NDB Cluster is not supported in standard MySQL Server releases. Instead, MySQL NDB Cluster is provided as a separate product. Available NDB Cluster release series include the following:

You can obtain and compile NDB Cluster from source (see Section 25.3.1.4, “Building NDB Cluster from Source on Linux”, and Section 25.3.2.2, “Compiling and Installing NDB Cluster from Source on Windows”), but for all but the most specialized cases, we recommend using one of the following installers provided by Oracle that is appropriate to your operating platform and circumstances:

Installation packages may also be available from your platform's package management system.

You can determine whether your MySQL Server has NDB support using one of the statements SHOW VARIABLES LIKE 'have_%', SHOW ENGINES, or SHOW PLUGINS.

“NDB” stands for “Network Database”. NDB and NDBCLUSTER are both names for the storage engine that enables clustering support with MySQL. NDB is preferred, but either name is correct.

In traditional MySQL replication, a source MySQL server updates one or more replicas. Transactions are committed sequentially, and a slow transaction can cause the replica to lag behind the source. This means that if the source fails, it is possible that the replica might not have recorded the last few transactions. If a transaction-safe engine such as InnoDB is being used, a transaction is either completed on the replica or not applied at all, but replication does not guarantee that all data on the source and the replica remains consistent at all times. In NDB Cluster, all data nodes are kept in synchrony, and a transaction committed by any one data node is committed for all data nodes. In the event of a data node failure, all remaining data nodes remain in a consistent state.

In short, whereas standard MySQL replication is asynchronous, NDB Cluster is synchronous.

Asynchronous replication is also available in NDB Cluster. NDB Cluster Replication (also sometimes known as “geo-replication”) includes the capability to replicate both between two NDB Clusters, and from an NDB Cluster to a non-Cluster MySQL server. See Section 25.7, “NDB Cluster Replication”.

NDB Cluster is intended to be used in a high-bandwidth environment, with computers connecting using TCP/IP. Its performance depends directly upon the connection speed between the cluster's computers. The minimum connectivity requirements for NDB Cluster include a typical 100-megabit Ethernet network or the equivalent. We recommend you use gigabit Ethernet whenever available.

A minimum of three computers is required to run a viable cluster. However, the minimum recommended number of computers in an NDB Cluster is four: one each to run the management and SQL nodes, and two computers to serve as data nodes. The purpose of the two data nodes is to provide redundancy; the management node must run on a separate machine to guarantee continued arbitration services in the event that one of the data nodes fails.

To provide increased throughput and high availability, you should use multiple SQL nodes (MySQL Servers connected to the cluster). It is also possible (although not strictly necessary) to run multiple management servers.

An NDB Cluster has both a physical and logical organization, with computers being the physical elements. The logical or functional elements of a cluster are referred to as nodes, and a computer housing a cluster node is sometimes referred to as a cluster host. There are three types of nodes, each corresponding to a specific role within the cluster. These are:

The simplest answer is, “It's not something you can control, and it's nothing that you need to worry about in any case, unless you're a software engineer writing or analyzing the NDB Cluster source code”.

If you don't find that answer satisfactory, here's a longer and more technical version:

A number of mechanisms in NDB Cluster require distributed coordination among the data nodes. These distributed algorithms and protocols include global checkpointing, DDL (schema) changes, and node restart handling. To make this coordination simpler, the data nodes “elect” one of their number to act as leader. There is no user-facing mechanism for influencing this selection, which is completely automatic; the fact that it is automatic is a key part of NDB Cluster's internal architecture.

When a node acts as the “leader” for any of these mechanisms, it is usually the point of coordination for the activity, and the other nodes act as “followers”, carrying out their parts of the activity as directed by the leader. If the node acting as leader fails, then the remaining nodes elect a new leader. Tasks in progress that were being coordinated by the old leader may either fail or be continued by the new leader, depending on the actual mechanism involved.

It is possible for some of these different mechanisms and protocols to have different leader nodes, but in general the same leader is chosen for all of them. The node indicated as the leader in the output of SHOW in the management client is known internally as the DICT manager, responsible for coordinating DDL and metadata activity.

