Release Notes for MariaDB Enterprise Server 10.5.21-15

When a backup is created with mariadb-backup and aria_log_dir_path is configured, the Aria logs are not copied to the backup. (MDEV-30968)

When system versioning is enabled for a table without a primary key, changes to the table are not properly replicated. (MDEV-30430)

When a partitioned table contains a prefix index on a column that uses a NOPAD collation, queries with ORDER BY can return rows in the wrong order. (MDEV-30072)

For some collations, when a unique constraint is defined with UNIQUE(..) USING HASH, duplicate values are accepted. (MDEV-30034)

When an InnoDB table with ROW_FORMAT=REDUNDANT is being rebuilt due to a DDL statement, the server can crash while trying to apply cached DML operations to the rebuilt table. (MDEV-26198)

Long uniques don't work correctly with Unicode collations. Equal strings (in terms of the collation) are compared as unequal if the length of the strings are different. (MDEV-27653, MDEV-28190)

When innodb_buffer_pool_filename is set to the empty string, the server tries to delete the datadir during shutdown. (MDEV-30453)

• Starting with this release, the innodb_buffer_pool_filename system variable is read-only and can't be changed dynamically.

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.
  

When a UNIQUE index includes a PERIOD in its definition, a duplicate key error can be incorrectly raised when the table uses the utf8mb4_unicode_nopad_ci collation. (MDEV-30415)

When an UPDATE or DELETE is rolled back from an InnoDB table with ROW_FORMAT=COMPRESSED, the server can crash. (MDEV-30882)

When the LEFT() function is called on a string that has no character set defined, the server can crash. (MDEV-30351)

With Galera, when wsrep_sst_method=mariabackup is set and encrypt=4 is enabled for State Snapshot Transfers (SSTs), SSTs can fail if the version of socat installed on the donor node is 1.7.4.0 or later. (MDEV-30402)

• In previous releases, SSTs could fail with the following error in the donor node's MariaDB error log if the version of socat installed is 1.7.4.0 or later:

  E Failed to set SNI host ""
  

• Starting with this release, when the SST script starts the socat listener on the donor node, the error is prevented by setting no-sni=1 if the version of socat installed is 1.7.4.0 or later.

When optimizer trace is enabled, if a view is part of a multi-table update, the server can crash. (MDEV-31085)

When a view definition contains a UNION and the view is queried using server-side prepared statements, if the optimizer pushes down a condition into the execution of the view, the server can crash during character set conversions. (MDEV-31102)

When a replica server connects to a primary server with MASTER_USE_GTID=slave_pos, if the primary server has encrypted binary logs that it can no longer decrypt, the primary server crashes due to a segmentation fault. (MDEV-28798)

• In previous releases, the primary node would iterate over all of its binary logs to look for the requested GTID. When one of the binary logs could not be decrypted, the server would crash.

• Starting with this release, when the primary node fails to decrypt a binary log in this scenario, it stops iterating over the binary logs and raises an error with the ER_MASTER_FATAL_ERROR_READING_BINLOG error code with the following error message:

  Got fatal error 1236 from master when reading data from binary log: 'Could not set up decryption for binlog.'
  

With Galera, when streaming replicating is enabled by setting the wsrep_trx_fragment_size system variable, the server can crash when certain fragment sizes are specified. (MDEV-30838)

When EXPLAIN EXTENDED is executed for a single-table DELETE that contains an IN(..) predicand, the server can crash. (MDEV-31181)

When parallel replication is enabled by setting slave_parallel_threads greater than 0, the replica's parallel replication worker threads could hang after hitting an error. (MDEV-30780)

• In previous releases, when the server was hung in this scenario, the output of SHOW SLAVE STATUS would show that an error occurred, but the output would indicate that both the I/O and SQL threads were running.

  SHOW SLAVE STATUS\G
  

  *************************** 1. row ***************************
                  Slave_IO_State: Waiting for master to send event
  ..
                Slave_IO_Running: Yes
               Slave_SQL_Running: Yes
  ..
                      Last_Errno: 1062
                      Last_Error: Could not execute Write_rows_v1 event on table TABLE_NAME; Duplicate entry 'VALUE' for key 'KEY_NAME', Error_code: 1062; handler error HA_ERR_FOUND_DUPP_KEY; the event's master log LOG_FILE, end_log_pos END_LOG_POS
  ..
  

