more positive or even slightly exciting note I think I've managed to
> devise a way that ANTI JOINS can be used for NOT IN much more often. It
> seems that find_nonnullable_vars will analyse a quals list to find
> expressions that mean that the var cannot be NULL. This means we can perform
> ANTI JOINS for NOT IN with queries like:
>
> SELECT * FROM a WHERE id NOT IN(SELECT nullable_col FROM b WHERE
> nullable_col = 1);
> or
> SELECT * FROM a WHERE id NOT IN(SELECT nullable_col FROM b WHERE
> nullable_col IS NOT NULL);
>
> (The attached patch implements this)
>
> the nullable_col =1 will mean that nullable_col cannot be NULL, so the ANTI
> JOIN can be performed safely. I think this combined with the NOT NULL check
> will cover probably just about all valid uses of NOT IN with a subquery...
> unless of course I've assumed something wrongly about find_nonnullable_vars.
> I just need the correct RangeTblEntry in order to determine if the
> TargetEntry is from an out join.

This is the better way to go. It's much better to have explicit proof
its not null than a possibly long chain of metadata that might be
buggy.

> The attached patch is a broken implemention that still needs the lookup code
> fixed to reference the correct RTE. The failing regression tests show where
> the problems lie.
>
> Any help on this would be really appreciated.

I'd suggest we just drop the targetlist approach completely.

--
Simon Riggs http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services