Google Cloud and Anyscale introduced experimental gVisor sandboxing for distributed Ray clusters, available through Ray 2.58 APIs.

Google Cloud and Anyscale introduced an experimental Ray library that uses gVisor to provide isolated execution environments inside distributed Ray clusters. Ray is a distributed-computing runtime used to coordinate trainers, inference engines, rollout workers and other components in artificial-intelligence workflows.

Each high-level sandbox is represented as a Ray Actor. Ray’s scheduler chooses the node, reserves central processing unit and memory resources, and manages placement, while the Actor manages the sandbox lifecycle and gVisor supplies the isolated runtime environment.

Beginning with Ray 2.58, framework authors and researchers can manage sandboxes with the same Ray application programming interfaces used for other cluster resources. A sandbox can be created from an Open Container Initiative-compatible image, assigned CPU and memory limits, and accessed through normal Ray Actor calls from anywhere in the cluster.

The sandbox API supports environment creation, networking configuration, environment variables, working directories and command execution. Google positioned the capability for workloads such as dynamic rollouts, code generation and multi-turn tool interaction, while explicitly describing the library as experimental.