Valkey 9.2 Arrives with Distributed KV Caching and Forkless Backups to Cut AI Infrastructure Costs
The open-source in-memory datastore Valkey has officially launched version 9.2, introducing major performance upgrades designed to cut infrastructure and operational expenses for enterprise users. Governed by the Linux Foundation, the platform has rolled out features aimed at lowering AI inference costs, reducing cross-region network traffic, and eliminating memory doubling during backup processes.
Targeting AI Inference Spends and Latency
A major focus of the 9.2 release is addressing the financial burden of running large-scale artificial intelligence workloads. Distributed key-value caching, introduced alongside the core update as a separate open-source capability, tackles redundant prompt processing in agentic applications. By establishing a shared cache across an entire GPU fleet, inference engines can reuse computed work rather than repeatedly re-evaluating prompt instructions and conversation histories.
Early implementations illustrate significant operational benefits. Multiverse Computing reported handling fifty percent more requests on identical GPU configurations while meeting latency targets, alongside a peak-time reduction in time-to-first-token. Meanwhile, integration with engines such as vLLM and SGLang is supported through LMCache, utilizing Remote Direct Memory Access to move large chunks of data directly into GPU-serving nodes with minimal central processing unit overhead.
Enhancing Core Datastore Architecture
Version 9.2 also introduces Path Hash, a native data type functioning as a hierarchical index that points to cached AI data stored in host or GPU memory. Contributed by Alibaba Cloud, the feature replaces custom lookup code and acts as a general-purpose tool for matching hierarchical information.
To solve the memory-doubling issue traditionally associated with in-memory backups, the new release adopts a background thread approach for snapshots rather than relying on traditional Linux fork-based copy-on-write mechanisms. This keeps memory usage flat and allows operators to size hardware according to normal workloads rather than worst-case backup windows.
- Continuous stream compression reduces ongoing cross-region replication traffic significantly.
- Backup files are substantially smaller when utilizing the built-in LZ4 option.
- Rebuilt sorted sets on a B+ tree execute updates faster while improving memory efficiency for large datasets.
Security and Vendor-Neutral Governance
In addition to performance upgrades, the project has modernized its security operations by utilizing an adversarial testing pipeline powered by large language models to uncover potential vulnerabilities before human review. Furthermore, the platform shortened its default security disclosure embargo to accelerate patch deployment across supported branches.
With major contributors including Amazon Web Services, Alibaba Cloud, and Aiven, the project continues to champion open-source standards and vendor neutrality. As public sector platforms like cloud.gov transition away from proprietary alternatives to adopt Valkey, the community aims to provide a resilient, cost-effective standard for enterprise caching and data management.