Details
-
Bug
-
Status: Closed (View Workflow)
-
Major
-
Resolution: Fixed
-
5.3.12, 5.5(EOL), 10.0(EOL), 10.1(EOL)
-
None
-
mysql Ver 15.1 Distrib 10.0.18-MariaDB, for osx10.10 (x86_64) using readline 5.1
MariaDB installed with homebrew
-
5.5.45, 10.1.7-1
Description
Turning semijoin optimization on/off can affect number of rows returned by a query.
Enabling semojoin optimisation can cause a query like that:
SELECT * FROM manufacturers WHERE (...); |
to return more rows than:
SELECT * FROM manufacturers; |
Please, find a full test case setup attached.
Just to sum up the test case:
It executes the query below: (manufacturers table size: 2)
SELECT * FROM manufacturers WHERE manufacturers.id IN ( |
SELECT ships.manufacturers_id FROM owners JOIN ships ON owners.id = ships.owners_id WHERE owners.id = 1 |
);
|
Then the query returns:
- 2 rows ( SET optimizer_switch='semijoin=off'; )
- 3 rows ( SET optimizer_switch='semijoin=on'; )
Attachments
Activity
EXPLAIN:
+------+-------------+---------------+--------+------------------------------+------------+---------+------------------------+------+----------------------------------+
|
| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
|
+------+-------------+---------------+--------+------------------------------+------------+---------+------------------------+------+----------------------------------+
|
| 1 | PRIMARY | owners | system | PRIMARY | NULL | NULL | NULL | 1 | |
|
| 1 | PRIMARY | ships | range | owners_idx,manufacturers_idx | owners_idx | 4 | NULL | 3 | Using index condition; LooseScan |
|
| 1 | PRIMARY | manufacturers | eq_ref | PRIMARY | PRIMARY | 4 | ships.manufacturers_id | 1 | Using index |
|
+------+-------------+---------------+--------+------------------------------+------------+---------+------------------------+------+----------------------------------+
|
if I use optimizer_switch='loosescan=off' , I get the correct result. It seems, Loose Scan strategy is the cause of the problem.
Indeed, why does it use range over owners_idx with LooseScan. The subquery is
manufacturers.id IN ( SELECT ships.manufacturers_id ...)
|
and the query doesn't have an equality between manufacturers_id and owners_idx.
The query plan seems to be incorrect.
Debugging, I see the following to happen:
1. Join optimization is run. It chooses to use LooseScan for manufacturers table, the index is manufacturers_idx. Correct.
None of the keyparts is bound, so we will be doing a full table scan. tab->type=JT_ALL, there is no quick select.
2. We end up here:
|
(gdb) wher
|
#0 SQL_SELECT::test_quick_select (this=0x7fff68022d50, thd=0x59ff350, keys_to_use=..., prev_tables=13835058055282163714, limit=18446744073709551615, force_quick_range=false, ordered_output=false) at /home/psergey/dev-git/10.0/sql/opt_range.cc:2994
|
#1 0x00000000006b6fd6 in make_join_select (join=0x7fff68009ed8, select=0x7fff6801f970, cond=0x7fff6800ba28) at /home/psergey/dev-git/10.0/sql/sql_select.cc:9707
|
#2 0x00000000006a092d in JOIN::optimize_inner (this=0x7fff68009ed8) at /home/psergey/dev-git/10.0/sql/sql_select.cc:1533
|
#3 0x000000000069ebbe in JOIN::optimize (this=0x7fff68009ed8) at /home/psergey/dev-git/10.0/sql/sql_select.cc:1022
|
#4 0x00000000006a6a53 in mysql_select (thd=0x59ff350, rref_pointer_array=0x5a03610, tables=0x7fff68005570, wild_num=1, fields=..., conds=0x7fff680095a0, og_num=0, order=0x0, group=0x0, having=0x0, proc_param=0x0, select_options=2147748612, result=0x7fff68009818, unit=0x5a02ca8, select_lex=0x5a03398) at /home/psergey/dev-git/10.0/sql/sql_select.cc:3294
|
#5 0x00000000006d9a38 in mysql_explain_union (thd=0x59ff350, unit=0x5a02ca8, result=0x7fff68009818) at /home/psergey/dev-git/10.0/sql/sql_select.cc:24052
|
#6 0x000000000066fe35 in execute_sqlcom_select (thd=0x59ff350, all_tables=0x7fff68005570) at /home/psergey/dev-git/10.0/sql/sql_parse.cc:5237
|
#7 0x0000000000668021 in mysql_execute_command (thd=0x59ff350) at /home/psergey/dev-git/10.0/sql/sql_parse.cc:2562
|
#8 0x0000000000672d58 in mysql_parse (thd=0x59ff350, rawbuf=0x7fff68005228 "explain SELECT * FROM manufacturers WHERE manufacturers.id IN (", ' ' <repeats 16 times>, "SELECT ships.manufacturers_id FROM owners JOIN ships ON owners.id = ships.owners_id WHERE owners.id = 1 )", length=184, parser_state=0x7ffff7f39520) at /home/psergey/dev-git/10.0/sql/sql_parse.cc:6531
|
Here, the code choses to switch to a quick select. It totally ignores the fact that LooseScan doesn't work for the index that the quick select is using.
The first reaction is to make the code in make_join_select to not invoke SQL_SELECT::test_quick_select when the table in question is used for LooseScan.
It is a bit difficult to do: At this point in execution, information about whether Loose Scan is used is only stored in JOIN::best_positions. (It is moved into JOIN_TAB by setup_semijoin_dups_elimination which is called by make_join_readinfo which is called by JOIN::optimize_inner after the call to make_join_select.
Having fixed that, I get this (incorrect) plan (and a wrong query result):
+------+-------------+---------------+--------+------------------------------+---------+---------+--------------------------------+------+------------------------+
|
| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
|
+------+-------------+---------------+--------+------------------------------+---------+---------+--------------------------------+------+------------------------+
|
| 1 | PRIMARY | owners | system | PRIMARY | NULL | NULL | NULL | 1 | |
|
| 1 | PRIMARY | ships | ALL | owners_idx,manufacturers_idx | NULL | NULL | NULL | 4 | Using where; LooseScan |
|
| 1 | PRIMARY | manufacturers | eq_ref | PRIMARY | PRIMARY | 4 | testing.ships.manufacturers_id | 1 | Using index |
|
+------+-------------+---------------+--------+------------------------------+---------+---------+--------------------------------+------+------------------------+
|
http://lists.askmonty.org/pipermail/commits/2015-August/008227.html.
elenst, I would need testing for this commit. The test should check semi-join subqueries for wrong query results and crashes.
I have not got any regressions from 10.0 to 10.0 + your patch. Please push.
I used 10.0 because that's what the patch said in branch nick. If you want separate tests on 5.5, please let me know.
Thanks for the report and the test case.