Skip to content

Latest commit

 

History

History
805 lines (635 loc) · 48.8 KB

user-guide.md

File metadata and controls

805 lines (635 loc) · 48.8 KB

User Guide

For a quick introduction on how to build and install the Kubernetes Operator for Apache Spark, and how to run some example applications, please refer to the Quick Start Guide. For a complete reference of the API definition of the SparkApplication and ScheduledSparkApplication custom resources, please refer to the API Specification.

The Kubernetes Operator for Apache Spark ships with a command-line tool called sparkctl that offers additional features beyond what kubectl is able to do. Documentation on sparkctl can be found in README. If you are running the Spark Operator 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.

Table of Contents

Using a SparkApplication

The operator runs Spark applications specified in Kubernetes objects of the SparkApplication custom resource type. The most common way of using a SparkApplication is store the SparkApplication specification in a YAML file and use the kubectl command or alternatively the sparkctl command to work with the SparkApplication. The operator automatically submits the application as configured in a SparkApplication to run on the Kubernetes cluster and uses the SparkApplication to collect and surface the status of the driver and executors to the user.

Writing a SparkApplication Spec

As with all other Kubernetes API objects, a SparkApplication needs the apiVersion, kind, and metadata fields. For general information about working with manifests, see object management using kubectl.

A SparkApplication also needs a .spec section. This section contains fields for specifying various aspects of an application including its type (Scala, Java, Python, or R), deployment mode (cluster or client), main application resource URI (e.g., the URI of the application jar), main class, arguments, etc. Node selectors are also supported via the optional field .spec.nodeSelector.

It also has fields for specifying the unified container image (to use for both the driver and executors) and the image pull policy, namely, .spec.image and .spec.imagePullPolicy respectively. If a custom init-container (in both the driver and executor pods) image needs to be used, the optional field .spec.initContainerImage can be used to specify it. If set, .spec.initContainerImage overrides .spec.image for the init-container image. Otherwise, the image specified by .spec.image will be used for the init-container. It is invalid if both .spec.image and .spec.initContainerImage are not set.

Below is an example showing part of a SparkApplication specification:

apiVersion: sparkoperator.k8s.io/v1beta2
kind: SparkApplication
metadata:
  name: spark-pi
  namespace: default
spec:
  type: Scala
  mode: cluster
  image: gcr.io/spark/spark:v3.0.0
  mainClass: org.apache.spark.examples.SparkPi
  mainApplicationFile: local:///opt/spark/examples/jars/spark-examples_2.12-3.0.0.jar

Specifying Application Dependencies

Often Spark applications need additional files additionally to the main application resource to run. Such application dependencies can include for example jars and data files the application needs at runtime. When using the spark-submit script to submit a Spark application, such dependencies are specified using the --jars and --files options. To support specification of application dependenies, a SparkApplication uses an optional field .spec.deps that in turn supports specifying jars and files, respectively. More specifically, the optional fields .spec.deps.jars and.spec.deps.files correspond to the --jars and --files options of the spark-submit script, respectively.

Additionally, .spec.deps also has fields for specifying the locations in the driver and executor containers where jars and files should be downloaded to, namely, .spec.deps.jarsDownloadDir and .spec.deps.filesDownloadDir. The optional fields .spec.deps.downloadTimeout and .spec.deps.maxSimultaneousDownloads are used to control the timeout and maximum parallelism of downloading dependencies that are hosted remotely, e.g., on an HTTP server, or in external storage such as HDFS, Google Cloud Storage, or AWS S3.

The following is an example specification with both container-local (i.e., within the container) and remote dependencies:

spec:
  deps:
    jars:
      - local:///opt/spark-jars/gcs-connector.jar
    files:
      - gs://spark-data/data-file-1.txt
      - gs://spark-data/data-file-2.txt

Specifying Spark Configuration

There are two ways to add Spark configuration: setting individual Spark configuration properties using the optional field .spec.sparkConf or mounting a special Kubernetes ConfigMap storing Spark configuration files (e.g. spark-defaults.conf, spark-env.sh, log4j.properties) using the optional field .spec.sparkConfigMap. If .spec.sparkConfigMap is used, additionally to mounting the ConfigMap into the driver and executors, the operator additionally sets the environment variable SPARK_CONF_DIR to point to the mount path of the ConfigMap.

spec:
  sparkConf:
    "spark.ui.port": "4045"
    "spark.eventLog.enabled": "true"
    "spark.eventLog.dir": "hdfs://hdfs-namenode-1:8020/spark/spark-events"

Specifying Hadoop Configuration

There are two ways to add Hadoop configuration: setting individual Hadoop configuration properties using the optional field .spec.hadoopConf or mounting a special Kubernetes ConfigMap storing Hadoop configuration files (e.g. core-site.xml) using the optional field .spec.hadoopConfigMap. The operator automatically adds the prefix spark.hadoop. to the names of individual Hadoop configuration properties in .spec.hadoopConf. If .spec.hadoopConfigMap is used, additionally to mounting the ConfigMap into the driver and executors, the operator additionally sets the environment variable HADOOP_CONF_DIR to point to the mount path of the ConfigMap.

