For a large cloud service provider that manages multiple data centers globally and has a large inventory of bare metal, it often has its own hardware inventory database and its own inventory management (IM) system. However current baremetal solution for OpenStack requires that all the nodes be registered in Ironic and each node is a "dummy hypervisor" internally for the Nova scheduler. Obviously It is less than optimal for a global deployment on the size and scale of a service provider cannot afford to present it's inventory in this way and requires a specific solution tailored to manage this fleet of hardware and provide nodes from this fleet to Ironic.
A reference architecture will provide a standard interface to work with the external IM system. It will work with Nova scheduler to provide matching nodes. The matching node will be added to Ironic before the deployment and be removed from Ironic after it is returned. This also allows the service provider to take advantage of a specialized matching algorithm they might have.
The original blueprint seems to be referencing this library - https:/