Re: BUG #18652: Planner can not find pathkey item to sort for query with expression and expression index

On 11.10.2024 11:02, Alena Rybakina wrote:
> On 10.10.2024 14:57, Richard Guo wrote:
>> On Thu, Oct 10, 2024 at 12:34 PM Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>>> So it's specific to MergeAppend and it's been wrong from day zero.
>>> That makes me think it's probably not find_computable_ec_member's
>>> fault directly.  Fixing it there might be the most expedient answer,
>>> but I feel like first we should drill down a bit further to understand
>>> the root problem.  I'm too tired to do more tonight though.
>> Usually a relation's targetlist should include only Vars and PHVs
>> during this phase.  I think this may be the rationale behind
>> find_computable_ec_member matching only Vars and quasi-Vars. However,
>> for a child rel, when we copy the parent's targetlist with appropriate
>> substitution, we may generate arbitrary expressions, such as an OpExpr
>> for 'i + 1' in this case.
>>
>> We have a note in the comments in set_append_rel_size saying that
>>
>> * Code that might be looking at an appendrel child must cope with
>> * such.
>>
>> It seems to me that find_computable_ec_member does not get this memo.
>>
>> Alternatively, can we wrap non-var expressions in a childrel's
>> targetlist into PHVs, so that we do not need special handling for an
>> appendrel child?  I have not tried this though.
>>
> To be honest, looking at the MergeAppend processing plan, it’s not
> entirely clear to me what the functional difference is between
> set_rel_pathlist and
>
> find_computable_ec_member. Both prepare pathkeys , but what the first
> function doesn't do is what the other one needs to do?Could you
> explain it to me?
>
Sorry, I meant add_child_rel_equivalences function I've just understood
it. Sorry for noise

--
Regards,
Alena Rybakina
Postgres Professional