Kern: A 1.5 MB Daemonless Container and Resource Runtime for Untrusted Code

Introduction to Kern

Kern is a new runtime designed for containerization and resource management, packaged as a single 1.52 MB binary. It operates without a daemon, socket, or any background processes, aiming for minimal overhead and fast execution. The tool provides a rootless sandbox and virtual resource runtime for various workloads, including untrusted and AI-generated code.

Core Functionality and Features

Kern functions as a container runtime, sandbox, resource slicer, and stack runner. It supports real OCI images, allowing users to pull, build, commit, push, save, and load images. A container from an image starts in approximately 3.5 milliseconds. The sandboxing features include user, PID, mount, network, UTS, and IPC namespaces, along with a deny-by-default seccomp allowlist and cgroup v2 limits. A single flag, `--security-profile untrusted`, activates a hardened security bundle.

Resource Management and Stacks

Beyond isolation, Kern offers resource profiles for CPU, memory, disk, and devices, declared in a `kern.toml` file. These profiles can be attached by name to workloads. Kern also supports stacks, using its own format or existing `docker-compose.yml` files without conversion. This allows for defining and running multi-service applications, with services communicating by name within a single pod.

Technical Implementation and Scope

The tool's small size is attributed to its minimal dependencies, with `libc` being its only Rust dependency. JSON and OCI manifests are parsed manually, and operations like `pull` shell out to existing `curl` and `tar` utilities rather than linking a TLS stack. Kern is not a hypervisor; its isolation boundary is the Linux kernel, similar to Docker and Podman, meaning a kernel privilege-escalation bug could lead to an escape.