Consistency versus availability

A database's consistency refers to the reliability of its functions' performance. A consistent system is one in which reads return the value of the last write, and reads at a given time epoch return the same value regardless of where they were initiated.

NoSQL databases support a range of consistency models, such as the following:

• Strong consistency: A system that is strongly consistent ensures that updates to a given key are ordered and reads reflect the latest update that has been accepted by the system
• Timeline consistency: A system that is timeline consistent ensures that updates to a given key are ordered in all the replicants, but reads at a given replicant might be stale and may not reflect the latest update that has been accepted by the system
• Eventual consistency: A system that is eventually consistent makes no guarantees about whether updates will be applied in order in all the replicants, nor does it make guarantees about when a read would reflect a prior update accepted by the system

A database's availability refers to the system's ability to complete a certain operation. Like consistency, availability is a spectrum. A system can be unavailable for writes while being available for reads. A system can be unavailable for admin operations while being available for data operations.

As is well known at this point, there's tension between consistency and availability. A system that is highly available needs to allow operations to succeed even if some nodes in the system are unreachable (either dead or partitioned off by the network). However, since it is unknown as to whether those nodes are still alive and are reachable by some clients or are dead and reachable by no one, there are no guarantees about whether those operations left the system in a consistent state or not.

So, a system that guarantees consistency must make sure that all of the nodes that contain data for a given key must be reachable and participate in the operation. The degenerate case is that a single node is responsible for operations on a given key. Since there is just a single node, there is no chance of inconsistency of the sort we've been discussing. The downside is that when a node goes down, there is a complete loss of availability for operations on that key.