From Siddharth Seth on review of
HIVE-13620, in regards to LlapBaseInputFormat.getServiceInstance() and how to handle the case of no LLAP service instance for the host specified in the LLAP input split:
This should really be a jira and TODO (post merge to master) - to either 1) go to an alternate random address from the available llap instances, or 2) have additional locations provided by HS2.
I'd lean towards 1. It's absolutely possible for an llap instance to go down, or the node to go down, which would end up causing failures.