You are viewing documentation for Kubeflow 1.4

This is a static snapshot from the time of the Kubeflow 1.4 release.
For up-to-date information, see the latest version.

Argo Workflow Executors

How to choose and configure the Argo Workflow Executor?

An Argo workflow executor is a process that conforms to a specific interface that allows Argo to perform certain actions like monitoring pod logs, collecting artifacts, managing container lifecycles, etc.

Kubeflow Pipelines runs on Argo Workflows as the workflow engine, so Kubeflow Pipelines users need to choose a workflow executor.

Choosing the Workflow Executor

  1. Some users may value stability and backward compatibility. For example, if you are running Kubeflow Pipelines in a production cluster or you maintain production pipelines that you don’t want to break or migrate.

    In this case, we recommend you use docker executor and configure your Kubernetes nodes to use docker container runtime.

    However, Kubernetes is deprecating docker as a container runtime, so we recommend starting to try out emissary and prepare for a migration when it’s stable.

  2. For users less concerned with stability and backwards compatibility, we recommend trying out the new emissary executor.

Note that Argo Workflows support other workflow executors, but the Kubeflow Pipelines team only recommend choosing between docker executor and emissary executor.

Docker Executor

Docker executor is the default workflow executor.

  • Container Runtime: docker only. However, Kubernetes is deprecating Docker as a container runtime after v1.20. On Google Kubernetes Engine (GKE) 1.19+, container runtime already defaults to containerd.
  • Reliability: most well-tested and most popular argo workflows executor
  • Security: least secure
    • It requires privileged access to docker.sock of the host to be mounted which. Often rejected by Open Policy Agent (OPA) or your Pod Security Policy (PSP). GKE autopilot mode also rejects it, because No privileged Pods.
    • It can escape the privileges of the pod’s service account.

Prepare a GKE cluster for Docker Executor

For GKE, the node image decides which container runtime is used. To use docker container runtime, you need to specify a node image with Docker.

You must use one of the following node images:

  • Container-Optimized OS with Docker (cos)
  • Ubuntu with Docker (ubuntu)

If your nodes are not using docker as container runtime, when you run pipelines you will always find error messages like:

This step is in Error state with this message: failed to save outputs: Error response from daemon: No such container: XXXXXX

Emissary Executor

Emissary executor is a new workflow executor. It was first released in Argo Workflows v3.1 (June 2021). However, the Kubeflow Pipelines team believe that its architectural and portability improvements can make it the default executor that most people should use in the future.

Therefore, the team makes a commitment to actively collect feedback and fix bugs for the emissary executor, so that we can stablize it faster. Submit your feedback in the Emissary Executor feedback github issue.

So far, Kubeflow Pipelines test infrastructure has been running stably with the emissary executor.

Migrate to Emissary Executor

Prerequisite: emissary executor is only available in Kubeflow Pipelines backend version 1.7+. To upgrade, refer to upgrading Kubeflow Pipelines.

Configure an existing Kubeflow Pipelines cluster to use emissary executor
  1. Install kubectl.

  2. Connect to your cluster via kubectl.

  3. Switch to the namespace you installed Kubeflow Pipelines:

    kubectl config set-context --current --namespace <your-kfp-namespace>
    

    Note, usually it’s kubeflow or default.

  4. Confirm current workflow executor:

    kubectl describe configmap workflow-controller-configmap | grep -A 2 containerRuntimeExecutor
    

    You’ll see output like the following when using docker executor:

    containerRuntimeExecutor:
    ----
    docker
    
  5. Configure workflow executor to emissary:

    kubectl patch configmap workflow-controller-configmap --patch '{"data":{"containerRuntimeExecutor":"emissary"}}'
    
  6. Confirm workflow executor is changed successfully:

    kubectl describe configmap workflow-controller-configmap | grep -A 2 containerRuntimeExecutor
    

    You’ll see output like the following:

    containerRuntimeExecutor:
    ----
    emissary
    
Deploy a new Kubeflow Pipelines cluster with emissary executor

For AI Platform Pipelines, check the “Use emissary executor” checkbox during installation.

For Kubeflow Pipelines Standalone, install env/platform-agnostic-emissary:

kubectl apply -k "github.com/kubeflow/pipelines/manifests/kustomize/env/platform-agnostic-emissary?ref=$PIPELINE_VERSION"

When in doubt, you can always deploy your Kubeflow Pipelines cluster first and configure workflow executor after installation using the instructions for existing clusters.

Migrate pipeline components to run on emissary executor

Some pipeline components require manual updates to run on emissary executor. For Kubeflow Pipelines component specification YAML, the command field must be specified.

Step by step component migration tutorial:

  1. There is a hello world component:

    name: hello-world
    implementation:
      container:
        image: hello-world
    
  2. We can run the container without command/args:

    $ docker run hello-world
    Hello from Docker!
    ...
    
  3. Find out what the default ENTRYPOINT and CMD is in the image:

    $ docker image inspect -f '{{.Config.Entrypoint}} {{.Config.Cmd}}' hello-world
    [] [/hello]
    

    So ENTRYPOINT is not specified, and CMD is ["/hello"]. Note, ENTRYPOINT roughly means command and CMD roughly means arguments. command and arguments are concatenated as the user command.

  4. Update the component YAML:

    name: hello-world
    implementation:
      container:
        image: hello-world
        command: ["/hello"]
    
  5. The updated component can run on emissary executor now.

Note: Kubeflow Pipelines SDK compiler always specifies a command for python function based components. Therefore, these components will continue to work on emissary executor without modifications.

References