This section discusses known problems or issues when using replication with NDB Cluster.

Loss of connection between source and replica.  A loss of connection can occur either between the source cluster SQL node and the replica cluster SQL node, or between the source SQL node and the data nodes of the source cluster. In the latter case, this can occur not only as a result of loss of physical connection (for example, a broken network cable), but due to the overflow of data node event buffers; if the SQL node is too slow to respond, it may be dropped by the cluster (this is controllable to some degree by adjusting the MaxBufferedEpochs and TimeBetweenEpochs configuration parameters). If this occurs, it is entirely possible for new data to be inserted into the source cluster without being recorded in the source SQL node's binary log. For this reason, to guarantee high availability, it is extremely important to maintain a backup replication channel, to monitor the primary channel, and to fail over to the secondary replication channel when necessary to keep the replica cluster synchronized with the source. NDB Cluster is not designed to perform such monitoring on its own; for this, an external application is required.

The source SQL node issues a “gap” event when connecting or reconnecting to the source cluster. (A gap event is a type of “incident event,” which indicates an incident that occurs that affects the contents of the database but that cannot easily be represented as a set of changes. Examples of incidents are server failures, database resynchronization, some software updates, and some hardware changes.) When the replica encounters a gap in the replication log, it stops with an error message. This message is available in the output of SHOW REPLICA STATUS (prior to NDB 8.0.22, use SHOW SLAVE STATUS), and indicates that the SQL thread has stopped due to an incident registered in the replication stream, and that manual intervention is required. See Section 25.7.8, “Implementing Failover with NDB Cluster Replication”, for more information about what to do in such circumstances.

Circular replication.  NDB Cluster Replication supports circular replication, as shown in the next example. The replication setup involves three NDB Clusters numbered 1, 2, and 3, in which Cluster 1 acts as the replication source for Cluster 2, Cluster 2 acts as the source for Cluster 3, and Cluster 3 acts as the source for Cluster 1, thus completing the circle. Each NDB Cluster has two SQL nodes, with SQL nodes A and B belonging to Cluster 1, SQL nodes C and D belonging to Cluster 2, and SQL nodes E and F belonging to Cluster 3.

Circular replication using these clusters is supported as long as the following conditions are met:

This type of circular replication setup is shown in the following diagram:

Figure 25.11 NDB Cluster Circular Replication With All Sources As Replicas

In this scenario, SQL node A in Cluster 1 replicates to SQL node C in Cluster 2; SQL node C replicates to SQL node E in Cluster 3; SQL node E replicates to SQL node A. In other words, the replication line (indicated by the curved arrows in the diagram) directly connects all SQL nodes used as sources and replicas.

It should also be possible to set up circular replication in which not all source SQL nodes are also replicas, as shown here:

Figure 25.12 NDB Cluster Circular Replication Where Not All Sources Are Replicas

In this case, different SQL nodes in each cluster are used as sources and replicas. However, you must not start any of the SQL nodes with the log_replica_updates or log_slave_updates system variable enabled. This type of circular replication scheme for NDB Cluster, in which the line of replication (again indicated by the curved arrows in the diagram) is discontinuous, should be possible, but it should be noted that it has not yet been thoroughly tested and must therefore still be considered experimental.

NDB Cluster replication and primary keys.  In the event of a node failure, errors in replication of NDB tables without primary keys can still occur, due to the possibility of duplicate rows being inserted in such cases. For this reason, it is highly recommended that all NDB tables being replicated have explicit primary keys.

NDB Cluster Replication and Unique Keys.  In older versions of NDB Cluster, operations that updated values of unique key columns of NDB tables could result in duplicate-key errors when replicated. This issue is solved for replication between NDB tables by deferring unique key checks until after all table row updates have been performed.

Deferring constraints in this way is currently supported only by NDB. Thus, updates of unique keys when replicating from NDB to a different storage engine such as InnoDB or MyISAM are still not supported.

The problem encountered when replicating without deferred checking of unique key updates can be illustrated using NDB table such as t, is created and populated on the source (and transmitted to a replica that does not support deferred unique key updates) as shown here:

CREATE TABLE t (
    p INT PRIMARY KEY,
    c INT,
    UNIQUE KEY u (c)
)   ENGINE NDB;

INSERT INTO t
    VALUES (1,1), (2,2), (3,3), (4,4), (5,5);

The following UPDATE statement on t succeeds on the source, since the rows affected are processed in the order determined by the ORDER BY option, performed over the entire table:

UPDATE t SET c = c - 1 ORDER BY p;

The same statement fails with a duplicate key error or other constraint violation on the replica, because the ordering of the row updates is performed for one partition at a time, rather than for the table as a whole.

GTIDs not supported.  Replication using global transaction IDs is not compatible with the NDB storage engine, and is not supported. Enabling GTIDs is likely to cause NDB Cluster Replication to fail.

Multithreaded replicas.  Previously, NDB Cluster did not support multithreaded replicas. This restriction was removed in NDB 8.0.33.

