Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China (PDF).

Cluster terminates with NO_SLAVE_LEFT and core nodes FAILED_BY_MASTER

Usually, this happens because termination protection is disabled, and all core nodes exceed disk storage capacity as specified by a maximum utilization threshold in the yarn-site configuration classification, which corresponds to the yarn-site.xml file. This value is 90% by default. When disk utilization for a core node exceeds the utilization threshold, the YARN NodeManager health service reports the node as UNHEALTHY. While it's in this state, Amazon EMR deny lists the node and does not allocate YARN containers to it. If the node remains unhealthy for 45 minutes, Amazon EMR marks the associated Amazon EC2 instance for termination as FAILED_BY_MASTER. When all Amazon EC2 instances associated with core nodes are marked for termination, the cluster terminates with the status NO_SLAVE_LEFT because there are no resources to execute jobs.

Exceeding disk utilization on one core node might lead to a chain reaction. If a single node exceeds the disk utilization threshold because of HDFS, other nodes are likely to be near the threshold as well. The first node exceeds the disk utilization threshold, so Amazon EMR deny lists it. This increases the burden of disk utilization for remaining nodes because they begin to replicate HDFS data among themselves that they lost on the deny-listed node. Each node subsequently goes UNHEALTHY in the same way, and the cluster eventually terminates.

Best practices and recommendations

Configure cluster hardware with adequate storage

When you create a cluster, make sure that there are enough core nodes and that each has an adequate instance store and EBS storage volumes for HDFS. For more information, see Calculating the required HDFS capacity of a cluster. You can also add core instances to existing instance groups manually or by using auto-scaling. The new instances have the same storage configuration as other instances in the instance group. For more information, see Use cluster scaling.

Enable termination protection

Enable termination protection. This way, if a core node is deny listed, you can connect to the associated Amazon EC2 instance using SSH to troubleshoot and recover data. If you enable termination protection, be aware that Amazon EMR does not replace the Amazon EC2 instance with a new instance. For more information, see Using termination protection.

Create an alarm for the MRUnhealthyNodes CloudWatch metric

This metric reports the number of nodes reporting an UNHEALTHY status. It's equivalent to the YARN metric mapred.resourcemanager.NoOfUnhealthyNodes. You can set up a notification for this alarm to warn you of unhealthy nodes before the 45-minute timeout is reached. For more information, see Monitoring Amazon EMR metrics with CloudWatch.

Tweak settings using yarn-site

The settings below can be adjusted according to your application requirements. For example, you may want to increase the disk utilization threshold where a node reports UNHEALTHY by increasing the value of yarn.nodemanager.disk-health-checker.max-disk-utilization-per-disk-percentage.

You can set these values when you create a cluster using the yarn-site configuration classification. For more information see Configuring applications in the Amazon EMR Release Guide. You can also connect to the Amazon EC2 instances associated with core nodes using SSH, and then add the values in /etc/hadoop/conf.empty/yarn-site.xml using a text editor. After making the change, you must restart hadoop-yarn-nodemanager as shown below.

sudo /sbin/stop hadoop-yarn-nodemanager
sudo /sbin/start hadoop-yarn-nodemanager

For more information about current yarn-site properties and default values, see YARN default settings in Apache Hadoop documentation.