Hermes replication
Hermes is a replication protocol that combines invalidations and per-key logical timestamps to provide linearizable operations on in-memory data.
In the lease-based design, local reads require a Valid key and an operational replica with a membership lease. An invalid key stalls reads. Any replica can coordinate a write, but it needs acknowledgements from every other member of the current live group, not just a majority. [2]
In the base failure-free path, the coordinator can reply after one round trip; validation messages still follow. Invalidation already carries the new value and timestamp. A blocked replica can replay that value with the original timestamp, preserving the write order. [1]
Failure handling: a missing acknowledgement is not permission to remove a replica locally. Reliable membership reconfiguration follows lease expiry; minority partitions stop serving requests. Ordinary writes avoid conflict aborts, but read-modify-write operations can abort. [1]
Timing scope: one round trip is measured from a replica, not an external client. The clock-free variant waits for majority-backed membership confirmation before returning speculative reads. [1]
Artifact boundary: the published implementation uses an RDMA cluster. Its protocol-action diagram includes optimizations beyond the paper, including starting writes from Invalid state; do not treat the diagram as the base state machine. [3] [4]
Relations
- related
- CRAQ, Linearizability, Quorum lease
Limitations
- Single-key linearizability does not supply multi-key transactions or disk persistence; the stated failure model is non-Byzantine crash-stop and network faults. [1]
Citations
- Katsarakis et al. Hermes: a Fast, Fault-Tolerant and Linearizable Replication Protocol. ASPLOS 2020, sections 2.4, 3, 5 and 8.
- Hermes research project: protocol properties and read authority
- Hermes artifact README, revision 949229c23881: implementation and hardware requirements
- Hermes formal-specification README, revision 949229c23881: diagram optimizations
Sources & caveats
- Source synthesis of the ASPLOS 2020 paper and pinned research artifact; no native RDMA experiment or model-checker run was performed.
- The paper's benchmark uses in-memory key-value stores and protocol-specific configurations; its throughput is not evidence of persistent-storage performance. [1]