Getting Started

Getting started with Spark Operator

For a more detailed guide on how to use, compose, and work with SparkApplications, please refer to the User Guide. If you are running the Kubernetes Operator for Apache Spark on Google Kubernetes Engine and want to use Google Cloud Storage (GCS) and/or BigQuery for reading/writing data, also refer to the GCP guide. The Kubernetes Operator for Apache Spark will simply be referred to as the operator for the rest of this guide.

Prerequisites

  • Helm >= 3
  • Kubernetes >= 1.16

Installation

Add Helm Repo

helm repo add spark-operator https://kubeflow.github.io/spark-operator

helm repo update

See helm repo for command documentation.

Install the chart

helm install [RELEASE_NAME] spark-operator/spark-operator

For example, if you want to create a release with name spark-operator in the spark-operator namespace:

helm install spark-operator spark-operator/spark-operator \
    --namespace spark-operator \
    --create-namespace

See helm install for command documentation.

Installing the chart will create a namespace spark-operator if it doesn’t exist, and helm will set up RBAC for the operator to run in the namespace. It will also set up RBAC in the default namespace for driver pods of your Spark applications to be able to manipulate executor pods. In addition, the chart will create a Deployment in the namespace spark-operator. The chart by default does not enable Mutating Admission Webhook for Spark pod customization. When enabled, a webhook service and a secret storing the x509 certificate called spark-webhook-certs are created for that purpose. To install the operator with the mutating admission webhook on a Kubernetes cluster, install the chart with the flag webhook.enable=true:

helm install my-release spark-operator/spark-operator \
    --namespace spark-operator \
    --create-namespace \
    --set webhook.enable=true

If you want to deploy the chart to GKE cluster, you will first need to grant yourself cluster-admin privileges before you can create custom roles and role bindings on a GKE cluster versioned 1.6 and up. Run the following command before installing the chart on GKE:

kubectl create clusterrolebinding <user>-cluster-admin-binding --clusterrole=cluster-admin --user=<user>@<domain>

Now you should see the operator running in the cluster by checking the status of the Helm release.

helm status --namespace spark-operator my-release

Upgrade the Chart

helm upgrade [RELEASE_NAME] spark-operator/spark-operator [flags]

See helm upgrade for command documentation.

Uninstall the Chart

helm uninstall [RELEASE_NAME]

This removes all the Kubernetes resources associated with the chart and deletes the release, except for the crds, those will have to be removed manually.

See helm uninstall for command documentation.

Running the Examples

To run the Spark PI example, run the following command:

kubectl apply -f examples/spark-pi.yaml

Note that spark-pi.yaml configures the driver pod to use the spark service account to communicate with the Kubernetes API server. You might need to replace it with the appropriate service account before submitting the job. If you installed the operator using the Helm chart and overrode sparkJobNamespaces, the service account name ends with -spark and starts with the Helm release name. For example, if you would like to run your Spark jobs to run in a namespace called test-ns, first make sure it already exists, and then install the chart with the command:

helm install my-release spark-operator/spark-operator --namespace spark-operator --set "sparkJobNamespaces={test-ns}"

Then the chart will set up a service account for your Spark jobs to use in that namespace.

See the section on the Spark Job Namespace for details on the behavior of the default Spark Job Namespace.

Running the above command will create a SparkApplication object named spark-pi. Check the object by running the following command:

kubectl get sparkapplication spark-pi -o=yaml

This will show something similar to the following:

apiVersion: sparkoperator.k8s.io/v1beta2
kind: SparkApplication
metadata:
  ...
spec:
  deps: {}
  driver:
    coreLimit: 1200m
    cores: 1
    labels:
      version: 2.3.0
    memory: 512m
    serviceAccount: spark
  executor:
    cores: 1
    instances: 1
    labels:
      version: 2.3.0
    memory: 512m
  image: gcr.io/ynli-k8s/spark:v3.1.1
  mainApplicationFile: local:///opt/spark/examples/jars/spark-examples_2.12-3.1.1.jar
  mainClass: org.apache.spark.examples.SparkPi
  mode: cluster
  restartPolicy:
      type: OnFailure
      onFailureRetries: 3
      onFailureRetryInterval: 10
      onSubmissionFailureRetries: 5
      onSubmissionFailureRetryInterval: 20
  type: Scala
status:
  sparkApplicationId: spark-5f4ba921c85ff3f1cb04bef324f9154c9
  applicationState:
    state: COMPLETED
  completionTime: 2018-02-20T23:33:55Z
  driverInfo:
    podName: spark-pi-83ba921c85ff3f1cb04bef324f9154c9-driver
    webUIAddress: 35.192.234.248:31064
    webUIPort: 31064
    webUIServiceName: spark-pi-2402118027-ui-svc
    webUIIngressName: spark-pi-ui-ingress
    webUIIngressAddress: spark-pi.ingress.cluster.com
  executorState:
    spark-pi-83ba921c85ff3f1cb04bef324f9154c9-exec-1: COMPLETED
  LastSubmissionAttemptTime: 2018-02-20T23:32:27Z

To check events for the SparkApplication object, run the following command:

kubectl describe sparkapplication spark-pi

This will show the events similarly to the following:

Events:
  Type    Reason                      Age   From            Message
  ----    ------                      ----  ----            -------
  Normal  SparkApplicationAdded       5m    spark-operator  SparkApplication spark-pi was added, enqueued it for submission
  Normal  SparkApplicationTerminated  4m    spark-operator  SparkApplication spark-pi terminated with state: COMPLETED

The operator submits the Spark Pi example to run once it receives an event indicating the SparkApplication object was added.

