[ https://issues.apache.org/jira/browse/YARN-1068?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13798043#comment-13798043 ]
Vinod Kumar Vavilapalli commented on YARN-1068: ----------------------------------------------- bq. IIUC, the suggestion is to use the RPC server from AdminService. AdminService currently is an Active service and not an Always-On service, so doesn't start until the RM transitions to Active. Moving the AdminService to Always-On requires defining the semantics when the RM is Standby. I'd do this instead of adding a new service. For now as well as long term, we want to deny all the existing AdminService operations on Standby. Doing it via not stopping the server or explicitly rejecting the requests is an implementation detail and not a big change in semantics. Repeating what I said, we originally added AdminService separately from client-service only for prioritizing admin operations. No need for a new server for this. bq. To do this, we need to have RMAdminCLI extend HAAdmin, and augment the run() method to call super.run() when applicable, and the usage needs to be augmented to include the HAAdmin usage. Yes. I guess there is no argument here other than stating the obvious. bq. YARN expects the actual PB/PBImpl files to be at a particular location, and can't find the corresponding files when using HAServiceProtocol from common. Hence, had to use PB interfaces. HAServiceProtocolPB is the PB interface and there seems to be no PBImpl as Common/HDFS follow a different pattern from YARN's and the last I heard they liked YARN's PB impl stuff. In any case, +1 to skip using YARNRPC for that reason. bq. The patch primarily adds command line support for HA transitions. Have tested this manually several times on a real cluster. We need junit tests for everything. We can skip down to manual tests for hard-to-test race conditions or security features that are not possible to address otherwise. Manual testing is not a substitute for junit tests. > Add admin support for HA operations > ----------------------------------- > > Key: YARN-1068 > URL: https://issues.apache.org/jira/browse/YARN-1068 > Project: Hadoop YARN > Issue Type: Sub-task > Components: resourcemanager > Affects Versions: 2.1.0-beta > Reporter: Karthik Kambatla > Assignee: Karthik Kambatla > Labels: ha > Attachments: yarn-1068-10.patch, yarn-1068-11.patch, > yarn-1068-1.patch, yarn-1068-2.patch, yarn-1068-3.patch, yarn-1068-4.patch, > yarn-1068-5.patch, yarn-1068-6.patch, yarn-1068-7.patch, yarn-1068-8.patch, > yarn-1068-9.patch, yarn-1068-prelim.patch > > > Support HA admin operations to facilitate transitioning the RM to Active and > Standby states. -- This message was sent by Atlassian JIRA (v6.1#6144)