> In this setup, UML is essentially a userspace process that cleverly employs concepts like files and sockets to launch a new Linux kernel instance capable of running its own processes. The exact mapping of these processes to the host — specifically, how the CPU is virtualized — is something I’m not entirely clear on, and I’d welcome insights in the comments. One could envision an implementation where guest threads and processes map to host counterparts but with restricted system visibility, akin to containers, yet still operating within a nested Linux kernel.

At least in the first generation of UML, the guest processes are in fact host processes. The guest kernel (a userland process) essentially runs them under ptrace() and catches all of the system calls made by the guest process and rewires them so they do operations inside of the guest kernel. They otherwise run like host processes on host CPU, though.

Completing the illusion, however, the guest kernel also skillfully rewires the guest ptrace() calls so you can still use strace or gdb inside of the guest!

It's good enough that you can go deeper and run UML inside of UML.

> What’s the real-world utility here? Is UML suitable for running isolated workloads? My educated guess is: probably not for most production scenarios.

Back in the day there were hosts offering UML VMs for rent. This is actually how Linode got its start!