Flink Kubernetes Deployment¶

Flink offers a k8s Operator to deploy and manage applications. This note summarize how to use this operator, with basic getting started yaml files.

The operator takes care of submitting, savepointing, upgrading and generally managing Flink jobs using the built-in Flink Kubernetes integration. I

The operator fully automates the entire lifecycle of job manager, task managers, and applications. Failures of Job Manager pods are handled by the Deployment Controller which will take care of spawning a new Job Manager.

As other operator it can run namespace-scoped, to get multiple versions of the operator in the same Kubernetes cluster, or cluster-scoped for highly distributed deployment.

The following figure represents a simple deployment view of a Flink and a Kafka clusters on kubernetes platform:

The custom resource definition that describes the schema of a FlinkDeployment is a cluster wide resource. The Operator continuously tracks cluster events relating to the FlinkDeployment and FlinkSessionJob custom resources. The operator control flow is described in this note.

Confluent Platform for Flink has also an operator compatible with the open-source one.

Deploy Flink Kubernetes Operator¶

Flink Kubernetes Operator acts as a control plane to manage the complete deployment lifecycle of Apache Flink applications.

For hands-on instructions and Makefile to deploy Flink, Confluent Platform on k8s (local colima or minikube) see this readme and the Confluent product instructions.

Custom Resources¶

Once the operator is running, we can submit jobs using FlinkDeployment (for Flink Application) and FlinkSessionJob Custom Resources for Session (Confluent Managed for Flink supports application mode only and is using a new CRD for FlinkApplication).

The FlinkDeployment spec is here and is used to define Flink application (will have a job section) or session cluster (only job and task managers configuration).

It is important to note that FlinkDeployment and FlinkApplication CRD have a podTemplate, so config map and secrets can be used to configure environment variables for the flink app. (Be sure to keep the name flink-main-container)

spec:
  podTemplate:
    spec:
      containers:
        - name: flink-main-container
          envFrom:
            - configMapRef:
                name: flink-app-cm

A RWX, shared PersistentVolumeClaim (PVC) for the Flink JobManagers and TaskManagers provides persistence for stateful checkpoint and savepoint of Flink jobs.

A flow is a packaged as a jar, so developers need to define a docker image with the Flink API and any connector jars. Example of Dockerfile and FlinkApplication manifest.

Also one solution includes using MinIO to persist application jars.

HA configuration¶

Within Kubernetes, we can enable Flink HA in the ConfigMap of the cluster configuration:

  flinkConfiguration:
    taskmanager.numberOfTaskSlots: "2"
    state.backend: rockdb
    state.savepoints.dir: file:///flink-data/savepoints
    state.checkpoints.dir: file:///flink-data/checkpoints
    high-availability.type: kubernetes
    high-availability.storageDir: file:///flink-data/ha

JobManager metadata is persisted in the file system high-availability.storageDir . This storageDir stores all metadata needed to recover a JobManager failure.

JobManager Pod that crashed are restarted automatically by the kubernetes scheduler, and as Flink persists metadata and the job artifacts, it is important to mount pv to the expected paths.

podTemplate:
  spec:
    containers:
      - name: flink-main-container
        volumeMounts:
        - mountPath: /flink-data
          name: flink-volume
    volumes:
    - name: flink-volume
      hostPath:
        # directory location on host
        path: /tmp/flink
        # this field is optional
        type: Directory

Recall that podTemplate is a base declaration common for job and task manager pods. Can be overridden by the jobManager and taskManager pod template sub-elements. The previous declaration will work for minikube with hostPath access, for Kubernetes cluster with separate storage class then the volume declaration is:

volumes:
  - name: flink-volume
    persistenceVolumeClaim:
      claimName: flink-pvc

For Flink job or application, it is important to enable checkpointing and savepointing:

job:
  jarURI: local:///opt/flink/examples/streaming/StateMachineExample.jar

  parallelism: 2
  upgradeMode: savepoint
  #savepointTriggerNonce: 0
  # initialSavepointPath: file:///

Flink Config Update¶

• If a write operation fails when the pod creates a folder or updates the Flink config, verify the following:

  • Assess PVC and R/W access. Verify PVC configuration. Some storage classes or persistent volume types may have restrictions on directory creation
  • Verify security context for the pod. Modify the pod's security context to allow necessary permissions.
  • The podTemplate can be configured at the same level as the task and job managers so any mounted volumes will be available to those pods. See basic-reactive.yaml from Flink Operator examples.

See PVC and PV declarations

Flink Session¶

For Session cluster, there is no jobSpec. See this deployment definition. Once a cluster is defined, it has a name and can be referenced to submit SessionJobs.

A SessionJob is executed as a long-running Kubernetes Deployment. We may run multiple Flink jobs on a Session cluster. Each job needs to be submitted to the cluster after the cluster has been deployed. To deploy a job, we need at least three components:

• a Deployment which runs a JobManager
• a Deployment for a pool of TaskManagers
• a Service exposing the JobManager’s REST and UI ports

For a deployment select the execution mode: application, or session. For production it is recommended to deploy in application mode for better isolation, and using a cloud native approach. We can just build a dockerfile for our application using the Flink jars.

Session Deployment¶

Flink has a set of examples like the Car top speed computation with simulated record. As this code is packaged in a jar available in maven repository, we can declare a job session.

