Using managed scaling in Amazon EMR

In the following AWS Regions, Amazon EMR managed scaling is available with Amazon EMR 6.14.0 and higher:

In the following AWS Regions, Amazon EMR managed scaling is available with Amazon EMR 5.30.0 and 6.1.0 and higher:

Amazon EMR managed scaling only works with YARN applications, such as Spark, Hadoop, Hive, and Flink. It doesn't support applications that are not based on YARN, such as Presto and HBase.

Minimum (MinimumCapacityUnits) – The lower boundary of allowed EC2 capacity in a cluster. It is measured through virtual central processing unit (vCPU) cores or instances for instance groups. It is measured through units for instance fleets.

Maximum (MaximumCapacityUnits) – The upper boundary of allowed EC2 capacity in a cluster. It is measured through virtual central processing unit (vCPU) cores or instances for instance groups. It is measured through units for instance fleets.

On-Demand limit (MaximumOnDemandCapacityUnits) (Optional) – The upper boundary of allowed EC2 capacity for On-Demand market type in a cluster. If this parameter is not specified, it defaults to the value of MaximumCapacityUnits.

Maximum core nodes (MaximumCoreCapacityUnits) (Optional) – The upper boundary of allowed EC2 capacity for core node type in a cluster. If this parameter is not specified, it defaults to the value of MaximumCapacityUnits.

Managed scaling is supported in limited AWS Regions and Amazon EMR releases. For more information, see Managed scaling availability.

You must configure the required parameters for Amazon EMR managed scaling. For more information, see Managed scaling parameters.

To use managed scaling, the metrics-collector process must be able to connect to the public API endpoint for managed scaling in API Gateway. If you use a private DNS name with Amazon Virtual Private Cloud, managed scaling won't function properly. To ensure that managed scaling works, we recommend that you take one of the following actions:

If your YARN jobs are intermittently slow during scale down, and YARN Resource Manager logs show that most of your nodes were deny-listed during that time, you can adjust the decommissioning timeout threshold.

Reduce the spark.blacklist.decommissioning.timeout from one hour to one minute to make the node available for other pending containers to continue task processing.

You should also set YARN.resourcemanager.nodemanager-graceful-decommission-timeout-secs to a larger value to ensure Amazon EMR doesn’t force terminate the node while the longest “Spark Task” is still running on the node. The current default is 60 minutes, which means YARN force-terminates the container after 60 minutes once the node enters the decomissioning state.

The following example YARN Resource Manager Log line shows nodes added to the decomissioning state:

2021-10-20 15:55:26,994 INFO org.apache.hadoop.YARN.server.resourcemanager.DefaultAMSProcessor (IPC Server handler 37 on default port 8030): blacklist are updated in Scheduler.blacklistAdditions: [ip-10-10-27-207.us-west-2.compute.internal, ip-10-10-29-216.us-west-2.compute.internal, ip-10-10-31-13.us-west-2.compute.internal, ... , ip-10-10-30-77.us-west-2.compute.internal], blacklistRemovals: []

See more details on how Amazon EMR integrates with YARN deny listing during decommissioning of nodes, cases when nodes in Amazon EMR can be deny listed, and configuring Spark node-decommissioning behavior.

Over-utilization of EBS volumes can cause Managed Scaling issues. We recommend that you maintain EBS volume below 90% utilization. For more information, see Instance storage options and behavior in Amazon EMR.

Amazon CloudWatch metrics are critical for Amazon EMR managed scaling to operate. We recommend that you closely monitor Amazon CloudWatch metrics to make sure data is not missing. For more information about how you can configure CloudWatch alarms to detect missing metrics, see Using Amazon CloudWatch alarms.

Managed scaling operations on 5.30.0 and 5.30.1 clusters without Presto installed may cause application failures or cause a uniform instance group or instance fleet to stay in the ARRESTED state, particularly when a scale down operation is followed quickly by a scale up operation.

