When the compute resource tracker checks nodes, the ironic driver checks the node against a list of states that it should return resources for. This is to prevent nodes in various ironic states, like our cleaning process, that are not available from being scheduled to by nova.
The logic around this check ( https://github.com/openstack/nova/blob/master/nova/virt/ironic/driver.py#L334-L351 ) looks for existing instances on the node, and if they aren't found it then looks at the conditions for returning the node as unavailable.
The problem is when you have an orphaned instance on your node, one which ironic sees as present but nova does not (usually nova lists it as having been deleted).
The instance detection will return true, causing the memory_mb_used and memory_mb values to be set to the retrieved value from instance_info['memory_mb'].
The check for _node_resources_unavailable will not run as it is an elif. This means that even if this node is in maintenance state, we won't notice and return all zeros for resources as we normally would.
Once the resource tracker calls _update_usage_from_instance, it will not find an instance associated with the node from nova's point of view and will return all of the memory as available instead, causing builds to be scheduled to this node.
Ironic will then fail the build attempt due to it showing an instance already associated with the node.
My initial thought was to swap the logical order of the _node_resources
It looks like if the ironic driver implements the get_per_
Trying to see what the best way to do this is.