It is possible to use NDB Cluster in multi-master replication, including circular replication between a number of NDB Clusters.
Circular replication example. In the next few paragraphs we consider the example of a replication setup involving three NDB Clusters numbered 1, 2, and 3, in which Cluster 1 acts as the replication master for Cluster 2, Cluster 2 acts as the master for Cluster 3, and Cluster 3 acts as the master for Cluster 1. Each 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:
The SQL nodes on all masters and slaves are the same.
All SQL nodes acting as replication masters and slaves are started with the
log_slave_updatessystem variable enabled.
This type of circular replication setup is shown in the following diagram:
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 replication masters and slaves.
It is also possible to set up circular replication in such a way that not all master SQL nodes are also slaves, as shown here:
In this case, different SQL nodes in each cluster are used as
replication masters and slaves. However, you must
not start any of the SQL nodes with the
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.
Using NDB-native backup and restore to initialize a slave NDB Cluster.
When setting up circular replication, it is possible to
initialize the slave cluster by using the management client
BACKUP command on one NDB Cluster to create a
backup and then applying this backup on another NDB Cluster
using ndb_restore. However, this does not
automatically create binary logs on the second NDB
Cluster's SQL node acting as the replication slave. In
order to cause the binary logs to be created, you must issue a
SHOW TABLES statement on that SQL
node; this should be done prior to running
This is a known issue which we intend to address in a future release.
Multi-master failover example. In this section, we discuss failover in a multi-master NDB Cluster replication setup with three NDB Clusters having server IDs 1, 2, and 3. In this scenario, Cluster 1 replicates to Clusters 2 and 3; Cluster 2 also replicates to Cluster 3. This relationship is shown here:
In other words, data replicates from Cluster 1 to Cluster 3 through 2 different routes: directly, and by way of Cluster 2.
Not all MySQL servers taking part in multi-master replication must act as both master and slave, and a given NDB Cluster might use different SQL nodes for different replication channels. Such a case is shown here:
system variable has no effect on servers not being run as
The need for failover arises when one of the replicating clusters goes down. In this example, we consider the case where Cluster 1 is lost to service, and so Cluster 3 loses 2 sources of updates from Cluster 1. Because replication between NDB Clusters is asynchronous, there is no guarantee that Cluster 3's updates originating directly from Cluster 1 are more recent than those received through Cluster 2. You can handle this by ensuring that Cluster 3 catches up to Cluster 2 with regard to updates from Cluster 1. In terms of MySQL servers, this means that you need to replicate any outstanding updates from MySQL server C to server F.
On server C, perform the following queries:
mysqlC> SELECT @latest:=MAX(epoch) -> FROM mysql.ndb_apply_status -> WHERE server_id=1; mysqlC> SELECT -> @file:=SUBSTRING_INDEX(File, '/', -1), -> @pos:=Position -> FROM mysql.ndb_binlog_index -> WHERE orig_epoch >= @latest -> AND orig_server_id = 1 -> ORDER BY epoch ASC LIMIT 1;
You can improve the performance of this query, and thus likely
speed up failover times significantly, by adding the appropriate
index to the
ndb_binlog_index table. See
Section 18.6.4, “NDB Cluster Replication Schema and Tables”, for more
Copy over the values for
@pos manually from server C to server F
(or have your application perform the equivalent). Then, on server
F, execute the following
mysqlF> CHANGE MASTER TO -> MASTER_HOST = 'serverC' -> MASTER_LOG_FILE='@file', -> MASTER_LOG_POS=@pos;
Once this has been done, you can issue a
START SLAVE statement on MySQL
server F, and any missing updates originating from server B will
be replicated to server F.
CHANGE MASTER TO statement also
IGNORE_SERVER_IDS option which
takes a comma-separated list of server IDs and causes events
originating from the corresponding servers to be ignored. For more
information, see Section 184.108.40.206, “CHANGE MASTER TO Statement”, and
Section 220.127.116.11, “SHOW SLAVE STATUS Statement”. For information about how
this option intereacts with the
see Section 18.6.8, “Implementing Failover with NDB Cluster Replication”.