The following is an example showing the use of individual Hadoop configuration properties:

spec:
  hadoopConf:
    "fs.gs.project.id": spark
    "fs.gs.system.bucket": spark
    "google.cloud.auth.service.account.enable": true
    "google.cloud.auth.service.account.json.keyfile": /mnt/secrets/key.json

Writing Driver Specification

The .spec section of a SparkApplication has a .spec.driver field for configuring the driver. It allows users to set the memory and CPU resources to request for the driver pod, and the container image the driver should use. It also has fields for optionally specifying labels, annotations, and environment variables for the driver pod. By default, the driver pod name of an application is automatically generated by the Spark submission client. If instead you want to use a particular name for the driver pod, the optional field .spec.driver.podName can be used. The driver pod by default uses the default service account in the namespace it is running in to talk to the Kubernetes API server. The default service account, however, may or may not have sufficient permissions to create executor pods and the headless service used by the executors to connect to the driver. If it does not and a custom service account that has the right permissions should be used instead, the optional field .spec.driver.serviceAccount can be used to specify the name of the custom service account. When a custom container image is needed for the driver, the field .spec.driver.image can be used to specify it. This overrides the image specified in .spec.image if it is also set. It is invalid if both .spec.image and .spec.driver.image are not set.

For applications that need to mount Kubernetes Secrets or ConfigMaps into the driver pod, fields .spec.driver.secrets and .spec.driver.configMaps can be used. For more details, please refer to Mounting Secrets and Mounting ConfigMaps.

The following is an example driver specification:

spec:
  driver:
    cores: 1
    coreLimit: 200m
    memory: 512m
    labels:
      version: 3.0.0
    serviceAccount: spark

Writing Executor Specification

The .spec section of a SparkApplication has a .spec.executor field for configuring the executors. It allows users to set the memory and CPU resources to request for the executor pods, and the container image the executors should use. It also has fields for optionally specifying labels, annotations, and environment variables for the executor pods. By default, a single executor is requested for an application. If more than one executor are needed, the optional field .spec.executor.instances can be used to specify the number of executors to request. When a custom container image is needed for the executors, the field .spec.executor.image can be used to specify it. This overrides the image specified in .spec.image if it is also set. It is invalid if both .spec.image and .spec.executor.image are not set.

For applications that need to mount Kubernetes Secrets or ConfigMaps into the executor pods, fields .spec.executor.secrets and .spec.executor.configMaps can be used. For more details, please refer to Mounting Secrets and Mounting ConfigMaps.

An example executor specification is shown below:

spec:
  executor:
    cores: 1
    instances: 1
    memory: 512m
    labels:
      version: 3.0.0

Specifying Extra Java Options

A SparkApplication can specify extra Java options for the driver or executors, using the optional field .spec.driver.javaOptions for the driver and .spec.executor.javaOptions for executors. Below is an example:

spec:
  executor:
    javaOptions: "-XX:+UnlockExperimentalVMOptions -XX:+UseCGroupMemoryLimitForHeap"

Values specified using those two fields get converted to Spark configuration properties spark.driver.extraJavaOptions and spark.executor.extraJavaOptions, respectively. Prefer using the above two fields over configuration properties spark.driver.extraJavaOptions and spark.executor.extraJavaOptions as the fields work well with other fields that might modify what gets set for spark.driver.extraJavaOptions or spark.executor.extraJavaOptions.

Specifying Environment Variables

There are two fields for specifying environment variables for the driver and/or executor containers, namely .spec.driver.env (or .spec.executor.env for the executor container) and .spec.driver.envFrom (or .spec.executor.envFrom for the executor container). Specifically, .spec.driver.env (and .spec.executor.env) takes a list of EnvVar, each of which specifies an environment variable or the source of an environment variable, e.g., a name-value pair, a ConfigMap key, a Secret key, etc. Alternatively, .spec.driver.envFrom (and .spec.executor.envFrom) takes a list of EnvFromSource and allows using all key-value pairs in a ConfigMap or Secret as environment variables. The SparkApplication snippet below shows the use of both fields:

