Details
-
Bug
-
Status: Closed (View Workflow)
-
Blocker
-
Resolution: Fixed
-
10.5.25, 10.6.18, 10.11.8, 11.4.2
-
Debian 11 Bullseye on amd64, with the MariaDB-provided Maria/Galera packages.
Description
In a 10.11.8 Galera cluster (Galera 26.4.18), the following causes the cluster to become inconsistent and a node kicked out:
CREATE TABLE `t1` (
|
`id` int(10) unsigned NOT NULL,
|
`other_id` int(10) unsigned NOT NULL,
|
PRIMARY KEY (`id`,`other_id`)
|
) ENGINE=InnoDB
|
PARTITION BY LIST (`id` MOD 2)
|
(PARTITION `p0` VALUES IN (0) ENGINE = InnoDB,
|
PARTITION `p1` VALUES IN (1) ENGINE = InnoDB);
|
|
|
INSERT INTO t1 VALUES (1, 0);
|
|
|
CREATE TABLE t2 LIKE t1;
|
|
|
START TRANSACTION;
|
|
|
INSERT INTO t2(SELECT * FROM t1 WHERE id = 1);
|
DELETE FROM t1 WHERE id = 1;
|
|
|
SELECT SLEEP(1);
|
|
|
COMMIT;
|
The partitioning can be based on either id or other_id, it doesn't seem to matter. It also shouldn't matter what particular values are used in the INSERT.
Without the SLEEP call, the other node hangs forever executing the commit. When I discovered this problem in the real world, the symptom was the inconsistency and not the hang, so I've put the SLEEP in the example. Of course either failure is really bad.
This problem does not occur on 10.11.7 (Galera 26.4.16).
Attachments
Issue Links
- is caused by
-
-
- Closed
-
- is duplicated by
-
-
- Open
-
- relates to
-
-
- Closed
-
Activity
The fix has been merged with the head revision: https://github.com/MariaDB/server/commit/0172887980cd6297cbda1dda79ecbde0568c4664
+post-fix change: https://github.com/MariaDB/server/commit/0d85c905c44f9d88e4a17f56594cc389d777abaf
I've modified the MariaDB-provided 10.11.8 Debian sources to include the fix, and it seems to work perfectly. Thanks very much, Jan and Julius.
James, you probably want to close out MDEV-34267.
"bug is present if you use INSERT INTO ... SELECT ... for partitioned tables." means to me that bug not present if database doesn't have partitioned tables. Therefore I have ONLY downgraded clusters having partitioned tables. I downgraded because I do not expect a fix immediately.