[ https://issues.apache.org/jira/browse/HBASE-9918?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13817715#comment-13817715 ]
Ted Yu commented on HBASE-9918: ------------------------------- {code} + TEST_UTIL.startMiniDFSCluster(1); TEST_UTIL.startMiniZKCluster(); conf.setBoolean("dfs.support.append", true); conf.setInt(HConstants.ZK_SESSION_TIMEOUT, 1000); conf.setClass(HConstants.HBASE_MASTER_LOADBALANCER_CLASS, MockLoadBalancer.class, LoadBalancer.class); - TEST_UTIL.startMiniCluster(2); {code} Is the above change in number of slaves intentional ? > MasterAddressTracker & ZKNamespaceManager ZK listeners are missed after > master recovery > --------------------------------------------------------------------------------------- > > Key: HBASE-9918 > URL: https://issues.apache.org/jira/browse/HBASE-9918 > Project: HBase > Issue Type: Bug > Reporter: Jeffrey Zhong > Assignee: Jeffrey Zhong > Attachments: HBase-9918.patch > > > TestZooKeeper#testRegionAssignmentAfterMasterRecoveryDueToZKExpiry always > failed at the following verification for me in my dev env(you have to run the > single test not the whole TestZooKeeper suite to reproduce) > {code} > assertEquals("Number of rows should be equal to number of puts.", > numberOfPuts, numberOfRows); > {code} > We missed two ZK listeners after master recovery MasterAddressTracker & > ZKNamespaceManager. > My current patch is to fix the JIRA issue while I'm wondering if we should > totally remove the master failover implementation when ZK session expired > because this causes reinitialize HMaster partially which is error prone and > not a clean state to start from. > -- This message was sent by Atlassian JIRA (v6.1#6144)