• In this scenario, this issue causes one of the parallel replication worker threads to hang in the closing tables state, so the output of SHOW PROCESSLIST would show one worker thread in that state indefinitely:

  SHOW PROCESSLIST;
  

  +------+--------------+--------------------+------+--------------+-------+-----------------------------------------------+------------------+----------+
  | Id   | User         | Host               | db   | Command      | Time  | State                                         | Info             | Progress |
  +------+--------------+--------------------+------+--------------+-------+-----------------------------------------------+------------------+----------+
  ..
  | 2394 | system user  |                    | NULL | Slave_worker | 50852 | closing tables                                | NULL             |    0.000 |
  ..
  +------+--------------+--------------------+------+--------------+-------+-----------------------------------------------+------------------+----------+
  

When the optimizer chooses how to split a semi-join, the server can crash. (MDEV-31403)

With Galera, when CREATE TEMPORARY SEQUENCE is executed on a cluster node and binary logging is enabled, the server crashes. (MDEV-25045)

With Galera, when a write set fails certification and binary logging is enabled, the WSREP sequence numbers (cluster-wide transaction IDs) used by a WSREP applier thread can become out of sync with the node's XIDs (internal transaction IDs) due to a race condition, which can cause the node to crash. (MDEV-27317)

• In previous releases, when a write set failed certification and binary logging was enabled, a WSREP applier thread could sync the WSREP sequence number out-of-order, because the commit order could be released too early.

With Galera, when streaming replicating is enabled by setting the wsrep_trx_fragment_size system variable and CREATE TABLE .. SELECT is executed, the server can crash. The following assertion is written to the MariaDB error log during the crash: (MDEV-30862)

Assertion `mode_ == m_high_priority' failed in void wsrep::client_state::after_applying()

• Starting with this release, the server prohibits CREATE TABLE .. SELECT in this scenario and raises the ER_NOT_ALLOWED_COMMAND error code with the following error message:

  ERROR 42000: CREATE TABLE AS SELECT is not supported with streaming replication
  

With Galera, when a connection uses the handler interface to start a transaction on a table, the server can crash when the client disconnects. (MDEV-30955)

In previous releases, when the client disconnected, the server would rollback the transaction and release all locks, including the locks that the handler interface expected to survive after the transaction ended, which would cause the server to crash.

• In previous releases, the following assertion is written to the MariaDB error log during the crash:

  void close_thread_table(THD*, TABLE**): Assertion `thd->mdl_context.is_lock_owner(MDL_key::TABLE, table->s->db.str, table->s->table_name.str, MDL_SHARED)' failed.
  

With Galera, when an SST donor changes to the non-primary state, the SST is not terminated properly, and the donor node crashes. (MENT-1708)

• In previous releases, the following error message and assertion is written to the MariaDB error log during the crash:

  [Warning] WSREP: server: NODE_NAME unallowed state transition: connected -> joined
  void wsrep::server_state::state(wsrep::unique_lock<wsrep::mutex>&, wsrep::server_state::state): Assertion `0' failed.
  

When a query is executed that uses DISTINCT and an aggregate function on a group, the server can crash. (MDEV-31113)

When a server-side prepared statement is used to execute a query that references views and contains a HAVING clause, the server can crash upon second execution of the query. (MDEV-31189)

When the InnoDB purge thread tries to use the change buffer for an uncommitted index, the server aborts with an assertion. (MDEV-30076)

When the rowid_filtering optimization is used with a partitioned table, the server aborts with an assertion. (MDEV-30596)

With Galera, a hang can occur in "starting" commit state due a deadlock between a KILL command and an abort issued by an applier. (MENT-1855)

• Starting with this release, Total Order Isolation (TOI) is not used for the KILL command.

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)

When a backup is prepared with mariadb-backup, the utility can hang due to a race condition between the thread flushing the buffer pool and the thread deleting the redo log file. (MDEV-30860)

With Galera, when a transaction changes multiple tables that use different storage engines, some of which support the server's internal 2-phase commit protocol and some of which don't support it, the node crashes with an assertion failure. (MDEV-30804)

• In previous releases, the following assertion failure is written to the MariaDB error log during the crash in this scenario:

  int wsrep::transaction::ordered_commit(): Assertion `state() == s_committing' failed.
  

• Starting with this release, when Galera is enabled, mixed transactions are rejected in with the following error message:

  ERROR HY000: Transactional commit not supported by involved engine(s)
  

When InnoDB uses a lot of memory during crash recovery, the server can hang due to a race condition between the thread flushing the buffer pool and the thread performing the recovery. (MDEV-30551)

With Galera, when streaming replicating is enabled by setting the wsrep_trx_fragment_size system variable, the server can crash when backup locks or user-level locks are released. (MDEV-25037)

With Galera, when one session is terminated using the KILL statement and another session's transaction is aborted by the WSREP applier threads due to a conflict, the two operations use a conflicting mutex locking order, which can cause the node to hang. (MDEV-28460)

With Galera, when many sessions are aborted by a WSREP applier thread due to conflicts, the sessions can hang due to a deadlock. (MENT-1693)

• In previous releases, the output of SHOW PROCESSLIST could show sessions hang in the acquiring total order isolation state with queries like KILL QUERY n:

  12345678	myuser	127.0.0.1:45071	mydb	Query	1058	acquiring total order isolation	KILL QUERY 20254887	0.000
  

When EXPLAIN EXTENDED is executed for a multi-table UPDATE that uses the system join type, the output can be incorrect. (MDEV-31224)

