Обсуждение: ERROR: PlaceHolderVar found where not expected

Поиск
Список
Период
Сортировка

ERROR: PlaceHolderVar found where not expected

От
Richard Guo
Дата:
I came across an ERROR as $subject with query below.

create table t (a int, b int);
create statistics if not exists t_s0 (dependencies, ndistinct) on a, b from t;
insert into t values(1,1);
analyze t;

SELECT * FROM t LEFT JOIN (select true as c0) s0 ON true INNER JOIN (select true as c0) s1 ON s0.c0 INNER JOIN (select true as c0) s2 ON s0.c0;

This is due to that we use 0 flags for pull_var_clause in
dependency_is_compatible_expression, assuming that the 'clause_expr'
cannot contain Aggrefs, WindowFuncs or PlaceHolderVars.  This should be
an oversight as we can see that it's possible to find PHVs here.  We can
fix it by

--- a/src/backend/statistics/dependencies.c
+++ b/src/backend/statistics/dependencies.c
@@ -1316,7 +1316,7 @@ dependency_is_compatible_expression(Node *clause, Index relid, List *statlist, N
    if (IsA(clause_expr, RelabelType))
        clause_expr = (Node *) ((RelabelType *) clause_expr)->arg;

-   vars = pull_var_clause(clause_expr, 0);
+   vars = pull_var_clause(clause_expr, PVC_RECURSE_PLACEHOLDERS);

But I'm not sure if Aggrefs and WindowFuncs are possible to be found
here.

This issue can be seen on 14, 15 and master.

Thanks
Richard

Re: ERROR: PlaceHolderVar found where not expected

От
Tom Lane
Дата:
Richard Guo <guofenglinux@gmail.com> writes:
> I came across an ERROR as $subject with query below.
> ...
> This is due to that we use 0 flags for pull_var_clause in
> dependency_is_compatible_expression, assuming that the 'clause_expr'
> cannot contain Aggrefs, WindowFuncs or PlaceHolderVars.  This should be
> an oversight as we can see that it's possible to find PHVs here.

Nice catch.

> But I'm not sure if Aggrefs and WindowFuncs are possible to be found
> here.

WindowFuncs should be disallowed in qual clauses, so I think it's okay
to leave those flags out.  An Aggref could occur in a HAVING qual though.
I'm not sure if this code could get applied to HAVING ... but it's
not immediately clear that it can't.  I'd be inclined to add
PVC_RECURSE_AGGREGATES, as that seems more likely to be okay than not.

            regards, tom lane

Re: ERROR: PlaceHolderVar found where not expected

От
Tom Lane
Дата:
I wrote:
> WindowFuncs should be disallowed in qual clauses, so I think it's okay
> to leave those flags out.  An Aggref could occur in a HAVING qual though.
> I'm not sure if this code could get applied to HAVING ... but it's
> not immediately clear that it can't.  I'd be inclined to add
> PVC_RECURSE_AGGREGATES, as that seems more likely to be okay than not.

Actually, on closer look: why don't we just nuke that pull_var_clause
call entirely, along with the following loop inspecting its result?

The subsequent loop that looks for a matching StatisticExtInfo
expression will do just fine at rejecting any expression that
contains Vars of the wrong relation.  Maybe there is some performance
argument why the pull_var_clause precheck is worth the trouble,
but I'm inclined to bet that it's actually a net loss.

            regards, tom lane

Re: ERROR: PlaceHolderVar found where not expected

От
Richard Guo
Дата:

On Tue, Mar 14, 2023 at 11:39 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
Actually, on closer look: why don't we just nuke that pull_var_clause
call entirely, along with the following loop inspecting its result?

The subsequent loop that looks for a matching StatisticExtInfo
expression will do just fine at rejecting any expression that
contains Vars of the wrong relation.  Maybe there is some performance
argument why the pull_var_clause precheck is worth the trouble,
but I'm inclined to bet that it's actually a net loss.

Yes agreed.  The pull_var_clause precheck is not necessary considering
we would match the 'clause_expr' to StatisticExtInfo expressions with
equal().  +1 to remove it.

Thanks
Richard

Re: ERROR: PlaceHolderVar found where not expected

От
Tom Lane
Дата:
Richard Guo <guofenglinux@gmail.com> writes:
> Yes agreed.  The pull_var_clause precheck is not necessary considering
> we would match the 'clause_expr' to StatisticExtInfo expressions with
> equal().  +1 to remove it.

Done that way.

            regards, tom lane