NDB Cluster is designed in such a way that the choice of leader has no discernible effect outside the cluster itself. For example, the current leader does not have significantly higher CPU or resource usage than the other data nodes, and failure of the leader should not have a significantly different impact on the cluster than the failure of any other data node.

NDB Cluster is supported on most Unix-like operating systems. NDB Cluster is also supported in production settings on Microsoft Windows operating systems.

For more detailed information concerning the level of support which is offered for NDB Cluster on various operating system versions, operating system distributions, and hardware platforms, please refer to https://www.mysql.com/support/supportedplatforms/cluster.html.

NDB Cluster should run on any platform for which NDB-enabled binaries are available. For data nodes and API nodes, faster CPUs and more memory are likely to improve performance, and 64-bit CPUs are likely to be more effective than 32-bit processors. There must be sufficient memory on machines used for data nodes to hold each node's share of the database (see How much RAM do I Need? for more information). For a computer which is used only for running the NDB Cluster management server, the requirements are minimal; a common desktop PC (or the equivalent) is generally sufficient for this task. Nodes can communicate through the standard TCP/IP network and hardware. They can also use the high-speed SCI protocol; however, special networking hardware and software are required to use SCI (see Section 25.4.4, “Using High-Speed Interconnects with NDB Cluster”).

NDB Cluster was originally implemented as in-memory only, but all versions currently available also provide the ability to store NDB Cluster on disk. See Section 25.6.11, “NDB Cluster Disk Data Tables”, for more information.

For in-memory NDB tables, you can use the following formula for obtaining a rough estimate of how much RAM is needed for each data node in the cluster:

(SizeofDatabase × NumberOfReplicas × 1.1 ) / NumberOfDataNodes

To calculate the memory requirements more exactly requires determining, for each table in the cluster database, the storage space required per row (see Section 13.7, “Data Type Storage Requirements”, for details), and multiplying this by the number of rows. You must also remember to account for any column indexes as follows:

Creating NDB Cluster tables with USING HASH for all primary keys and unique indexes generally causes table updates to run more quickly—in some cases by a much as 20 to 30 percent faster than updates on tables where USING HASH was not used in creating primary and unique keys. This is due to the fact that less memory is required (because no ordered indexes are created), and that less CPU must be utilized (because fewer indexes must be read and possibly updated). However, it also means that queries that could otherwise use range scans must be satisfied by other means, which can result in slower selects.

When calculating Cluster memory requirements, you may find useful the ndb_size.pl utility which is available in recent MySQL 8.4 releases. This Perl script connects to a current (non-Cluster) MySQL database and creates a report on how much space that database would require if it used the NDBCLUSTER storage engine. For more information, see Section 25.5.29, “ndb_size.pl — NDBCLUSTER Size Requirement Estimator”.

It is especially important to keep in mind that every NDB Cluster table must have a primary key. The NDB storage engine creates a primary key automatically if none is defined; this primary key is created without USING HASH.

You can determine how much memory is being used for storage of NDB Cluster data and indexes at any given time using the REPORT MEMORYUSAGE command in the ndb_mgm client; see Section 25.6.1, “Commands in the NDB Cluster Management Client”, for more information. In addition, warnings are written to the cluster log when 80% of available DataMemory or (prior to NDB 7.6) IndexMemory is in use, and again when usage reaches 90%, 99%, and 100%.

Generally, any file system that is native to the host operating system should work well with NDB Cluster. If you find that a given file system works particularly well (or not so especially well) with NDB Cluster, we invite you to discuss your findings in the NDB Cluster Forums.

For Windows, we recommend that you use NTFS file systems for NDB Cluster, just as we do for standard MySQL. We do not test NDB Cluster with FAT or VFAT file systems. Because of this, we do not recommend their use with MySQL or NDB Cluster.

NDB Cluster is implemented as a shared-nothing solution; the idea behind this is that the failure of a single piece of hardware should not cause the failure of multiple cluster nodes, or possibly even the failure of the cluster as a whole. For this reason, the use of network shares or network file systems is not supported for NDB Cluster. This also applies to shared storage devices such as SANs.

