--source include/count_sessions.inc # This is the essense of the scenario from the report in Bug#33724166. # There are two rows: # c0000 | c0003 | c0004 | partition # ------+-------+-------+---------- # 123 | 12 | 0 | 5 # 122 | 13 | 0 | 2 # # It's important that 5 is the last of 6 partitions. # A buggy implementation, would keep the internal state after the loop # over all partitions, and reuse this state when handling subsequent # unlock request from Server layer - in other words, will use state # of the locks of 5th partition 5, when reasoning about 2nd partition locks. # The con1 is only used to create an X lock on row from 2nd partition, so that # the con2 is forced to use semi-consistent read on it without a lock. # The InnoDB should correctly track which row needed a lock and which didn't, # during the initial phase of filling up the priority queue with one record from # each partition, and then correctly handle unlock requests from Server layer, # ignoring request for 2nd partition, and performing it for 5th partition only. CREATE TABLE t ( `c0000` INT NOT NULL, `c0003` TINYINT NOT NULL, `c0004` INT NOT NULL, PRIMARY KEY (`c0000`,`c0003`,`c0004`) ) PARTITION BY KEY (c0003) PARTITIONS 6; INSERT INTO t VALUES (123,12,0); INSERT INTO t VALUES (122,13,0); --connect (con1, localhost, root,,) BEGIN; SELECT * FROM t WHERE `c0000` = 122 AND `c0003` = 13 FOR UPDATE; --connect (con2, localhost, root,,) SET TRANSACTION ISOLATION LEVEL READ UNCOMMITTED; BEGIN; # The WHERE condition intentionally doesn't match any record, but in a way # which InnoDB can't determine (no condition push down) thus reporting the # record to the Server layer UPDATE t SET `c0003` = 16 WHERE `c0004`; COMMIT; --connection con1 COMMIT; --connection default --disconnect con1 --disconnect con2 DROP TABLE t; --source include/wait_until_count_sessions.inc # Here again we have the same two rows: # c0000 | c0003 | c0004 | partition # ------+-------+-------+---------- # 123 | 12 | 0 | 5 # 122 | 13 | 0 | 2 # # The transaction first locks the record from 2nd partition, which can be reused # in the subsequent UPDATE query. Such reused lock should not be released even # if a row doesn't match the UPDATE's WHERE condition, because this lock was # explicitly needed for a previous SELECT. InnoDB should correctly track that # it reused a lock for record in 2nd partition, but has created a new lock for # the record in 5th partition - and the latter can be released. CREATE TABLE t ( `c0000` INT NOT NULL, `c0003` TINYINT NOT NULL, `c0004` INT NOT NULL, PRIMARY KEY (`c0000`,`c0003`,`c0004`) ) PARTITION BY KEY (c0003) PARTITIONS 6; INSERT INTO t VALUES (121,13,0); # partition 2 INSERT INTO t VALUES (123,12,0); # partition 5 SET TRANSACTION ISOLATION LEVEL READ UNCOMMITTED; BEGIN; # put an X REC_NOT_GAP lock on the first row in partition#2 SELECT * FROM t WHERE `c0000` = 121 AND `c0003` = 13 AND `c0004` = 0 FOR UPDATE; SELECT LOCK_DATA FROM performance_schema.data_locks WHERE LOCK_TYPE = 'RECORD' AND OBJECT_NAME = 't'; # now try to iterate over rows from all partitions UPDATE t SET `c0003` = 16 WHERE `c0004`; SELECT LOCK_DATA FROM performance_schema.data_locks WHERE LOCK_TYPE = 'RECORD' AND OBJECT_NAME = 't'; COMMIT; DROP TABLE t; # Here we have one row in each of 6 partitions, except for the last one (5th). # InnoDB should properly track that it has created a lock in 5 of 6 partitions, # and then release them, as none of the records matches the WHERE condition. CREATE TABLE t ( `c0000` INT NOT NULL, `c0003` TINYINT NOT NULL, `c0004` INT NOT NULL, PRIMARY KEY (`c0000`,`c0003`,`c0004`) ) PARTITION BY KEY (c0003) PARTITIONS 6; INSERT INTO t VALUES (1,10,0); INSERT INTO t VALUES (2,11,0); #INSERT INTO t VALUES (3,12,0); this would be partition 5, but we omit it INSERT INTO t VALUES (4,13,0); INSERT INTO t VALUES (5,14,0); INSERT INTO t VALUES (6,15,0); INSERT INTO t VALUES (7,16,0); SET TRANSACTION ISOLATION LEVEL READ UNCOMMITTED; BEGIN; SELECT LOCK_DATA FROM performance_schema.data_locks WHERE LOCK_TYPE = 'RECORD' AND OBJECT_NAME = 't'; # now try to iterate over rows from all partitions UPDATE t SET `c0003` = 16 WHERE `c0004`; SELECT LOCK_DATA FROM performance_schema.data_locks WHERE LOCK_TYPE = 'RECORD' AND OBJECT_NAME = 't'; COMMIT; DROP TABLE t;