Description
When a cluster restarts quickly after it's crashed, although a new region server is reported in, the master still pulls in the dead region server from the zk.
2013-05-12 18:32:52,996 INFO [IPC Server handler 6 on 36000] org.apache.hadoop.hbase.master.ServerManager: Registering server=a1217.halxg.cloudera.com,36020,1368408767773 .... 2013-05-12 18:32:54,653 INFO [master-a1220.halxg.cloudera.com,36000,1368408767520] org.apache.hadoop.hbase.master.HMaster: Registering server found up in zk but who has not yet reported in: a1217.halxg.cloudera.com,36020,1368378273768 2013-05-12 18:32:54,653 INFO [master-a1220.halxg.cloudera.com,36000,1368408767520] org.apache.hadoop.hbase.master.ServerManager: Registering server=a1217.halxg.cloudera.com,36020,1368378273768
We should not pull in the second region server instance from zk. It is actually dead. We can figure this out by the hostname, and the port. We can assume no two region server instances can be alive on the same host, the same port. To be more cautious, we can check the timestamp as well. The live one should be that with the late timestamp, not pulled in from zk.