spec:
  driver:
    env:
      - name: ENV1
        value: VAL1
      - name: ENV2
        value: VAL2
      - name: ENV3
        valueFrom:
          configMapKeyRef:
            name: some-config-map
            key: env3-key
      - name: AUTH_KEY
        valueFrom:
          secretKeyRef:
            name: some-secret
            key: auth-key
    envFrom:
      - configMapRef:
          name: env-config-map
      - secretRef:
          name: env-secret
  executor:
    env:
      - name: ENV1
        value: VAL1
      - name: ENV2
        value: VAL2
      - name: ENV3
        valueFrom:
          configMapKeyRef:
            name: some-config-map
            key: env3-key
      - name: AUTH_KEY
        valueFrom:
          secretKeyRef:
            name: some-secret
            key: auth-key  
    envFrom:
      - configMapRef:
          name: my-env-config-map
      - secretRef:
          name: my-env-secret

Note: legacy field envVars that can also be used for specifying environment variables is deprecated and will be removed in a future API version.

Requesting GPU Resources

A SparkApplication can specify GPU resources for the driver or executor pod, using the optional field .spec.driver.gpu or .spec.executor.gpu. Below is an example:

spec:
  driver:
    cores: 0.1
    coreLimit: "200m"
    memory: "512m"
    gpu:
      name: "amd.com/gpu"   # GPU resource name
      quantity: 1           # number of GPUs to request
    labels:
      version: 3.0.0
    serviceAccount: spark
  executor:
    cores: 1
    instances: 1
    memory: "512m"
    gpu:
      name: "nvidia.com/gpu"
      quantity: 1

Note that the mutating admission webhook is needed to use this feature. Please refer to the Quick Start Guide on how to enable the mutating admission webhook.

Host Network

A SparkApplication can specify hostNetwork for the driver or executor pod, using the optional field .spec.driver.hostNetwork or .spec.executor.hostNetwork. When hostNetwork is true, the operator sets pods' spec.hostNetwork to true and sets pods' spec.dnsPolicy to ClusterFirstWithHostNet. Below is an example:

spec:
  driver:
    cores: 0.1
    coreLimit: "200m"
    memory: "512m"
    hostNetwork: true
    labels:
      version: 3.0.0
    serviceAccount: spark
  executor:
    cores: 1
    instances: 1
    memory: "512m"

Note that the mutating admission webhook is needed to use this feature. Please refer to the Quick Start Guide on how to enable the mutating admission webhook.

Mounting Secrets

As mentioned above, both the driver specification and executor specification have an optional field secrets for configuring the list of Kubernetes Secrets to be mounted into the driver and executors, respectively. The field is a map with the names of the Secrets as keys and values specifying the mount path and type of each Secret. For instance, the following example shows a driver specification with a Secret named gcp-svc-account of type GCPServiceAccount to be mounted to /mnt/secrets in the driver pod.

spec:
  driver:
    secrets:
      - name: gcp-svc-account
        path: /mnt/secrets
        secretType: GCPServiceAccount

The type of a Secret as specified by the secretType field is a hint to the operator on what extra configuration it needs to take care of for the specific type of Secrets. For example, if a Secret is of type GCPServiceAccount, the operator additionally sets the environment variable GOOGLE_APPLICATION_CREDENTIALS to point to the JSON key file stored in the secret. Please refer to Getting Started with Authentication for more information on how to authenticate with GCP services using a service account JSON key file. Note that the operator assumes that the key of the service account JSON key file in the Secret data map is key.json so it is able to set the environment variable automatically. Similarly, if the type of a Secret is HadoopDelegationToken, the operator additionally sets the environment variable HADOOP_TOKEN_FILE_LOCATION to point to the file storing the Hadoop delegation token. In this case, the operator assumes that the key of the delegation token file in the Secret data map is hadoop.token. The secretType field should have the value Generic if no extra configuration is required.

Mounting ConfigMaps

Both the driver specification and executor specifications have an optional field for configuring the list of Kubernetes ConfigMaps to be mounted into the driver and executors, respectively. The field is a map with keys being the names of the ConfigMaps and values specifying the mount path of each ConfigMap. For instance, the following example shows a driver specification with a ConfigMap named configmap1 to be mounted to /mnt/config-maps in the driver pod.

spec:
  driver:
    configMaps:
      - name: configmap1
        path: /mnt/config-maps

Note that the mutating admission webhook is needed to use this feature. Please refer to the Quick Start Guide on how to enable the mutating admission webhook.

