etcd

etcd is a strongly consistent distributed key-value store for data that must be shared across a cluster of machines. It handles leader elections during network partitions, tolerates machine failure including failure of the leader, distributes state using the Raft protocol, and exposes a simple interface that can be driven with ordinary HTTP tools. It is a CNCF project and is benchmarked at thousands of writes per second per instance.

Strong consistency is the deliberate choice that separates it from registries built for availability. etcd would rather refuse a write than return an answer that might be wrong, which is exactly what you want from the store holding cluster state and exactly the wrong trade for a cache. Knowing which of those two a system needs is the whole of the selection decision.

Its position under Kubernetes makes it consequential far beyond its own user base. Every Kubernetes cluster stores its entire state here, so etcd's backup, restore and disk latency characteristics become the cluster's characteristics, and an administrator who has never chosen it still has to operate it.

The feature set stays deliberately small. Optional TTLs for key expiry, watches on keys and directories and optional client certificate authentication are close to the whole surface, which is appropriate for something that must be dependable rather than capable.