[SPARK-16533] Spark application not handling preemption messages

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.6.0
• Fix Version/s: 2.0.1, 2.1.0
• Component/s: EC2, Input/Output, Optimizer, Scheduler, Spark Core, Spark Submit, YARN

Environment

Yarn version: Hadoop 2.7.1-amzn-0

AWS EMR Cluster running:
1 x r3.8xlarge (Master)
52 x r3.8xlarge (Core)

Spark version : 1.6.0
Scala version: 2.10.5
Java version: 1.8.0_51

Input size: ~10 tb
Input coming from S3

Queue Configuration:
Dynamic allocation: enabled
Preemption: enabled
Q1: 70% capacity with max of 100%
Q2: 30% capacity with max of 100%

Job Configuration:
Driver memory = 10g
Executor cores = 6
Executor memory = 10g
Deploy mode = cluster
Master = yarn
maxResultSize = 4g
Shuffle manager = hash

• Assignee: Angus Gerry
• Reporter: Lucas Winkelmann
• Created: 13/Jul/16 21:13
• Updated: 17/May/20 17:46
• Resolved: 01/Sep/16 17:37

Description

Here is the scenario:
I launch job 1 into Q1 and allow it to grow to 100% cluster utilization.
I wait between 15-30 mins ( for this job to complete with 100% of the cluster available takes about 1hr so job 1 is between 25-50% complete). Note that if I wait less time then the issue sometimes does not occur, it appears to be only after the job 1 is at least 25% complete.
I launch job 2 into Q2 and preemption occurs on the Q1 shrinking the job to allow 70% of cluster utilization.
At this point job 1 basically halts progress while job 2 continues to execute as normal and finishes. Job 2 either:

• Fails its attempt and restarts. By the time this attempt fails the other job is already complete meaning the second attempt has full cluster availability and finishes.
• The job remains at its current progress and simply does not finish ( I have waited ~6 hrs until finally killing the application ).

Looking into the error log there is this constant error message:
WARN NettyRpcEndpointRef: Error sending message [message = RemoveExecutor(454,Container container_1468422920649_0001_01_000594 on host: ip-NUMBERS.ec2.internal was preempted.)] in X attempts

My observations have led me to believe that the application master does not know about this container being killed and continuously asks the container to remove the executor until eventually failing the attempt or continue trying to remove the executor.

I have done much digging online for anyone else experiencing this issue but have come up with nothing.

Issue Links

• is duplicated by: SPARK-16702 Driver hangs after executors are lost
• links to: [Github] Pull Request #14710 (angolon)
• links to: [Github] Pull Request #14925 (vanzin)
• links to: [Github] Pull Request #14933 (angolon)