When a query specifies DISTINCT and contains expressions using the SUM() function, the wrong results are returned. (MDEV-20057)

When a view's definition contains a HAVING clause, selecting from the view can fail with an error. (MDEV-28570)

• In previous releases, queries could raise an error with the ER_VIEW_INVALID error code and the following error message:

  View 'test.v1' references invalid table(s) or column(s) or function(s) or definer/invoker of view lack rights to use them
  

When a view's definition contains an aggregate function, selecting from the view can fail with an error. (MDEV-28571)

• In previous releases, queries could raise an error with the ER_INVALID_GROUP_FUNC_USE error code and the following error message:

  ERROR 1111 (HY000): Invalid use of group function
  

When a view's definition contains a table-value constructor (TVC) as a single-value subquery, selecting from the view can fail with an error. (MDEV-28603)

• In previous releases, queries could raise an error with the ER_VIEW_INVALID error code and the following error message:

  View 'test.v1' references invalid table(s) or column(s) or function(s) or definer/invoker of view lack rights to use them
  

When a view's definition contains an aggregate function in an uncorrelated subquery, the wrong result is returned. (MDEV-29224)

When a DELETE statement contains a subquery with a HAVING clause or an aggregate function in the WHERE clause, the statement can fail with an error. (MDEV-30586)

• In previous releases, queries could raise an error with the ER_INVALID_GROUP_FUNC_USE error code and the following error message:

  ERROR 1111 (HY000): Invalid use of group function
  

When an InnoDB tablespace has been discarded, selecting from information_schema.INNODB_SYS_INDEXES fails with an error. (MDEV-30615)

• In previous releases, an error with the ER_UNKNOWN_ERROR error code is raised:

  ERROR 1105 (HY000): Unknown error
  

• Starting with this release, no error is returned, and the results contain NULL for the PAGE_NO and SPACE columns for discarded tablespaces.

When innodb_undo_directory is set to a relative path, the path is not properly used by mariadb-backup --copy-back. (MDEV-28187)

• In previous releases, the undo logs would be copied to the relative path compared to the current working directory.

• Starting with this release, the undo logs are copied to the relative path compared to the datadir.

When UNIX_TIMESTAMP(CURRENT_TIME()) is executed, the incorrect value is returned. (MDEV-26765)

• In previous releases, NULL is returned.

With Galera, when wsrep_sst_method='mariabackup' is set, systemd raises an error about a mismatched PID. (MDEV-25887)

• In previous releases, systemd could raise the following error, where BACKUP_PID is the PID of MariaDB Enterprise Backup and SERVER_PID is the PID of MariaDB Enterprise Server:

  Got notification message from PID BACKUP_PID, but reception only permitted for main PID SERVER_PID
  

When an UPDATE contains a WHERE clause that contains a range condition over a non-indexed VARCHAR column, an error is raised. (MDEV-20773)

• In previous releases, an error with the ER_DATA_TOO_LONG error code is raised with the following error message:

  ERROR 1406 (22001): Data too long for column 'COLUMN_NAME' at row 1
  

When slave_parallel_threads is greater than 0 and SHOW SLAVE STATUS is executed, the connection can try to acquire an uninitialized mutex. (MDEV-30620)

• In previous releases, a race condition could cause the mutexes of parallel replication worker threads to be acquired before they are initialized.

The ucs2_general_mysql500_ci collation, which is intended for compatibility with older versions of MySQL, incorrectly sorts 'ß' after 's'. (MDEV-30746)

When EXPLAIN EXTENDED is executed with an INSERT, UPDATE, DELETE, or REPLACE, a warning containing the query text is not printed. (MDEV-30539)

The rowid_filtering optimization is applied incorrectly in some cases. (MDEV-30218)

When preparing a partial backup, MariaDB Enterprise Backup raises error messages about missing InnoDB tablespace files that are expected to be missing, because they were excluded from the backup. (MDEV-29050)

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

When InnoDB writes data from the doublewrite buffer to the redo log file, the Innodb_data_written status variable is not properly incremented. (MDEV-31124)

Parallel replication breaks if XA PREPARE fails updating replica GTID State (MDEV-31038)

When InnoDB has opened more data files than innodb_open_files, opening additional data files takes longer than expected due to a performance regression. (MDEV-30775)

When the JSON_OBJECTAGG function is called with a key value that includes a double quote, the double quote character is not escaped. (MDEV-30412)

With optimizer_switch='not_null_range_scan=on', when a LEFT JOIN is executed on an empty table, the results can be incorrect. (MDEV-30333)

When a query contains a GROUP BY clause and the query calls an aggregate function on a table's primary key, the results can be incorrect if the GROUP BY clause is evaluated using an index. (MDEV-30605)

With Galera, when a cluster node has the query cache enabled and the node has regular MariaDB replication configured, query cache entries are not properly invalidated when tables are changed due to replication. (MDEV-28641)

When a backup is created with mariadb-backup, the utility opens the aria_log_control file in read/write mode instead of in read-only mode. (MENT-1794)