Deploy a config map to define the log4j-console.properties and other parameters for Flink (flink-conf.yaml)

The diagram below illustrates the standard deployment of a job on k8s with session mode:

src: apache Flink site

apiVersion: flink.apache.org/v1beta1
kind: FlinkSessionJob
metadata:
  name: car-top-speed-job
spec:
  deploymentName: flink-session-cluster
  job:
    jarURI: https://repo1.maven.org/maven2/org/apache/flink/flink-examples-streaming_2.12/1.17.2/flink-examples-streaming_2.12-1.17.2-TopSpeedWindowing.jar
    parallelism: 4
    upgradeMode: stateless

Before deploying this job, be sure to deploy a session cluster using the following command:

# under deployment/k8s
kubectl apply -f basic-job-task-mgrs.yaml 

Once the job is deployed we can see the pod and then using the user interface the job continuously running:

• Example of deploying Java based SQL Runner to interpret a Flink SQL script: package it as docker images, and deploy it with a Session Job. There is a equivalent for Python using Pyflink.

  • See the ported code for Java
  • And for the Python implementation

Flink State Snapshot¶

To help managing snapshots, there is another CR called FlinkStateSnapshot

Flink Application deployment¶

An application deployment must define the job (JobSpec) field with the jarURI, parallelism, upgradeMode one of (stateless/savepoint/last-state) and the desired state of the job (running/suspended). See this sample app or the cmf_app_deployment.yaml in the e-com-sale demonstration.

job:
    jarURI: local:///opt/flink/examples/streaming/StateMachineExample.jar
    # your own deployment
    jarURI: local:///opt/flink/usrlib/yourapp01.0.0.jar
    parallelism: 2
    upgradeMode: stateless
    state: running

flinkConfiguration is a hash map used to define the Flink configuration, such as the task slot, HA and checkpointing parameters.

  flinkConfiguration:
    high-availability.type: org.apache.flink.kubernetes.highavailability.KubernetesHaServicesFactory
    high-availability.storageDir: 'file:///opt/flink/volume/flink-ha'
    restart-strategy: failure-rate
    restart-strategy.failure-rate.max-failures-per-interval: '10'
    restart-strategy.failure-rate.failure-rate-interval: '10 min'
    restart-strategy.failure-rate.delay: '30 s'
    execution.checkpointing.interval: '5000'
    execution.checkpointing.unaligned: 'false'
    state.backend.type: rocksdb
    state.backend.incremental: 'true'
    state.backend.rocksdb.use-bloom-filter: 'true'
    state.checkpoints.dir: 'file:///opt/flink/volume/flink-cp'
    state.checkpoints.num-retained: '3'
    state.savepoints.dir: 'file:///opt/flink/volume/flink-sp'
    taskmanager.numberOfTaskSlots: '10'
    table.exec.source.idle-timeout: '30 s'

The application jar needs to be in a custom Flink docker image built using the Dockerfile as in e-com-sale-demo, or uploaded to MinIO.

The following Dockerfile is used for deploying a solution in application mode, which packages the Java Flink jars with the app, and any connector jars needed for the integration and starts the main() function.

FROM flink
RUN mkdir -p $FLINK_HOME/usrlib
COPY /path/of/my-flink-job-*.jar $FLINK_HOME/usrlib/my-flink-job.jar

• With CP for Flink:

    # First be sure the service is expose
    kubectl port-forward svc/cmf-service 8080:80 -n flink
    # Deploy the app given its deployment
    confluent flink application create k8s/cmf_app_deployment.yaml  --environment $(ENV_NAME) --url http://localhost:8080 
  

kubectl port-forward ${flink-jobmanager-pod} 8081:8081 
# Or using CP for Flink
confluent flink application web-ui-forward $(APP_NAME) --environment $(ENV_NAME) --port 8081 --url http://localhost:8080

Using MinIO¶

MinIO is an object storage solution that provides an Amazon Web Services S3-compatible API and supports all core S3 features, on k8s.

• First be sure the MinIO is installed on k8s. See minio-dev.yaml

sh brew install minio/stable/mc # Verify installation mc --help

• Config minio under minio-dev namespace

  # under deploymenet/k8s/cpf
  kubectl apply -f minio-dev.yaml
  kubectl get pods -n minio-dev
  

• Access MinIO S3 API and Console

  kubectl port-forward pod/minio 9000 9090 -n minio-dev
  

• Log in to the Console with the credentials minioadmin | minioadmin

• Setup a minio client with credential saved to $HOME/.mc/config.json

  mc alias set dev-minio http://localhost:9000 minioadmin minioadmin
  # make a bucket
  mc mb dev-minio/flink
  

• Upload an application to minio bucket:

  mc cp ./target/flink-app-0.1.0.jar dev-minio/flink/flink-app-0.1.0.jar
  mc ls dev-minio/flink
  

• Start the application:

  confluent flink application create --environment env1 --url http://localhost:8080 app-deployment.json
  

• Open Flink UI:

  confluent flink application web-ui-forward --environment env1 flink-app --url http://localhost:8080
  

• Produce messages to kafka topic

  echo 'message1' | kubectl exec -i -n confluent kafka-0 -- /bin/kafka-console-producer --bootstrap-server kafka.confluent.svc.cluster.local:9092 --topic in
  

• Cleanup

  # the Flink app
  confluent flink application delete kafka-reader-writer-example --environment development --url http://localhost:8080
  # the Kafka cluster
  # the operators
  

Flink SQL processing¶

There are multiple choices to run Flink SQL, using the SQL client, or package the SQL scripts and get a java SQL runner executing the SQL statements from a file, so the application deployment is Java based even if SQL scripts are used for stream processing.

Practices¶

• It is not recommended to host a Flink Cluster across multiple Kubernetes clusters. Flink node exchanges data between task managers and so better to run in same region, and within same k8s.