Mounting a ConfigMap storing Spark Configuration Files

A SparkApplication can specify a Kubernetes ConfigMap storing Spark configuration files such as spark-env.sh or spark-defaults.conf using the optional field .spec.sparkConfigMap whose value is the name of the ConfigMap. The ConfigMap is assumed to be in the same namespace as that of the SparkApplication. The operator mounts the ConfigMap onto path /etc/spark/conf in both the driver and executors. Additionally, it also sets the environment variable SPARK_CONF_DIR to point to /etc/spark/conf in the driver and executors.

Note that the mutating admission webhook is needed to use this feature. Please refer to the Quick Start Guide on how to enable the mutating admission webhook.

Mounting a ConfigMap storing Hadoop Configuration Files

A SparkApplication can specify a Kubernetes ConfigMap storing Hadoop configuration files such as core-site.xml using the optional field .spec.hadoopConfigMap whose value is the name of the ConfigMap. The ConfigMap is assumed to be in the same namespace as that of the SparkApplication. The operator mounts the ConfigMap onto path /etc/hadoop/conf in both the driver and executors. Additionally, it also sets the environment variable HADOOP_CONF_DIR to point to /etc/hadoop/conf in the driver and executors.

Note that the mutating admission webhook is needed to use this feature. Please refer to the Quick Start Guide on how to enable the mutating admission webhook.

Mounting Volumes

The operator also supports mounting user-specified Kubernetes volumes into the driver and executors. A SparkApplication has an optional field .spec.volumes for specifying the list of volumes the driver and the executors need collectively. Then both the driver and executor specifications have an optional field volumeMounts that specifies the volume mounts for the volumes needed by the driver and executors, respectively. The following is an example showing a SparkApplication with both driver and executor volume mounts.

spec:
  volumes:
    - name: spark-data
      persistentVolumeClaim:
        claimName: my-pvc
    - name: spark-work
      emptyDir: {}
  driver:
    volumeMounts:
      - name: spark-work
        mountPath: /mnt/spark/work
  executor:
    volumeMounts:
      - name: spark-data
        mountPath: /mnt/spark/data
      - name: spark-work
        mountPath: /mnt/spark/work

Note that the mutating admission webhook is needed to use this feature. Please refer to the Quick Start Guide on how to enable the mutating admission webhook.

Using Secrets As Environment Variables

Note: envSecretKeyRefs is deprecated and will be removed in a future API version.

A SparkApplication can use secrets as environment variables, through the optional field .spec.driver.envSecretKeyRefs for the driver pod and the optional field .spec.executor.envSecretKeyRefs for the executor pods. A envSecretKeyRefs is a map from environment variable names to pairs consisting of a secret name and a secret key. Below is an example:

spec:
  driver:
    envSecretKeyRefs:
      SECRET_USERNAME:
        name: mysecret
        key: username
      SECRET_PASSWORD:
        name: mysecret
        key: password

Using Image Pull Secrets

Note that this feature requires an image based on the latest Spark master branch.

For images that need image-pull secrets to be pulled, a SparkApplication has an optional field .spec.imagePullSecrets for specifying a list of image-pull secrets. Below is an example:

spec:
  imagePullSecrets:
    - secret1
    - secret2

Using Pod Affinity

A SparkApplication can specify an Affinity for the driver or executor pod, using the optional field .spec.driver.affinity or .spec.executor.affinity. Below is an example:

spec:
  driver:
    affinity:
      podAffinity:
        requiredDuringSchedulingIgnoredDuringExecution:
          ...   
  executor:
    affinity:
      podAntiAffinity:
        requiredDuringSchedulingIgnoredDuringExecution:
          ...

Note that the mutating admission webhook is needed to use this feature. Please refer to the Quick Start Guide on how to enable the mutating admission webhook.

Using Tolerations

A SparkApplication can specify an Tolerations for the driver or executor pod, using the optional field .spec.driver.tolerations or .spec.executor.tolerations. Below is an example:

spec:
  driver:
    tolerations:
    - key: Key
      operator: Exists
      effect: NoSchedule

  executor:
    tolerations:
    - key: Key
      operator: Equal
      value: Value
      effect: NoSchedule    

Note that the mutating admission webhook is needed to use this feature. Please refer to the Quick Start Guide on how to enable the mutating admission webhook.

Using Pod Security Context

A SparkApplication can specify a PodSecurityContext for the driver or executor pod, using the optional field .spec.driver.securityContext or .spec.executor.securityContext. Below is an example:

spec:
  driver:
    securityContext:
      runAsUser: 1000    
  executor:
    securityContext:
      runAsUser: 1000

Note that the mutating admission webhook is needed to use this feature. Please refer to the Quick Start Guide on how to enable the mutating admission webhook.

