Argo Workflow + Spark operator + App logs not generated

Question

Am in very early stages of exploring Argo with Spark operator to run Spark samples on the minikube setup on my EC2 instance.

Following are the resources details, not sure why am not able to see the spark app logs.

WORKFLOW.YAML

kind: Workflow
metadata:
  name: spark-argo-groupby
spec:
  entrypoint: sparkling-operator
  templates:
  - name: spark-groupby
    resource:
      action: create
      manifest: |
        apiVersion: "sparkoperator.k8s.io/v1beta2"
        kind: SparkApplication
        metadata:
          generateName: spark-argo-groupby
        spec:
          type: Scala
          mode: cluster
          image: gcr.io/spark-operator/spark:v3.0.3
          imagePullPolicy: Always
          mainClass: org.apache.spark.examples.GroupByTest
          mainApplicationFile:  local:///opt/spark/spark-examples_2.12-3.1.1-hadoop-2.7.jar
          sparkVersion: "3.0.3"
          driver:
            cores: 1
            coreLimit: "1200m"
            memory: "512m"
            labels:
              version: 3.0.0
          executor:
            cores: 1
            instances: 1
            memory: "512m"
            labels:
              version: 3.0.0
  - name: sparkling-operator
    dag:
      tasks:
      - name: SparkGroupBY
        template: spark-groupby

ROLES

# Role for spark-on-k8s-operator to create resources on cluster
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
metadata:
  name: spark-cluster-cr
  labels:
    rbac.authorization.kubeflow.org/aggregate-to-kubeflow-edit: "true"
rules:
  - apiGroups:
      - sparkoperator.k8s.io
    resources:
      - sparkapplications
    verbs:
      - '*'
---
# Allow airflow-worker service account access for spark-on-k8s
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: argo-spark-crb
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: spark-cluster-cr
subjects:
  - kind: ServiceAccount
    name: default
    namespace: argo

ARGO UI

To dig deep I tried all the steps that's listed on https://dev.to/crenshaw_dev/how-to-debug-an-argo-workflow-31ng yet could not get app logs.

Basically when I run these examples am expecting spark app logs to be printed - in this case output of following Scala example

https://github.com/apache/spark/blob/master/examples/src/main/scala/org/apache/spark/examples/GroupByTest.scala

Interesting when I list PODS, I was expecting to see driver pods and executor pods but always see only one POD and it has above logs as in the image attached. Please help me to understand why logs are not generated and how can I get it?

RAW LOGS
$ kubectl logs spark-pi-dag-739246604 -n argo

time="2021-12-10T13:28:09.560Z" level=info msg="Starting Workflow Executor" version="{v3.0.3 2021-05-11T21:14:20Z 02071057c082cf295ab8da68f1b2027ff8762b5a v3.0.3 clean go1.15.7 gc linux/amd64}"
time="2021-12-10T13:28:09.581Z" level=info msg="Creating a docker executor"
time="2021-12-10T13:28:09.581Z" level=info msg="Executor (version: v3.0.3, build_date: 2021-05-11T21:14:20Z) initialized (pod: argo/spark-pi-dag-739246604) with template:\n{\"name\":\"sparkpi\",\"inputs\":{},\"outputs\":{},\"metadata\":{},\"resource\":{\"action\":\"create\",\"manifest\":\"apiVersion: \\\"sparkoperator.k8s.io/v1beta2\\\"\\nkind: SparkApplication\\nmetadata:\\n  generateName: spark-pi-dag\\nspec:\\n  type: Scala\\n  mode: cluster\\n  image: gjeevanm/spark:v3.1.1\\n  imagePullPolicy: Always\\n  mainClass: org.apache.spark.examples.SparkPi\\n  mainApplicationFile: local:///opt/spark/spark-examples_2.12-3.1.1-hadoop-2.7.jar\\n  sparkVersion: 3.1.1\\n  driver:\\n    cores: 1\\n    coreLimit: \\\"1200m\\\"\\n    memory: \\\"512m\\\"\\n    labels:\\n      version: 3.0.0\\n  executor:\\n    cores: 1\\n    instances: 1\\n    memory: \\\"512m\\\"\\n    labels:\\n      version: 3.0.0\\n\"},\"archiveLocation\":{\"archiveLogs\":true,\"s3\":{\"endpoint\":\"minio:9000\",\"bucket\":\"my-bucket\",\"insecure\":true,\"accessKeySecret\":{\"name\":\"my-minio-cred\",\"key\":\"accesskey\"},\"secretKeySecret\":{\"name\":\"my-minio-cred\",\"key\":\"secretkey\"},\"key\":\"spark-pi-dag/spark-pi-dag-739246604\"}}}"
time="2021-12-10T13:28:09.581Z" level=info msg="Loading manifest to /tmp/manifest.yaml"
time="2021-12-10T13:28:09.581Z" level=info msg="kubectl create -f /tmp/manifest.yaml -o json"
time="2021-12-10T13:28:10.348Z" level=info msg=argo/SparkApplication.sparkoperator.k8s.io/spark-pi-daghhl6s
time="2021-12-10T13:28:10.348Z" level=info msg="Starting SIGUSR2 signal monitor"
time="2021-12-10T13:28:10.348Z" level=info msg="No output parameters"

Answer 1

As Michael mentioned in his answer, Argo Workflows does not know how other CRDs (such as SparkApplication that you used) work and thus could not pull the logs from the pods created by that particular CRD.

However, you can add the label workflows.argoproj.io/workflow: {{workflow.name}} to the pods generated by SparkApplication to let Argo Workflows know and then use argo logs -c <container-name> to pull the logs from those pods.

You can find an example here but Kubeflow CRD but in your case you'll want to add labels to the executor and driver to your SparkApplication CRD in the resource template: https://github.com/argoproj/argo-workflows/blob/master/examples/k8s-kubeflow-jobs.yaml

Answer 2

Argo Workflows' resource templates (like your spark-groupby template) are simplistic. The Workflow controller is running kubectl create , and that's where its involvement in the SparkApplication ends.

The logs you're seeing from the Argo Workflow pod describe the kubectl create process. Your resource is written to a temporary yaml file and then applied to the cluster.

time="2021-12-10T13:28:09.581Z" level=info msg="Loading manifest to /tmp/manifest.yaml"
time="2021-12-10T13:28:09.581Z" level=info msg="kubectl create -f /tmp/manifest.yaml -o json"
time="2021-12-10T13:28:10.348Z" level=info msg=argo/SparkApplication.sparkoperator.k8s.io/spark-pi-daghhl6s

Old answer:

To view the logs generated by your SparkApplication, you'll need to follow the Spark docs. I'm not familiar, but I'm guessing the application gets run in a Pod somewhere. If you can find that pod, you should be able to view the logs with kubectl logs .

It would be really cool if Argo Workflows could pull Spark logs into its UI. But building a generic solution would probably be prohibitively difficult.

Update:

Check Yuan's answer. There's a way to pull the Spark logs into the Workflows CLI!