[CALCITE-3170] ANTI join on conditions push down generates wrong plan

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.19.0
• Fix Version/s: 1.21.0
• Component/s: core
• Labels: pull-request-available
• Assignee: Danny Chen
• Reporter: godfrey he
• Created: 03/Jul/19 07:27
• Updated: 10/Sep/19 21:37
• Resolved: 09/Jul/19 01:31

Description

create table A (a int, b int);
create table B (c int, d int);
insert into A values (1, 1);
insert into A values (2, 2);

sql:

select * from A where A.a not in (select B.c from B where A.b > 1);

the equivalent logical tree:

LogicalProject(a=[$0], b=[$1])
+- LogicalProject(a=[$0], b=[$1])
   +- LogicalJoin(condition=[AND(OR(=($0, $3), IS NULL($0), IS NULL($3)), $2)], joinType=[anti])
      :- LogicalProject(a=[$0], b=[$1], $f2=[>($1, 1)])
      :  +- LogicalTableScan(table=[[A]])
      +- LogicalProject(c=[$0])
         +- LogicalFilter(condition=[true])
            +- LogicalTableScan(table=[[B]])

the correct result is: (1, 1), (2, 2)
while if the predicate ($2 in join condition) is pushed into left side, the result is (2, 2) which is incorrect.

Issue Links

• contains: CALCITE-3169 RelDecorrelator should return for SEMI/ANTI join decorrelate
• links to: GitHub Pull Request #1303