On Wed, Apr 29, 2015 at 12:33 PM, Tomas Vondra
<tomas.vondra@2ndquadrant.com> wrote:
> On 04/29/15 18:26, Tom Lane wrote:
>> Tomas Vondra <tomas.vondra@2ndquadrant.com> writes:
>>> OK, let me explain the context a bit more. When working on the
>>> multivariate statistics patch, I need to choose which stats to use for
>>> estimating the clauses. I do that in clauselist_selectivity(), although
>>> there might be other places where to do that.
>>
>> Hm, well, clauselist_selectivity and clause_selectivity already contain
>> extensive special-casing for RestrictInfos, so I'm not sure that I see why
>> you're having difficulty working within that structure for this change.
>
> So let's I'm in the clause_selectivity(), and have AND or OR clause to
> estimate. I need to get varnos/varattnos for the clause(s). What should I
> do? I can't call pull_varnos() or pull_varattnos() because there might be
> nested RestrictInfos, as demonstrated by the query I posted.
>
>> But there are basic reasons why expression_tree_walker should not try
>> to deal with RestrictInfos; the most obvious one being that it's not
>> clear whether it should descend into both the basic and OR-clause
>> subtrees. Also, if a node has expression_tree_walker support then it
>> should logically have expression_tree_mutator support as well, but
>> that's got multiple issues. RestrictInfos are not supposed to be
>> copied once created; and the mutator couldn't detect whether their
>> derived fields are still valid.
>
> OK, I do understand that. So what about pull_varnos_walker and
> pull_varattnos_walker - what about teaching them about RestrictInfos?
Tom:
This patch has become part 1 of many under the "multivariate
statistics vNNN" family of threads, but I guess it would be helpful if
you could opine on the reasonable-ness of this approach. I don't want
to commit anything over your objections, but this kind of tailed off
without a conclusion.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company