[jira] [Commented] (SPARK-15725) Dynamic allocation hangs YARN app when executors time out

["jin xing (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-15725?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15769384#comment-15769384 ] 

jin xing commented on SPARK-15725:
----------------------------------

[~blue@cloudera.com]
May I ask two questions?
1. "a large stage will cause a lot of executors to get killed around the same time": 
    Lots of executors are killed because they are idle for over "spark.dynamicAllocation.executorIdleTimeout" seconds, am I right?

2. " That results in a call to the AM to find out why the executor died, followed by a blocking and retrying `RemoveExecutor` RPC call that results in a second `KillExecutor` call to the AM.":
```
    val future = am.ask[ExecutorLossReason](lossReasonRequest, askTimeout)
    future onSuccess {
        case reason: ExecutorLossReason => {
            driverEndpoint.askWithRetry[Boolean](RemoveExecutor(executorId, reason))
        }
    }
    future onFailure {
        case NonFatal(e) => {
            logWarning(s"Attempted to get executor loss reason" +
                s" for executor id ${executorId} at RPC address ${executorRpcAddress}," +
                s" but got no response. Marking as slave lost.", e)
            driverEndpoint.askWithRetry[Boolean](RemoveExecutor(executorId, SlaveLost()))
        }
        case t => throw t
    }
```
    If ExecutorLossReason request are timeout, it will go to ````driverEndpoint.askWithRetry[Boolean](RemoveExecutor(executorId, SlaveLost()))````, My question is how it will result in a second 'KillExecutor'

> Dynamic allocation hangs YARN app when executors time out
> ---------------------------------------------------------
>
>                 Key: SPARK-15725
>                 URL: https://issues.apache.org/jira/browse/SPARK-15725
>             Project: Spark
>          Issue Type: Bug
>          Components: Spark Core
>    Affects Versions: 1.6.1, 2.0.0
>            Reporter: Ryan Blue
>            Assignee: Ryan Blue
>             Fix For: 2.0.0
>
>
> We've had a problem with a dynamic allocation and YARN (since 1.6) where a large stage will cause a lot of executors to get killed around the same time, causing the driver and AM to lock up and wait forever. This can happen even with a small number of executors (~100).
> When executors are killed by the driver, the [network connection to the driver disconnects|https://github.com/apache/spark/blob/master/yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnSchedulerBackend.scala#L201]. That results in a call to the AM to find out why the executor died, followed by a [blocking and retrying `RemoveExecutor` RPC call|https://github.com/apache/spark/blob/master/yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnSchedulerBackend.scala#L227] that results in a second `KillExecutor` call to the AM. When a lot of executors are killed around the same time, the driver's AM threads are all taken up blocking and waiting on the AM (see the stack trace below, which was the same for 42 threads). I think this behavior, the network disconnect and subsequent cleanup, is unique to YARN.
> {code:title=Driver AM thread stack}
> sun.misc.Unsafe.park(Native Method)
> java.util.concurrent.locks.LockSupport.parkNanos(LockSupport.java:215)
> java.util.concurrent.locks.AbstractQueuedSynchronizer.doAcquireSharedNanos(AbstractQueuedSynchronizer.java:1037)
> java.util.concurrent.locks.AbstractQueuedSynchronizer.tryAcquireSharedNanos(AbstractQueuedSynchronizer.java:1328)
> scala.concurrent.impl.Promise$DefaultPromise.tryAwait(Promise.scala:208)
> scala.concurrent.impl.Promise$DefaultPromise.ready(Promise.scala:218)
> scala.concurrent.impl.Promise$DefaultPromise.result(Promise.scala:223)
> scala.concurrent.Await$$anonfun$result$1.apply(package.scala:190)
> scala.concurrent.BlockContext$DefaultBlockContext$.blockOn(BlockContext.scala:53)
> scala.concurrent.Await$.result(package.scala:190)
> org.apache.spark.rpc.RpcTimeout.awaitResult(RpcTimeout.scala:81)
> org.apache.spark.rpc.RpcEndpointRef.askWithRetry(RpcEndpointRef.scala:102)
> org.apache.spark.rpc.RpcEndpointRef.askWithRetry(RpcEndpointRef.scala:78)
> org.apache.spark.scheduler.cluster.YarnSchedulerBackend$YarnSchedulerEndpoint$$anonfun$receiveAndReply$1$$anonfun$applyOrElse$2.apply$mcV$sp(YarnSchedulerBackend.scala:286)
> org.apache.spark.scheduler.cluster.YarnSchedulerBackend$YarnSchedulerEndpoint$$anonfun$receiveAndReply$1$$anonfun$applyOrElse$2.apply(YarnSchedulerBackend.scala:286)
> org.apache.spark.scheduler.cluster.YarnSchedulerBackend$YarnSchedulerEndpoint$$anonfun$receiveAndReply$1$$anonfun$applyOrElse$2.apply(YarnSchedulerBackend.scala:286)
> scala.concurrent.impl.Future$PromiseCompletingRunnable.liftedTree1$1(Future.scala:24)
> scala.concurrent.impl.Future$PromiseCompletingRunnable.run(Future.scala:24)
> java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
> java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
> java.lang.Thread.run(Thread.java:745)
> {code}
> The RPC calls to the AM aren't returning because the `YarnAllocator` is spending all of its time in the `allocateResources` method. That class's public methods are synchronized so only one RPC can be satisfied at a time. The reason why it is constantly calling `allocateResources` is because [its thread|https://github.com/apache/spark/blob/master/yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala#L467] is [woken up|https://github.com/apache/spark/blob/master/yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala#L686] by calls to get the failure reason for an executor -- which is part of the chain of events in the driver for each executor that goes down.
> The final result is that the `YarnAllocator` doesn't respond to RPC calls for long enough that calls time out and replies for non-blocking calls are dropped. Then the application is unable to do any work because everything retries or exits and the application *hangs for 24+ hours*, until enough errors accumulate that it dies.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

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