Настройка YuniKorn для Spark в Kubernetes с помощью CLI

apiVersion: adc.arenadata.io/v1alpha1
kind: YuniKorn
metadata:
  name: yunikorn
  namespace: yunikorn (1)
spec:
  image: hub.arenadata.io/adc-enterprise/yunikorn-k8shim:<tag> (2)
  web:
    image: hub.arenadata.io/adc-enterprise/yunikorn-web:<tag> (3)
    #resources:
    #  limits:
    #    cpu: 200m
    #    memory: 500Mi
    #  requests:
    #    cpu: 100m
    #    memory: 100Mi

    goMemoryLimitPercentage: 60 (4)
    goGC: 100 (5)

    ## Environment variables passed to the component container.
    #envs:
    #  - name: YUNIKORN_LDAP_BIND_DN
    #    valueFrom:
    #      secretKeyRef:
    #        key: adminDN
    #        name: yunikorn-ldap-credentials
    #  - name: YUNIKORN_LDAP_BIND_PASSWORD
    #    valueFrom:
    #      secretKeyRef:
    #        key: adminPW
    #        name: yunikorn-ldap-credentials
    #  - name: YUNIKORN_LDAP_BASE_DN
    #    value: cn=accounts,dc=example,dc=com
    #  - name: YUNIKORN_LDAP_USER_BASE_DN
    #    value: cn=users,cn=accounts,dc=example,dc=com
    #  - name: YUNIKORN_LDAP_ADMIN_GROUPS
    #    value: admins
    #  - name: YUNIKORN_LDAP_ALLOWED_GROUPS
    #    value: ipausers
    #  - name: YUNIKORN_AUTH_SHARED_SECRET
    #    valueFrom:
    #      secretKeyRef:
    #        key: shared-secret
    #        name: yunikorn-web-auth

  ## LDAP authentication configuration.
  ## Uncomment and fill url and userBindPattern.
  ## For ldaps:// URLs the ssl: or ca: section must also be configured (depends on product)
  #ldap:
  #  # LDAP service url.
  #  url: ldap://ldap.example.com:389
  #
  #  # LDAP user Bind pattern.
  #  userBindPattern: uid=#UID,cn=users,dc=example,dc=com

  ## Kerberos configuration for authentication.
  #kerberos:
  #  realm: EXAMPLE.COM
  #
  #  # Service name in the Kerberos principal. Defaults to the product name.
  #  service: HTTP
  #
  #  # Hostname in the Kerberos principal.
  #  # Required for a fixed service principal; leave it empty only to derive one principal per pod from the cluster domain.
  #  hostname: yunikorn.example.com
  #  keytab:
  #    # true - kerberos-operator creates the keytab Secret.
  #    # false (default) - reference an existing keytab Secret with name keytab.secretName.
  #    create: false
  #
  #    # Name of the keytab Secret.
  #    # Optional when create: true - names the generated Secret (default: <name>-keytab).
  #    # Required when create: false - must reference an existing Secret.
  #    secretName: yunikorn-keytab
  #
  #    # Label selector for the Pod that generates the keytab.
  #    # Required when create: true; ignored when create: false.
  #    labelSelector:
  #      env: prod
  #    rotation:
  #      interval: 24h
  #      checkInterval: 1h

  ## HTTPS on the user-facing YuniKorn Web UI endpoint on port 9889.
  ## Together with monitoring.exportMetrics it also serves the scheduler /metrics over HTTPS
  ## using the same certificate and key, keeping the fixed "metrics" port name.
  ## The Web UI to k8shim unix socket is not changed.
  ## Set externalSecretName to reference an existing Secret,
  ## or set files and optional secretName to have ADC create it.
  #webTLS:
  #  ## Optional name of the Secret ADC creates from local files.
  #  #secretName: custom-yunikorn-web-tls
  #  externalSecretName: existing-yunikorn-web-tls
  #
  #  # Key in the Secret containing the web TLS certificate.
  #  certificateKey: tls.crt
  #
  #  # Key in the Secret containing the web TLS private key.
  #  privateKey: tls.key
  #
  #  ## Local files 'adc apply' puts into the Secret named by webTLS.secretName.
  #  ## Relative paths are resolved against the config file.
  #  #files:
  #  #  certificatePath: /path/to/tls.crt
  #  #  privateKeyPath: /path/to/tls.key

  # Whether the CLI creates the product namespace.
  # The namespace name is set in metadata.namespace.
  namespace:
    create: true

  # ServiceAccount used by the YuniKorn scheduler. The CLI creates it, plus a ClusterRole, Role
  # and their bindings, by default (create: true). Set create: false to skip all of them and only
  # reference a ServiceAccount (and RBAC) managed elsewhere. name is optional; when empty, it
  # defaults to the release name.
  serviceAccount: (6)
    create: true
    name: yunikorn

  ## Image pull secret for a private registry.
  ## Set 'externalSecretName' to reference an existing Secret,
  ## or set 'credentials' and optionally 'secretName' to let the CLI create one.
  #imagePullSecret:
  #  # Use a Secret managed outside ADC.
  #  externalSecretName: existing-registry-secret
  #
  #  ## Or let ADC create the Secret.
  #  #secretName: custom-registry-secret
  #
  #  #credentials:
  #  #  registry: registry.example.com
  #  #  username: user
  #  #  password: pass

  #resources:
  #  limits:
  #    cpu: "4"
  #    memory: 2Gi
  #  requests:
  #    cpu: 200m
  #    memory: 1Gi

  goMemoryLimitPercentage: 80 (7)
  goGC: 100 (8)

  ## Bootstrap config rendered into the yunikorn-defaults ConfigMap.
  ## Put the queue hierarchy under the "queues.yaml" key as a YAML string;
  ## k8shim reads it and creates the queues. Leave empty to use the built-in root queue.
  yunikornDefaults: (9)
    queues.yaml: | (10)
      partitions:
        - name: default
          queues:
            - name: root
              submitacl: '*'
              queues:
                - name: engineering
                - name: analytics
                - name: ad-hoc
    service.exposeMetricsOnly: "true"
    service.placeholderImage: hub.arenadata.io/adc-enterprise/pause:<tag> (11)

  ## Controls whether Secret/ConfigMap changes restart pods.
  ## Set enabled: false to update referenced Secrets without restarting
  ## the workload; pods keep running the previous configuration until
  ## the policy is re-enabled. Defaults to enabled.
  #configurationRollout:
  #  enabled: false

  ## Monitoring configuration.
  #monitoring:
  #  # Enables Prometheus metrics export from this product's pods.
  #  exportMetrics: true
  #
  #  ## Renders a namespace-scoped vmagent that sends metrics to an external ADM.
  #  #vmagent:
  #  #  remoteWrite:
  #  #    url: http://vminsert.example.com:8480/insert/0/prometheus/api/v1/write
  #  #  scrapeInterval: 30s
  #  #
  #  #  ## HTTPS settings used by vmagent when scraping product metrics.
  #  #  #tls:
  #  #  #  ## CA certificate source used to verify the metrics endpoint.
  #  #  #  #ca:
  #  #  #  #  # Use a Secret managed outside ADC.
  #  #  #  #  externalSecretName: existing-product-metrics-ca
  #  #  #  #
  #  #  #  #  ## Or let ADC create the Secret.
  #  #  #  #  #secretName: product-metrics-ca
  #  #  #  #
  #  #  #  #  # Key containing the CA certificate in the referenced Secret.
  #  #  #  #  certificateKey: ca.crt
  #  #  #  #
  #  #  #  #  ## Local CA certificate read by ADC to create the configured Secret.
  #  #  #  #  #files:
  #  #  #  #  #  certificatePath: /path/to/ca.crt
  #  #  #
  #  #  #  ## Server name used to verify the metrics endpoint certificate hostname.
  #  #  #  #serverName: metrics.example.com
  #  #  #
  #  #  #  # Skip verification of the metrics endpoint certificate. Do not use together with ca.
  #  #  #  insecureSkipVerify: true

