[ https://issues.apache.org/jira/browse/HBASE-13937?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14592928#comment-14592928 ]
Duo Zhang commented on HBASE-13937: ----------------------------------- {quote} In theory we should not have isServerReachable() at all. It is a parallel cluster membership mechanism to the already existing zk based one {quote} I'm not familiar with the code of AM and SM, but I think zk is not enough to keep things consistency. The EPHEMERAL node on zookeeper disappeared does not mean the server is really dead. We still need a method like isServerReachable to decide whether the server is really dead. One example is in HDFS HA, a fencing is needed before transforming standby namenode to active namenode. {quote} If we get a connection exception or smt, we cannot assume the server is dead. {quote} Agree(and excuse me, what is smt?), in general only the server tells us it is dead then we can make sure the server is dead. And after googling I think connection refused is also not that stable(a firewall or backlog queue full can also cause connection refused). So I think we need fencing like what HDFS HA does? Yeah, you may challenge that if the machine is crashed, how can we make sure the server is dead...Honestly I do not have perfect solution. Maybe we could introduce a DeadServerManager and make several levels of consistency, the lowest level does not do fencing at all, the medium level do fencing with a timeout, and the highest level will do fencing for ever(let a person stop it maybe) Thanks. > Partially revert HBASE-13172 > ----------------------------- > > Key: HBASE-13937 > URL: https://issues.apache.org/jira/browse/HBASE-13937 > Project: HBase > Issue Type: Sub-task > Components: Region Assignment > Reporter: Enis Soztutar > Assignee: Enis Soztutar > Fix For: 0.98.14, 1.2.0, 1.1.1, 1.3.0 > > Attachments: hbase-13937_v1.patch, hbase-13937_v2.patch > > > HBASE-13172 is supposed to fix a UT issue, but causes other problems that > parent jira (HBASE-13605) is attempting to fix. > However, HBASE-13605 patch v4 uncovers at least 2 different issues which are, > to put it mildly, major design flaws in AM / RS. > Regardless of 13605, the issue with 13172 is that we catch > {{ServerNotRunningYetException}} from {{isServerReachable()}} and return > false, which then puts the Server to the {{RegionStates.deadServers}} list. > Once it is in that list, we can still assign and unassign regions to the RS > after it has started (because regular assignment does not check whether the > server is in {{RegionStates.deadServers}}. However, after the first assign > and unassign, we cannot assign the region again since then the check for the > lastServer will think that the server is dead. > It turns out that a proper patch for 13605 is very hard without fixing rest > of broken AM assumptions (see HBASE-13605, HBASE-13877 and HBASE-13895 for a > colorful history). For 1.1.1, I think we should just revert parts of > HBASE-13172 for now. -- This message was sent by Atlassian JIRA (v6.3.4#6332)