As a workaround, choose Presto as an application to install when you create a cluster with Amazon EMR releases 5.30.0 and 5.30.1, even if your job does not require Presto.

When you set the maximum core node and the On-Demand limit for Amazon EMR managed scaling, consider the differences between instance groups and instance fleets. Each instance group consists of the same instance type and the same purchasing option for instances: On-Demand or Spot. For each instance fleet, you can specify up to five instance types, which can be provisioned as On-Demand and Spot Instances. For more information, see Create a cluster with instance fleets or uniform instance groups, Instance fleet options, and Node allocation scenarios.

With Amazon EMR 5.30.0 and higher, if you remove the default Allow All outbound rule to 0.0.0.0/ for the master security group, you must add a rule that allows outbound TCP connectivity to your security group for service access on port 9443. Your security group for service access must also allow inbound TCP traffic on port 9443 from the master security group. For more information about configuring security groups, see Amazon EMR-managed security group for the primary instance (private subnets).

You can use AWS CloudFormation to configure Amazon EMR managed scaling. For more information, see AWS::EMR::Cluster in the AWS CloudFormation User Guide.

If you're using Spot nodes, consider using node labels to prevent Amazon EMR from removing application processes when Amazon EMR removes Spot nodes. For more information about node labels, see Task nodes.

Node labeling is not supported by default in Amazon EMR releases 6.15 or lower. For more information, see Understand node types: primary, core, and task nodes.

If you're using Amazon EMR releases 6.15 or lower, you can only assign node labels by node type, such as core and task nodes. However, if you're using Amazon EMR release 7.0 or higher, you can configure node labels by node type and market type, such as On-Demand and Spot.

If application process demand increases and executor demand decreases when you restricted the application process to core nodes, you can add back core nodes and remove task nodes in the same resize operation. For more information, see Understanding node allocation strategy and scenarios.

Amazon EMR doesn't label task nodes, so you can't set the YARN properties to restrict application processes only for task nodes. However, if you want to use market types as node labels, you can use the ON_DEMAND or SPOT labels for application process placement. We don't recommend using Spot nodes for application primary processes.

When using node labels, the total running units in the cluster can temporarily exceed the max compute set in your managed scaling policy while Amazon EMR decommissions some of your instances. Total requested units will always stay at or below your policy’s max compute.

Managed scaling only supports the node labels ON_DEMAND and SPOT or CORE and TASK. Custom node labels aren't supported.

Amazon EMR creates node labels when creating the cluster and provisioning resources. Amazon EMR doesn't support adding node labels when you reconfigure the cluster. You also can't modify the node labels when configuring managed scaling after launching the cluster.

Managed scaling scales core and task nodes independently based on application process and executor demand. To prevent HDFS data loss issues during core scale down, follow standard practice for core nodes. To learn more about best practices about core nodes and HDFS replication, see Considerations and best practices.

You can't place both the application process and executors on only the core or the ON_DEMAND node. If you want to add both the application process and executors on one of the nodes, don't use the yarn.node-labels.am.default-node-label-expression configuration.

For example, to place both the application process and executors in ON_DEMAND nodes, set max compute to the same as the maximum in the ON_DEMAND node. Also remove the yarn.node-labels.am.default-node-label-expression configuration.

To add both the application process and executors on core nodes, remove the yarn.node-labels.am.default-node-label-expression configuration.

When you use managed scaling with node labels, set the property yarn.scheduler.capacity.maximum-am-resource-percent: 1 if you plan to run multiple applications in parallel. Doing so ensures that your application processes fully utilize the available CORE or ON_DEMAND nodes.

If you use managed scaling with node labels, set the property yarn.resourcemanager.decommissioning.timeout to a value that's longer than the longest running application on your cluster. Doing so reduces the chance that Amazon EMR managed scalling needs to reschedule your applications to recommission CORE or ON_DEMAND nodes.