Configuration

The operator is typically deployed and run using the Helm chart. However, users can still run it outside a Kubernetes cluster and make it talk to the Kubernetes API server of a cluster by specifying path to kubeconfig, which can be done using the -kubeconfig flag.

The operator uses multiple workers in the SparkApplication controller. The number of worker threads are controlled using command-line flag -controller-threads which has a default value of 10.

The operator enables cache resynchronization so periodically the informers used by the operator will re-list existing objects it manages and re-trigger resource events. The resynchronization interval in seconds can be configured using the flag -resync-interval, with a default value of 30 seconds.

By default, the operator will install the CustomResourceDefinitions for the custom resources it manages. This can be disabled by setting the flag -install-crds=false, in which case the CustomResourceDefinitions can be installed manually using kubectl apply -f manifest/spark-operator-crds.yaml.

The mutating admission webhook is an optional component and can be enabled or disabled using the -enable-webhook flag, which defaults to false.

By default, the operator will manage custom resource objects of the managed CRD types for the whole cluster. It can be configured to manage only the custom resource objects in a specific namespace with the flag -namespace=<namespace>

Upgrade

To upgrade the operator, e.g., to use a newer version container image with a new tag, run the following command with updated parameters for the Helm release:

helm upgrade <YOUR-HELM-RELEASE-NAME> --set image.repository=org/image --set image.tag=newTag

Refer to the Helm documentation for more details on helm upgrade.

About Spark Job Namespaces

The Spark Job Namespaces value defines the namespaces where SparkApplications can be deployed. The Helm chart value for the Spark Job Namespaces is sparkJobNamespaces, and its default value is []. When the list of namespaces is empty the Helm chart will create a service account in the namespace where the spark-operator is deployed.

If you installed the operator using the Helm chart and overrode the sparkJobNamespaces to some other, pre-existing namespace, the Helm chart will create the necessary service account and RBAC in the specified namespace.

The Spark Operator uses the Spark Job Namespace to identify and filter relevant events for the SparkApplication CRD. If you specify a namespace for Spark Jobs, and then submit a SparkApplication resource to another namespace, the Spark Operator will filter out the event, and the resource will not get deployed. If you don’t specify a namespace, the Spark Operator will see only SparkApplication events for the Spark Operator namespace.

About the Service Account for Driver Pods

A Spark driver pod need a Kubernetes service account in the pod’s namespace that has permissions to create, get, list, and delete executor pods, and create a Kubernetes headless service for the driver. The driver will fail and exit without the service account, unless the default service account in the pod’s namespace has the needed permissions. To submit and run a SparkApplication in a namespace, please make sure there is a service account with the permissions in the namespace and set .spec.driver.serviceAccount to the name of the service account. Please refer to spark-rbac.yaml for an example RBAC setup that creates a driver service account named spark in the default namespace, with a RBAC role binding giving the service account the needed permissions.

About the Service Account for Executor Pods

A Spark executor pod may be configured with a Kubernetes service account in the pod namespace. To submit and run a SparkApplication in a namespace, please make sure there is a service account with the permissions required in the namespace and set .spec.executor.serviceAccount to the name of the service account.

Enable Metric Exporting to Prometheus

The operator exposes a set of metrics via the metric endpoint to be scraped by Prometheus. The Helm chart by default installs the operator with the additional flag to enable metrics (-enable-metrics=true) as well as other annotations used by Prometheus to scrape the metric endpoint. If podMonitor.enable is enabled, the helm chart will submit a pod monitor for the operator’s pod. To install the operator without metrics enabled, pass the appropriate flag during helm install:

helm install my-release spark-operator/spark-operator \
    --namespace spark-operator \
    --create-namespace \
    --set metrics.enable=false

If enabled, the operator generates the following metrics:

Spark Application Metrics

MetricDescription
spark_app_countTotal number of SparkApplication handled by the Operator.
spark_app_submit_countTotal number of SparkApplication spark-submitted by the Operator.
spark_app_success_countTotal number of SparkApplication which completed successfully.
spark_app_failure_countTotal number of SparkApplication which failed to complete.
spark_app_running_countTotal number of SparkApplication which are currently running.
spark_app_success_execution_time_microsecondsExecution time for applications which succeeded.
spark_app_failure_execution_time_microsecondsExecution time for applications which failed.
spark_app_start_latency_microsecondsStart latency of SparkApplication as type of Prometheus Summary.
spark_app_start_latency_secondsStart latency of SparkApplication as type of Prometheus Histogram.
spark_app_executor_success_countTotal number of Spark Executors which completed successfully.
spark_app_executor_failure_countTotal number of Spark Executors which failed.
spark_app_executor_running_countTotal number of Spark Executors which are currently running.

