The .IP() call that was previously used for sorting resulted in a call to netutil to parse an IP out of an IP:Port string. This was very slow and resulted in this sort taking up ~50% of total CPU util for kube-proxy.