Argo Workflow demo to launch kubernetes dask distributed job

dask-test-workflow.yaml

apiVersion: argoproj.io/v1alpha1
kind: Workflow
metadata:
  generateName: dask-test-
spec:
  entrypoint: dask
  activeDeadlineSeconds: 1800  # Safety first, kids!
  templates:
  - name: dask
    script:
      image: daskdev/dask:2.30.0
      env:
      - name: EXTRA_PIP_PACKAGES
        value: "dask-kubernetes dask[distributed]"
      command: ["python"]
      source: |
        # Hack to get the dask container entrypoint to fire-off, otherwise
        # container skips EXTRA_PIP_PACKAGES install
        import subprocess

        subprocess.run(
            ["bash", "/usr/bin/prepare.sh"], stdout=subprocess.PIPE, universal_newlines=True
        )

        # ...And now for something completely different...
        from dask.distributed import Client
        from dask_kubernetes import KubeCluster, make_pod_spec
        import dask.array as da

        pod_spec = make_pod_spec(
            image="daskdev/dask:latest",
            memory_limit="500Mi",
            memory_request="500m",
            cpu_limit=1,
            cpu_request=1,
            env={"EXTRA_PIP_PACKAGES": "dask[distributed]"},
        )
        cluster = KubeCluster(pod_spec)
        cluster.scale(10)  # specify number of workers explicitly

        client = Client(cluster)

        array = da.ones((1000, 1000, 1000))
        print(f"OUR ANSWER IS: {array.mean().compute()} (...it should be 1.0 ...)")
      resources:
        limits:
          memory: 500Mi
          cpu: 500m

@emileten Yeah, and if you're working on a fresh un-configured cluster you also need to ensure the ServiceAccount has permissions like here https://argoproj.github.io/argo-workflows/workflow-rbac/, in addition to those you pointed to above in your link.

thank you for this, this helped me on so many levels