Wrong result (missing rows, wrong values) with materialization+semijoin+join_cache_hashed, join_cache_level>2, RIGHT JOIN, IN subquery

[Elena Stepanova]

The following query

SELECT *
FROM t1 RIGHT JOIN t2 ON ( a = c )
WHERE ( b, c ) IN
  ( SELECT b, c FROM t2 )

on the test data returns wrong result (less rows and in one row a NULL instead of a value) with join_cache_level>2, and correct result otherwise.

bzr version-info
revision-id: <email address hidden>
date: 2012-02-16 20:15:57 +0400
build-date: 2012-02-17 21:10:34 +0400
revno: 3424

EXPLAIN with join_cache_level=3 (wrong result):

id select_type table type possible_keys key key_len ref rows filtered Extra
1 PRIMARY t2 ALL b NULL NULL NULL 3 100.00
1 PRIMARY <subquery2> eq_ref distinct_key distinct_key 10 func,func 1 100.00
1 PRIMARY t1 hash_index a #hash#a:a 4:4 test.t2.c 4 50.00 Using where; Using join buffer (flat, BNLH join)
2 MATERIALIZED t2 ALL b NULL NULL NULL 3 100.00
Warnings:
Note 1003 select `test`.`t1`.`a` AS `a`,`test`.`t2`.`b` AS `b`,`test`.`t2`.`c` AS `c` from `test`.`t2` semi join (`test`.`t2`) left join `test`.`t1` on(((`test`.`t1`.`a` = `test`.`t2`.`c`) and (`test`.`t2`.`c` is not null))) where 1

EXPLAIN with join_cache_level=2 (correct result):

d select_type table type possible_keys key key_len ref rows filtered Extra
1 PRIMARY t2 ALL b NULL NULL NULL 3 100.00
1 PRIMARY <subquery2> eq_ref distinct_key distinct_key 10 func,func 1 100.00
1 PRIMARY t1 ref a a 4 test.t2.c 2 100.00 Using where; Using index
2 MATERIALIZED t2 ALL b NULL NULL NULL 3 100.00
Warnings:
Note 1003 select `test`.`t1`.`a` AS `a`,`test`.`t2`.`b` AS `b`,`test`.`t2`.`c` AS `c` from `test`.`t2` semi join (`test`.`t2`) left join `test`.`t1` on(((`test`.`t1`.`a` = `test`.`t2`.`c`) and (`test`.`t2`.`c` is not null))) where 1

Minimal optimizer_switch: materialization=on,semijoin=on,join_cache_hashed=on
(and join_cache_level>=3)
Full optimizer_switch: index_merge=on,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on,index_merge_sort_intersection=off,index_condition_pushdown=on,derived_merge=on,derived_with_keys=on,firstmatch=on,loosescan=on,materialization=on,in_to_exists=on,semijoin=on,partial_match_rowid_merge=on,partial_match_table_scan=on,subquery_cache=on,mrr=off,mrr_cost_based=off,mrr_sort_keys=off,outer_join_with_cache=on,semijoin_with_cache=on,join_cache_incremental=on,join_cache_hashed=on,join_cache_bka=on,optimize_join_buffer_size=off,table_elimination=on

Test case:

SET optimizer_switch = 'materialization=on,semijoin=on,join_cache_hashed=on';
# join_cache_level >= 3
SET SESSION join_cache_level = 3;

CREATE TABLE t1 ( a VARCHAR(1), KEY(a) );
INSERT INTO t1 VALUES ('v'),('v'),('s'),('j');

CREATE TABLE t2 ( b VARCHAR(1), c VARCHAR(1), KEY(b) );
INSERT INTO t2 VALUES ('v','v'),('w','w'),('t','t');

SELECT *
FROM t1 RIGHT JOIN t2 ON ( a = c )
WHERE ( b, c ) IN
  ( SELECT b, c FROM t2 );

# End of test case

# Expected result:
# a b c
# -----------------
# v v v
# v v v
# NULL w w
# NULL t t

# Actual result:
# a b c
# -----------------
# NULL v v
# NULL w w
# NULL t t

[Igor Babaev (igorb-seattle)]

The bug is reproducible with LEFT JOIN as well.

With Elena's settings and
set join_cache_level = 6
I get the following execution plan:

MariaDB [test]> EXPLAIN EXTENDED SELECT * FROM t2 LEFT JOIN t1 ON ( a = c ) WHERE ( b, c ) IN ( SELECT b, c FROM t2 t );
+----+--------------+-------------+--------+---------------+--------------+---------+-----------+------+----------+--------------------------+
| id | select_type | table | type | possible_keys | key | key_len | ref | rows | filtered | Extra |
+----+--------------+-------------+--------+---------------+--------------+---------+-----------+------+----------+--------------------------+
| 1 | PRIMARY | t2 | ALL | b | NULL | NULL | NULL | 3 | 100.00 | |
| 1 | PRIMARY | <subquery2> | eq_ref | distinct_key | distinct_key | 10 | func,func | 1 | 100.00 | |
| 1 | PRIMARY | t1 | ref | a | a | 4 | test.t.c | 2 | 100.00 | Using where; Using index |
| 2 | MATERIALIZED | t | ALL | b | NULL | NULL | NULL | 3 | 100.00 | |
+----+--------------+-------------+--------+---------------+--------------+---------+-----------+------+----------+--------------------------+

I get a wrong result set when executing by this plan.

If I use
set join_cache_level = 0;
I get the execution plan:

MariaDB [test]> EXPLAIN EXTENDED SELECT * FROM t2 LEFT JOIN t1 ON ( a = c ) WHERE ( b, c ) IN ( SELECT b, c FROM t2 t );
+----+--------------+-------------+--------+---------------+--------------+---------+-----------+------+----------+--------------------------+
| id | select_type | table | type | possible_keys | key | key_len | ref | rows | filtered | Extra |
+----+--------------+-------------+--------+---------------+--------------+---------+-----------+------+----------+--------------------------+
| 1 | PRIMARY | t2 | ALL | b | NULL | NULL | NULL | 3 | 100.00 | |
| 1 | PRIMARY | <subquery2> | eq_ref | distinct_key | distinct_key | 10 | func,func | 1 | 100.00 | |
| 1 | PRIMARY | t1 | ref | a | a | 4 | test.t2.c | 2 | 100.00 | Using where; Using index |
| 2 | MATERIALIZED | t | ALL | b | NULL | NULL | NULL | 3 | 100.00 | |
+----+--------------+-------------+--------+---------------+--------------+---------+-----------+------+----------+--------------------------+
4 rows in set, 1 warning (0.00 sec)

I get the right result set when executing by this plan.

The plans differ only in line 3:
the first plan uses test.t.c to build the key to access table t1,
while the second plan uses test.t2.c for this purpose.

[Igor Babaev (igorb-seattle)]

The first plan shows that the key is copied from table t, rather then from the materialized table. The last record in table t is ('t','t'). So the key to access table t1 always is built over this record. This explains why we have a wrong result for the first plan.

[Elena Stepanova (elenst)]

Fix released with 5.3.5-ga.