Devaraj Das created HBASE-9451:
----------------------------------
Summary: Meta remains unassigned when the meta server crashes with
the ClusterStatusListener set
Key: HBASE-9451
URL: https://issues.apache.org/jira/browse/HBASE-9451
Project: HBase
Issue Type: Bug
Reporter: Devaraj Das
While running tests described in HBASE-9338, ran into this problem. The
hbase.status.listener.class was set to
org.apache.hadoop.hbase.client.ClusterStatusListener$MultiCastListener.
1. I had the meta server coming down
2. The metaSSH got triggered. The call chain:
2.1 verifyAndAssignMetaWithRetries
2.2 verifyMetaRegionLocation
2.3 waitForMetaServerConnection
2.4 getMetaServerConnection
2.5 getCachedConnection
2.6 HConnectionManager.getAdmin(serverName, false)
2.7 isDeadServer(serverName) -> This is hardcoded to return 'false' when the
clusterStatusListener field is null. If clusterStatusListener is not null (in
my test), then it could return true in certain cases (and in this case, indeed
it should return true since the server is down). I am trying to understand why
it's hardcoded to 'false' for former case.
3. When isDeadServer returns true, the method
HConnectionManager.getAdmin(ServerName, boolean) throws
RegionServerStoppedException.
4. Finally, after the retries are over verifyAndAssignMetaWithRetries gives up
and the master aborts.
The methods in the above call chain don't handle RegionServerStoppedException.
Maybe something to look at...
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
