Quorum and epsilon are important measures of cluster health and function that together indicate how clusters address potential communications and connectivity challenges.
Quorum is a precondition for a fully functioning cluster. When a cluster is in quorum, a simple majority of nodes are healthy and can communicate with each other. When quorum is lost, the cluster loses the ability to accomplish normal cluster operations. Only one collection of nodes can have quorum at any one time because all of the nodes collectively share a single view of the data. Therefore, if two non-communicating nodes are permitted to modify the data in divergent ways, it is no longer possible to reconcile the data into a single data view.
Each node in the cluster participates in a voting protocol that elects one node master; each remaining node is a secondary. The master node is responsible for synchronizing information across the cluster. When quorum is formed, it is maintained by continual voting; if the master node goes offline, a new master is elected by the nodes that remain online.
Because there is the possibility of a tie in a cluster that has an even number of nodes, one node has an extra fractional voting weight called epsilon. When the connectivity between two equal portions of a large cluster fails, the group of nodes containing epsilon maintains quorum, assuming that all of the nodes are healthy. For example, if a single link is established between 12 nodes in one row and 12 nodes in another row to compose a 24-node cluster and the link fails, then the group of nodes that holds epsilon would maintain quorum and continue to serve data while the other 12 nodes would stop serving data. However, if the node holding epsilon was unhealthy or offline, then quorum would not be formed, and all of the nodes would stop serving data.
Epsilon is automatically assigned to the first node when the cluster is created. As long as the cluster is in quorum, if the node that holds epsilon becomes unhealthy or is taken over by its high-availability partner, epsilon is automatically reassigned to a healthy node.
In general, assuming reliable connectivity among the nodes of the cluster, a larger cluster is more stable than a smaller cluster. The quorum requirement of a simple majority of half the nodes plus epsilon is easier to maintain in a cluster of 24 nodes than in a cluster of two nodes.
A two-node cluster presents some unique challenges for maintaining quorum. In a two-node cluster, neither node holds epsilon; instead, both nodes are continuously polled to ensure that if one node fails, the other has full read-write access to data, as well as access to logical interfaces and management functions.