It depends on your application you're running and if you will benefit from the features presented with virtualizing your underlying atomic host.
Live migration in Docker is much more complicated than live migrating an entire atomic host that is itself a VM (Via KVM live migration or vMotion, for example). Virtual machines are easier to deploy than physical machines, and are easier to delegate network connections to. It's also much easier to create demarkation of responsibility within an organization when using VMs. For example, the infrastructure team might end its responsibility at the hypervisor level, while the ops team picks up that responsibility at the guest level. The same applies to deployment tools and configuration management.
However, if your application demands for performance outweigh those benefits (or if those benefits don't have anything to do with your application), and baremetal deployments help in this way for you then by all means stick it on bare metal. In all likelihood, whatever you're doing needs to be continuously reproducible, so be careful with potential "special snowflake" configurations.