[jira] [Commented] (FLINK-5542) YARN client incorrectly uses local YARN config to check vcore capacity

["ASF GitHub Bot (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-5542?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16641033#comment-16641033 ] 

ASF GitHub Bot commented on FLINK-5542:
---------------------------------------

tillrohrmann commented on a change in pull request #6775: [FLINK-5542] use YarnCluster vcores setting to do MaxVCore validation
URL: https://github.com/apache/flink/pull/6775#discussion_r223209376
 
 

 ##########
 File path: flink-yarn/src/main/java/org/apache/flink/yarn/AbstractYarnClusterDescriptor.java
 ##########
 @@ -282,18 +282,29 @@ private void isReadyForDeployment(ClusterSpecification clusterSpecification) thr
 		}
 
 		// Check if we don't exceed YARN's maximum virtual cores.
-		// The number of cores can be configured in the config.
-		// If not configured, it is set to the number of task slots
-		int numYarnVcores = yarnConfiguration.getInt(YarnConfiguration.NM_VCORES, YarnConfiguration.DEFAULT_NM_VCORES);
+		// Fetch numYarnMaxVcores from all the RUNNING nodes via yarnClient
+		int numYarnMaxVcores = Integer.MIN_VALUE;
+		try {
+			List<NodeReport> nodes = yarnClient.getNodeReports(NodeState.RUNNING);
+			for (NodeReport rep : nodes) {
+				final Resource res = rep.getCapability();
+				if (res.getVirtualCores() > numYarnMaxVcores) {
+					numYarnMaxVcores = res.getVirtualCores();
+				}
+			}
+		} catch (Exception e) {
+			throw new YarnDeploymentException("Couldn't get cluster description, please check on the YarnConfiguration", e);
+		}
+
 		int configuredVcores = flinkConfiguration.getInteger(YarnConfigOptions.VCORES, clusterSpecification.getSlotsPerTaskManager());
 		// don't configure more than the maximum configured number of vcores
-		if (configuredVcores > numYarnVcores) {
+		if (configuredVcores > numYarnMaxVcores) {
 			throw new IllegalConfigurationException(
 				String.format("The number of virtual cores per node were configured with %d" +
-						" but Yarn only has %d virtual cores available. Please note that the number" +
+						" already exceeds the max vcores %d set on Yarn Cluster. Please note that the number" +
 
 Review comment:
   Maybe reword: "The number of requested virtual cores per node %d exceeds the maximum number virtual cores %d available in the Yarn cluster."

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


> YARN client incorrectly uses local YARN config to check vcore capacity
> ----------------------------------------------------------------------
>
>                 Key: FLINK-5542
>                 URL: https://issues.apache.org/jira/browse/FLINK-5542
>             Project: Flink
>          Issue Type: Bug
>          Components: YARN
>    Affects Versions: 1.1.4, 1.5.3, 1.6.0, 1.7.0
>            Reporter: Shannon Carey
>            Priority: Major
>              Labels: pull-request-available
>             Fix For: 1.7.0
>
>
> See http://apache-flink-user-mailing-list-archive.2336050.n4.nabble.com/1-1-4-on-YARN-vcores-change-td11016.html
> When using bin/yarn-session.sh, AbstractYarnClusterDescriptor line 271 in 1.1.4 is comparing the user's selected number of vcores to the vcores configured in the local node's YARN config (from YarnConfiguration eg. yarn-site.xml and yarn-default.xml). It incorrectly prevents Flink from launching even if there is sufficient vcore capacity on the cluster.
> That is not correct, because the application will not necessarily run on the local node. For example, if running the yarn-session.sh client from the AWS EMR master node, the vcore count there may be different from the vcore count on the core nodes where Flink will actually run.
> A reasonable way to fix this would probably be to reuse the logic from "yarn-session.sh -q" (FlinkYarnSessionCli line 550) which knows how to get vcore information from the real worker nodes.  Alternatively, perhaps we could remove the check entirely and rely on YARN's Scheduler to determine whether sufficient resources exist.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)