Details
-
Bug
-
Status: Closed (View Workflow)
-
Critical
-
Resolution: Fixed
-
10.5, 10.6, 10.4(EOL)
Description
Queries involving INFORMATION_SCHEMA.OPTIMIZER_TRACE table (even an empty one and/or with disabled optimizer_trace) take quite long time and make the server memory consumption temporarily jump ~4G up for each query and instance of the table. That is, if the query joins OPTIMIZER_TRACE table with itself, it's 8G, if it joins twice, it's 12G. After the job is done, the consumption drops back, but these peaks can apparently cause OOM on small machines.
|
10.4 ad0f0d2b1 non-debug |
MariaDB [test]> SELECT COUNT(*) FROM INFORMATION_SCHEMA.OPTIMIZER_TRACE; |
+----------+ |
| COUNT(*) | |
+----------+ |
| 0 |
|
+----------+ |
1 row in set (0.998 sec) |
|
|
MariaDB [test]> SELECT COUNT(*) FROM INFORMATION_SCHEMA.OPTIMIZER_TRACE; |
+----------+ |
| COUNT(*) | |
+----------+ |
| 0 |
|
+----------+ |
1 row in set (0.998 sec) |
|
|
MariaDB [test]> SELECT COUNT(*) FROM INFORMATION_SCHEMA.OPTIMIZER_TRACE; |
+----------+ |
| COUNT(*) | |
+----------+ |
| 0 |
|
+----------+ |
1 row in set (0.987 sec) |
4109823 elenst 20 0 1758144 80088 19976 S 0.0 0.1 0:05.14 mysqld
|
4109823 elenst 20 0 5952452 3.6g 19976 S 73.5 3.8 0:05.89 mysqld
|
4109823 elenst 20 0 5952452 3.6g 19976 S 101.0 3.8 0:06.90 mysqld
|
4109823 elenst 20 0 5952452 3.7g 19976 S 99.0 4.0 0:07.91 mysqld
|
4109823 elenst 20 0 1758144 80088 19976 S 19.8 0.1 0:08.11 mysqld
|
MariaDB [test]> SELECT 1 FROM INFORMATION_SCHEMA.OPTIMIZER_TRACE t1, INFORMATION_SCHEMA.OPTIMIZER_TRACE t2, INFORMATION_SCHEMA.OPTIMIZER_TRACE t3; |
Empty set (2.792 sec) |
|
|
MariaDB [test]> SELECT 1 FROM INFORMATION_SCHEMA.OPTIMIZER_TRACE t1, INFORMATION_SCHEMA.OPTIMIZER_TRACE t2, INFORMATION_SCHEMA.OPTIMIZER_TRACE t3; |
Empty set (2.765 sec) |
|
|
MariaDB [test]> SELECT 1 FROM INFORMATION_SCHEMA.OPTIMIZER_TRACE t1, INFORMATION_SCHEMA.OPTIMIZER_TRACE t2, INFORMATION_SCHEMA.OPTIMIZER_TRACE t3; |
Empty set (3.185 sec) |
4109823 elenst 20 0 1758144 80088 19976 S 0.0 0.1 0:08.15 mysqld
|
4109823 elenst 20 0 9908.9m 4.3g 19976 S 86.1 4.5 0:09.02 mysqld
|
4109823 elenst 20 0 13.7g 9.4g 19976 S 100.0 9.9 0:10.04 mysqld
|
4109823 elenst 20 0 5952452 609764 19976 S 100.0 0.6 0:11.05 mysqld
|
4109823 elenst 20 0 9908.9m 5.5g 19976 S 98.0 5.8 0:12.05 mysqld
|
4109823 elenst 20 0 13.7g 10.6g 19976 S 102.0 11.3 0:13.07 mysqld
|
4109823 elenst 20 0 5952452 1.7g 19976 S 98.0 1.8 0:14.07 mysqld
|
4109823 elenst 20 0 9908.9m 5.9g 19976 S 100.0 6.2 0:15.08 mysqld
|
4109823 elenst 20 0 13.7g 10.4g 19976 S 100.0 11.0 0:16.09 mysqld
|
4109823 elenst 20 0 1758144 80196 19976 S 75.5 0.1 0:16.86 mysqld
|
4109823 elenst 20 0 1758144 80196 19976 S 0.0 0.1 0:16.86 mysqld
|
Reproducible on 10.4-10.6.
The test case is not applicable to 10.3 due to the use of optimizer_trace.
Most likely the problem was introduced by this commit:
commit e64084d5a3a72462fa6263d1d0a86e72c0ba0d47
|
Author: Sergei Golubchik
|
Date: Sat Aug 1 13:12:50 2020 +0200
|
|
MDEV-21201 No records produced in information_schema query, depending on projection
|
I can only confirm that it's reproducible on 10.4 after the merge containing the commit, and not reproducible before; but since the commit was in 10.3, 10.3 doesn't have optimizer trace, and I couldn't reproduce it on other information_schema tables, I couldn't check the exact commit.