The general problem is a when a node pushes its IP addresses to S3, it has no way to tell whether one of its IP address will be usable by other nodes. On the consumer side, we are unable to determine at runtime which of the addresses will work. So we end up doing discovery using worthless IP addresses, and then need to suffer timeouts to sort this out, slowing down startup.
Our fix is to allow configuration of a set of exclusion patterns (regex) on TCPCommunicationSpi, so we could exclude "192.168.*" for example. https://jira.apache.org/jira/browse/IGNITE-10791
