The Pulse
Valkey 9.2 adds backup and replication features alongside AI cache
Valkey 9.2 is generally available with forkless snapshots, replication compression and a new data type for finding cached AI context. A separate distributed caching capability released alongside it is designed to reuse computed work across GPU fleets.

Valkey 9.2 is generally available with new backup, replication and data-structure features aimed at reducing infrastructure demands. A separate open-source distributed key-value caching capability released alongside it lets AI inference engines reuse computed work across GPU fleets, the Linux Foundation said on Nov. 19.
Source linuxfoundation.org
The release updates an in-memory datastore used for workloads including caching, message queues and real-time data. The accompanying cache is intended for AI applications that repeatedly process material such as prompt instructions, tool definitions and conversation history.
Sharing computed AI context
The Linux Foundation says the distributed cache can reduce AI inference costs by up to 40% and time to first token by up to 80%. The capability creates a shared cache across a GPU fleet so inference engines can reuse computed work rather than recompute it.
Builders running open-weight models with self-managed inference can connect the cache through LMCache, which plugs into inference engines including vLLM and SGLang. The foundation says the setup uses a single configuration change and requires no application code updates.
Multiverse Computing, which provides compressed models for inference, said it processed 50% more requests on the same GPUs within its one-second target for median time to first token after adding distributed caching on Valkey. It reported a 23% reduction in median time to first token at normal traffic and a 95% reduction at peak traffic.
For finding cached context, Valkey 9.2 adds Path Hash, a native data type that serves as an index. A lookup can find the longest matching prefix and the location of its cached results, while the larger AI data stays in GPU or host memory.
Path Hash also supports uses beyond AI, including hierarchical data lookups. Alibaba Cloud contributed the data type, according to the supporting statement from Guoyun Zhu, head of Tair R&D at the company.
Backups without a separate process
Valkey 9.2 also introduces forkless snapshots, which save data on a background thread instead of creating a separate process. The change targets the temporary memory overhead that snapshots can cause on busy systems.
The feature is opt-in, according to the project's release post. Enabling its infrastructure costs 4 additional bytes per key and must be configured at server startup. A prefer-forkless policy falls back to a fork when an operation cannot run without one.
Full synchronization to replicas can also run without forking, using a background-thread snapshot for disk-based and diskless full syncs. The release post says the approach avoids copy-on-write memory overhead that a fork can incur under write-heavy load.
Compression and smaller data structures
Continuous stream compression is available for replication traffic. The foundation says ongoing traffic is reduced by roughly 50% to 65%, while backup files shrink by up to 80% using the built-in LZ4 option. It also reports new replicas syncing about 35% faster, at a cost of about 2% to 4% of a CPU core per replica.
In a separate benchmark described in the project post, LZ4 reduced replication bandwidth by 52% with less than 1% throughput overhead across 3 million keys of about 315 bytes each. Compression is negotiated per replica, so older replicas or replicas that opt out continue receiving plaintext.
Large sorted sets now use a B+ tree in place of a skiplist. The project reports a 36% to 43% reduction in memory overhead per member in a test of 5 million members with 20-byte elements; it says the tree also compacts sparse leaves when members are deleted, so a sorted set that shrinks gives memory back.
The release also adds reusable access-control roles, server-side hot key detection and optional replication throttling. The throttling feature slows client writes when a replica falls behind, rather than allowing it to disconnect and require a full synchronization.
Sources
- Original source · Press release The Linux Foundation linuxfoundation.org
Other pages read
- Valkey valkey.io
Related stories
- CleanStart report flags risks in AI-recommended software dependencies
- Open Agentic Web Challenge offers $13,000 for interoperable AI tools
- Cerebras puts split AI inference into production with AMD GPUs
- AMD details GPU-initiated DMA transfers for multi-GPU workloads
- NVIDIA adds AI workload inventory to GPU platform runtime