NDB Cluster is supported for use in virtual machines. We currently support and test using Oracle VM.

Some NDB Cluster users have successfully deployed NDB Cluster using other virtualization products; in such cases, Oracle can provide NDB Cluster support, but issues specific to the virtual environment must be referred to that product's vendor.

The cause is very likely to be that your setup does not provide sufficient RAM for all table data and all indexes, including the primary key required by the NDB storage engine and automatically created in the event that the table definition does not include the definition of a primary key.

It is also worth noting that all data nodes should have the same amount of RAM, since no data node in a cluster can use more memory than the least amount available to any individual data node. For example, if there are four computers hosting Cluster data nodes, and three of these have 3GB of RAM available to store Cluster data while the remaining data node has only 1GB RAM, then each data node can devote at most 1GB to NDB Cluster data and indexes.

In some cases it is possible to get Table is full errors in MySQL client applications even when ndb_mgm -e "ALL REPORT MEMORYUSAGE" shows significant free DataMemory. You can force NDB to create extra partitions for NDB Cluster tables and thus have more memory available for hash indexes by using the MAX_ROWS option for CREATE TABLE. In general, setting MAX_ROWS to twice the number of rows that you expect to store in the table should be sufficient.

For similar reasons, you can also sometimes encounter problems with data node restarts on nodes that are heavily loaded with data. The MinFreePct parameter can help with this issue by reserving a portion (5% by default) of DataMemory and (prior to NDB 7.6) IndexMemory for use in restarts. This reserved memory is not available for storing NDB tables or data.

It is very unlikely that a cluster would perform reliably under such conditions, as NDB Cluster was designed and implemented with the assumption that it would be run under conditions guaranteeing dedicated high-speed connectivity such as that found in a LAN setting using 100 Mbps or gigabit Ethernet—preferably the latter. We neither test nor warrant its performance using anything slower than this.

Also, it is extremely important to keep in mind that communications between the nodes in an NDB Cluster are not secure; they are neither encrypted nor safeguarded by any other protective mechanism. The most secure configuration for a cluster is in a private network behind a firewall, with no direct access to any Cluster data or management nodes from outside. (For SQL nodes, you should take the same precautions as you would with any other instance of the MySQL server.) For more information, see Section 25.6.21, “NDB Cluster Security Issues”.

No. Although some specialized commands are used to manage and configure the cluster itself, only standard (My)SQL statements are required for the following operations:

Some specialized configuration parameters and files are required to set up an NDB Cluster—see Section 25.4.3, “NDB Cluster Configuration Files”, for information about these.

A few simple commands are used in the NDB Cluster management client (ndb_mgm) for tasks such as starting and stopping cluster nodes. See Section 25.6.1, “Commands in the NDB Cluster Management Client”.

NDB Cluster supports the same programming APIs and languages as the standard MySQL Server, including ODBC, .Net, the MySQL C API, and numerous drivers for popular scripting languages such as PHP, Perl, and Python. NDB Cluster applications written using these APIs behave similarly to other MySQL applications; they transmit SQL statements to a MySQL Server (in the case of NDB Cluster, an SQL node), and receive responses containing rows of data. For more information about these APIs, see Chapter 31, Connectors and APIs.

NDB Cluster also supports application programming using the NDB API, which provides a low-level C++ interface to NDB Cluster data without needing to go through a MySQL Server. See The NDB API. In addition, many NDBCLUSTER management functions are exposed by the C-language MGM API; see The MGM API, for more information.

NDB Cluster also supports Java application programming using ClusterJ, which supports a domain object model of data using sessions and transactions. See Java and NDB Cluster, for more information.

NDB Cluster 8.0 also includes adapters supporting NoSQL applications written against Node.js, with NDB Cluster as the data store. See MySQL NoSQL Connector for JavaScript, for more information.

NDB Cluster includes a command line client for performing basic management functions. See Section 25.5.5, “ndb_mgm — The NDB Cluster Management Client”, and Section 25.6.1, “Commands in the NDB Cluster Management Client”.

