Partition tolerance is a distributed system’s ability to handle disconnected nodes or network partitions. Under the CAP theorem, partition tolerance is treated as necessary, so a system trades between consistency and availability. section 1.2.1 section 8.1
ScyllaDB is classified as an AP system: during a partition, it favors availability over strict consistency. Replication and tunable consistency levels let requests succeed with some replicas unavailable, while eventual consistency allows data to converge later. section 1.2.1 section 1.4
Consistency levels determine how many replicas must respond:
- ALL requires every replica and tolerates no replica failure. section 1.2.1
- QUORUM requires a majority, balancing consistency and availability while tolerating one failure in a three-replica cluster. section 1.2.1
- ONE requires a single replica, maximizing availability while accepting potentially weaker consistency. section 1.2.1
A three-replica cluster
If one node becomes disconnected, QUORUM can still succeed because two replicas respond. ONE can continue with either available replica, while ALL fails until every replica responds. section 1.2.1
Table 1.1 Sample of consistency level options, assuming a cluster with three replicas
|
Consistency level
|
Description
|
Number required to succeed
|
Failures tolerated
|
|---|---|---|---|
ALL
|
Requires all replicas to succeed
|
3
|
0
|
QUORUM
|
Requires a majority of replicas to succeed
|
2
|
1
|
ONE
|
Requires a single replica to succeed
|
1
|
2
|
Compare replica requirements and tolerated failures for ALL, QUORUM, and ONE. section 1.2.1
