[ https://issues.apache.org/jira/browse/FLINK-9567?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Shimin Yang updated FLINK-9567: ------------------------------- Description: After restart the Job Manager in Yarn Cluster mode, Flink does not release task manager containers in some specific case. In the first log I posted, the container with id 24 is the reason why Yarn did not release resources. Although the Task Manager in the container with id 24 was released before restart. But in line 347, 2018-06-14 22:50:47,846 WARN akka.remote.ReliableDeliverySupervisor - Association with remote system [akka.tcp://flink@bd-r1hdp69:30609] has failed, address is now gated for [50] ms. Reason: [Disassociated] this problem caused flink to request for one more container more than need. As the excessive container return id determined by the *numPendingContainerRequests* variable in *YarnResourceManager*, I think it's the *onContainersCompleted* in *YarnResourceManager* called the method *requestYarnContainer* which leads to the increase of *numPendingContainerRequests.* However, the restart logic has already allocated enough containers for Task Managers, Flink will possess the extra container for a long time for nothing. In the worst case, I had a job configured to 5 task managers, but possess more than 100 containers in the end. ps: Another strange thing I found is that when sometimes request for a yarn container, it will return much more than requested. Is it a normal scenario for AMRMAsyncClient? was: After restart the Job Manager in Yarn Cluster mode, Flink does not release task manager containers in some specific case. According to my observation, the reason is the instance variable *numPendingContainerRequests* in *YarnResourceManager* class does not decrease since it has not received the containers. And after restart of job manager, it will make increase the *numPendingContainerRequests* by the number of task executors. Since the callback function *onContainersAllocated* will return the excessive container immediately only if the *numPendingContainerRequests* <= 0, so the number of container grows bigger and bigger while only a few are acting as task manager. I think it is important to clear the *numPendingContainerRequests* variable after restart the Job Manager, but not very clear at how to do that. There's no other way to decrease the *numPendingContainerRequests* except the *onContainersAllocated*. Is it fine to add a method to operate on the *numPendingContainerRequests* variable? And meanwhile, there's no handle of YarnResourceManager in the *ExecutionGraph* restart logic. ps: Another strange thing I found is that when sometimes request for a yarn container, it will return much more than requested. Is it a normal scenario for AMRMAsyncClient? > Flink does not release resource in Yarn Cluster mode > ---------------------------------------------------- > > Key: FLINK-9567 > URL: https://issues.apache.org/jira/browse/FLINK-9567 > Project: Flink > Issue Type: Bug > Components: Cluster Management, YARN > Affects Versions: 1.5.0 > Reporter: Shimin Yang > Priority: Major > Attachments: jobmanager.log > > > After restart the Job Manager in Yarn Cluster mode, Flink does not release > task manager containers in some specific case. > In the first log I posted, the container with id 24 is the reason why Yarn > did not release resources. Although the Task Manager in the container with id > 24 was released before restart. > But in line 347, > 2018-06-14 22:50:47,846 WARN akka.remote.ReliableDeliverySupervisor - > Association with remote system [akka.tcp://flink@bd-r1hdp69:30609] has > failed, address is now gated for [50] ms. Reason: [Disassociated] > this problem caused flink to request for one more container more than need. > As the excessive container return id determined by the > *numPendingContainerRequests* variable in *YarnResourceManager*, I think it's > the *onContainersCompleted* in *YarnResourceManager* called the method > *requestYarnContainer* which leads to the increase of > *numPendingContainerRequests.* However, the restart logic has already > allocated enough containers for Task Managers, Flink will possess the extra > container for a long time for nothing. In the worst case, I had a job > configured to 5 task managers, but possess more than 100 containers in the > end. > ps: Another strange thing I found is that when sometimes request for a yarn > container, it will return much more than requested. Is it a normal scenario > for AMRMAsyncClient? -- This message was sent by Atlassian JIRA (v7.6.3#76005)