Sub‑ms sandboxes via copy‑on‑write
Open‑source work on copy‑on‑write forking sandboxes promises sub‑millisecond isolation for AI agent workloads, letting teams iterate faster without spinning heavy VMs — useful for simulation and strategy testing but not yet battle‑proven for live matching engines. The approach lowers overhead for rapid experimentation while keeping production order paths on deterministic hardware. (opensourceprojects.dev)
ZeroBoot is an open‑source project by adammiribyan on GitHub released under the Apache‑2.0 license with visible recent commits and repository metadata showing active development. (github.com) The project's published benchmarks report a 0.79 ms spawn latency at p50 and 1.74 ms at p99, with an average memory footprint of ~265 KB per sandbox and a 1,000‑fork burst completing in 815 ms on a single host. (github.com) ZeroBoot implements a Firecracker snapshot → mmap(MAP_PRIVATE) → KVM restore flow that recreates CPU state into a new guest in roughly 0.8 ms by leveraging OS copy‑on‑write semantics. (github.com) Key operational constraints called out in the project and independent write‑ups include single vCPU forks, no networking inside forked sandboxes, and a template update path that requires a full re‑snapshot taking on the order of ~15 seconds. (rywalker.com) The repository ships language SDKs (Python and TypeScript) and a REST API example using api.zeroboot.dev with sample curl and SDK snippets in the README. (github.com) ZeroBoot’s README and coverage contrast its ~0.8 ms p50 against alternatives cited in the same materials — E2B at ~150 ms and Daytona at ~27 ms — positioning the approach as an order‑of‑magnitude density/latency experiment versus snapshot‑oriented microVM systems. (github.com) Authors and analysts mark the project as a working prototype that is not production‑hardened and flag practical concerns (entropy/CSPRNG reseeding and lack of networking) that would need remediation before being applied to latency‑critical matching or live execution paths. (github.com)
