Vertica uses the Spread service to broadcast control messages between database nodes. This service can limit the growth of a Vertica database cluster. As you increase the number of cluster nodes, the load on the Spread service also increases as more participants exchange messages. This increased load can slow overall cluster performance. Also, network addressing limits the maximum number of participants in the Spread service to 120 (and often far less). In this case, you can use large cluster to overcome these Spread limitations.
When large cluster is enabled, a subset of cluster nodes, called control nodes, exchange messages using the Spread service. Other nodes in the cluster are assigned to one of these control nodes, and depend on them for cluster-wide communication. Each control node passes messages from the Spread service to its dependent nodes. When a dependent node needs to broadcast a message to other nodes in the cluster, it passes the message to its control node, which in turn sends the message out to its other dependent nodes and the Spread service.
By setting up dependencies between control nodes and other nodes, you can grow the total number of database nodes, and remain in compliance with the Spread limit of 120 nodes.
Technically, when large cluster is disabled, all of the nodes in the cluster are control nodes. In this case, all nodes connect to Spread. When large cluster is enabled, some nodes become dependent on control nodes.
A downside of the large cluster feature is that if a control node fails, its dependent nodes are cut off from the rest of the database cluster. These nodes cannot participate in database activities, and Vertica considers them to be down as well. When the control node recovers, it re-establishes communication between its dependent nodes and the database, so all of the nodes rejoin the cluster.
The Spread service demon runs as an independent process on the control node host. It is not part of the Vertica process. If the Vertica process goes down on the node—for example, you use admintools to stop the Vertica process on the host—Spread continues to run. As long as the Spread demon runs on the control node, the node's dependents can communicate with the database cluster and participate in database activity. Normally, the control node only goes down if the node's host has an issue—or example, you shut it down, it becomes disconnected from the network, or a hardware failure occurs.
Large Cluster and Database Growth
When your database has large cluster enabled, Vertica decides whether to make a newly added node into a control or a dependent node as follows:
- In Enterprise Mode, if the number of control nodes configured for the database cluster is greater than the current number of nodes it contains, Vertica makes the new node a control node. In Eon Mode, the number of control nodes is set at the subcluster level. If the number of control nodes set for the subcluster containing the new node is less than this setting, Vertica makes the new node a control node.
- If the Enterprise Mode cluster or Eon Mode subcluster has reached its limit on control nodes, a new node becomes a dependent of an existing control node.
When a newly-added node is a dependent node, Vertica automatically assigns it to a control node. Which control node it chooses is guided by the database mode:
- Enterprise Mode database: Vertica assigns the new node to the control node with the least number of dependents. If you created fault groups in your database, Vertica chooses a control node in the same fault group as the new node. This feature lets you use fault groups to organize control nodes and their dependents to reflect the physical layout of the underlying host hardware. For example, you might want dependent nodes to be in the same rack as their control nodes. Otherwise, a failure that affects the entire rack (such as a power supply failure) will not only cause nodes in the rack to go down, but also nodes in other racks whose control node is in the affected rack. See Fault Groups for more information.
- Eon Mode database: Vertica always adds new nodes to a subcluster. Vertica assigns the new node to the control node with the fewest dependent nodes in that subcluster. Every subcluster in an Eon Mode database with large cluster enabled has at least one control node. Keeping dependent nodes in the same subcluster as their control node maintains subcluster isolation.
In versions of Vertica prior to 10.0.1, nodes in an Eon Mode database with large cluster enabled were not necessarily assigned a control node in their subcluster. If you have upgraded your Eon Mode database from a version of Vertica earlier than 10.0.1 and have large cluster enabled, realign the control nodes in your database. This process reassigns dependent nodes and fixes any cross-subcluster control node dependencies. See Realigning Control Nodes and Reloading Spread for more information.
Spread's upper limit of 120 participants can cause errors when adding a subcluster to an Eon Mode database. If your database cluster has 120 control nodes, attempting to create a subcluster fails with an error. Every subcluster must have at least one control node. When your cluster has 120 control nodes , Vertica cannot create a control node for the new subcluster. If this error occurs, you must reduce the number of control nodes in your database cluster before adding a subcluster.
When To Enable Large Cluster
Vertica automatically enables large cluster in two cases:
- The cluster contains 120 or more nodes.
- You create an Eon Mode subcluster with more than 16 nodes.
Typically, you manually enable large cluster before you reach either limit, when one of the following occurs:
- A cloud-based database cluster reaches 16 nodes in your database cluster.
- An on-premises database reaches 50 to 80 nodes.
- You begin to notice Spread-related performance issues in your database cluster, one of the following:
- The load on the spread service begins to cause performance issues. Because Vertica uses Spread for cluster-wide control messages, Spread performance issues can adversely affect database performance. This is particularly true for cloud-based databases, where Spread performance problems becomes a bottleneck sooner, due to the nature of network broadcasting in the cloud infrastructure. In on-premises databases, broadcast messages are usually less of a concern because messages usually remain within the local subnet. Even so, eventually, Spread usually becomes a bottleneck before Vertica automatically enables large cluster.
- The compressed list of addresses in your cluster is too large to fit in a maximum transmission unit (MTU) packet (1478 bytes). The MTU packet has to contain all of the addresses for the nodes participating in the Spread service. Under ideal circumstances (when your nodes have the IP addresses 220.127.116.11, 18.104.22.168 and so on) 120 addresses can fit in this packet. This is why Vertica automatically enables large cluster if your database cluster reaches 120 nodes. In practice, the compressed list of IP addresses will reach the MTU packet size limit at 50 to 80 nodes.