[ https://issues.apache.org/jira/browse/YARN-4676?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15404555#comment-15404555 ]
Junping Du commented on YARN-4676: ---------------------------------- Thanks for sharing these details, Daniel. bq. In the typical EMR cluster scenario, daemon like NM will be configured to auto-start if killed/shutdown, however RM will reject such NM if it appear in the exclude list. In today's YARN (community version), if RM reject NM's register request, NM should get terminated directly. I think we should follow existing behavior or it could be incompatible issues there. bq. 1, DECOMMISSIONED NM, will try to register to RM but will be rejected. It continue such loop until either: 1) the host being terminated; 2) the host being recommissioned. It was likely the DECOMMISSIONED->LOST transition is defensive coding — without it invalid event throws. I can understand we want to gain the scale in and out capability here for cluster's elasticity. However, I am not sure how much benefit we can gain from this hacking behavior - it sounds like we just saving NM daemon start time which is several seconds in most cases which is trivial comparing with container launching ad running. Do I miss other benefit here? bq. It was likely the DECOMMISSIONED->LOST transition is defensive coding — without it invalid event throws. As I mentioned above, we should remove watching DECOMMISSIONED node which is unnecessary to consume RM resource to take care of it. If EXPIRE event get throw in your case, then we should check something wrong there (like race-condition, etc.) and fix there. bq. CLEANUP_CONTAINER and CLEANUP_APP were for sure added to prevent otherwise invalid event exception at the DECOMMISSIONED state I can understand we want to get rid of any annoy invalid transition in our logs. However, similar to what I mentioned above, we need to find out where we send these events and check if these case are valid or belongs to bug due to race condition, etc. Even if we really sure some of events are hard to get rid of, we should empty the transition here as any logic in transition is not necessary. > Automatic and Asynchronous Decommissioning Nodes Status Tracking > ---------------------------------------------------------------- > > Key: YARN-4676 > URL: https://issues.apache.org/jira/browse/YARN-4676 > Project: Hadoop YARN > Issue Type: Sub-task > Components: resourcemanager > Affects Versions: 2.8.0 > Reporter: Daniel Zhi > Assignee: Daniel Zhi > Labels: features > Attachments: GracefulDecommissionYarnNode.pdf, > GracefulDecommissionYarnNode.pdf, YARN-4676.004.patch, YARN-4676.005.patch, > YARN-4676.006.patch, YARN-4676.007.patch, YARN-4676.008.patch, > YARN-4676.009.patch, YARN-4676.010.patch, YARN-4676.011.patch, > YARN-4676.012.patch, YARN-4676.013.patch, YARN-4676.014.patch, > YARN-4676.015.patch, YARN-4676.016.patch, YARN-4676.017.patch, > YARN-4676.018.patch, YARN-4676.019.patch > > > YARN-4676 implements an automatic, asynchronous and flexible mechanism to > graceful decommission > YARN nodes. After user issues the refreshNodes request, ResourceManager > automatically evaluates > status of all affected nodes to kicks out decommission or recommission > actions. RM asynchronously > tracks container and application status related to DECOMMISSIONING nodes to > decommission the > nodes immediately after there are ready to be decommissioned. Decommissioning > timeout at individual > nodes granularity is supported and could be dynamically updated. The > mechanism naturally supports multiple > independent graceful decommissioning “sessions” where each one involves > different sets of nodes with > different timeout settings. Such support is ideal and necessary for graceful > decommission request issued > by external cluster management software instead of human. > DecommissioningNodeWatcher inside ResourceTrackingService tracks > DECOMMISSIONING nodes status automatically and asynchronously after > client/admin made the graceful decommission request. It tracks > DECOMMISSIONING nodes status to decide when, after all running containers on > the node have completed, will be transitioned into DECOMMISSIONED state. > NodesListManager detect and handle include and exclude list changes to kick > out decommission or recommission as necessary. -- This message was sent by Atlassian JIRA (v6.3.4#6332) --------------------------------------------------------------------- To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org