AWS Compute Blog

Optimizing Apache Flink on Amazon EKS using Amazon EC2 Spot Instances

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

Apache Flink is a distributed data processing engine for stateful computations for both batch and stream data sources. Flink supports event time semantics for out-of-order events, exactly-once semantics, backpressure control, and optimized APIs. Flink has connectors for third-party data sources and AWS Services, such as Apache Kafka, Apache NiFi, Amazon Kinesis, and Amazon MSK. Flink can be used for Event Driven (Fraud Detection), Data Analytics (Ad-Hoc Analysis), and Data Pipeline (Continuous ETL) applications. Amazon Elastic Kubernetes Service (Amazon EKS) is the chosen deployment option for many AWS customers for Big Data frameworks such as Apache Spark and Apache Flink. Flink has native integration with Kubernetes allowing direct deployment and dynamic resource allocation.

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


Amazon EC2 Spot Instances

Amazon EC2 Spot Instances let you take advantage of spare EC2 capacity in the AWS Cloud and are available at up to a 90% discount compared to On-Demand Instances. Spot Instances receive a two-minute warning when these instances are about to be reclaimed by Amazon EC2. There are many graceful ways to handle the interruption. Recently EC2 Instance rebalance recommendation has been added to send proactive notifications when a Spot Instance is at elevated risk of interruption. Spot Instances are a great way to scale up and increase throughput of Big Data workloads and has been adopted by many customers.

Apache Flink and Kubernetes

Apache Flink is an adaptable framework and it allows multiple deployment options and one of them being Kubernetes. Flink framework has a couple of key building blocks.

  • Job Client submits the job in form of a JobGraph to the Job Manager.
  • Job Manager plays the role of central work coordinator which distributes the job to the Task Managers.
  • Task Managers are the worker component, which runs the operators for source, transformations and sinks.
  • External components which are optional such as Resource Provider, HA Service Provider, Application Data Source, Sinks etc., and this varies with the deployment mode and options.

Image shows Flink application deployment architecture with Job Manager, Task Manager, Scheduler, Data Flow Graph, and client.

Flink supports different deployment (Resource Provider) modes when running on Kubernetes. In this blog we will use the Standalone Deployment mode, as we just want to showcase the functionality. We recommend first-time users however to deploy Flink on Kubernetes using the Native Kubernetes Deployment.

Flink can be run in different modes such as Session, Application, and Per-Job. The modes differ in cluster lifecycle, resource isolation and execution of the main() method. Flink can run jobs on Kubernetes via Application and Session Modes only.

  • Application Mode: This is a lightweight and scalable way to submit an application on Flink and is the preferred way to launch application as it supports better resource isolation. Resource isolation is achieved by running a cluster per job. Once the application shuts down all the Flink components are cleaned up.
  • Session Mode: This is a long running Kubernetes deployment of Flink. Multiple applications can be launched on a cluster and the applications competes for the resources. There may be multiple jobs running on a TaskManager in parallel. Its main advantage is that it saves time on spinning up a new Flink cluster for new jobs, however if one of the Task Managers fails it may impact all the jobs running on that.

Amazon EKS

Amazon EKS is a fully managed Kubernetes service. EKS supports creating and managing 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. EKS-managed node groups require less operational effort compared to using self-managed nodes. You can learn more in the blog “Amazon EKS now supports provisioning and managing EC2 Spot Instances in managed node groups.”

Apache Flink and Spot

Big Data frameworks like Spark and Flink are distributed to manage and process high volumes of data. Designed for failure, they can run on machines with different configurations, inherently resilient and flexible. Spot Instances can optimize runtimes by increasing throughput, while spending the same (or less). Flink can tolerate interruptions using restart and failover strategies.

Fault Tolerance

Fault tolerance is implemented in Flink with the help of check-pointing the state. Checkpoints allow Flink to recover state and positions in the streams. There are two per-requisites for check-pointing a persistent data source (Apache Kafka, Amazon Kinesis) which has the ability to replay data and a persistent distributed storage to store state (Amazon Simple Storage Service (Amazon S3), HDFS).

Cost Optimization

Job Manager and Task Manager are key building blocks of Flink. The Task Manager is the compute intensive part and Job Manager is the orchestrator. We would be running Task Manager on Spot Instances and Job Manager on On Demand Instances.


Flink supports elastic scaling via Reactive Mode, Task Managers can be added/removed based on metrics monitored by an external service monitor like Horizontal Pod Autoscaling (HPA). When scaling up new pods would be added, if the cluster has resources they would be scheduled it not then they will go in pending state. Cluster Autoscaler (CA) detects pods in pending state and new nodes will be added by EC2 Auto Scaling. This is ideal with Spot Instances as it implements elastic scaling with higher throughput in a cost optimized way.

Tutorial: Running Flink applications in a cost optimized way

In this tutorial, I review steps, which help you launch cost optimized and resilient Flink workloads running on EKS via Application mode. The streaming application will read dummy Stock ticker prices send to an Amazon Kinesis Data Stream by Amazon Kinesis Data Generator, try to determine the highest price within a per-defined window, and output will be written onto Amazon S3 files.

Image shows Flink application pipeline with data flowing from Amazon Kinesis Data Generator to Kinesis Data Stream, processed in Apache Flink and output being written in Amazon S3

The configuration files can be found in this github location. To run the workload on Kubernetes, make sure you have eksctl and kubectl command line utilities 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 Administrator Access 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 EKS Managed node group for Spot Instances.