---
apiVersion: v1
kind: Namespace
metadata:
  name: yunikorn
spec: {}
status: {}
---
apiVersion: v1
kind: ServiceAccount
metadata:
  name: yunikorn
  namespace: yunikorn
---
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
metadata:
  name: yunikorn-scheduler
rules:
- apiGroups:
  - ""
  resources:
  - namespaces
  verbs:
  - get
  - watch
  - list
- apiGroups:
  - scheduling.k8s.io
  resources:
  - priorityclasses
  verbs:
  - get
  - watch
  - list
- apiGroups:
  - storage.k8s.io
  resources:
  - volumeattachments
  verbs:
  - get
  - watch
  - list
- apiGroups:
  - ""
  resources:
  - pods
  verbs:
  - get
  - watch
  - list
  - create
  - patch
  - update
  - delete
---
apiVersion: rbac.authorization.k8s.io/v1
kind: Role
metadata:
  name: yunikorn-scheduler
  namespace: yunikorn
rules:
- apiGroups:
  - ""
  resources:
  - secrets
  verbs:
  - get
  - watch
  - list
  - update
- apiGroups:
  - ""
  resources:
  - configmaps
  verbs:
  - get
  - watch
  - list
---
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: yunikorn-rbac
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: yunikorn-scheduler
subjects:
- kind: ServiceAccount
  name: yunikorn
  namespace: yunikorn
---
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: yunikorn-rbac-kube-scheduler
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: system:kube-scheduler
subjects:
- kind: ServiceAccount
  name: yunikorn
  namespace: yunikorn
---
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: yunikorn-rbac-volume-scheduler
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: system:volume-scheduler
subjects:
- kind: ServiceAccount
  name: yunikorn
  namespace: yunikorn
---
apiVersion: rbac.authorization.k8s.io/v1
kind: RoleBinding
metadata:
  name: yunikorn-rbac
  namespace: yunikorn
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: Role
  name: yunikorn-scheduler
subjects:
- kind: ServiceAccount
  name: yunikorn
  namespace: yunikorn
---
apiVersion: apps/v1
kind: Deployment
metadata:
  labels:
    app.kubernetes.io/component: yunikorn-scheduler
    app.kubernetes.io/instance: yunikorn
    app.kubernetes.io/managed-by: adc-cli
    app.kubernetes.io/name: yunikorn
  name: yunikorn-scheduler
  namespace: yunikorn
spec:
  replicas: 1
  selector:
    matchLabels:
      app.kubernetes.io/component: yunikorn-scheduler
      app.kubernetes.io/instance: yunikorn
      app.kubernetes.io/managed-by: adc-cli
      app.kubernetes.io/name: yunikorn
  strategy: {}
  template:
    metadata:
      labels:
        app.kubernetes.io/component: yunikorn-scheduler
        app.kubernetes.io/instance: yunikorn
        app.kubernetes.io/managed-by: adc-cli
        app.kubernetes.io/name: yunikorn
    spec:
      containers:
      - env:
        - name: NAMESPACE
          valueFrom:
            fieldRef:
              fieldPath: metadata.namespace
        - name: YUNIKORN_K8SHIM_SOCKET_PATH
          value: /var/run/yunikorn/k8shim.sock
        - name: GOMEMLIMIT
          value: "1717986918"
        - name: GOGC
          value: "100"
