With ROOT table gone, we no longer cache the location of the meta table (in MetaCache) in 96+. I've checked 94 code, and there we cache meta, but not root.
However, not caching the metas own location means that we are doing a zookeeper request every time we want to look up a regions location from meta. This means that there is a significant spike in zk requests whenever a region server goes down.
This affects trunk,0.98 and 0.96 as well as hbase-10070 branch. I've discovered the issue in hbase-10070 because of the integration test (
HBASE-10572) results in 150K requests to zk in 10min.
A thread dump from one of the runs have 100+ threads from client in this stack trace: