Running cost optimized Spark workloads on Kubernetes using EC2 Spot Instances

This post is written by Kinnar Sen, Senior Solutions Architect, EC2 Spot 

Apache Spark is an open-source, distributed processing system used for big data workloads. It provides API operations to perform multiple tasks such as streaming, extract transform load (ETL), query, machine learning (ML), and graph processing. Spark supports four different types of cluster managers (Spark standalone, Apache Mesos, Hadoop YARN, and Kubernetes), which are responsible for scheduling and allocation of resources in the cluster. Spark can run with native Kubernetes support since 2018 (Spark 2.3). AWS customers that have already chosen Kubernetes as their container orchestration tool can also choose to run Spark applications in Kubernetes, increasing the effectiveness of their operations and compute resources.

In this post, I illustrate the deployment of scalable, resilient, and cost optimized Spark application using Kubernetes via Amazon Elastic Kubernetes Service (Amazon EKS) and Amazon EC2 Spot Instances. Learn how to save money on big data workloads by implementing this solution.

Overview

Amazon EC2 Spot Instances

Amazon EC2 Spot Instances let you take advantage of unused EC2 capacity in the AWS Cloud. Spot Instances are available at up to a 90% discount compared to On-Demand Instance prices. Capacity pools are a group of EC2 instances that belong to particular instance family, size, and Availability Zone (AZ). If EC2 needs capacity back for On-Demand Instance usage, Spot Instances can be interrupted by EC2 with a two-minute notification. There are many graceful ways to handle the interruption to ensure that the application is well architected for resilience and fault tolerance. This can be automated via the application and/or infrastructure deployments. Spot Instances are ideal for stateless, fault tolerant, loosely coupled and flexible workloads that can handle interruptions.

Amazon Elastic Kubernetes Service

Amazon EKS is a fully managed Kubernetes service that makes it easy for you to run Kubernetes on AWS without needing to install, operate, and maintain your own Kubernetes control plane. It provides a highly available and scalable managed control plane. It also provides managed worker nodes, which let you create, update, or terminate worker nodes for your cluster with a single command. It is a great choice for deploying flexible and fault tolerant containerized applications. Amazon EKS supports creating and managing Amazon EC2 Spot Instances using Amazon EKS-managed node groups following Spot best practices. This enables you to take advantage of the steep savings and scale that Spot Instances provide for interruptible workloads running in your Kubernetes cluster. Using EKS-managed node groups with Spot Instances requires less operational effort compared to using self-managed nodes. In addition to launching Spot Instances in managed node groups, it is possible to specify multiple instance types in EKS managed node groups. You can find more in this blog.

Apache Spark and Kubernetes

When a spark application is submitted to the Kubernetes cluster the following happens:

• A Spark driver is created.
• The driver and the run within pods.
• The Spark driver then requests for executors, which are scheduled to run within pods. The executors are managed by the driver.
• The application is launched and once it completes, the executor pods are cleaned up. The driver pod persists the logs and remains in a completed state until the pod is cleared by garbage collection or manually removed. The driver in a completed stage does not consume any memory or compute resources.

When a spark application runs on clusters managed by Kubernetes, the native Kubernetes scheduler is used. It is possible to schedule the driver/executor pods on a subset of available nodes. The applications can be launched either by a vanilla ‘spark submit’, a workflow orchestrator like Apache Airflow or the spark operator. I use vanilla ‘spark submit’ in this blog. Amazon EMR is is also able to schedule Spark applications on EKS clusters as described in this launch blog, but Amazon EMR on EKS is out of scope for this post.

Cost optimization

For any organization running big data workloads there are three key requirements: scalability, performance, and low cost. As the size of data increases, there is demand for more compute capacity and the total cost of ownership increases. It is critical to optimize the cost of big data applications. Big Data frameworks (in this case, Spark) are distributed to manage and process high volumes of data. These frameworks are designed for failure, can run on machines with different configurations, and are inherently resilient and flexible.

If Spark deploys on Kubernetes, the executor pods can be scheduled on EC2 Spot Instances and driver pods on On-Demand Instances. This reduces the overall cost of deployment – Spot Instances can save up to 90% over On-Demand Instance prices. This also enables faster results by scaling out executors running on Spot Instances. Spot Instances, by design, can be interrupted when EC2 needs the capacity back. If a driver pod is running on a Spot Instance, which is interrupted then the application fails and the application must be re-submitted. To avoid this situation, the driver pod can be scheduled on On-Demand Instances only. This adds a layer of resiliency to the Spark application running on Kubernetes. To cost optimize the deployment, all the executor pods are scheduled on Spot Instances as that’s where the bulk of compute happens. Spark’s inherent resiliency has the driver launch new executors to replace the ones that fail due to Spot interruptions.

There are a couple of key points to note here.

• The idea is to start with minimum number of nodes for both On-Demand and Spot Instances (one each) and then auto-scale usingCluster Autoscaler and EC2 Auto Scaling  Cluster Autoscaler for AWS provides integration with Auto Scaling groups. If there are not sufficient resources, the driver and executor pods go into pending state. The Cluster Autoscaler detects pods in pending state and scales worker nodes within the identified Auto Scaling group in the cluster using EC2 Auto Scaling.

• The scaling for On-Demand and Spot nodes is exclusive of one another. So, if multiple applications are launched the driver and executor pods can be scheduled in different node groups independently per the resource requirements. This helps reduce job failures due to lack of resources for the driver, thus adding to the overall resiliency of the system.
• Using EKS Managed node groups
  • This requires significantly less operational effort compared to using self-managed nodegroup and enables:
    • Auto enforcement of Spot best practices like Capacity Optimized allocation strategy, Capacity Rebalancing and use multiple instances types.
    • Proactive replacement of Spot nodes using rebalance notifications.
    • Managed draining of Spot nodes via re-balance recommendations.
  • The nodes are auto-labeled so that the pods can be scheduled with NodeAffinity.
    • eks.amazonaws.com/capacityType: SPOT
    • eks.amazonaws.com/capacityType: ON_DEMAND

Now that you understand the products and best practices of used in this tutorial, let’s get started.

Tutorial: running Spark in EKS managed node groups with Spot Instances

In this tutorial, I review steps, which help you launch cost optimized and resilient Spark jobs inside Kubernetes clusters running on EKS. I launch a word-count application counting the words from an Amazon Customer Review dataset and write the output to an Amazon S3 folder. To run the Spark workload on Kubernetes, make sure you have eksctl and kubectl installed on your computer or on an AWS Cloud9 environment. You can run this by using an AWS IAM user or role that has the AdministratorAccess policy attached to it, or check the minimum required permissions for using eksctl. The spot node groups in the Amazon EKS cluster can be launched both in a managed or a self-managed way, in this post I use the former. The config files for this tutorial can be found here. The job is finally launched in cluster mode.

Create Amazon S3 Access Policy

First, I must create an Amazon S3 access policy to allow the Spark application to read/write from Amazon S3. Amazon S3 Access is provisioned by attaching the policy by ARN to the node groups. This associates Amazon S3 access to the NodeInstanceRole and, hence, the node groups then have access to Amazon S3. Download the Amazon S3 policy file from here and modify the <<output folder>> to an Amazon S3 bucket you created. Run the following to create the policy. Note the ARN.

aws iam create-policy --policy-name spark-s3-policy --policy-document file://spark-s3.json

Cluster and node groups deployment

Create an EKS cluster using the following command:

eksctl create cluster –name= sparkonk8 --node-private-networking  --without-nodegroup --asg-access –region=<<AWS Region>>

The cluster takes approximately 15 minutes to launch.

Create the nodegroup using the nodeGroup config file. Replace the <<Policy ARN>> string using the ARN string from the previous step.

eksctl create nodegroup -f managedNodeGroups.yml

Scheduling driver/executor pods

The driver and executor pods can be assigned to nodes using affinity. PodTemplates can be used to configure the detail, which is not supported by Spark launch configuration by default. This feature is available from Spark 3.0.0, requiredDuringScheduling node affinity is used to schedule the driver and executor jobs. Sample podTemplates have been uploaded here.

Launching a Spark application

Create a service account. The spark driver pod uses the service account to create and watch executor pods using Kubernetes API server.

kubectl create serviceaccount spark
kubectl create clusterrolebinding spark-role --clusterrole='edit'  --serviceaccount=default:spark --namespace=default

Download the Cluster Autoscaler and edit it to add the cluster-name. 

curl -LO https://raw.githubusercontent.com/kubernetes/autoscaler/master/cluster-autoscaler/cloudprovider/aws/examples/cluster-autoscaler-autodiscover.yaml

Install the Cluster AutoScaler using the following command:

kubectl apply -f cluster-autoscaler-autodiscover.yaml

Get the details of Kubernetes master to get the head URL.

kubectl cluster-info 

Use the following instructions to build the docker image.

Download the application file (script.py) from here and upload into the Amazon S3 bucket created.

Download the pod template files from here. Submit the application.

bin/spark-submit \
--master k8s://<<MASTER URL>> \
--deploy-mode cluster \
--name 'Job Name' \
--conf spark.eventLog.dir=s3a:// <<S3 BUCKET>>/logs \
--conf spark.eventLog.enabled=true \
--conf spark.history.fs.inProgressOptimization.enabled=true \
--conf spark.history.fs.update.interval=5s \
--conf spark.kubernetes.container.image=<<ECR Spark Docker Image>> \
--conf spark.kubernetes.container.image.pullPolicy=IfNotPresent \
--conf spark.kubernetes.driver.podTemplateFile='../driver_pod_template.yml' \
--conf spark.kubernetes.executor.podTemplateFile='../executor_pod_template.yml' \
--conf spark.kubernetes.authenticate.driver.serviceAccountName=spark \
--conf spark.dynamicAllocation.enabled=true \
--conf spark.dynamicAllocation.shuffleTracking.enabled=true \
--conf spark.dynamicAllocation.maxExecutors=100 \
--conf spark.dynamicAllocation.executorAllocationRatio=0.33 \
--conf spark.dynamicAllocation.sustainedSchedulerBacklogTimeout=30 \
--conf spark.dynamicAllocation.executorIdleTimeout=60s \
--conf spark.driver.memory=8g \
--conf spark.kubernetes.driver.request.cores=2 \
--conf spark.kubernetes.driver.limit.cores=4 \
--conf spark.executor.memory=8g \
--conf spark.kubernetes.executor.request.cores=2 \
--conf spark.kubernetes.executor.limit.cores=4 \
--conf spark.hadoop.fs.s3a.impl=org.apache.hadoop.fs.s3a.S3AFileSystem \
--conf spark.hadoop.fs.s3a.connection.ssl.enabled=false \
--conf spark.hadoop.fs.s3a.fast.upload=true \
s3a://<<S3 BUCKET>>/script.py \
s3a://<<S3 BUCKET>>/output 

A couple of key points to note here

• podTemplateFile is used here, which enables scheduling of the driver pods to On-Demand Instances and executor pods to Spot Instances.
• Spark provides a mechanism to allocate resources dynamically based on workloads. In the latest release of Spark (3.0.0), dynamicAllocation can be used with Kubernetes cluster manager. The executors that do not store, active, shuffled files can be removed to free up the resources. DynamicAllocation works well in tandem with Cluster Autoscaler for resource allocation and optimizes resource for jobs. We are using dynamicAllocation here to enable optimized resource sharing.
• The application file and output are both in Amazon S3.

• Spark Event logs are redirected to Amazon S3. Spark on Kubernetes creates local temporary files for logs and removes them once the application completes. The logs are redirected to Amazon S3 and Spark History Server can be used to analyze the logs. Note, you can create more instrumentation using tools like Prometheus and Grafana to monitor and manage the cluster.

Observations

EC2 Spot Interruptions

The following diagram and log screenshot details from Spark History server showcases the behavior of a Spark application in case of an EC2 Spot interruption.

Four Spark applications launched in parallel in a cluster and one of the Spot nodes was interrupted. A couple of executor pods were terminated in three of the four applications, but due to the resilient nature of Spark new executors were launched and the applications finished almost around the same time.
The Spark Driver identified the shut down executors, which handled the shuffle files and relaunched the tasks running on those executors.

The Spark Driver identified the shut down executors, which handled the shuffle files and relaunched the tasks running on those executors.

Dynamic Allocation

Dynamic Allocation works with the caveat that it is an experimental feature.

Cost Optimization

Cost Optimization is achieved in several different ways from this tutorial.

• Use of 100% Spot Instances for the Spark executors
• Use of dynamicAllocation along with cluster autoscaler does make optimized use of resources and hence save cost
• With the deployment of one driver and executor nodes to begin with and then scaling up on demand reduces the waste of a continuously running cluster

Cluster Autoscaling

Cluster Autoscaling is triggered as it is designed when there are pending (Spark executor) pods.

The Cluster Autoscaler logs can be fetched by:

kubectl logs -f deployment/cluster-autoscaler -n kube-system —tail=10  

Cleanup

If you are trying out the tutorial, run the following steps to make sure that you don’t encounter unwanted costs.

Delete the EKS cluster and the nodegroups with the following command:

eksctl delete cluster --name sparkonk8

Delete the Amazon S3 Access Policy with the following command:

aws iam delete-policy --policy-arn <<POLICY ARN>>

Delete the Amazon S3 Output Bucket with the following command:

aws s3 rb --force s3://<<S3_BUCKET>>

Conclusion

In this blog, I demonstrated how you can run Spark workloads on a Kubernetes Cluster using Spot Instances, achieving scalability, resilience, and cost optimization. To cost optimize your Spark based big data workloads, consider running spark application using Kubernetes and EC2 Spot Instances.