Thanks! The snapshot approach is exactly what we're using — UFFD (userfaultfd) to restore from a pre-booted snapshot, with on-demand page faulting so the VM responds before all memory is loaded. That's how we get to ~300ms.