Important Change; NDB Replication: The replica_allow_batching system variable affects how efficiently the replica applies epoch transactions. When this is set to OFF, by default, every discrete replication event in the binary log is applied and executed separately, which generally leads to poor performance.

Beginning with this release, the default value for replica_allow_batching is changed from OFF to ON.

NDB Cluster APIs: Removed a number of potential memory leaks by using std::uniqe_ptr for managing any Event returned by Dictionary::getEvent().

As part of this fix, we add a releaseEvent() method to Dictionary to clean up events created with getEvent() after they are no longer needed. (Bug #33855045)

NDB Cluster APIs: The Node.JS package included with NDB Cluster has been updated to version 16.5.0. (Bug #33770627)

Empty lines in CSV files are now accepted as valid input by ndb_import. (Previously, empty lines in such files were always rejected.) Now, if an empty value can be used as the value for a single imported column, ndb_import uses it in the same manner as LOAD DATA. (Bug #34119833)

NDB stores blob column values differently from other types; by default, only the first 256 bytes of the value are stored in the table (“inline”), with any remainder kept in a separate blob parts table. This is true for columns of MySQL type BLOB, MEDIUMBLOB, LONGBLOB, TEXT, MEDIUMTEXT, and LONGTEXT. (TINYBLOB and TINYTEXT are exceptions, since they are always inline-only.) NDB handles JSON column values in a similar fashion, the only difference being that, for a JSON column, the first 4000 bytes of the value are stored inline.

Previously, it was possible to control the inline size for blob columns of NDB tables only by using the NDB API (Column::setInlineSize() method). This now can be done in the mysql client (or other application supplying an SQL interface) using a column comment which consists of an NDB_COLUMN string containing a BLOB_INLINE_SIZE specification, as part of a CREATE TABLE statement like this one:

CREATE TABLE t (
    a BIGINT NOT NULL AUTO_INCREMENT PRIMARY KEY,
    b BLOB COMMENT 'NDB_COLUMN=BLOB_INLINE_SIZE=3000'
) ENGINE NDBCLUSTER;

In table t created by the statement just shown, column b (emphasized text in the preceding example) is a BLOB column whose first 3000 bytes are stored in t itself, rather than just the first 256 bytes. This means that, if no value stored in b exceeds 3000 bytes in length, no extra work is required to read or write any excess data from the NDB blob parts table when storing or retrieving the column value. This can improve performance significantly when performing many operations on blob columns.

You can see the effects of this option by querying the ndbinfo.blobs table, or examining the output of ndb_desc.

The maximum supported value for BLOB_INLINE_SIZE is 29980. Setting it to any value less than 1 causes the default inline size to be used for the column.

You can also alter a column as part of a copying ALTER TABLE; ALGORITHM=INPLACE is not supported for such operations.

BLOB_INLINE_SIZE can be used alone, or together with MAX_BLOB_PART_SIZE in the same NDB_COMMENT string. Unlike the case with MAX_BLOB_PART_SIZE, setting BLOB_INLINE_SIZE is supported for JSON columns of NDB tables.

For more information, see NDB_COLUMN Options, as well as String Type Storage Requirements. (Bug #33755627, WL #15044)

A new --missing-ai-column option is added to ndb_import. This enables ndb_import to accept a CSV file from which the data for an AUTO_INCREMENT column is missing and to supply these values itself, much as LOAD DATA does. This can be done with one or more tables for which the CSV representation contains no values for such a column.

This option works only when the CSV file contains no nonempty values for the AUTO_INCREMENT column to be imported. (Bug #102730, Bug #32553029)

This release adds Performance Schema instrumentation for transaction batch memory used by NDBCLUSTER, making it possible to monitor memory used by transactions. For more information, see Transaction Memory Usage. (WL #15073)

Important Change: When using the ThreadConfig multithreaded data node parameter to specify the threads to be created on the data nodes, the receive thread (recv) in some cases was placed in the same worker thread as block threads such as DBLQH(0) and DBTC(0). This represented a regression from NDB 8.0.22 and earlier, where the receive thread is colocated only with THRMAN and TRPMAN, as expected in such cases.

Now, when setting the value of ThreadConfig, you must include main, rep, recv, and ldm explicitly; to avoid using one or more of the main, rep, or ldm thread types, you must set count=0 explicitly for each applicable thread type.

