Gryff
Gryff is a replication protocol that combines shared-register reads and writes with consensus for linearizable read-modify-write operations.
Gryff couples multi-writer ABD with EPaxos. Carstamps contain a logical timestamp, coordinator identifier and RMW counter. An RMW increments its base update's counter, placing it immediately after that update in the timestamp order; a concurrent write cannot be inserted between them. [1]
An RMW must execute at a quorum before completion, rather than at only one replica. Otherwise, a later register-path read could miss its effect. RMW execution still resolves EPaxos dependencies. [1]
With n = 2f + 1 replicas and reliable delivery, the paper proves read/write wait-freedom under up to f crashes. RMW completion additionally needs partial synchrony. Per-object linearizability does not turn separate reads of different keys into one atomic scan. [1] [2]
Relations
- related
- Pineapple, Linearizability, Compare-and-swap
Limitations
- The evaluation trades lower read-tail latency against higher write latency in its geo-replicated workloads; it is not an all-operation speedup. [1]
Citations
- Burke, Cheng and Lloyd. Gryff: Unifying Consensus and Shared Registers. NSDI 2020, pp. 591-617; sections 2-6 and Appendix B.
- Bantikyan et al. Pineapple: Unifying Multi-Paxos and Atomic Shared Registers. NSDI 2025, section 2.3.
Sources & caveats
- Source synthesis of the published protocol and its later comparison. No distributed experiment, benchmark replication or deployment claim.