Bootstrapping currently waits to observe 2 endpoints in the "kubernetes" service in HA topologies. The bootstrap kube-apiserver instance itself appears to be included in that number. Soon after observing 2 (bootstrap instance plus one permanent instance), the bootstrap instance is torn down and leaves the cluster with only 1 instance. Each rollout to that instance causes disruption to kube-apiserver availability until the second permanent instance is started for the first time, easily totaling multiple minutes of 0% kube-apiserver availability.

    1.
    2.
    3.