        image: hub.arenadata.io/adc-enterprise/yunikorn-k8shim:<tag>
        imagePullPolicy: Always
        name: app
        ports:
        - containerPort: 9080
          name: metrics
          protocol: TCP
        resources:
          limits:
            cpu: "4"
            memory: 2Gi
          requests:
            cpu: 200m
            memory: 1Gi
        securityContext:
          allowPrivilegeEscalation: false
          capabilities:
            drop:
            - ALL
          privileged: false
          readOnlyRootFilesystem: true
          runAsGroup: 4444
          runAsNonRoot: true
          runAsUser: 4444
        volumeMounts:
        - mountPath: /var/run/yunikorn
          name: k8shim-socket
      - env:
        - name: YUNIKORN_K8SHIM_URL
          value: unix:///var/run/yunikorn/k8shim.sock
        - name: GOMEMLIMIT
          value: "314572800"
        - name: GOGC
          value: "100"
        image: hub.arenadata.io/adc-enterprise/yunikorn-web:<tag>
        imagePullPolicy: Always
        name: yunikorn-scheduler-web
        ports:
        - containerPort: 9889
          name: web
          protocol: TCP
        resources:
          limits:
            cpu: 200m
            memory: 500Mi
          requests:
            cpu: 100m
            memory: 100Mi
        securityContext:
          allowPrivilegeEscalation: false
          capabilities:
            drop:
            - ALL
          privileged: false
          readOnlyRootFilesystem: true
          runAsGroup: 4444
          runAsNonRoot: true
          runAsUser: 4444
        volumeMounts:
        - mountPath: /var/run/yunikorn
          name: k8shim-socket
      securityContext:
        fsGroup: 4444
        fsGroupChangePolicy: OnRootMismatch
        runAsGroup: 4444
        runAsNonRoot: true
        runAsUser: 4444
        supplementalGroups:
        - 4444
      serviceAccountName: yunikorn
      volumes:
      - emptyDir: {}
        name: k8shim-socket
status: {}
---
apiVersion: v1
kind: Service
metadata:
  labels:
    app.kubernetes.io/component: yunikorn-scheduler
    app.kubernetes.io/instance: yunikorn
    app.kubernetes.io/managed-by: adc-cli
    app.kubernetes.io/name: yunikorn
  name: yunikorn-service
  namespace: yunikorn
spec:
  ports:
  - name: metrics
    port: 9080
    protocol: TCP
    targetPort: metrics
  - name: web
    port: 9889
    protocol: TCP
    targetPort: web
  selector:
    app.kubernetes.io/component: yunikorn-scheduler
    app.kubernetes.io/instance: yunikorn
    app.kubernetes.io/managed-by: adc-cli
    app.kubernetes.io/name: yunikorn
  type: ClusterIP
status:
  loadBalancer: {}
---
apiVersion: v1
data:
  queues.yaml: |
    partitions:
      - name: default
        queues:
          - name: root
            submitacl: '*'
            queues:
              - name: engineering
              - name: analytics
              - name: ad-hoc
  service.exposeMetricsOnly: "true"
  service.placeholderImage: hub.arenadata.io/adc-enterprise/pause:<tag>
kind: ConfigMap
metadata:
  name: yunikorn-defaults
  namespace: yunikorn