NDB Cluster is also supported by MySQL Cluster Manager, a separate product providing an advanced command line interface that can automate many NDB Cluster management tasks such as rolling restarts and configuration changes. For more information about MySQL Cluster Manager, see MySQL Cluster Manager 8.4.0 User Manual.

There are two ways in which this can be done:

Yes. For tables created with the NDB storage engine, transactions are supported. Currently, NDB Cluster supports only the READ COMMITTED transaction isolation level.

NDB Cluster requires the NDB storage engine. That is, in order for a table to be shared between nodes in an NDB Cluster, the table must be created using ENGINE=NDB (or the equivalent option ENGINE=NDBCLUSTER).

It is possible to create tables using other storage engines (such as InnoDB or MyISAM) on a MySQL server being used with NDB Cluster, but since these tables do not use NDB, they do not participate in clustering; each such table is strictly local to the individual MySQL server instance on which it is created.

NDB Cluster is quite different from InnoDB clustering with regard to architecture, requirements, and implementation; despite any similarity in their names, the two are not compatible. For more information about InnoDB clustering, see MySQL AdminAPI. See also Section 25.2.6, “MySQL Server Using InnoDB Compared with NDB Cluster”, for information about the differences between the NDB and InnoDB storage engines.

All committed transactions are logged. Therefore, although it is possible that some data could be lost in the event of a catastrophe, this should be quite limited. Data loss can be further reduced by minimizing the number of operations per transaction. (It is not a good idea to perform large numbers of operations per transaction in any case.)

FULLTEXT indexing is currently supported only by the InnoDB and MyISAM storage engines. See Section 14.9, “Full-Text Search Functions”, for more information.

It is possible but not always advisable. One of the chief reasons to run a cluster is to provide redundancy. To obtain the full benefits of this redundancy, each node should reside on a separate machine. If you place multiple nodes on a single machine and that machine fails, you lose all of those nodes. For this reason, if you do run multiple data nodes on a single machine, it is extremely important that they be set up in such a way that the failure of this machine does not cause the loss of all the data nodes in a given node group.

Given that NDB Cluster can be run on commodity hardware loaded with a low-cost (or even no-cost) operating system, the expense of an extra machine or two is well worth it to safeguard mission-critical data. It also worth noting that the requirements for a cluster host running a management node are minimal. This task can be accomplished with a 300 MHz Pentium or equivalent CPU and sufficient RAM for the operating system, plus a small amount of overhead for the ndb_mgmd and ndb_mgm processes.

It is acceptable to run multiple cluster data nodes on a single host that has multiple CPUs, cores, or both. The NDB Cluster distribution also provides a multithreaded version of the data node binary intended for use on such systems. For more information, see Section 25.5.3, “ndbmtd — The NDB Cluster Data Node Daemon (Multi-Threaded)”.

It is also possible in some cases to run data nodes and SQL nodes concurrently on the same machine; how well such an arrangement performs is dependent on a number of factors such as number of cores and CPUs as well as the amount of disk and memory available to the data node and SQL node processes, and you must take these factors into account when planning such a configuration.

It is possible to add new data nodes to a running NDB Cluster without taking the cluster offline. For more information, see Section 25.6.7, “Adding NDB Cluster Data Nodes Online”.

For other types of NDB Cluster nodes, a rolling restart is all that is required (see Section 25.6.5, “Performing a Rolling Restart of an NDB Cluster”).

Limitations on NDB tables in MySQL NDB Cluster include the following:

For a complete listing of limitations in NDB Cluster, see Section 25.2.7, “Known Limitations of NDB Cluster”. See also Section 25.2.7.11, “Previous NDB Cluster Issues Resolved in NDB Cluster 8.4”.

NDB Cluster provides support for foreign key constraints which is comparable to that found in the InnoDB storage engine; see Section 1.7.3.2, “FOREIGN KEY Constraints”, for more detailed information, as well as Section 15.1.20.5, “FOREIGN KEY Constraints”. Applications requiring foreign key support should use NDB Cluster 7.3, 7.4, 7.5, or later.

