If a cluster is fragmented into two (or more) parts, the data (disk and tape) integrity must be protected. For instance, a database application may be running on separate CPUs at two data centers several miles apart for safety reasons. The common database is stored in a volume shadow set configuration, RAID 1, with some disks at each site. In normal operation, both sites interrogate and update the database, and the changes to the database are duplicated (or mirrored) in both places. If the link between the sites is disrupted, however, one data center must shut itself down so that the composite database does not diverge.
Deciding which data center (or which CPU) is to control the database is done with a quorum mechanism. This is how it works: Each file server node in the clusters has a vote, but satellites do not. The system manager determines the minimum number of votes (the quorum) possible for continued cluster operation. A cluster component called the Quorum Watcher on each node monitors the cluster quorum and signals a shutdown when the cluster drops below the minimum number of votes. It is essential that the manager understand this algorithm. It is detailed in all of the cluster manuals.