Re: Oversight in reparameterize_path_by_child leading to executor crash

Well, the way it's working is that for child rels all parameterized
paths arriving at try_nestloop_path (or try_partial_nestloop_path) are
parameterized by top parents at first.  Then our current code (without
this patch) applies reparameterize_path_by_child to the inner path
before calling create_nestloop_path().  So in create_nestloop_path() the
inner path is parameterized by child rel.  That's why we check against
outer rel itself not its top parent there.

With this patch, the reparameterize_path_by_child work is postponed
until createplan time, so in create_nestloop_path() the inner path is
still parameterized by top parent.  So we have to check against the top
parent of outer rel.

I think the query shown upthread is sufficient to verify this theory.

regression=# explain (costs off)
select * from prt1 t1 left join lateral
    (select t1.a as t1a, t2.* from prt1 t2) s on t1.a = s.a;
                       QUERY PLAN
---------------------------------------------------------
 Append
   ->  Nested Loop Left Join
         ->  Seq Scan on prt1_p1 t1_1
         ->  Index Scan using iprt1_p1_a on prt1_p1 t2_1
               Index Cond: (a = t1_1.a)
   ->  Nested Loop Left Join
         ->  Seq Scan on prt1_p2 t1_2
         ->  Index Scan using iprt1_p2_a on prt1_p2 t2_2
               Index Cond: (a = t1_2.a)
   ->  Nested Loop Left Join
         ->  Seq Scan on prt1_p3 t1_3
         ->  Index Scan using iprt1_p3_a on prt1_p3 t2_3
               Index Cond: (a = t1_3.a)
(13 rows)

 

Done here https://commitfest.postgresql.org/44/4477/

Thanks
Richard