On Tue, 5 Oct 2021 at 14:21, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> Elvis Pranskevichus <elprans@gmail.com> writes:
> > On Monday, October 4, 2021 5:14:47 PM PDT David Rowley wrote:
> >>> It appears a combination of Merge Semi Join and Memoize in
> >>> PostgreSQL 14 produces incorrect results on a particular query.
>
> > It seems like a very particular set of stats is causing the plan there,
> > as running ANALYZE turns the query into a Nested Loop Semi Join.
>
> Surely this plan tree is broken on its face? The inner side of
> a mergejoin has to be able to support mark/restore, and nestloop
> doesn't (cf ExecSupportsMarkRestore, as well as assertions in
> nodeNestloop). It looks to me like somebody removed an essential
> plan-time check.
Couldn't that just be because it's a Merge Semi Join and there's not
really any need to rewind the inner side if all join clauses are merge
join clauses, aka:
if ((path->jpath.jointype == JOIN_SEMI ||
path->jpath.jointype == JOIN_ANTI ||
extra->inner_unique) &&
(list_length(path->jpath.joinrestrictinfo) ==
list_length(path->path_mergeclauses)))
path->skip_mark_restore = true;
?
David