You can import databases into NDB Cluster much as you would with any other version of MySQL. Other than the limitations mentioned elsewhere in this FAQ, the only other special requirement is that any tables to be included in the cluster must use the NDB storage engine. This means that the tables must be created with ENGINE=NDB or ENGINE=NDBCLUSTER.

It is also possible to convert existing tables that use other storage engines to NDBCLUSTER using one or more ALTER TABLE statement. However, the definition of the table must be compatible with the NDBCLUSTER storage engine prior to making the conversion. In MySQL 8.4, an additional workaround is also required; see Section 25.2.7, “Known Limitations of NDB Cluster”, for details.

Cluster nodes can communicate through any of three different transport mechanisms: TCP/IP, SHM (shared memory), and SCI (Scalable Coherent Interface). Where available, SHM is used by default between nodes residing on the same cluster host; however, this is considered experimental. SCI is a high-speed (1 gigabit per second and higher), high-availability protocol used in building scalable multi-processor systems; it requires special hardware and drivers. See Section 25.4.4, “Using High-Speed Interconnects with NDB Cluster”, for more about using SCI as a transport mechanism for NDB Cluster.

If one or more data nodes in a cluster fail, it is possible that not all cluster data nodes are able to “see” one another. In fact, it is possible that two sets of data nodes might become isolated from one another in a network partitioning, also known as a “split-brain” scenario. This type of situation is undesirable because each set of data nodes tries to behave as though it is the entire cluster. An arbitrator is required to decide between the competing sets of data nodes.

When all data nodes in at least one node group are alive, network partitioning is not an issue, because no single subset of the cluster can form a functional cluster on its own. The real problem arises when no single node group has all its nodes alive, in which case network partitioning (the “split-brain” scenario) becomes possible. Then an arbitrator is required. All cluster nodes recognize the same node as the arbitrator, which is normally the management server; however, it is possible to configure any of the MySQL Servers in the cluster to act as the arbitrator instead. The arbitrator accepts the first set of cluster nodes to contact it, and tells the remaining set to shut down. Arbitrator selection is controlled by the ArbitrationRank configuration parameter for MySQL Server and management server nodes. You can also use the ArbitrationRank configuration parameter to control the arbitrator selection process. For more information about these parameters, see Section 25.4.3.5, “Defining an NDB Cluster Management Server”.

The role of arbitrator does not in and of itself impose any heavy demands upon the host so designated, and thus the arbitrator host does not need to be particularly fast or to have extra memory especially for this purpose.

NDB Cluster supports all of the usual MySQL data types, including those associated with MySQL's spatial extensions; however, the NDB storage engine does not support spatial indexes. (Spatial indexes are supported only by MyISAM; see Section 13.4, “Spatial Data Types”, for more information.) In addition, there are some differences with regard to indexes when used with NDB tables.

See Section 25.2.7, “Known Limitations of NDB Cluster”, for more information about these issues.

It is necessary to start each node in the cluster separately, in the following order:

Each of these commands must be run from a system shell on the machine housing the affected node. (You do not have to be physically present at the machine—a remote login shell can be used for this purpose.) You can verify that the cluster is running by starting the NDB management client ndb_mgm on the machine housing the management node and issuing the SHOW or ALL STATUS command.

To shut down a running cluster, issue the command SHUTDOWN in the management client. Alternatively, you may enter the following command in a system shell:

$> ndb_mgm -e "SHUTDOWN"

(The quotation marks in this example are optional, since there are no spaces in the command string following the -e option; in addition, the SHUTDOWN command, like other management client commands, is not case-sensitive.)

Either of these commands causes the ndb_mgm, ndb_mgm, and any ndbd processes to terminate gracefully. MySQL servers running as SQL nodes can be stopped using mysqladmin shutdown.

For more information, see Section 25.6.1, “Commands in the NDB Cluster Management Client”, and Section 25.3.6, “Safe Shutdown and Restart of NDB Cluster”.

MySQL Cluster Manager provides additional ways to handle starting ansd stopping of NDB Cluster nodes. See MySQL Cluster Manager 8.4.0 User Manual, for more information about this tool.