apiVersion: adc.arenadata.io/v1alpha1
kind: SparkApplication
metadata:
  name: spark-application
  namespace: spark-applications
spec:
  image: hub.adsw.io/adc-develop-enterprise/spark3:<tag>

  ## Image pull secret for a private registry.
  ## Set 'externalSecretName' to reference an existing Secret,
  ## or set 'credentials' and optionally 'secretName' to let the CLI create one.
  #imagePullSecret:
  #
  #  # Use a Secret managed outside ADC.
  #  externalSecretName: existing-registry-secret
  #
  #  ## Or let ADC create the Secret.
  #  #secretName: custom-registry-secret
  #  #credentials:
  #  #  registry: registry.example.com
  #  #  username: user
  #  #  password: pass
  # hadoop:
  #   core:
  #   hive:


  ## Kerberos configuration for authentication.
  #kerberos:
  #  principal: user@EXAMPLE.COM
  #
  #  # CLI reads the local files and creates the kerberos-ccache Secret on 'adc apply'.
  #  # Alternative - keytab mode: replace this block with:
  #  #   keytab:
  #  #     secretName: <name-of-existing-keytab-secret>
  #  ticketCache:
  #    #secretName: custom-ticket-cache
  #    externalSecretName: existing-ticket-cache
  #    #ticketPath: /tmp/krb5cc_1000
  #    #krb5ConfPath: /etc/krb5.conf

  ## Ranger plugin configuration.
  ## Uncomment and fill the lines below. adc apply derives the rest.
  #ranger:
  #  # fill ranger.plugin.spark.policy.rest.url below with Ranger endpoint, e.g. https://adps-adc.ru-central1.internal:6182
  #  # fill ranger.plugin.spark.service.name below with Ranger service name you want to use for product, e.g. adc_spark_id_1
  #  security:
  #    ranger.plugin.spark.policy.rest.url: ""
  #    ranger.plugin.spark.service.name: ""
  #
  #  # fill xasecure.audit.destination.solr.zookeepers below with Zookeepers endpoints to resolve solr service, e.g. adps-adc.ru-central1.internal:2181/Arenadata.Hadoop-2.solr.server
  #  audit:
  #    xasecure.audit.destination.solr.zookeepers: ""
  #
  #  # Local Ranger files 'adc apply' writes into the configs Secret.
  #  # Relative paths are resolved against the config file.
  #  files:
  #    jceksStorePath: /path/to/ranger.jceks

  ## Java KeyStore/TrustStore certificate configuration.
  ## Set externalSecretName to reference an existing Secret,
  ## or set files and optional secretName to have ADC create it.
  #ssl:
  #  ## Name of the Secret containing Java keystores.
  #  #secretName: custom-ssl-secret
  #  externalSecretName: existing-ssl-secret
  #
  #  # Key in the Secret containing the truststore file.
  #  trustStoreKey: truststore.jks
  #
  #  ## Password for the truststore (optional).
  #  #trustStorePassword: bigdata
  #
  #  ## Key in the Secret containing the keystore file (optional).
  #  #keyStoreKey: keystore.jks
  #
  #  ## Password for the keystore (optional).
  #  #keyStorePassword: bigdata
  #
  #  ## Local files 'adc apply' puts into the Secret named by ssl.secretName.
  #  ## Relative paths are resolved against the config file.
  #  #files:
  #  #  trustStorePath: /path/to/truststore.jks
  #  #  #keyStorePath: /path/to/keystore.jks

  ## Use an external Hadoop configs Secret instead of the one rendered by ADC.
  #hadoopConfigsSecret:
  #  # Use a Secret managed outside ADC.
  #  externalSecretName: existing-spark-hadoop-configs
  #
  #  ## Or let ADC create the Secret.
  #  #secretName: custom-spark-hadoop-configs

  ## Use an external Ranger configs Secret instead of the one rendered by ADC.
  #rangerConfigsSecret:
  #  # Use a Secret managed outside ADC.
  #  externalSecretName: existing-spark-ranger-configs
  #
  #  ## Or let ADC create the Secret.
  #  #secretName: custom-spark-ranger-configs

  # Spark application main resource (e.g. local:///opt/spark/examples/jars/spark-examples.jar).
  mainApplicationFile: "local:///opt/spark/examples/jars/spark-examples_2.13-3.5.4.4-4.3.0-2.jar"

  ## HDFS or local directory for the Spark event log.
  ## When set, the CLI adds spark.eventLog.enabled=true, spark.eventLog.dir,
  ## spark.eventLog.rolling.enabled=true and spark.eventLog.rolling.interval=30s to sparkConf.
  #eventLogDir: ""

  ## Fully-qualified main class name. Required for Java/Scala applications.
  mainClass: "org.apache.spark.examples.SparkPi"

