I have a google cloud composer environment. In my DAG I want to create a pod in GKE. When I come to deploy a simple app based on a docker container that doesn't need any volume configuration or secrets, everything works fine, for example:

kubernetes_max = GKEStartPodOperator(
    # The ID specified for the task.
    task_id="python-simple-app",
    # Name of task you want to run, used to generate Pod ID.
    name="python-demo-app",
    project_id=PROJECT_ID,
    location=CLUSTER_REGION,
    cluster_name=CLUSTER_NAME,
    # Entrypoint of the container, if not specified the Docker container's
    # entrypoint is used. The cmds parameter is templated.
    cmds=["python", "app.py"],
    namespace="production",
    image="gcr.io/path/to/lab-python-job:latest",
)

But when I have an application that need to access to my GKE cluster volumes, I need to configure volumes in my pod. The issue is the documentation is not clear regarding this. The only example that I ever foud is this:

volume = k8s.V1Volume(
    name='test-volume',
    persistent_volume_claim=k8s.V1PersistentVolumeClaimVolumeSource(claim_name='test-volume'),
)

While the volumes in the my manifest file (I use it to deploy my app from local) looks like this:

volumes:
  - name: volume-prod
    secret:
      secretName: volume-prod
      items:
        - key: config
          path: config.json
        - key: another_config
          path: another_config.conf
        - key: random-ca
          path: random-ca.pem

So when I compare how both volumes looks like in the console (when I manually deploy the manifest file that successfully run, and when I deploy the pod using clod composer that fails):

• The successful run - Manifest file:

  volume-prod
  Name: volume-prod
  Type: secret
  Source volume identifier: volume-prod

• The failed run - Composer GKEStartPodOperator:

  volume-prod
  Name: volume-prod
  Type: emptyDir
  Source volume identifier: Node's default medium

How I can configure my pod from cloud composer in a way it can read the volume of my cluster?

CodePudding user response：

The KubernetesPodOperator/GKEStartOperator is just a wrapper around the python Kubernetes sdk - I agree that it isn't well documented in the Airflow/Cloud Composer documentation but the Python SDK for Kubernetes itself is well documented.

Start here with the kubernetes python sdk documentation: https://github.com/kubernetes-client/python/blob/master/kubernetes/docs/V1PodSpec.md

You'll notice that the arguments the KubernetesPodOperator/GKEStartOperator take match this spec. If you dig into the source code of the operators you'll see that the operator is nothing more than a builder that creates a kubernetes.client.models.V1Pod object and uses the API to deploy the pod.

The operator takes a volumes parameter which should be of type List[V1Volume], where the documentation for V1Volume is here.

So in your case you would need to provide:

from kubernetes.client import models as k8s

kubernetes_max = GKEStartPodOperator(
    # The ID specified for the task.
    task_id="python-simple-app",
    # Name of task you want to run, used to generate Pod ID.
    name="python-demo-app",
    project_id=PROJECT_ID,
    location=CLUSTER_REGION,
    cluster_name=CLUSTER_NAME,
    # Entrypoint of the container, if not specified the Docker container's
    # entrypoint is used. The cmds parameter is templated.
    cmds=["python", "app.py"],
    namespace="production",
    image="gcr.io/path/to/lab-python-job:latest",
    volumes=[
        k8s.V1Volume(
            secret=k8s.V1SecretVolumeSource(
                secret_name="volume-prod",
                items=[
                    k8s.V1KeyToPath(key="config", path="config.json"),
                    k8s.V1KeyToPath(key="another_config", path="another_config.conf"),
                    k8s.V1KeyToPath(key="random-ca", path="random-ca.pem"),
                ],
            )
        )
    ]
)

Alternatively, you can provide your manifest to the pod_template_file argument in GKEStartPodOperator - this will need to be available to the workers inside airflow.

There are 3 ways to create pods in Airflow using this Operator:

1. Use the arguments of the operator to specify what you need and have the operator build the V1Pod for you.
2. Provide a manifest by passing in pod_template_file argument.
3. Use the Kubernetes sdk to create a V1Pod object yourself and pass this to the full_pod_spec argument.