The host-etcd endpoint IPs do not change very often. This makes them more stable than relying on the "real" service endpoints, which reduces rollouts. Using the IPs directly reduces the the dependency stack for working openshift-apiserver and should improve availability during upgrades.
Until we work out a restart mechanism if all IPs for master nodes have changed, the kube-apiserver using the dns name is safer for the kube-apiserver. the openshift-apiserver doesn't have a bootstrapping problem, so it can use IPs.
*** This bug has been marked as a duplicate of bug 1805254 ***