  # ServiceAccount used by the Spark driver, also injected into
  # spark.kubernetes.authenticate.driver.serviceAccountName. The CLI creates it, plus a Role
  # and RoleBinding for Spark pods, by default (create: true). Set create: false to skip that
  # and only reference a ServiceAccount managed elsewhere.
  # The Role rules are managed by the CLI and cannot be customized.
  serviceAccount:
    create: true
    name: spark-application
  job:
    ## true (default) deletes the spark-submit Job pod after it finishes; set false to keep it for debugging.
    deleteOnTermination: false

    #resources:
    #  limits:
    #    cpu: 500m
    #    memory: 512Mi
    #  requests:
    #    cpu: 250m
    #    memory: 256Mi

    ## Component arguments. Key-value pairs passed to the component configuration.
    args:
      driver-memory: 4g
      executor-memory: 4g

  ## Application arguments appended after mainApplicationFile.
  #args:
  #  - "100"

  # Spark configuration entries (spark.*).
  sparkConf:
    spark.artifactory.dir.path: /tmp/artifacts
    spark.jars.ivy: /tmp/ivy
    spark.local.dir: /tmp/data
    spark.sql.catalog.spark_catalog: org.apache.iceberg.spark.SparkSessionCatalog
    spark.sql.extensions: org.apache.iceberg.spark.extensions.IcebergSparkSessionExtensions
    spark.sql.security.confblacklist: spark.sql.extensions


  ## Seconds after the application finishes (Succeeded, or Failed with no
  ## retries left) before the SparkApplication is deleted. Omit to keep it
  ## until explicit deletion.
  #ttlSecondsAfterFinished: 3600

  ## Celeborn remote shuffle service for Spark.
  ## tiers mirror the cluster's storage.tiers: local (SSD/HDD) and MEMORY advertise their name only,
  ## remote (S3/HDFS) also set dir (and, for S3, credentials). Ozone is an HDFS tier with an ofs:// dir.
  #celeborn:
    #  # Storage tiers mirroring the cluster's storage.tiers so the client advertises the same layers.
    #  # type: SSD, HDD, S3, HDFS, MEMORY. SSD/HDD and MEMORY are advertise-only here (no dir; worker-only
    #  # fields ignored); remote S3/HDFS set dir (s3a:// for S3; hdfs:// or ofs:// for Ozone on HDFS). Example:
    #  #   - type: MEMORY   # cache tier; pair with a durable tier below, no dir
    #  #   - type: SSD      # advertise-only on the client, no dir needed
    #  #   - dir: s3a://bucket/celeborn
    #  #     s3: { endpoint: https://s3:9878, region: us-east-1 }
    #  #     type: S3
    #  #   - dir: hdfs://nn/celeborn   # or ofs://om/volume/bucket/celeborn for Ozone
    #  #     type: HDFS
    #  tiers:
    #    - dir: s3a://shuffle/my-cluster
    #      s3:
    #        accessKey: <access-key>
    #        endpoint: https://s3.endpoint:443
    #        pathStyleAccess: true
    #        region: <region>
    #        secretKey: <secret-key>
    #      type: S3
    #  masterEndpoint: ""
    #  extraSparkConf:
    #    spark.sql.adaptive.enabled: "true"

    # Enable TLS on the client's RPC connection to the Celeborn cluster.
    # When true the CLI renders spark.celeborn.ssl.* into the Spark conf
    # and requires the ssl section with trustStoreKey.
    rpcEncryption: false

  ## YuniKorn scheduler configuration for queue selection and Gang scheduling.
  ## Uncomment the block to route the Spark job into a YuniKorn queue; the CLI renders the
  ## scheduler name, queue labels and gang annotations into sparkConf.
  yunikorn:
    queue: root.analytics
    taskGroups:
      - minMember: 1
        minResource:
          cpu: "1"
          memory: 1433Mi
        name: spark-driver
      - minMember: 2
        minResource:
          cpu: "1"
          memory: 1433Mi
        name: spark-executor

  ## Monitoring configuration.
  monitoring:
    # Enables Prometheus metrics export from this product's pods.
    exportMetrics: true