In addition, a minimum value of 1 is now enforced for the recv count; setting the replication thread (rep) count to 1 also requires setting count=1 for the main thread.

These changes can have serious implications for upgrades from previous NDB Cluster releases. For more information, see Upgrading and Downgrading NDB Cluster, as well as the description of the ThreadConfig parameter, in the MySQL Manual. (Bug #33869715)

References: See also: Bug #34038016, Bug #34025532.

macOS: ndb_import could not be compiled on MacOS/ARM because the ndbgeneral library was not explicitly included in LINK_LIBRARIES. (Bug #33931512)

NDB Disk Data: The LGMAN kernel block did not initialize its local encrypted filesystem state, and did not check EncryptedFileSystem for undo log files, so that their encryption status was never actually set.

This meant that, for release builds, it was possible for the undo log files to be encrypted on some systems, even though they should not have been; in debug builds, undo log files were always encrypted. This could lead to problems when using Disk Data tables and upgrading to or from NDB 8.0.29. (A workaround is to perform initial restarts of the data nodes when doing so.)

This issue could also cause unexpected CPU load for I/O threads when there were a great many Disk Data updates to write to the undo log, or at data node startup while reading the undo log.

(Bug #34185524)

NDB Cluster APIs: The internal function NdbThread_SetThreadPrio() sets the thread priority (thread_prio) for a given thread type when applying the setting of the ThreadConfig configuration parameter. It was possible for this function in some cases to return an error when it had actually succeeded, which could have a an unfavorable impact on the performance of some NDB API applications. (Bug #34038630)

NDB Cluster APIs: The following NdbInterpretedCode methods did not function correctly when a nonzero value was employed for the label argument:

(Bug #33888962)

MySQL NDB ClusterJ: ClusterJ support for systems with ARM-based Apple silicon is now enabled by default. (Bug #34148474)

Compilation of NDB Cluster on Debian 11 and Ubuntu 22.04 halted during the link time optimization phase due to source code warnings being treated as errors. (Bug #34252425)

NDB does not support in-place changes of default values for columns; such changes can be made only by using a copying ALTER TABLE. Changing of the default value in such cases was already detected, but the additional or removal of default value was not.

We fix this issue by detecting when default value is added or removed during ALTER TABLE, and refusing to perform the operation in place. (Bug #34224193)

After creating a user on SQL node A and granting it the NDB_STORED_USER privilege, dropping this user from SQL node B led to inconsistent results. In some cases, the drop was not distributed, so that after the drop the user still existed on SQL node A.

The cause of this issue is that NDB maintains a cache of all local users with NDB_STORED_USER, but when a user was created on SQL node B, this cache was not updated. Later, when executing DROP USER, this led SQL node B to determine that the drop did not have to be distributed. We fix this by ensuring that this cache is updated whenever a new distributed user is created. (Bug #34183149)

When the internal ndbd_exit() function was invoked on a data node, information and error messages sent to the event logger just prior to the ndbd_exit() call were not printed in the log as expected. (Bug #34148712)

NDB Cluster did not compile correctly on Ubuntu 22.04 due to changes in OpenSSL 3.0. (Bug #34109171)

NDB Cluster would not compile correctly using GCC 8.4 due to a change in Bison fallthrough handling. (Bug #34098818)

Compiling the ndbcluster plugin or the libndbclient library required a number of files kept under directories specific to data nodes (src/kernel) and management servers (src/mgmsrv). These have now been moved to more suitable locations. Files moved that may be of interest are listed here:

(Bug #34045289)

When an error occurred during the begin schema transaction phase, an attempt to update the index statistics automatically was made without releasing the transaction handle, leading to a leak. (Bug #34007422)

References: See also: Bug #34992370.

Path lengths were not always calculated correctly by the data nodes. (Bug #33993607)

When ndb_restore performed an NDB API operation with any concurrent NDB API events taking place, contention could occur in the event of limited resources in DBUTIL. This led to temporary errors in NDB. In such cases, ndb_restore now attempts to retry the NDB API operation which failed. (Bug #33984717)

References: See also: Bug #33982499.

