Valkey, the Linux Foundation-backed open-source Redis alternative, has announced the general availability of Valkey 9.2, introducing forkless snapshots, faster sorted-set operations, and replication compression.
One of the most significant changes is the introduction of forkless snapshots. Traditionally, creating snapshots on busy instances involves spawning a separate process, causing substantial temporary memory increases due to copy-on-write operations. In some cases, memory usage even nearly doubles, forcing administrators to provision additional RAM to accommodate backups.
Valkey 9.2 deals with this by performing snapshots through a background thread instead of a separate process. According to the project, this eliminates temporary memory spikes from traditional snapshots, letting operators size infrastructure based on actual workloads rather than worst-case backup needs. The improvement also applies when preparing snapshots to synchronize new replicas.
Another important change concerns sorted sets, commonly used for leaderboards, rankings, and priority queues. Valkey 9.2 uses a B+ tree-based architecture for large sorted sets, delivering up to 105% faster updates and 36% to 43% better memory usage efficiency in tests with millions of entries.
On the replication side, new continuous stream compression reduces ongoing replication traffic by 50% to 65%, while the built-in LZ4 compression option can shrink backup files by up to 80%. The project also reports that new replicas synchronize about 35% faster, with ongoing compression requiring 2% to 4% of a CPU core per replica.
In addition to these improvements, Valkey 9.2 introduces Path Hash, a new native data type designed for efficient hierarchical lookups. It allows applications to find the longest matching prefix and identify the location of cached data through a single command.
The feature is especially relevant to AI inference systems, where previously computed context can be reused instead of processed repeatedly. Instead of storing large AI cache objects directly in Valkey, Path Hash serves as a lightweight index pointing to data maintained in GPU or host memory.
Alongside Valkey 9.2, the project has also announced a new open-source distributed key-value caching capability for AI workloads. It enables inference engines running across multiple GPU servers to share previously computed context, cutting down on redundant processing.
The integration initially involves LMCache, which works with inference engines such as vLLM and SGLang. According to the announcement, distributed KV caching can reduce AI inference costs by up to 40% and time-to-first-token latency by up to 80%, depending on the workload.
Importantly, this distributed caching capability is a separate development announced alongside Valkey 9.2, rather than simply another built-in feature of the datastore.
For more information about Valkey 9.2 and its availability, visit the release notes.
