Skip navigation links
17.5.11.2 Adding MySQL Cluster Data Nodes Online: Basic procedure »
Section Navigation      [Toggle]

17.5.11.1. Adding MySQL Cluster Data Nodes Online: General Issues

This section provides general information about the behavior of and current limitations in adding MySQL Cluster nodes online.

Redistribution of Data.  The ability to add new nodes online includes a means to reorganize NDBCLUSTER table data and indexes so that they are distributed across all data nodes, including the new ones. Table reorganization of both in-memory and Disk Data tables is supported. This redistribution does not currently include unique indexes (only ordered indexes are redistributed) or BLOB table data, but we are working to add redistibution of these in the near future. The redistribution for NDBCLUSTER tables already existing before the new data nodes were added is not automatic, but can be accomplished using simple SQL statements in mysql or another MySQL client application. However, all data and indexes added to tables created after a new node group has been added are distributed automatically among all cluster data nodes, including those added as part of the new node group.

Partial starts.  It is possible to add a new node group without all of the new data nodes being started. It is also possible to add a new node group to a degraded cluster — that is, a cluster that is only partially started, or where one or more data nodes are not running. In the latter case, the cluster must have enough nodes running to be viable before the new node group can be added.

Effects on ongoing operations.  Normal DML operations using MySQL Cluster data are not prevented by the creation or addition of a new node group, or by table reorganization. However, it is not possible to perform DDL concurrently with table reorganization — that is, no other DDL statements can be issued while an ALTER TABLE ... REORGANIZE PARTITION statement is executing. In addition, during the execution of ALTER TABLE ... REORGANIZE PARTITION (or the execution of any other DDL statement), it is not possible to restart cluster data nodes.

Failure handling.  Failures of data nodes during node group creation and table reorganization are handled as hown in the following table:

Failure occurs during: Failure occurs in:
“Old” data nodes “New” data nodes System
Node group creation

  • If a node other than the master fails:  The creation of the node group is always rolled forward.

  • If the master fails: 

    • If the internal commit point has been reached:  The creation of the node group is rolled forward.

    • If the internal commit point has not yet been reached.  The creation of the node group is rolled back

  • If a node other than the master fails:  The creation of the node group is always rolled forward.

  • If the master fails: 

    • If the internal commit point has been reached:  The creation of the node group is rolled forward.

    • If the internal commit point has not yet been reached.  The creation of the node group is rolled back

  • If the execution of CREATE NODEGROUP has reached the internal commit point:  When restarted, the cluster includes the new node group. Otherwise it without.

  • If the execution of CREATE NODEGROUP has not yet reached the internal commit point:  When restarted, the cluster does not include the new node group.

Table reorganization

  • If a node other than the master fails:  The table reorganization is always rolled forward.

  • If the master fails: 

    • If the internal commit point has been reached:  The table reorganization is rolled forward.

    • If the internal commit point has not yet been reached.  The table reorganization is rolled back.

  • If a node other than the master fails:  The table reorganization is always rolled forward.

  • If the master fails: 

    • If the internal commit point has been reached:  The table reorganization is rolled forward.

    • If the internal commit point has not yet been reached.  The table reorganization is rolled back.

  • If the execution of an ALTER ONLINE TABLE table REORGANIZE PARTITION statement has reached the internal commit point:  When the cluster is restarted, the data and indexes belonging to table are distributed using the “new” data nodes.

  • If the execution of an ALTER ONLINE TABLE table REORGANIZE PARTITION statement has not yet reached the internal commit point:  When the cluster is restarted, the data and indexes belonging to table are distributed using only the “old” data nodes.

Dropping node groups.  The ndb_mgm client supports a DROP NODEGROUP command, but it is possible to drop a node group only when no data nodes in the node group contain any data. Since there is currently no way to “empty” a specific data node or node group, this command works only the following two cases:

  1. After issuing CREATE NODEGROUP in the ndb_mgm client, but before issuing any ALTER ONLINE TABLE ... REORGANIZE PARTITION statements in the mysql client.

  2. After dropping all NDBCLUSTER tables using DROP TABLE.

    TRUNCATE TABLE does not work for this purpose because the data nodes continue to store the table definitions.


User Comments

Posted by Matthew Montgomery on May 2 2009 5:31am[Delete] [Edit]

Currently to reorganize BLOB or UNIQUE INDEXES you can do a "NULL" ALTER TABLE operation. (i.e. ALTER TABLE `tbl_name` ENGINE=NDBCLUSTER;)

This will not be an online operation. This table will be locked while the ALTER is in progress. This operation also requires some free space on the existing nodes. You will need 1/N't the size of the table of free space where N is the number of node groups in the cluster.

Add your own comment.