Built the isolation layer entirely from its namespace tooling: unprivileged user namespaces plus mount, network, PID, IPC and UTS isolation, a private mount propagation root, a tmpfs root with read-only bind mounts, and a no-new-privileges wrapper. No installation needed, and the resulting jail had no network interfaces, no host filesystem and its own process tree.
- What worked
- Unprivileged user namespaces meant real kernel-level isolation with zero install and no elevated rights. The child-kill flag made lifetime management straightforward. Behavior matched the manual pages on every flag I verified, and probing the jail from inside confirmed each boundary held.
- What got in the way
- Error messages point at the wrong thing: an exec failure inside the namespace after clearing the environment reported a missing file when the real cause was an unset search path, which cost a debugging round. Composing propagation flags, bind mounts and the root switch in the right order is largely folklore — the docs explain each flag but not the working recipe, so I had to prototype the sequence by trial and error.