Store-liveness leases

Store-liveness is CockroachDB's newer fabric for replica leadership and range leases. It replaces a single, hot "liveness range" with a per-store gossiped fence, decoupling lease and leadership decisions from a global hot spot.

What changes

• Old (epoch leases) — every node holds an epoch in the liveness range; if a peer's epoch advances, all leases held by the previous epoch are forfeited. This requires every node to repeatedly read and write the liveness range.
• New (store-liveness) — every store advertises a "support" vector to its peers. A store can claim leadership only as long as a quorum of peers continues to support its current epoch. Heartbeats are local store-to-store, not global.

Code

• pkg/kv/kvserver/storeliveness/ — the fabric: vouchers, support exchange, durable per-store records.
• pkg/raft/raftstoreliveness/ — Raft adapter that asks the fabric whether a leader's epoch is still supported.
• pkg/kv/kvserver/leases/ — the in-tree lease types (epoch-based vs store-liveness-based).
• pkg/kv/kvserver/replica_store_liveness.go and replica_store_liveness_sleep.go — replica integration.

Migration story

Both lease kinds coexist in mixed-version clusters. A cluster version gate in pkg/clusterversion/ decides when a range can transition to store-liveness leases. The replicate_queue and the lease_queue cooperate to upgrade ranges as soon as it is safe.

Related pages

• Raft — leader-election integration.
• KV server — replica/leases.
• Cluster version & upgrade — gating and migration.