OpenDsec: self-hosted sandboxes for RL rollouts and code-running agents

说明与用途

OpenDsec 是一个自托管的沙箱平台，支持快速启动微VM、内存快照复制和任务级暂停，简化了代理强化学习的训练流程。

问题

Agentic reinforcement learning training requires thousands of isolated code execution environments that can be paused, resumed, and forked with memory, but hosted sandbox services are limited and costly.

结果

OpenDsec delivers microVMs that boot in ~0.1 s, fork with memory in ~10 ms, and support task-level pause, enabling efficient RL rollouts on bare-metal or cloud VMs.

工作流

• Run the install script to build and deploy OpenDsec components
• Import a base image (e.g., python:3.12) from Docker Hub
• Create a sandbox via the SDK (DSecClient.run_sandbox)
• Execute code inside the sandbox with run_shell
• Fork the sandbox with memory using fork(count=N)
• Stop all sandboxes with client.stop()

证据

If you train a model with agentic reinforcement learning, every rollout needs somewhere to run code: install a package, run the tests, hit the network, write files.

来源

The program is the whole client side: ... print(6*7) ...

来源

Fast microVMs. About 0.1 s for the VM to boot, about 55 ms to create one from a memory snapshot, about 50 ms to resume one from a pause.

来源

One 384-thread node took a burst of 5120 microVMs in 37 s. A two-node cluster took 8192 containers in one burst with no failures.

来源