To enable multithreading on the replica in NDB 8.0.33 and later, it is necessary to perform the following steps:

Prior to NDB 8.0.26, setting any system variables relating to multithreaded replicas such as replica_parallel_workers or slave_parallel_workers, and replica_checkpoint_group or slave_checkpoint_group (or the equivalent mysqld startup options) was completely ignored, and had no effect.

In NDB 8.0.27 through NDB 8.0.32, replica_parallel_workers must be set to 0. In these versions, if this is set to any other value on startup, NDB changes it to 0, and writes a message to the mysqld server log file. This restriction is also lifted in NDB 8.0.33.

Restarting with --initial.  Restarting the cluster with the --initial option causes the sequence of GCI and epoch numbers to start over from 0. (This is generally true of NDB Cluster and not limited to replication scenarios involving Cluster.) The MySQL servers involved in replication should in this case be restarted. After this, you should use the RESET MASTER and RESET REPLICA (prior to NDB 8.0.22, use RESET SLAVE) statements to clear the invalid ndb_binlog_index and ndb_apply_status tables, respectively.

Replication from NDB to other storage engines.  It is possible to replicate an NDB table on the source to a table using a different storage engine on the replica, taking into account the restrictions listed here:

The next few paragraphs provide additional information about each of the issues just described.

Multiple sources not supported when replicating NDB to other storage engines.  For replication from NDB to a different storage engine, the relationship between the two databases must be one-to-one. This means that bidirectional or circular replication is not supported between NDB Cluster and other storage engines.

In addition, it is not possible to configure more than one replication channel when replicating between NDB and a different storage engine. (An NDB Cluster database can simultaneously replicate to multiple NDB Cluster databases.) If the source uses NDB tables, it is still possible to have more than one MySQL Server maintain a binary log of all changes, but for the replica to change sources (fail over), the new source-replica relationship must be explicitly defined on the replica.

Replicating NDB tables to a storage engine that does not perform binary logging.  If you attempt to replicate from an NDB Cluster to a replica that uses a storage engine that does not handle its own binary logging, the replication process aborts with the error Binary logging not possible ... Statement cannot be written atomically since more than one engine involved and at least one engine is self-logging (Error 1595). It is possible to work around this issue in one of the following ways:

Replication from NDB to a nontransactional storage engine.  When replicating from NDB to a nontransactional storage engine such as MyISAM, you may encounter unnecessary duplicate key errors when replicating INSERT ... ON DUPLICATE KEY UPDATE statements. You can suppress these by using --ndb-log-update-as-write=0, which forces updates to be logged as writes, rather than as updates.

NDB Replication and File System Encryption (TDE).  The use of an encrypted filesystem does not have any effect on NDB Replication. All of the following scenarios are supported:

Replication and binary log filtering rules with replication between NDB Clusters.  If you are using any of the options --replicate-do-*, --replicate-ignore-*, --binlog-do-db, or --binlog-ignore-db to filter databases or tables being replicated, you must take care not to block replication or binary logging of the mysql.ndb_apply_status, which is required for replication between NDB Clusters to operate properly. In particular, you must keep in mind the following:

You should also remember that each replication rule requires the following:

If you are replicating an NDB Cluster to a replica that uses a storage engine other than NDB, the considerations just given previously may not apply, as discussed elsewhere in this section.

NDB Cluster Replication and IPv6.  Beginning with NDB 8.0.22, all types of NDB Cluster nodes support IPv6; this includes management nodes, data nodes, and API or SQL nodes.

Prior to NDB 8.0.22, the NDB API and MGM API (and thus data nodes and management nodes) do not support IPv6, although MySQL Servers—including those acting as SQL nodes in an NDB Cluster —can use IPv6 to contact other MySQL Servers. In versions of NDB Cluster prior to 8.0.22, you can replicate between clusters using IPv6 to connect the SQL nodes acting as source and replica as shown by the dotted arrow in the following diagram:

Figure 25.13 Replication Between SQL Nodes Connected Using IPv6

Prior to NDB 8.0.22, all connections originating within the NDB Cluster —represented in the preceding diagram by solid arrows—must use IPv4. In other words, all NDB Cluster data nodes, management servers, and management clients must be accessible from one another using IPv4. In addition, SQL nodes must use IPv4 to communicate with the cluster. In NDB 8.0.22 and later, these restrictions no longer apply; in addition, any applications written using the NDB and MGM APIs can be written and deployed assuming an IPv6-only environment.

Attribute promotion and demotion.  NDB Cluster Replication includes support for attribute promotion and demotion. The implementation of the latter distinguishes between lossy and non-lossy type conversions, and their use on the replica can be controlled by setting the global value of the system variable replica_type_conversions (NDB 8.0.26 and later) or slave_type_conversions (prior to NDB 8.0.26).

For more information about attribute promotion and demotion in NDB Cluster, see Row-based replication: attribute promotion and demotion.

NDB, unlike InnoDB or MyISAM, does not write changes to virtual columns to the binary log; however, this has no detrimental effects on NDB Cluster Replication or replication between NDB and other storage engines. Changes to stored generated columns are logged.