Underneath every DAG is a Windows failover cluster. Failover clusters use the concept of quorum, which uses a consensus of voters to ensure that only one subset of the cluster members (which could mean all members or a majority of members) is functioning at one time. Quorum isn't a new concept for Exchange 2016. Highly available Mailbox servers in previous versions of Exchange also use failover clustering and its concept of quorum. Quorum represents a shared view of members and resources, and the term quorum is also used to describe the physical data that represents the configuration within the cluster that's shared between all cluster members. As a result, all DAGs require their underlying failover cluster to have quorum. If the cluster loses quorum, all DAG operations terminate and all mounted databases hosted in the DAG dismount. In this event, administrator intervention is required to correct the quorum problem and restore DAG operations.
Quorum is important to ensure consistency, to act as a tie-breaker to avoid partitioning, and to ensure cluster responsiveness:
- Ensuring consistency A primary requirement for a Windows failover cluster is that each of the members always has a view of the cluster that's consistent with the other members. The cluster hive acts as the definitive repository for all configuration information relating to the cluster. If the cluster hive can't be loaded locally on a DAG member, the Cluster service doesn't start, because it isn't able to guarantee that the member meets the requirement of having a view of the cluster that's consistent with the other members.
- Acting as a tie-breaker A quorum witness resource is used in DAGs with an even number of members to avoid split brain syndrome scenarios and to make sure that only one collection of the members in the DAG is considered official. When the witness server is needed for quorum, any member of the DAG that can communicate with the witness server can place a Server Message Block (SMB) lock on the witness server's witness.log file. The DAG member that locks the witness server (referred to as the locking node) retains an additional vote for quorum purposes. The DAG members in contact with the locking node are in the majority and maintain quorum. Any DAG members that can't contact the locking node are in the minority and therefore lose quorum.
- Ensuring responsiveness To ensure responsiveness, the quorum model makes sure that, whenever the cluster is running, enough members of the distributed system are operational and communicative, and at least one replica of the cluster's current state can be guaranteed. No additional time is required to bring members into communication or to determine whether a specific replica is guaranteed.
DAGs with an even number of members use the failover cluster's Node and File Share Majority quorum mode, which employs an external witness server that acts as a tie-breaker. In this quorum mode, each DAG member gets a vote. In addition, the witness server is used to provide one DAG member with a weighted vote (for example, it gets two votes instead of one). The cluster quorum data is stored by default on the system disk of each member of the DAG, and is kept consistent across those disks. However, a copy of the quorum data isn't stored on the witness server. A file on the witness server is used to keep track of which member has the most updated copy of the data, but the witness server doesn't have a copy of the cluster quorum data. In this mode, a majority of the voters (the DAG members plus the witness server) must be operational and able to communicate with each other to maintain quorum. If a majority of the voters can't communicate with each other, the DAG's underlying cluster loses quorum, and the DAG will require administrator intervention to become operational again.
DAGs with an odd number of members use the failover cluster's Node Majority quorum mode. In this mode, each member gets a vote, and each member's local system disk is used to store the cluster quorum data. If the configuration of the DAG changes, that change is reflected across the different disks. The change is only considered to have been committed and made persistent if that change is made to the disks on half the members (rounding down) plus one. For example, in a five-member DAG, the change must be made on two plus one members, or three members total.
Quorum requires a majority of voters to be able to communicate with each other. Consider a DAG that has four members. Because this DAG has an even number of members, an external witness server is used to provide one of the cluster members with a fifth, tie-breaking vote. To maintain a majority of voters (and therefore quorum), at least three voters must be able to communicate with each other. At any time, a maximum of two voters can be offline without disrupting service and data access. If three or more voters are offline, the DAG loses quorum, and service and data access will be disrupted until you resolve the problem.