6.3.8. Elastic Cluster Operations
When a set of Virtuoso servers configured as a cluster is first started, the elastic cluster setup is automatically performed if the cluster.ini of each host contains the [ELASTIC] section as described below. This is the normal way in which an elastic cluster is initialized.
The __elastic option of the create cluster statement is used for creating an elastic logical cluster.
create cluster ELASTIC __elastic 2048 4 group (Host1), group (Host2), group (Host3), group (Host4);
create cluster ELASTIC __elastic 2048 4 ALL;
This creates an elastic cluster called ELASTIC which initially consists of 4 host groups. Each host group has a single copy of the data allocated to the host group. The data is partitioned among the host groups so that each host group has a separate fraction of the data. The second variant defaults the host groups from the cluster.ini, so that each host mentioned is in a host group of its own. The first number after the __elastic clause is the maximum number of physical slices in the cluster. The second number is the number of physical slices initially created per host group. In general each host group should have as many slices as there are hardware threads on the CPU (s) running the host(s) in the host group.
The name ELASTIC is specially known. If a logical cluster of this name exists, it is the default logical cluster instead of the non-elastic __ALL cluster in create or alter statements that specify partitioning.
Database files for an elastic cluster are created in a directory and striping set that is specified in the cluster.ini file of each host. The cluster.ini file of each participating host must have this section at the time of executing the create cluster statement. The section of the ini file has the name as in the create cluster statement, e.g.:
[ELASTIC] Slices = 16 Segment1 = 1024, /data6/dbs/btc1-el.db = q1 Segment2 = 1024, /data2/dbs/btc1-el.db = q2 Segment3 = 1024, /data8/btc1-1/btc1-el.db = q3 [ELASTIC_1] Segment1 = 1024, /data2/dbs/btc1-d-el.db = q2
Note that an elastic cluster may specify multiple segments. Each segment has an independent striping, as in other segments. When initialized, one segment is created for each of the slices hostted by the host in question. The segment may consist of multiple files if striping is specified. The database files for the segments are named as per the path and file name in the segment clause, with the slice number appended at the end, prefixed with a '.'. The files are initially created in the first segment of the logical cluster. If one wishes to move the files, one can do so between the segments declared for the cluster when the server is not running. At the next startup, it will look in all the segments for slices matching the path name plus .[0-9]*.
Specifying distinct IO queues for distinct devices is useful, as with other file system layouts.
To move slices between servers, more steps are necessary, see following sections.
In the example at hand, we specify two sets of files, one labeled ELASTIC_1. This is useful if one wishes to divide objects among different types of storage, e.g. disk and ssd. If two cluster names differ only by having _[0-9]* tagged at the end or having _[0-9]* with different numbers at the end while the rest of the name is the same, the system will manage these as twins so as to keep the slices belonging to either collocated. So slice movement and splitting will be mirrored between the two. If an If two objects share a partitioning key between such twin clusters operations are assumed to be colocatable.
The Slices setting specifies how many physical slices are initially created per host group. For example if a server has 2 CPU sockets with a 8 core 16 thread CPU on each, and two Virtuoso processes are run on each server, one per NUMA node (CPU socket), it is practical to specify 16 slices per host group. In this way there will be up to 32 threads per query on the machine, leading to full platform utilization.
After creation, the cluster can be used in the cluster clause of create index and alter index for specifying that the object in question is to be created in the elastic cluster.
create table TE ( row_no int, string1 varchar, string2 varchar, primary key (row_no)); alter index TE on TE partition cluster ELASTIC (row_no int (0hexffff00));
The table TE is now allocated in the cluster ELASTIC. Due to default behavior, the cluster ELASTIC is redundant since objects preferentially get created in it if it exists. The table is now ready for use. Note that the bit mask in the partitioning clause of the partitioning key row_no must specify enough bits to cover the number of logical slices in the cluster, 2048 in the present case. If this is not the case, some logical slices will never be used, as there will be less distinct partitioning hash values than logical slices.
Adding and Removing Hosts
The number of server processes (hosts) used for managing an elastic cluster may vary during the cluster's lifetime. If the elastic cluster keeps partitions in multiple copies, i.e. host groups in the initial create cluster statement have more than one host, any new host groups that are added must have the same number of hosts. Each new host of a new host group must be identically initialized to contain a copy of the REPLICATED logical cluster of the cluster into which the hosts are added. This is done by simply copying the non-elastic database file(s) of any host of the existing cluster. The copy can be made while the cluster is running.
cl_add_host_group (in cl_name varchar, in hosts any array) returns int
This function adds a host group to an elastic cluster. The host group must contain the same number of hosts as all the other host groups and the hosts must not be part of any other host group in this cluster.
The new hosts must be assigned IP addresses in the cluster.ini file of all presently running hosts. At the time of being added, the hosts in the new host group must be running and must be initialized to share a copy of the REPLICATED cluster shared by all hosts of the physical cluster.
After successful addition:
cl_remove_host_group (in cl_name varchar, in host varchar)