Google Cloud launched Memorystore for Valkey 9.1, claiming up to 3× queries per second versus Memorystore for Redis Cluster.

Memorystore for Valkey 9.1 is generally available, with Google claiming up to three times the queries per second (QPS) at microsecond latency compared with Memorystore for Redis Cluster. The supplied excerpt does not give the workload, hardware configuration, or exact latency measurements underlying that comparison, so the result should retain its “up to” qualification.

Google traces its Valkey support to 2024, when Redis Inc. moved away from the permissive BSD open-source license to a dual-license model. Google and other technology companies backed Valkey as a Linux Foundation-governed open-source alternative, and the announcement positions the service for AI and microservices applications serving millions of concurrent users.

Valkey 9.1 replaces static socket assignment and list polling with lock-free, multi-queue communication for input/output (I/O) work. A single-producer, multiple-consumer queue distributes read and write jobs to available workers; a multiple-producer, single-consumer queue returns completed work to the main thread; and dedicated single-producer, single-consumer queues handle thread-specific memory cleanup and epoll offloading.

Thread activation now uses a two-stage scaling mechanism. The first background I/O thread activates when main-thread CPU utilization exceeds 30%, after which the number of active I/O workers scales up or down according to the dispatch queue’s backlog, with idle workers parked when unnecessary.