Removed a duplicate check of a table pointer found in the internal method Dbtc::execSCAN_TABREQ(). (Bug #33945967)

The internal function NdbReceiver::unpackRecAttr(), which unpacks attribute values from a buffer from a GSN_TRANSID_AI signal, did not check to ensure that attribute sizes fit within the buffer. This could corrupt the buffer which could in turn lead to reading beyond the buffer and copying beyond destination buffers. (Bug #33941167)

Improved formatting of log messages such that the format string verification employed by some compilers is no longer bypassed. (Bug #33930738)

Some NDB internal signals were not always checked properly. (Bug #33896428)

Fixed a number of issues in the source that raised -Wunused-parameter warnings when compiling NDB Cluster with GCC 11.2. (Bug #33881953)

When an SQL node was not yet connected to NDBCLUSTER, an excessive number of warnings were written to the MySQL error log when the SQL node could not discover an NDB table. (Bug #33875273)

The NDB API statistics variables Ndb_api_wait_nanos_count, Ndb_api_wait_nanos_count_replica, and Ndb_api_wait_nanos_count_session are used for determining CPU times and wait times for applications. These counters are intended to show the time spent waiting for responses from data nodes, but they were not entirely accurate because time spent waiting for key requests was not included.

For more information, see NDB API Statistics Counters and Variables. (Bug #33840016)

References: See also: Bug #33850590.

It was possible in some cases for a duplicate engine-se_private_id entry to be installed in the MySQL data dictionary for an NDB table, even when the previous table definition should have been dropped.

When data nodes drop out of the cluster and need to rejoin, each SQL node starts to synchronize the schema definitions in its own data dictionary. The se_private_id for an NDB table installed in the data dictionary is the same as its NDB table ID. It is common for tables to be updated with different IDs, such as when executing an ALTER TABLE, DROP TABLE, or CREATE TABLE statement. The previous table definition, obtained by referencing the table in schema.table format, is usually sufficient for a drop and thus for the new table to be installed with a new ID, since it is assumed that no other installed table definition uses that ID. An exception to this could occur during synchronization, if a data node shutdown allowed the previous table definition of a table having the same ID other than the one to be installed to remain, then the old definition was not dropped.

To correct this issue, we now check whether the ID of the table to be installed in the data dictionary differs from that of the previous one, in which case we also check whether an old table definition already exists with that ID, and, if it does, we drop the old table before continuing. (Bug #33824058)

After receiving a COPY_FRAGREQ signal, DBLQH sometimes places the signal in a queue by copying the signal object into a stored object. Problems could arise when this signal object was used to send another signal before the incoming COPY_FRAGREQ was stored; this led to saving a corrupt signal that, when sent, prevented a system restart from completing. We fix this by using a static copy of the signal for storage and retrieval, rather than the original signal object. (Bug #33581144)

When the mysqld binary supplied with NDB Cluster was run without NDB support enabled, the ndbinfo and ndb_transid_mysql_connection_map plugins were still enabled, and for example, still shown with status ACTIVE in the output of SHOW PLUGINS. (Bug #33473346)

Attempting to seize a redo log page could in theory fail due to a wrong bound error. (Bug #32959887)

When a data node was started using the --foreground option, and with a node ID not configured to connect from a valid host, the data node underwent a forced shutdown instead of reporting an error. (Bug #106962, Bug #34052740)

NDB tables were skipped in the MySQL Server upgrade phase and were instead migrated by the ndbcluster plugin at a later stage. As a result, triggers associated with NDB tables were not created during upgrades from 5.7 based versions.

This occurred because it is not possible to create such triggers when the NDB tables are migrated by the ndbcluster plugin, since metadata about the triggers is lost in the upgrade finalization phase of the MySQL Server upgrade in which all .TRG files are deleted.

To fix this issue, we make the following changes:

(Bug #106883, Bug #34058984)

When initializing a file, NDBFS enabled autosync but never called do_sync_after_write() (then called sync_on_write()), so that the file was never synchronized to disk until it was saved. This meant that, for a system whose network disk was stalled for some time, the file could use up system memory on buffered file data.

We fix this by calling do_sync_after_write() each time NDBFS writes to a file.

As part of this work, we increase the autosync size from 1 MB to 16 MB when initializing files.

(Bug #106697, Bug #33946801, Bug #34131456)