Using Sidecar Containers

A SparkApplication can specify one or more optional sidecar containers for the driver or executor pod, using the optional field .spec.driver.sidecars or .spec.executor.sidecars. The specification of each sidecar container follows the Container API definition. Below is an example:

spec:
  driver:
    sidecars:
    - name: "sidecar1"
      image: "sidecar1:latest"
      ...  
  executor:
    sidecars:
    - name: "sidecar1"
      image: "sidecar1:latest"
      ...

Using Init-Containers

A SparkApplication can optionally specify one or more init-containers for the driver or executor pod, using the optional field .spec.driver.initContainers or .spec.executor.initContainers, respectively. The specification of each init-container follows the Container API definition. Below is an example:

spec:
  driver:
    initContainers:
    - name: "init-container1"
      image: "init-container1:latest"
      ...  
  executor:
    initContainers:
    - name: "init-container1"
      image: "init-container1:latest"
      ...

Note that the mutating admission webhook is needed to use this feature. Please refer to the Quick Start Guide on how to enable the mutating admission webhook.

Using DNS Settings

A SparkApplication can define DNS settings for the driver and/or executor pod, by adding the standart DNS kubernetes settings. Fields to add such configuration are .spec.driver.dnsConfig and .spec.executor.dnsConfig. Example:

spec:
  driver:
    dnsConfig:
      nameservers:
        - 1.2.3.4
      searches:
        - ns1.svc.cluster.local
        - my.dns.search.suffix
      options:
        - name: ndots
          value: "2"
        - name: edns0

Note that the mutating admission webhook is needed to use this feature. Please refer to the Quick Start Guide on how to enable the mutating admission webhook.

Using Volume For Scratch Space

By default, Spark uses temporary scratch space to spill data to disk during shuffles and other operations. The scratch directory defaults to /tmp of the container. If that storage isn't enough or you want to use a specific path, you can use one or more volumes. The volume names should start with spark-local-dir-.

spec:
  volumes:
    - name: "spark-local-dir-1"
      hostPath:
        path: "/tmp/spark-local-dir"
  executor:
    volumeMounts:
      - name: "spark-local-dir-1"
        mountPath: "/tmp/spark-local-dir"
    ...

Then you will get SPARK_LOCAL_DIRS set to /tmp/spark-local-dir in the pod like below.

Environment:
  SPARK_USER:                 root
  SPARK_DRIVER_BIND_ADDRESS:  (v1:status.podIP)
  SPARK_LOCAL_DIRS:           /tmp/spark-local-dir
  SPARK_CONF_DIR:             /opt/spark/conf

Note: Multiple volumes can be used together

spec:
  volumes:
    - name: "spark-local-dir-1"
      hostPath:
        path: "/mnt/dir1"
    - name: "spark-local-dir-2"
      hostPath:
        path: "/mnt/dir2"
  executor:
    volumeMounts:
      - name: "spark-local-dir-1"
        mountPath: "/tmp/dir1"
      - name: "spark-local-dir-2"
        mountPath: "/tmp/dir2"
    ...

Note: Besides hostPath, persistentVolumeClaim can be used as well.

spec:
  volumes:
    - name: "spark-local-dir-1"
      persistentVolumeClaim:
        claimName: network-file-storage
  executor:
    volumeMounts:
      - name: "spark-local-dir-1"
        mountPath: "/tmp/dir1"

Using Termination Grace Period

A Spark Application can optionally specify a termination grace Period seconds to the driver and executor pods. More info

spec:
  driver:
    terminationGracePeriodSeconds: 60

Using Container LifeCycle Hooks

A Spark Application can optionally specify a Container Lifecycle Hooks for a driver. It is useful in cases where you need a PreStop or PostStart hooks to driver.

spec:
  driver:
    lifecycle:
      preStop:
        exec:
          command:
          - /bin/bash
          - -c
          - touch /var/run/killspark && sleep 65

In cases like Spark Streaming or Spark Structured Streaming applications, you can test if a file exists to start a graceful shutdown and stop all streaming queries manually.

Python Support

Python support can be enabled by setting .spec.mainApplicationFile with path to your python application. Optionaly, the .spec.pythonVersion field can be used to set the major Python version of the docker image used to run the driver and executor containers. Below is an example showing part of a SparkApplication specification:

spec:
  type: Python
  pythonVersion: 2
  mainApplicationFile: local:///opt/spark/examples/src/main/python/pyfiles.py

Some PySpark applications need additional Python packages to run. Such dependencies are specified using the optional field .spec.deps.pyFiles, which translates to the --py-files option of the spark-submit command.

spec:
  deps:
    pyFiles:
       - local:///opt/spark/examples/src/main/python/py_container_checks.py
       - gs://spark-data/python-dep.zip

In order to use the dependencies that are hosted remotely, the following PySpark code can be used in Spark 2.4.

python_dep_file_path = SparkFiles.get("python-dep.zip")
spark.sparkContext.addPyFile(dep_file_path)

Note that Python binding for PySpark is available in Apache Spark 2.4.