Work Queue Metrics

MetricDescription
spark_application_controller_depthCurrent depth of workqueue
spark_application_controller_addsTotal number of adds handled by workqueue
spark_application_controller_latencyLatency for workqueue
spark_application_controller_work_durationHow long processing an item from workqueue takes
spark_application_controller_retriesTotal number of retries handled by workqueue
spark_application_controller_unfinished_work_secondsUnfinished work in seconds
spark_application_controller_longest_running_processor_microsecondsLongest running processor in microseconds

The following is a list of all the configurations the operators supports for metrics:

-enable-metrics=true
-metrics-port=10254
-metrics-endpoint=/metrics
-metrics-prefix=myServiceName
-metrics-label=label1Key
-metrics-label=label2Key

All configs except -enable-metrics are optional. If port and/or endpoint are specified, please ensure that the annotations prometheus.io/port, prometheus.io/path and containerPort in spark-operator-with-metrics.yaml are updated as well.

A note about metrics-labels: In Prometheus, every unique combination of key-value label pairs represents a new time series, which can dramatically increase the amount of data stored. Hence, labels should not be used to store dimensions with high cardinality with potentially a large or unbounded value range.

Additionally, these metrics are best-effort for the current operator run and will be reset on an operator restart. Also, some of these metrics are generated by listening to pod state updates for the driver/executors and deleting the pods outside the operator might lead to incorrect metric values for some of these metrics.

Driver UI Access and Ingress

The operator, by default, makes the Spark UI accessible by creating a service of type ClusterIP which exposes the UI. This is only accessible from within the cluster.

The operator also supports creating an optional Ingress for the UI. This can be turned on by setting the ingress-url-format command-line flag. The ingress-url-format should be a template like {{$appName}}.{ingress_suffix}/{{$appNamespace}}/{{$appName}}. The {ingress_suffix} should be replaced by the user to indicate the cluster’s ingress url and the operator will replace the {{$appName}} & {{$appNamespace}} with the appropriate value. Please note that Ingress support requires that cluster’s ingress url routing is correctly set-up. For e.g. if the ingress-url-format is {{$appName}}.ingress.cluster.com, it requires that anything *ingress.cluster.com should be routed to the ingress-controller on the K8s cluster.

The operator also sets both WebUIAddress which is accessible from within the cluster as well as WebUIIngressAddress as part of the DriverInfo field of the SparkApplication.

The operator generates ingress resources intended for use with the Ingress NGINX Controller. Include this in your application spec for the controller to ensure it recognizes the ingress and provides appropriate routes to your Spark UI.

spec:
  sparkUIOptions:
    ingressAnnotations:
        kubernetes.io/ingress.class: nginx

About the Mutating Admission Webhook

The Kubernetes Operator for Apache Spark comes with an optional mutating admission webhook for customizing Spark driver and executor pods based on the specification in SparkApplication objects, e.g., mounting user-specified ConfigMaps and volumes, and setting pod affinity/anti-affinity, and adding tolerations.

The webhook requires a X509 certificate for TLS for pod admission requests and responses between the Kubernetes API server and the webhook server running inside the operator. For that, the certificate and key files must be accessible by the webhook server. The location of these certs is configurable and they will be reloaded on a configurable period. The Kubernetes Operator for Spark ships with a tool at hack/gencerts.sh for generating the CA and server certificate and putting the certificate and key files into a secret named spark-webhook-certs in the namespace spark-operator. This secret will be mounted into the operator pod.

Run the following command to create the secret with a certificate and key files using a batch Job, and install the operator Deployment with the mutating admission webhook:

kubectl apply -f manifest/spark-operator-with-webhook.yaml

This will create a Deployment named sparkoperator and a Service named spark-webhook for the webhook in namespace spark-operator.

Mutating Admission Webhooks on a private GKE or EKS cluster

If you are deploying the operator on a GKE cluster with the Private cluster setting enabled, or on an enterprise AWS EKS cluster and you wish to deploy the cluster with the Mutating Admission Webhook, then make sure to change the webhookPort to 443. Alternatively you can choose to allow connections to the default port (8080).

By default, firewall rules restrict your cluster master to only initiate TCP connections to your nodes on ports 443 (HTTPS) and 10250 (kubelet). For some Kubernetes features, you might need to add firewall rules to allow access on additional ports. For example, in Kubernetes 1.9 and older, kubectl top accesses heapster, which needs a firewall rule to allow TCP connections on port 8080. To grant such access, you can add firewall rules. For GCP, refer to this link

To install the operator with a custom port, pass the appropriate flag during helm install:

helm install my-release spark-operator/spark-operator \
   --namespace spark-operator  \
   --create-namespace \
   --set "sparkJobNamespaces={spark}" \
   --set webhook.enable=true \
   --set webhook.port=443

Feedback

Was this page helpful?