[jira] [Created] (YARN-6678) Committer thread crashes with IllegalStateException in async-scheduling mode of CapacityScheduler

["Tao Yang (JIRA)" <ji...@apache.org>]

Tao Yang created YARN-6678:
------------------------------

             Summary: Committer thread crashes with IllegalStateException in async-scheduling mode of CapacityScheduler
                 Key: YARN-6678
                 URL: https://issues.apache.org/jira/browse/YARN-6678
             Project: Hadoop YARN
          Issue Type: Bug
          Components: capacityscheduler
    Affects Versions: 3.0.0-alpha3, 2.9.0
            Reporter: Tao Yang


Error log:
{noformat}
java.lang.IllegalStateException: Trying to reserve container container_e10_1495599791406_7129_01_001453 for application appattempt_1495599791406_7129_000001 when currently reserved container container_e10_1495599791406_7123_01_001513 on node host: node0123:45454 #containers=40 available=... used=...
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode.reserveResource(FiCaSchedulerNode.java:81)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp.reserve(FiCaSchedulerApp.java:1079)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp.apply(FiCaSchedulerApp.java:795)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler.tryCommit(CapacityScheduler.java:2770)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler$ResourceCommitterService.run(CapacityScheduler.java:546)
{noformat}

Reproduce this problem:
1. nm1 re-reserved app-1/container-X1 and generated reserved proposal-1
2. nm2 has enough resource for app-1, un-reserved app-1/container-X1 and allocated app-1/container-X2
3. nm1 reserved app-2/container-Y
4. proposal-1 was accepted but throw IllegalStateException when applying

Currently the check code for reserve proposal in FiCaSchedulerApp#accept as follows:
{code}
          // Container reserved first time will be NEW, after the container
          // accepted & confirmed, it will become RESERVED state
          if (schedulerContainer.getRmContainer().getState()
              == RMContainerState.RESERVED) {
            // Set reReservation == true
            reReservation = true;
          } else {
            // When reserve a resource (state == NEW is for new container,
            // state == RUNNING is for increase container).
            // Just check if the node is not already reserved by someone
            if (schedulerContainer.getSchedulerNode().getReservedContainer()
                != null) {
              if (LOG.isDebugEnabled()) {
                LOG.debug("Try to reserve a container, but the node is "
                    + "already reserved by another container="
                    + schedulerContainer.getSchedulerNode()
                    .getReservedContainer().getContainerId());
              }
              return false;
            }
          }
{code}

The reserved container on the node of reserve proposal will be checked only for first-reserve container, not for the re-reserve container.
I think FiCaSchedulerApp#accept should do this check for all reserve proposal not matter if the container is re-reserve or not.




--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org