Monitoring

The operator supports using the Spark metric system to expose metrics to a variety of sinks. Particularly, it is able to automatically configure the metric system to expose metrics to Prometheus. Specifically, the field .spec.monitoring specifies how application monitoring is handled and particularly how metrics are to be reported. The metric system is configured through the configuration file metrics.properties, which gets its content from the field .spec.monitoring.metricsProperties. The content of metrics.properties will be used by default if .spec.monitoring.metricsProperties is not specified. .spec.monitoring.metricsPropertiesFile overwrite the value spark.metrics.conf in spark.properties, and will not use content from .spec.monitoring.metricsProperties. You can choose to enable or disable reporting driver and executor metrics using the fields .spec.monitoring.exposeDriverMetrics and .spec.monitoring.exposeExecutorMetrics, respectively.

Further, the field .spec.monitoring.prometheus specifies how metrics are exposed to Prometheus using the Prometheus JMX exporter. When .spec.monitoring.prometheus is specified, the operator automatically configures the JMX exporter to run as a Java agent. The only required field of .spec.monitoring.prometheus is jmxExporterJar, which specified the path to the Prometheus JMX exporter Java agent jar in the container. If you use the image gcr.io/spark-operator/spark:v3.0.0-gcs-prometheus, the jar is located at /prometheus/jmx_prometheus_javaagent-0.11.0.jar. The field .spec.monitoring.prometheus.port specifies the port the JMX exporter Java agent binds to and defaults to 8090 if not specified. The field .spec.monitoring.prometheus.configuration specifies the content of the configuration to be used with the JMX exporter. The content of prometheus.yaml will be used by default if .spec.monitoring.prometheus.configuration is not specified.

Below is an example that shows how to configure the metric system to expose metrics to Prometheus using the Prometheus JMX exporter. Note that the JMX exporter Java agent jar is listed as a dependency and will be downloaded to where .spec.dep.jarsDownloadDir points to in Spark 2.3.x, which is /var/spark-data/spark-jars by default. Things are different in Spark 2.4 as dependencies will be downloaded to the local working directory instead in Spark 2.4. A complete example can be found in examples/spark-pi-prometheus.yaml.

spec:
  deps:
    jars:
    - http://central.maven.org/maven2/io/prometheus/jmx/jmx_prometheus_javaagent/0.11.0/jmx_prometheus_javaagent-0.11.0.jar
  monitoring:
    exposeDriverMetrics: true
    prometheus:
      jmxExporterJar: "/var/spark-data/spark-jars/jmx_prometheus_javaagent-0.11.0.jar"    

The operator automatically adds the annotations such as prometheus.io/scrape=true on the driver and/or executor pods (depending on the values of .spec.monitoring.exposeDriverMetrics and .spec.monitoring.exposeExecutorMetrics) so the metrics exposed on the pods can be scraped by the Prometheus server in the same cluster.

Dynamic Allocation

The operator supports a limited form of Spark Dynamic Resource Allocation through the shuffle tracking enhancement introduced in Spark 3.0.0 without needing an external shuffle service (not available in the Kubernetes mode). See this issue for detais on the enhancement. To enable this limited form of dynamic allocation, follow the example below:

spec:
  dynamicAllocation:
    enabled: true
    initialExecutors: 2
    minExecutors: 2
    maxExecutors: 10

Note that if dynamic allocation is enabled, the number of executors to request initially is set to the bigger of .spec.dynamicAllocation.initialExecutors and .spec.executor.instances if both are set.

Working with SparkApplications

Creating a New SparkApplication

A SparkApplication can be created from a YAML file storing the SparkApplication specification using either the kubectl apply -f <YAML file path> command or the sparkctl create <YAML file path> command. Please refer to the sparkctl README for usage of the sparkctl create command. Once a SparkApplication is successfully created, the operator will receive it and submits the application as configured in the specification to run on the Kubernetes cluster.

Deleting a SparkApplication

A SparkApplication can be deleted using either the kubectl delete <name> command or the sparkctl delete <name> command. Please refer to the sparkctl README for usage of the sparkctl delete command. Deleting a SparkApplication deletes the Spark application associated with it. If the application is running when the deletion happens, the application is killed and all Kubernetes resources associated with the application are deleted or garbage collected.

