Capacity Management and Amazon EMR Managed Scaling improvements for Amazon EMR on EC2 clusters
In 2022, we told you about the new enhancements we made in Amazon EMR Managed Scaling, which helped improve cluster utilization as well as reduced cluster costs. In 2023, we are happy to report that the Amazon EMR team has been hard at work. We worked backward from customer requirements and launched multiple new features to enhance your Amazon EMR on EC2 clusters capacity management and scaling experience.
Amazon EMR is the cloud big data solution for petabyte-scale data processing, interactive analytics, and machine learning (ML) using open-source frameworks such as Apache Spark, Apache Hive, and Presto. Customers asked us for features that would further improve the capacity management and scaling experience of their EMR on EC2 clusters, including their large, long-running clusters. We have been hard at work to meet those needs. The following are some of the key enhancements:
- Enhanced customer transparency and flexibility with provisioning timeout for Spot Instances
- Optimized task nodes scale-up for Amazon EMR on EC2 clusters launched with instance groups
- Improved job resiliency with enhanced protection for Spark Drivers
Let’s dive deeper and discuss the new Amazon EMR on EC2 features in detail.
Enhanced customer transparency and flexibility with provisioning timeout for Spot Instances
Many Amazon EMR customers use EC2 Spot Instances for their EMR on EC2 clusters to reduce costs. Spot Instances are spare Amazon Elastic Compute Cloud (Amazon EC2) compute capacity offered at discounts of up to 90% compared to On-Demand pricing. Amazon EMR offers you the capability to scale your cluster either manually or by using Automatic Scaling. You can also use the Amazon EMR Managed Scaling feature to automatically resize your cluster based on workload and utilization.
To enhance the customer experience when scaling up using Spot Instances, for EMR on EC2 clusters launched using instance fleets, you can now specify a provisioning timeout for Spot Instances. A provisioning timeout will tell Amazon EMR to stop provisioning Spot Instance capacity if the cluster exceeds a specified time threshold during cluster scaling operations. You can configure the Spot instance provisioning timeout for clusters getting resized manually or using Amazon EMR Managed Scaling and Auto Scaling.
Additionally, to provide better transparency, when the timeout period expires, Amazon EMR will also automatically send events to an Amazon CloudWatch Events stream. With these CloudWatch events, you can create rules that match events according to a specified pattern, and then route the events to targets to take action. To learn more, please refer to Customize a provisioning timeout period for cluster resize in Amazon EMR.
Please find summarized below the experience for different scenario’s when you configure a provisioning timeout period during resize for your Amazon EMR on EC2 cluster
|Amazon EMR is able to provision the desired Spot capacity before expiration of the provisioning timeout||Amazon EMR automatically scales-up the cluster to the desired capacity and no action is needed from the customer|
|Amazon EMR is not able to provision any Spot capacity or only able to provision partial Spot capacity and the provisioning timeout has expired||If Amazon EMR can’t provision the required Spot capacity and the provisioning timeout has expired, Amazon EMR will cancel the resize request and stops it’s attempts to provision additional Spot capacity. Amazon EMR will also publish events to an Amazon CloudWatch Events stream. Customers can use these events to create rules and take appropriate actions|
|If the Spot instances in your Amazon EMR on EC2 clusters are interrupted as Amazon EC2 needs them back||Amazon EMR will automatically trigger a new resize request to rebalance your clusters by replacing instances with any of the available types in your cluster. Amazon EMR will also use the same provisioning resize timeout which was configured on the cluster. No action is needed from the customer.|
You should consider the criticality of capacity availability when specifying the provisioning timeout value:
- When your workload capacity availability is critical – To ensure the desired capacity is available, we recommend configuring the resize provisioning timeout based on the time it takes to run the application and application SLAs. For example, if application SLA is 60 minutes and it takes 30 minutes for the application to complete, you should set the resize provisioning timeout to 30 minutes or less. Amazon EMR will try to provision to get Spot capacity until the timeout expires (30 minutes or less) and publish a CloudWatch event so that you can take appropriate actions.
- When your workload is time flexible and capacity availability is not a factor – If the workload is time flexible and capacity availability is not a factor, to ensure the highest likelihood for getting the desired Spot capacity, you can configure a higher timeout value for the resize provisioning timeout.
Optimized task nodes scale-up for Amazon EMR on EC2 clusters launched with Instance groups
Instance groups offer a simpler setup to launch EMR on EC2 clusters. Each cluster launched using instance groups can include up to 50 instance groups: one primary instance group that contains one EC2 instance, a core instance group that contains one or more EC2 instances, and up to 48 optional task instance groups. You can scale each instance group by adding and removing EC2 instances manually, or you can set up automatic scaling. You can also use the Amazon EMR Managed Scaling feature to automatically resize your cluster based on workload and utilization.
To enhance the customer experience for instance groups on EMR on EC2 clusters when scaling up task nodes using Amazon EMR Managed Scaling, we have enhanced the managed scaling algorithm to choose the task instance groups that have the highest likelihood of acquiring capacity. Furthermore, when managed scaling is not able to acquire capacity with a single task instance group, to reduce any scale-up delays, Amazon EMR will automatically switch to another task group and fulfill the capacity by using multiple task instance groups. Consequently, the more flexible you are about your instance types, the higher the chances of provisioning capacity. To learn more, refer to Best practices for instance and Availability Zone flexibility.
Improved job resiliency with enhanced protection for Spark Drivers
In 2022, to improve the job resiliency when using Amazon EMR Managed Scaling, we enhanced managed scaling to be Spark shuffle data aware, which prevents scale-down of instances that store intermediate shuffle data for Apache Spark. This helps prevents job reattempts and recomputations, which leads to better performance and lower cost.
To further improve job resiliency when using Amazon EMR Managed Scaling, we have further enhanced managed scaling to be Spark Driver aware, which ensures that during cluster scale-down, Amazon EMR Managed Scaling prioritizes the scale-down of nodes that don’t have an active Spark Driver running on them. This helps minimize job failures and job retries, helping further improve performance and reduce costs. This enhancement is enabled by default for EMR clusters using Amazon EMR versions 5.34.0 and later, and Amazon EMR versions 6.4.0 and later.
To confirm which nodes in your cluster are running Spark Driver, you can visit the Spark History Server and filter for the driver on the Executors tab of your Spark application ID.
In this post, we highlighted the improvements that we made in capacity management and Amazon EMR Managed Scaling for EMR on EC2 clusters. We focused on improving job resiliency, enhanced flexibility and transparency when provisioning Spot Instances, and optimizing the scale-up experience when using managed scaling with instance groups on Amazon EMR on EC2 clusters. Although we have launched multiple features so far in 2023 and the pace of innovation continues to accelerate, it remains day 1 and we look forward to hearing from you on how these features help you unlock more value for your organizations. We invite you to try these new features and get in touch with us through your AWS account team if you have further comments.
About the authors
Sushant Majithia is a Principal Product Manager for EMR at AWS.
Ankur Goyal is a SDM with Amazon EMR Big Data Platform team. He builds large scale distributed applications and cluster optimization algorithms. Ankur is interested in topics of Analytics, Machine Learning and Forecasting.
Matthew Liem is a Senior Solution Architecture Manager at AWS.
Tarun Chanana is an SDM with Amazon EMR Big Data Platform team.