For read operations, the following three consistency levels are possible:
- One: The lowest latency, but dirty reads are possible
- Quorum: The middle path
- All: The slowest, but data is always good
For write operations, the following three consistency levels are possible:
- One: The fastest
- Quorum: The middle path
- All: The slowest, but data is guaranteed to be written to all the replicas
Cassandra has more fine-grained consistency levels for both read and write, but the aforementioned three are mostly used.
Consistency is a combination of the write and read, and you can have any combination of the two, for example:
- write one, read one
- write one, read quorum
- write quorum, read quorum (recommended)