Updating a SparkApplication

A SparkApplication can be updated using the kubectl apply -f <updated YAML file> command. When a SparkApplication is successfully updated, the operator will receive both the updated and old SparkApplication objects. If the specification of the SparkApplication has changed, the operator submits the application to run, using the updated specification. If the application is currently running, the operator kills the running application before submitting a new run with the updated specification. There is planned work to enhance the way SparkApplication updates are handled. For example, if the change was to increase the number of executor instances, instead of killing the currently running application and starting a new run, it is a much better user experience to incrementally launch the additional executor pods.

Checking a SparkApplication

A SparkApplication can be checked using the kubectl describe sparkapplications <name> command. The output of the command shows the specification and status of the SparkApplication as well as events associated with it. The events communicate the overall process and errors of the SparkApplication.

Configuring Automatic Application Restart and Failure Handling

The operator supports automatic application restart with a configurable RestartPolicy using the optional field .spec.restartPolicy. The following is an example of a sample RestartPolicy:

 restartPolicy:
    type: OnFailure
    onFailureRetries: 3
    onFailureRetryInterval: 10
    onSubmissionFailureRetries: 5
    onSubmissionFailureRetryInterval: 20

The valid types of restartPolicy include Never, OnFailure, and Always. Upon termination of an application, the operator determines if the application is subject to restart based on its termination state and the RestartPolicy in the specification. If the application is subject to restart, the operator restarts it by submitting a new run of it. For OnFailure, the Operator further supports setting limits on number of retries via the onFailureRetries and onSubmissionFailureRetries fields. Additionally, if the submission retries has not been reached, the operator retries submitting the application using a linear backoff with the interval specified by onFailureRetryInterval and onSubmissionFailureRetryInterval which are required for both OnFailure and Always RestartPolicy. The old resources like driver pod, ui service/ingress etc. are deleted if it still exists before submitting the new run, and a new driver pod is created by the submission client so effectively the driver gets restarted.

Setting TTL for a SparkApplication

The v1beta2 version of the SparkApplication API starts having TTL support for SparkApplications through a new optional field named .spec.timeToLiveSeconds, which if set, defines the Time-To-Live (TTL) duration in seconds for a SparkAplication after its termination. The SparkApplication object will be garbage collected if the current time is more than the .spec.timeToLiveSeconds since its termination. The example below illustrates how to use the field:

spec:
  timeToLiveSeconds: 3600

Note that this feature requires that informer cache resync to be enabled, which is true by default with a resync internal of 30 seconds. You can change the resync interval by setting the flag -resync-interval=<interval>.

Running Spark Applications on a Schedule using a ScheduledSparkApplication

The operator supports running a Spark application on a standard cron schedule using objects of the ScheduledSparkApplication custom resource type. A ScheduledSparkApplication object specifies a cron schedule on which the application should run and a SparkApplication template from which a SparkApplication object for each run of the application is created. The following is an example ScheduledSparkApplication:

apiVersion: "sparkoperator.k8s.io/v1beta2"
kind: ScheduledSparkApplication
metadata:
  name: spark-pi-scheduled
  namespace: default
spec:
  schedule: "@every 5m"
  concurrencyPolicy: Allow
  successfulRunHistoryLimit: 1
  failedRunHistoryLimit: 3
  template:
    type: Scala
    mode: cluster
    image: gcr.io/spark/spark:v3.0.0
    mainClass: org.apache.spark.examples.SparkPi
    mainApplicationFile: local:///opt/spark/examples/jars/spark-examples_2.12-2.3.0.jar
    driver:
      cores: 1
      memory: 512m
    executor:
      cores: 1
      instances: 1
      memory: 512m
    restartPolicy:
      type: Never

The concurrency of runs of an application is controlled by .spec.concurrencyPolicy, whose valid values are Allow, Forbid, and Replace, with Allow being the default. The meanings of each value is described below:

  • Allow: more than one run of an application are allowed if for example the next run of the application is due even though the previous run has not completed yet.
  • Forbid: no more than one run of an application is allowed. The next run of the application can only start if the previous run has completed.
  • Replace: no more than one run of an application is allowed. When the next run of the application is due, the previous run is killed and the next run starts as a replacement.

A scheduled ScheduledSparkApplication can be temporarily suspended (no future scheduled runs of the application will be triggered) by setting .spec.suspend to true. The schedule can be resumed by removing .spec.suspend or setting it to false. A ScheduledSparkApplication can have names of SparkApplication objects for the past runs of the application tracked in the Status section as discussed below. The numbers of past successful runs and past failed runs to keep track of are controlled by field .spec.successfulRunHistoryLimit and field .spec.failedRunHistoryLimit, respectively. The example above allows 1 past successful run and 3 past failed runs to be tracked.

