With your changes, my experiments in my vm seem to succeed. However, as with
HIVE-5431, we really do need to have a unit test for this. I pushed for that one making it through without a test at that time because it was difficult to re-create as a test, and we could clearly demonstrate an obvious problem and prove we fixed it. Now that we aren't looking at 0.13 release deadlines yet, I think we should spend some time trying to create a test for this.
At the base, it looks like we can create a test for this by writing a junit unit test that manages mulitple hive client instances, each of which resets the conf at the every access.
HIVE-5431, I was considering something like writing out a hbase table directly using hbase/hfiles and then trying to read it as a hbase table from hive. Also possible is a change to the QTestUtil framework so as to have a param that allows us to reset confs between calls.