The data that was held in memory by the cluster's data nodes is written to disk, and is reloaded into memory the next time that the cluster is started.

It can be helpful as a fail-safe. Only one management node controls the cluster at any given time, but it is possible to configure one management node as primary, and one or more additional management nodes to take over in the event that the primary management node fails.

See Section 25.4.3, “NDB Cluster Configuration Files”, for information on how to configure NDB Cluster management nodes.

Yes, as long as all machines and operating systems have the same “endianness” (all big-endian or all little-endian).

It is also possible to use software from different NDB Cluster releases on different nodes. However, we support such use only as part of a rolling upgrade procedure (see Section 25.6.5, “Performing a Rolling Restart of an NDB Cluster”).

Yes, it is possible to do this. In the case of multiple data nodes, it is advisable (but not required) for each node to use a different data directory. If you want to run multiple SQL nodes on one machine, each instance of mysqld must use a different TCP/IP port.

Running data nodes and SQL nodes together on the same host is possible, but you should be aware that the ndbd or ndbmtd processes may compete for memory with mysqld.

Yes, it is possible to use DNS and DHCP for cluster hosts. However, if your application requires “five nines” availability, you should use fixed (numeric) IP addresses, since making communication between Cluster hosts dependent on services such as DNS and DHCP introduces additional potential points of failure.

IPv6 is supported for connections between SQL nodes (MySQL servers), but connections between all other types of NDB Cluster nodes must use IPv4.

In practical terms, this means that you can use IPv6 for replication between NDB Clusters, but connections between nodes in the same NDB Cluster must use IPv4. For more information, see Section 25.7.3, “Known Issues in NDB Cluster Replication”.

MySQL user accounts and privileges are normally not automatically propagated between different MySQL servers accessing the same NDB Cluster. MySQL NDB Cluster provides support for shared and synchronized users and privileges using the NDB_STORED_USER privilege; see Distributed Privileges Using Shared Grant Tables, for more information. You should be aware that this implementation was introduced in NDB 8.0, and is not compatible with the shared privileges mechanism employed in earlier versions of NDB Cluster. The older implementation is no longer supported in NDB 8.0.

MySQL NDB Cluster does not provide any sort of automatic failover between SQL nodes. Your application must be prepared to handle the loss of SQL nodes and to fail over between them.

You can use the NDB Cluster native backup and restore functionality in the NDB management client and the ndb_restore program. See Section 25.6.8, “Online Backup of NDB Cluster”, and Section 25.5.23, “ndb_restore — Restore an NDB Cluster Backup”.

You can also use the traditional functionality provided for this purpose in mysqldump and the MySQL server. See Section 6.5.4, “mysqldump — A Database Backup Program”, for more information.

This process monitors and, if necessary, attempts to restart the data node process. If you check the list of active processes on your system after starting ndbd, you can see that there are actually 2 processes running by that name, as shown here (we omit the output from ndb_mgmd and ndbd for brevity):

$> ./ndb_mgmd

$> ps aux | grep ndb
me      23002  0.0  0.0 122948  3104 ?        Ssl  14:14   0:00 ./ndb_mgmd
me      23025  0.0  0.0   5284   820 pts/2    S+   14:14   0:00 grep ndb

$> ./ndbd -c 127.0.0.1 --initial

$> ps aux | grep ndb
me      23002  0.0  0.0 123080  3356 ?        Ssl  14:14   0:00 ./ndb_mgmd
me      23096  0.0  0.0  35876  2036 ?        Ss   14:14   0:00 ./ndbmtd -c 127.0.0.1 --initial
me      23097  1.0  2.4 524116 91096 ?        Sl   14:14   0:00 ./ndbmtd -c 127.0.0.1 --initial
me      23168  0.0  0.0   5284   812 pts/2    R+   14:15   0:00 grep ndb

The ndbd process showing 0.0 for both memory and CPU usage is the angel process (although it actually does use a very small amount of each). This process merely checks to see if the main ndbd or ndbmtd process (the primary data node process which actually handles the data) is running. If permitted to do so (for example, if the StopOnError configuration parameter is set to false), the angel process tries to restart the primary data node process.