Some of the problems: * Node taints are ignored * A run with `cpu: 2..4` requirement doesn't match a node with 8 CPUs (this makes sense for VM backends, but not for K8s) * Resources allocated for other pods are not taken into account