The Status section of a ScheduledSparkApplication object shows the time of the last run and the proposed time of the next run of the application, through .status.lastRun and .status.nextRun, respectively. The names of the SparkApplication object for the most recent run (which may or may not be running) of the application are stored in .status.lastRunName. The names of SparkApplication objects of the past successful runs of the application are stored in .status.pastSuccessfulRunNames. Similarly, the names of SparkApplication objects of the past failed runs of the application are stored in .status.pastFailedRunNames.

Note that certain restart policies (specified in .spec.template.restartPolicy) may not work well with the specified schedule and concurrency policy of a ScheduledSparkApplication. For example, a restart policy of Always should never be used with a ScheduledSparkApplication. In most cases, a restart policy of OnFailure may not be a good choice as the next run usually picks up where the previous run left anyway. For these reasons, it's often the right choice to use a restart policy of Never as the example above shows.

Enabling Leader Election for High Availability

The operator supports a high-availability (HA) mode, in which there can be more than one replicas of the operator, with only one of the replicas (the leader replica) actively operating. If the leader replica fails, the leader election process is engaged again to determine a new leader from the replicas available. The HA mode can be enabled through an optional leader election process. Leader election is disabled by default but can be enabled via a command-line flag. The following table summarizes the command-line flags relevant to leader election:

Flag Default Value Description
leader-election false Whether to enable leader election (or the HA mode) or not.
leader-election-lock-namespace spark-operator Kubernetes namespace of the lock resource used for leader election.
leader-election-lock-name spark-operator-lock Name of the lock resource used for leader election.
leader-election-lease-duration 15 seconds Leader election lease duration.
leader-election-renew-deadline 14 seconds Leader election renew deadline.
leader-election-retry-period 4 seconds Leader election retry period.

Enabling Resource Quota Enforcement

The Spark Operator provides limited support for resource quota enforcement using a validating webhook. It will count the resources of non-terminal-phase SparkApplications and Pods, and determine whether a requested SparkApplication will fit given the remaining resources. ResourceQuota scope selectors are not supported, any ResourceQuota object that does not match the entire namespace will be ignored. Like the native Pod quota enforcement, current usage is updated asynchronously, so some overscheduling is possible.

If you are running Spark applications in namespaces that are subject to resource quota constraints, consider enabling this feature to avoid driver resource starvation. Quota enforcement can be enabled with the command line arguments -enable-resource-quota-enforcement=true. It is recommended to also set -webhook-fail-on-error=true.

Running Multiple Instances Of The Operator Within The Same K8s Cluster

If you need to run multiple instances of the operator within the same k8s cluster. Therefore, you need to make sure that the running instances should not compete for the same custom resources or pods. You can achieve this:

Either:

  • By specifying a different namespace flag for each instance of the operator.

Or if you want your operator to watch specific resources that may exist in different namespaces:

  • You need to add custom labels on resources by defining for each instance of the operator a different set of labels in -label-selector-filter (e.g. env=dev,app-type=spark).
  • Run different webhook instances by specifying different -webhook-config-name flag for each deployment of the operator.
  • Specify different webhook-svc-name and/or webhook-svc-namespace for each instance of the operator.
  • Edit the job that generates the certificates webhook-init by specifying the namespace and the service name of each instance of the operator, e.g. command: ["/usr/bin/gencerts.sh", "-n", "ns-op1", "-s", "spark-op1-webhook", "-p"]. Where spark-op1-webhook should match what you have specified in webhook-svc-name. For instance, if you use the following helm chart to deploy the operator you may specify for each instance of the operator a different --namespace and --name-template arguments to make sure you generate a different certificate for each instance, e.g:
helm install spark-op1 incubator/sparkoperator --namespace ns-op1
helm install spark-op2 incubator/sparkoperator --namespace ns-op2

Will run 2 webhook-init jobs. Each job executes respectively the command:

command: ["/usr/bin/gencerts.sh", "-n", "ns-op1", "-s", "spark-op1-webhook", "-p"`]
command: ["/usr/bin/gencerts.sh", "-n", "ns-op2", "-s", "spark-op2-webhook", "-p"`]
  • Although resources are already filtered with respect to the specified labels on resources. You may also specify different labels in -webhook-namespace-selector and attach these labels to the namepsaces on which you want the webhook to listen to.

Customizing the Operator

To customize the operator, you can follow the steps below:

  1. Compile Spark distribution with Kubernetes support as per Spark documentation.
  2. Create docker images to be used for Spark with docker-image tool.
  3. Create a new operator image based on the above image. You need to modify the FROM tag in the Dockerfile with your Spark image.
  4. Build and push your operator image built above.
  5. Deploy the new image by modifying the /manifest/spark-operator.yaml file and specfiying your operator image.