Expand description
SandboxBackend trait — pluggable sandbox runtime abstraction
Extracts the VM lifecycle out of SandboxPool so multiple backends
(Kata, systemd-nspawn, etc.) can provide sandbox isolation.
§#508 / #635 — narrowing in progress
Epic #508 converges worker sandboxing onto a “namespace IS the OS surface,
backends-as-transport” model: a SandboxBackend boots a guest OS under an
isolation boundary and delivers the composed Namespace to it — it does
NOT own image-pull or run commands. deliver_namespace
is the narrowed method this trait is converging on.
The CRI-shaped methods below (exec_sync/supports_exec, the
PodSandboxConfig-driven image pull implicit in start, and
update_resources’s LinuxContainerResources) are kept on the trait for
now and marked TRANSITIONAL: they move off once the guest-OS
control-file service (kata-agent #344, Wanix /task #612) and the
universal filesystem-service Mount (#633/#641) are both wired end-to-end.
Until then, removing them here would break every existing caller
(SandboxPool, ExecMount, CRI-facing worker RPCs) for no working
replacement. See issue #635 for the full narrowing rationale.
Enums§
- Namespace
Delivery - Result of
SandboxBackend::deliver_namespace: what’s now available for the guest to consume, and any transitional bookkeeping the caller needs. - Namespace
Transport - How a composed
Namespaceshould be attached to a booted guest.
Traits§
- Sandbox
Backend - Pluggable sandbox runtime backend.
- Sandbox
Handle - Opaque handle stored on each
PodSandbox.