liyakun created YARN-9556:
-----------------------------
Summary: Most containers will go through the process of ending
twice when they are finished
Key: YARN-9556
URL: https://issues.apache.org/jira/browse/YARN-9556
Project: Hadoop YARN
Issue Type: Improvement
Components: resourcemanager
Reporter: liyakun
Assignee: liyakun
When AM reports a container is finished in heartbeat, the container id will be
added to
containersToBeRemovedFromNM of RMNodeImpl.
Next, when NM's heartbeat arrives, it will first execute the
setAndUpdateNodeHeartbeatResponse() method in RMNodeImpl. In this method, it
will execute:
```
this.completedContainers.removeAll(this.containersToBeRemovedFromNM);
...
this.containersToBeRemovedFromNM.clear();
```
The container id is removed in completedContainers so far, but heartbeat
request still carries the state of the container id. In the subsequent
processing, the handleContainerStatus() method in the RMNodeImpl is entered. In
this method, the container id that was just deleted will be added to the
completedContainers and newlyCompletedContainers again.
Next, the container id will enter the updateCompletedContainers() method in the
AbstractYarnScheduler class. It will be added to the untrackedContainerIdList
and then go through the process of finish again by
RMNodeFinishedContainersPulledByAMEvent.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
