Release Notes for MariaDB Enterprise Server 10.6.12-8

When a prepared statement is used to query a view that contains a UNION, the server can crash. (MDEV-31102)

When binlog_row_image=FULL is set and slave_parallel_threads is greater than 0, replica servers can hang if data is inserted into tables with a sequence. (MDEV-29621)

• In previous releases, the replica server would treat full inserts into the sequence as DDL statements, which would cause it to acquire an exclusive lock on the sequence table. If another parallel replication thread is waiting on the lock at the same time, the two threads can deadlock.

• Starting with this release, replica servers manage their locks better to avoid deadlocks in this scenario.

When InnoDB splits or merges a B-tree page, the server can hang due to a race condition. (MDEV-29835)

When InnoDB undo log truncation is enabled by setting innodb_undo_log_truncate=ON, if some undo log tablespace exceeds innodb_max_undo_log_size, the server can hang. (MDEV-30863)

• In previous releases, InnoDB chooses an undo log tablespace for truncation and marks all the related rollback segments with the state "truncation in progress". When the truncation is completed, InnoDB fails to re-enable the rollback segments. Consequently, if there are further write transactions before the server is shut down, the transactions must wait for usable rollback segments in a busy-loop. Since InnoDB does not re-enable the rollback segments, the transactions wait indefinitely.

• Starting with this release, InnoDB properly manages the state of rollback segments when InnoDB undo log truncation is enabled.

For certain data distributions, the optimizer histogram code can produce wrong selectivity, which can lead to performance degradation. (MDEV-31067)

The optimizer does not take into account the selectivity of the equality conditions for Block Nested Hash (BNL-H) joins.

• In previous releases, this issue can cause mis-estimates and bad query plans when running with join_cache_level set to 3 or higher (the default is 2).

When the lateral derived optimization is used to execute a query, the derived table is re-filled on every incoming row combination, even if the parameter values have not changed. (MDEV-26301)

When slave_parallel_mode is optimistic and slave_parallel_threads is greater than 0, ALTER SEQUENCE can fail with an out-of-order binlog error if the sequence uses InnoDB. (MDEV-31077)

• In previous releases, the following error can be raised:

  Last_Error: Error 'An attempt was made to binlog GTID 0-1-100 which would create an out-of-order sequence number with existing GTID 0-1-100 and gtid stric mode is enabled' on query. Default database: 'test'. Query: 'alter sequence s1 restart with 1' will be shown.
  

InnoDB does not free undo logs when they are no longer needed. (MDEV-31234)

• In the previous release, when innodb_undo_tablespaces is set to 0, the undo logs would be stored in the InnoDB system tablespace (ibdata1 by default), so this issue could cause the InnoDB system tablespace to grow indefinitely.

• Starting with this release, InnoDB frees undo logs when they are no longer needed.

• For servers that previously ran ES 10.6.12-7, a special procedure is required to truncate the undo log tablespaces after the upgrade is complete.

InnoDB history list length increases faster in MariaDB Enterprise Server 10.6 compared to MariaDB Enterprise Server 10.5 for the same load. (MDEV-29401)