When we deploy Flink in Application Mode it runs as a single application. The cluster is exclusive for the job. We will be bundling the user code in the Flink image for that purpose and upload in Amazon Elastic Container Registry (Amazon ECR). Amazon ECR is a fully managed container registry that makes it easy to store, manage, share, and deploy your container images and artifacts anywhere.

1. Build the Amazon ECR Image

  • Login using the following cmd and don’t forget to replace the AWS_REGION and AWS_ACCOUNT_ID with your details.

aws ecr get-login-password --region ${AWS_REGION} | docker login --username AWS —password-stdin ${ACCOUNT_ID}.dkr.ecr.${AWS_REGION}

  • Create a repository

aws ecr create-repository --repository-name flink-demo --image-scanning-configuration scanOnPush=true —region ${AWS_REGION}

  • Build the Docker image:

Download the Docker file. I am using multistage docker build here. The sample code is from Github’s Amazon Kinesis Data Analytics Java examples. I modified the code to allow checkpointing and change the sliding window interval. Build and push the docker image using the following instructions.

docker build --tag flink-demo .

  • Tag and Push your image to Amazon ECR

docker tag flink-demo:latest ${ACCOUNT_ID}.dkr.ecr.${AWS_REGION}
docker push ${ACCOUNT_ID}.dkr.ecr.${AWS_REGION}.

2. Create Amazon S3/Amazon Kinesis Access Policy

First, I must create an access policy to allow the Flink application to read/write from Amazon fFS3 and read Kinesis data streams. Download the Amazon S3 policy file from here and modify the <<output folder>> to an Amazon S3 bucket which you have to create.

  • Run the following to create the policy. Note the ARN.

aws iam create-policy --policy-name flink-demo-policy --policy-document file://flink-demo-policy.json

3. Cluster and node groups deployment

  • Create an EKS cluster using the following command:

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

The cluster takes approximately 15 minutes to launch.

  • Create the node group using the nodeGroup config file. I am using multiple nodeGroups of different sizes to adapt Spot best practice of diversification.  Replace the <<Policy ARN>> string using the ARN string from the previous step.

eksctl create nodegroup -f managedNodeGroups.yml

  • Download the Cluster Autoscaler and edit it to add the cluster-name (flink-demo)

curl -LO

4. Install the Cluster AutoScaler using the following command:

kubectl apply -f cluster-autoscaler-autodiscover.yaml

  • Using EKS Managed node groups requires significantly less operational effort compared to using self-managed node group and enables:
    • Auto enforcement of Spot best practices.
    • Spot Instance lifecycle management.
    • Auto labeling of Pods.
  • eksctl has integrated amazon-ec2-instance-selector to enable auto selection of instances based on the criteria passed. This has multiple benefits
    • ‘instance diversification’ is implemented by enabling multiple instance types selection in the node group which works well with CA
    • Reduces manual effort of selecting the instances.
  • We can create node group manifests using ‘dryrun’ and then create node groups using that.

eksctl create cluster --name flink-demo --instance-selector-vcpus=2 --instance-selector-memory=4 --dry-run

eksctl create node group -f managedNodeGroups.yml

5. Create service accounts for Flink

$ kubectl create serviceaccount flink-service-account
$ kubectl create clusterrolebinding flink-role-binding-flink --clusterrole=edit --serviceaccount=default:flink-service-account

6. Deploy Flink

This install folder here has all the YAML files required to deploy a standalone Flink cluster. Run the file. This will deploy the cluster with a JobManager, a pool of TaskManagers and a Service exposing JobManager’s ports.

  • This is a High-Availability(HA) deployment of Flink with the use of Kubernetes high availability service.
  • The JobManager runs on OnDemand and TaskManager on Spot. As the cluster is launched in Application Mode, if a node is interrupted only one job will be restarted.
  • Autoscaling is enabled by the use of ‘Reactive Mode’. Horizontal Pod Autoscaler is used to monitor the CPU load and scale accordingly.
  • Check-pointing is enabled which allows Flink to save state and be fault tolerant.

Image shows the Flink dashboard highlighting checkpoints for a job

7. Create Amazon Kinesis data stream and send dummy data      

Log in to AWS Management Console and create a Kinesis data stream name ‘ExampleInputStream’. Kinesis Data Generator is used to send data to the data stream. The template of the dummy data can be found here. Once this sends data the Flink application starts processing.

Image shows Amazon Kinesis Data Generator console sending data to Kinesis Data Strea


Spot Interruptions

If there is an interruption then the Flick application will be restarted using check-pointed data. The JobManager will restore the job as highlighted in the following log. The node will be replaced automatically by the Managed Node Group.

mage shows logs from a Flink job highlighting job restart using checkpoints.

One will be able to observe the graceful restart in the Flink UI.

Image shows the Flink dashboard highlighting job restart after failure.


You can observe the elastic scaling using logs. The number of TaskManagers in the Flink UI will also reflect the scaling state.

Image shows kubectl output showing status of HPA during scale-out


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

  • Run the file.
  • Delete the EKS cluster and the node groups:
    • eksctl delete cluster --name flink-demo
  • Delete the Amazon S3 Access Policy:
    • aws iam delete-policy --policy-arn <<POLICY ARN>>
  • Delete the Amazon S3 Bucket:
    • aws s3 rb --force s3://<<S3_BUCKET>>
  • Delete the CloudFormation stack related to Kinesis Data Generator named ‘Kinesis-Data-Generator-Cognito-User’
  • Delete the Kinesis Data Stream.


In this blog, I demonstrated how you can run Flink workloads on a Kubernetes Cluster using Spot Instances, achieving scalability, resilience, and cost optimization. To cost optimize your Flink based big data workloads you should start thinking about using Amazon EKS and Spot Instances.