1. Administrator's Guide
  2. Administer
  3. Availability, Failover and Switchover
  4. Unplanned Network Partitions
  5. Master is in the Majority Node Partition

Master is in the Majority Node Partition

Suppose the shard is divided into two partitions. Partition A contains a simple majority of the Replication Nodes in primary zones, including the master. Partition B has the remaining nodes.

  • Partition A continues to service read and write requests as normal, but with a reduced read throughput from the loss of however many Replication Nodes are in Partition B. A caveat in this situation is what durability policy is in use at the time. If Partition A does not have enough replicas from primary zones to meet the durability policy requirements, it could be prevented from servicing write requests. If the durability policy requires a simple majority, or less, of replicas, then the shard will be able to service write requests.

  • Partition B continues to service read requests as normal, but with increasingly stale data. Depending on the consistency guarantee in place, Partition B might cease to service read requests. If a version-based consistency is in use, then Partition B will probably encounter ConsistencyException exceptions soon after the network partition occurs, due to its inability to obtain version tokens from the master. Similarly, if a time-based consistency policy is in use, then ConsistencyException exceptions will occur as soon as the replica lags too far behind the master, from which it is no longer receiving write updates. By default, a consistency guarantee is not required to service read requests. So unless you explicitly create and use a consistency policy, Partition B can continue to service read requests through the entire network outage.

    Partition B will attempt to elect a new master, but will be unable to do so because it does not contain the simple majority of Replication Nodes required to hold an election.

Further, if the partition is such that your client code can reach Partition A but not Partition B, then the shard will continue to service read and write requests as normal, but with a reduced read capacity.

However, if the partition is such that your client code can read Partition B but not Partition A, then the shard will be unable to service any write requests. This is because Partition A contains the master, and Partition B does not include enough Replication Nodes to elect a new master.