Tarun Parimi created YARN-9712:
----------------------------------
Summary: ResourceManager goes into a deadlock while transitioning
to standby
Key: YARN-9712
URL: https://issues.apache.org/jira/browse/YARN-9712
Project: Hadoop YARN
Issue Type: Bug
Components: resourcemanager, RM
Affects Versions: 2.9.0
Reporter: Tarun Parimi
We have observed RM go into a deadlock while transitioning to standby in a
heavily loaded production cluster which can observe random connection loss to a
zookeeper session and also has a large amount of RMDelegationToken requests due
to oozie jobs.
On analyzing the jstack and the logs, this seems to happen when the below
sequence of events occur.
1. Zookeeper session is lost and so the ActiveStandbyElector service will do
transitionToStandby . This transitionToStandby is a synchronized method and so
will acquire a lock on ResourceManager.
{code:java}
2019-07-25 14:31:24,497 INFO ha.ActiveStandbyElector
(ActiveStandbyElector.java:processWatchEvent(621)) - Session expired. Entering
neutral mode and rejoining...
2019-07-25 14:31:28,084 INFO resourcemanager.ResourceManager
(ResourceManager.java:transitionToStandby(1134)) - Transitioning to standby
state
{code}
2. While transitioning to standby, a java.lang.InterruptedException occurs in
RMStateStore while removing/storing RMDelegationToken. This is because
RMSecretManagerService will be stopped while transitioning to standby.
{code:java}
2019-07-25 14:31:28,576 ERROR recovery.RMStateStore
(RMStateStore.java:transition(373)) - Error While Removing RMDelegationToken
and SequenceNumber
java.lang.InterruptedException
2019-07-25 14:31:28,576 ERROR recovery.RMStateStore
(RMStateStore.java:notifyStoreOperationFailedInternal(992)) - State store
operation failed
java.lang.InterruptedException
{code}
3. When state store error occurs, a RMFatalEvent of type STATE_STORE_FENCED
will be sent.
{code:java}
2019-07-25 14:31:28,579 ERROR resourcemanager.ResourceManager
(ResourceManager.java:handle(767)) - Received RMFatalEvent of type
STATE_STORE_FENCED, caused by java.lang.InterruptedException
{code}
4. The problem occurs when the RMFatalEventDispatcher calls getConfig() . This
also needs a lock on ResourceManager since its a synchronized method. This will
cause the rmDispatcher eventHandlingThread to become blocked.
{code:java}
private class RMFatalEventDispatcher implements EventHandler<RMFatalEvent> {
@Override
public void handle(RMFatalEvent event) {
LOG.error("Received " + event);
if (HAUtil.isHAEnabled(getConfig())) {
// If we're in an HA config, the right answer is always to go into
// standby.
LOG.warn("Transitioning the resource manager to standby.");
handleTransitionToStandByInNewThread();
{code}
5. The transitionToStandby will wait forever as the eventHandlingThread of
rmDispatcher is blocked. This causes a deadlock and RM will not become active
until restarted.
Below are the relevant threads in the jstack captured.
The transitionToStandby thread that waits forever.
{code:java}
"main-EventThread" #138239 daemon prio=5 os_prio=0 tid=0x00007fea473b2800
nid=0x2f411 in Object.wait() [0x00007fda5bef5000]
java.lang.Thread.State: WAITING (on object monitor)
at java.lang.Object.wait(Native Method)
at java.lang.Thread.join(Thread.java:1245)
- locked <0x00007fdb6c5059a0> (a java.lang.Thread)
at java.lang.Thread.join(Thread.java:1319)
at
org.apache.hadoop.yarn.event.AsyncDispatcher.serviceStop(AsyncDispatcher.java:161)
at
org.apache.hadoop.service.AbstractService.stop(AbstractService.java:221)
- locked <0x00007fdb6c538ca0> (a java.lang.Object)
at
org.apache.hadoop.yarn.server.resourcemanager.ResourceManager.resetRMContext(ResourceManager.java:1323)
at
org.apache.hadoop.yarn.server.resourcemanager.ResourceManager.reinitialize(ResourceManager.java:1091)
at
org.apache.hadoop.yarn.server.resourcemanager.ResourceManager.transitionToStandby(ResourceManager.java:1139)
- locked <0x00007fdb33e418f0> (a
org.apache.hadoop.yarn.server.resourcemanager.ResourceManager)
at
org.apache.hadoop.yarn.server.resourcemanager.AdminService.transitionToStandby(AdminService.java:355)
- locked <0x00007fdb33e41828> (a
org.apache.hadoop.yarn.server.resourcemanager.AdminService)
at
org.apache.hadoop.yarn.server.resourcemanager.EmbeddedElectorService.becomeStandby(EmbeddedElectorService.java:147)
at
org.apache.hadoop.ha.ActiveStandbyElector.becomeStandby(ActiveStandbyElector.java:970)
at
org.apache.hadoop.ha.ActiveStandbyElector.processResult(ActiveStandbyElector.java:480)
- locked <0x00007fdb33e7bb88> (a
org.apache.hadoop.ha.ActiveStandbyElector)
at
org.apache.zookeeper.ClientCnxn$EventThread.processEvent(ClientCnxn.java:617)
at org.apache.zookeeper.ClientCnxn$EventThread.run(ClientCnxn.java:510)
Locked ownable synchronizers:
- None
{code}
The blocked rmDispatcher EventHandler.
{code:java}
"AsyncDispatcher event handler" #135565 daemon prio=5 os_prio=0
tid=0x00007fdb2107f000 nid=0x2484a waiting for monitor entry
[0x00007fda597cc000]
java.lang.Thread.State: BLOCKED (on object monitor)
at
org.apache.hadoop.service.AbstractService.getConfig(AbstractService.java:403)
- waiting to lock <0x00007fdb33e418f0> (a
org.apache.hadoop.yarn.server.resourcemanager.ResourceManager)
at
org.apache.hadoop.yarn.server.resourcemanager.ResourceManager$RMFatalEventDispatcher.handle(ResourceManager.java:769)
at
org.apache.hadoop.yarn.server.resourcemanager.ResourceManager$RMFatalEventDispatcher.handle(ResourceManager.java:764)
at
org.apache.hadoop.yarn.event.AsyncDispatcher.dispatch(AsyncDispatcher.java:184)
at
org.apache.hadoop.yarn.event.AsyncDispatcher$1.run(AsyncDispatcher.java:110)
at java.lang.Thread.run(Thread.java:745)
Locked ownable synchronizers:
- None
{code}
This scenario will happen only when having the changes introduced in YARN-3742
where RMFatalEventDispatcher handles ERROR scenarios such as STATE_STORE_FENCED
and tries to transitionToStandby.
--
This message was sent by Atlassian JIRA
(v7.6.14#76016)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
