5.7 XA with InnoDB allows a table to be dropped while an XA transaction has DML active in prepared transaction

[George Ormond Lorch III]

Consider the following sequence on two sessions:

mysql-1> create table t1 (a int) engine=innodb;
Query OK, 0 rows affected (0.09 sec)
mysql-1> insert t1 values (10);
Query OK, 1 row affected (0.01 sec)
mysql-1> insert t1 values (20);
Query OK, 1 row affected (0.00 sec)
mysql-1> select * from t1;
+------+
| a |
+------+
| 10 |
| 20 |
+------+
2 rows in set (0.00 sec)

mysql-2> xa start 'test';
Query OK, 0 rows affected (0.00 sec)
mysql-2> insert t1 values (40);
Query OK, 1 row affected (0.00 sec)
mysql-2> xa end 'test';
Query OK, 0 rows affected (0.00 sec)
mysql-2> xa prepare 'test';
Query OK, 0 rows affected (0.01 sec)
mysql-2> quit;
Bye

mysql-1> select * from t1;
+------+
| a |
+------+
| 10 |
| 20 |
+------+
2 rows in set (0.00 sec)
mysql-1> drop table t1;
Query OK, 0 rows affected (6.73 sec)

mysql-1> show tables;
Empty set (0.00 sec)

mysql-1> xa recover;
+----------+--------------+--------------+-------+
| formatID | gtrid_length | bqual_length | data |
+----------+--------------+--------------+-------+
| 1 | 4 | 0 | test |
+----------+--------------+--------------+-------+
1 row in set (0.00 sec)

mysql-1> xa commit 'test';
Query OK, 0 rows affected (0.01 sec)

mysql-1> xa recover;
Empty set (0.00 sec)

We can see here that if a session creates an XA transaction, does some work, prepares and terminates, it seems to release all locks on the tables referenced within the XA transaction. This allows another session to drop those tables when in fact the drops should wait for the table locks held by the XA transaction to be relinquished.

Later, when the XA transaction is committed, it receives no error or indication that the data was not committed, thus, potential data loss.

This seems to have been introduced in 5.7.7 for replication according to this blog : http://mysqlserverteam.com/the-mysql-5-7-7-release-candidate-is-available/
Referencing links:
http://dev.mysql.com/worklog/task/?id=6860
http://bugs.mysql.com/bug.php?id=12161

In 5.6, if the session terminated with an XA transaction in the prepared state, the transaction would be aborted and locks freed and thus there would be no way for any other session to drop or alter the table until all locks were released.

I believe this is an upstream bug, I am reporting this here first to track the upstream issue.

[George Ormond Lorch III (gl-az)]

reported upstream https://bugs.mysql.com/bug.php?id=79940

[Shahriyar Rzayev (rzayev-sehriyar)]

Percona now uses JIRA for bug reports so this bug report is migrated to: https://jira.percona.com/browse/PS-3358