[
https://issues.apache.org/jira/browse/YARN-8980?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17760897#comment-17760897
]
ASF GitHub Bot commented on YARN-8980:
--------------------------------------
slfan1989 commented on code in PR #5975:
URL: https://github.com/apache/hadoop/pull/5975#discussion_r1311542321
##########
hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-nodemanager/src/main/java/org/apache/hadoop/yarn/server/nodemanager/amrmproxy/FederationInterceptor.java:
##########
@@ -260,6 +261,16 @@ public class FederationInterceptor extends
AbstractRequestInterceptor {
private final MonotonicClock clock = new MonotonicClock();
+ /*
+ * For UAM, keepContainersAcrossApplicationAttempts is always true.
+ * When re-register to RM, RM will clear node set and regenerate NMToken for
transferred
+ * container. But If keepContainersAcrossApplicationAttempts of MA is false,
AM may not
+ * called getNMTokensFromPreviousAttempts, so the NMToken which is pass from
+ * RegisterApplicationMasterResponse will be missing. Here we cache these
NMToken,
+ * then pass to AM in allocate stage.
+ * */
+ private Set<NMToken> nmTokenMapFromRegisterSecondaryCluster;
Review Comment:
Using Set<NMToken> is feasible, but should we consider using
Map<SubClusterId, Set<NMToken>> for better differentiation of NMToken lists for
each subcluster?
> Mapreduce application container start fail after AM restart.
> -------------------------------------------------------------
>
> Key: YARN-8980
> URL: https://issues.apache.org/jira/browse/YARN-8980
> Project: Hadoop YARN
> Issue Type: Sub-task
> Reporter: Bibin Chundatt
> Assignee: zhengchenyu
> Priority: Major
> Labels: pull-request-available
>
> UAM to subclusters are always launched with keepContainers.
> On AM restart scenarios , UAM register again with RM . UAM receive running
> containers with NMToken. NMToken received by UAM in
> getPreviousAttemptContainersNMToken is never used by mapreduce application.
> Federation Interceptor should take care of such scenarios too. Merge NMToken
> received at registration to allocate response.
> Container allocation response on same node will have NMToken empty.
> issue credits : [~Nallasivan]
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
