Docker has always used microVMs (well since 2016)

There's a lot of buzz "microVMs", where a workload runs with a stripped down Linux kernel, on a minimalist VMM such as firecracker (announced at re:Invent 2018 ) on top of a hypervisor like KVM or Xen. MicroVMs are often contrasted to, and considered more secure than, traditional Linux Docker containers. What if I told you that Docker Desktop has always used microVMs?

When I joined Docker in 2015 the state of the art was Docker Toolbox . It used VirtualBox , which is a great product with lots of features. VirtualBox has its own GUI and its own update process; way more than we needed for Docker.

We wanted Docker to feel like a native app on Mac and Windows, rather than a bundle of components. Using our Mirage Unikernel libraries we started building a "library VMM": a VMM which could be embedded inside the Docker application, and which would be single purpose, minimal, secure and fast. The first version was called hyperkit and the most recent one is Docker VMM . 1

The VM kernel and root filesystem were minimal too, based on the LinuxKit project. The current version is an even more slimmed down and efficient variant but conceptually the same, similar to containerd/nerdbox .

For Docker for Mac (later renamed Docker Desktop) this allowed:

Running rootless on all platforms. Without requiring "rootless inside Linux": the whole Linux kernel is untrusted, so even a Linux CVE doesn't matter.

Minimal devices. We could carefully limit the host / VM interface, making it easy to understand and audit.

VPNs and network policy. It was easy to interoperate with VPNs, and to extend later to impose networking policy such as Registry Access Management .

The Docker microVM tech is the foundation of both Docker Desktop and also Docker Sandboxes ( why microVMs ). Docker Sandboxes enables you to do things like run your favourite coding agent and harness securely , with strict isolation and governance.

To see what Docker microVM's can do, install Docker Sandboxes (sbx) and try `sbx run claude`.

1 Although we were aiming to make everything a library and link into a static unikernel-like process, for technical reasons it makes sense to still have a single host process per VM. ↩