News

Valkey 9.2 adds forkless snapshots and replication compression

Self-managed Valkey deployments can opt into background backups without starting a child process. The new Path Hash data type can help inference engines locate cached attention state.

D
Nov 21, 2026 · 2 min read

The open-source Valkey project announced general availability of Valkey 9.2 on November 19. Operators can opt to write a point-in-time backup on a background thread, avoiding the extra copies of memory pages that writes can trigger during a forked backup. They can also enable compression for traffic to compatible replicas.

Forkless operations are off by default. Setting forkless-infrastructure-enabled to yes at startup adds four bytes of metadata per key. Operators can then set forkless-policy to prefer-forkless at runtime. The policy covers snapshots and full replica syncs, with a fallback to a fork if, for example, a loaded module lacks support or both ends enable replication compression for a full sync. Valkey says writes can briefly wait during forkless operations.

Replication compression is off by default. When repl-compression is enabled on both the primary and the replica, Valkey uses the strongest compression algorithm they share for ongoing replication and diskless full syncs. Traffic stays uncompressed if they share none. Setting repl-compression to yes selects LZ4. Zstandard requires build support. Disk-based full syncs use the separate rdbcompression setting.

In its replication benchmark, Valkey reported that LZ4 cut bandwidth by 52% with less than 1% throughput overhead using 3 million keys of about 315 bytes each. The release post does not detail the baseline.

Valkey 9.2 also adds Path Hash, a built-in data type for hierarchical keys. An inference engine can map the start of a prompt to the location of cached attention state in GPU or host memory. PHPREFIXES with COUNT 1 WITHVALUES returns the longest matching stored prefix and its fields when the engine queries a path built from the prompt’s token sequence and a match exists.

Alongside 9.2, Valkey announced a separate open-source distributed key-value cache integration that uses LMCache to share attention state across inference workers. For its remote direct memory access transport, LMCache documents lookups using EXISTS with the Valkey Large Object module. Valkey claims up to 40% lower inference costs and up to 80% less time to generate the first token with the integration. The announcement gives no test method or baseline for those figures.

AWS announced general availability of its managed distributed cache on Amazon ElastiCache on November 18, with the service available in 13 AWS Regions. The cache runs on node-based ElastiCache for Valkey 9.1 or later on cache.i8ge.48xlarge nodes and uses the Valkey Large Object module.

Valkey 9.2.0 changes how some sorted-set commands read score bounds. They now reject a bare opening parenthesis or an empty string as a bound. Valkey lists Docker images and Ubuntu binaries for 9.2.0.

More news