Since I recently configured and installed a MySQL-cluster, I thought I’d share the procedure. A lot of the examples around explains how to set it all up on the same machine for “testing purposes” — which, in theory, is the same as setting it up on different machines. I’ll be explaining the latter, that is, installing it onto different machines.
To achieve true redundancy in a MySQL-cluster, you need at least 3 seperate, physical machines; two data-nodes, and one management-node. The latter you can use a virtual machine for, as long as it doesn’t run on the two data-nodes (which means you still need at least 3 physical machines). You can also use the management-node as a mysql-proxy for transparent failover/load-balancing for the clients. My setup was done using two physical machines (db0 and db1) running Ubuntu 8.04 (Hardy Heron), and one virtual machine (mysql-mgmt) running Debian 6 (Squeeze). The VM is not running on the two physical machines. db0 and db1 is the actual data-nodes/servers, and mysql-mgmt is going to be used as the management-node for the cluster. In addition, mysql-mgmt is also going to be configured with mysql-proxy, so that we have transparent failover/load-balancing for the clients.
Update 2011-10-26: I?ve changed the setup a bit, compared to my original walkthrough. I hit some memory-limits when using the NDB-engine. This caused MySQL to fail inserting new rows (stating that the table was full). There are some variables that you can set (DataMemory and IndexMemory), to increase the memory-consumption for the ndb-process (which was what caused the issues). Since I had limited amount of memory available on the mysql-mgmt virtual machine (and lots on db0/1), I decided to run ndb_mgmd on db0 + db1. Apparently, you can do this, and it?s still redundant. The post has been changed to reflect this.
My setup was done using two physical machines (db0 and db1) running Ubuntu 8.04 (Hardy Heron), and one virtual machine (mysql-proxy) running Debian 6 (Squeeze). Previously, the virtual machine ran ndb_mgmd, but due to the above mentioned issues, both db0 and db1 runs their own ndb_mgmd-processes. The virtual machine is now only used to run mysql-proxy (and hence it’s hostname has changed to reflect this).
Update 2012-01-30: morphium pointed out that /etc/my.cnf needed it’s own [mysql_cluster]-section, so that ndbd and ndb_mgmd connects to something else than localhost (which is the default if no explicit hosts is defined). The post has been updated to reflect this.
Update 2013-01-03: Dave Weddell pointed out that newer versions of mysql-proxy had a different syntax for the proxy-backend-addresses parameter. Instead of having multiple proxy-backend-addresses-parameters (one for each backend), it wants all the backends in one parameter (comma separated). The post has been updated to reflect this.
Update 2013-05-15: Richard pointed out that he had to alter two more tables to use the ‘ndbcluster’-engine in order for it to work. It was not needed when I originally set this up, but recent versions might have introduced more tables. I’ve updated the post to reflect this.