You are viewing documentation for Kubeflow 0.4

This is a static snapshot from the time of the Kubeflow 0.4 release.
For up-to-date information, see the latest version.

Kubernetes Engine for Kubeflow

Get Kubeflow running on Google Cloud Platform

This guide is a quickstart to deploying Kubeflow on Google Kubernetes Engine (GKE).

Advantages of Kubeflow on GKE

Running Kubeflow on GKE brings the following advantages:

  • You use Deployment Manager to declaratively manage all non-Kubernetes resources (including the GKE cluster). Deployment Manager is easy to customize for your particular use case.
  • You can take advantage of GKE autoscaling to scale your cluster horizontally and vertically to meet the demands of machine learning (ML) workloads with large resource requirements.
  • Cloud Identity-Aware Proxy (Cloud IAP) makes it easy to securely connect to Jupyter and other web apps running as part of Kubeflow.
  • Stackdriver makes it easy to persist logs to aid in debugging and troubleshooting
  • You can use GPUs and TPUs to accelerate your workload.

Prerequisites

Ensure that the following APIs are enabled on your Google Cloud Platform (GCP) account:

If you plan to install Kubeflow on the command line:

  • Ensure you have installed ksonnet version 0.13.1 or later.
  • If you’re using Cloud Shell, enable boost mode.

You do not need a running GKE cluster. The deployment process will create a cluster for you.

Create OAuth client credentials

Create an OAuth client ID to be used to identify Cloud IAP when requesting access to a user’s email account. Kubeflow uses the email address to verify the user’s identity.

  1. Set up your OAuth consent screen:

    • In the Application name box, enter the name of your application. The example below uses the name “Kubeflow”.

    • Under Support email, select the email address that you want to display as a public contact. You must use either your email address or a Google Group that you own.

    • If you see Authorized domains, enter

      <project>.cloud.goog
      
      • where <project> is your Google Cloud Platform (GCP) project ID.
      • If you are using your own domain, such as acme.com, you should add that as well
      • The Authorized domains option appears only for certain project configurations. If you don’t see the option, then there’s nothing you need to set.
    • Click Save.

    • Here’s an example of the completed form:

  2. On the credentials tab:

    • Click Create credentials, and then click OAuth client ID.

    • Under Application type, select Web application.

    • In the Name box enter any name for your OAuth client ID. This is not the name of your application nor the name of your Kubeflow deployment. It’s just a way to help you identify the OAuth client ID.

    • In the Authorized redirect URIs box, enter the following:

      https://<deployment_name>.endpoints.<project>.cloud.goog/_gcp_gatekeeper/authenticate
      
      • <deployment_name> must be the name of your Kubeflow deployment. It must have the same value as the deployment name used in the next step when you deploy Kubeflow from the UI or by running the deployment script. Deployment name must be 4-20 characters long. The default value for the deployment name is the KFAPP value used when initializing your Kubeflow app, but you can configure this with the environment variable DEPLOYMENT_NAME.
      • <project> is your GCP project. It must have the same value as used in the next step when you deploy Kubeflow from the UI or by running the deployment script.
    • Here’s an example of the completed form:

  3. Press Enter/Return to add the URI. Check that the URI now appears as a confirmed item under Authorized redirect URIs. (It should no longer be editable.)

  4. Make note of the client ID and client secret that appear in the OAuth client window. You need them later to enable Cloud IAP.

Deploy Kubeflow on GKE using the UI

The simplest way to deploy Kubeflow is to use the Kubeflow deployment web interface:

  1. Open https://deploy.kubeflow.cloud/ in your web browser.
  2. Sign in using a GCP account that has administrator privileges for your GCP project.
  3. Complete the form, following the instructions on the left side of the form. In particular, ensure that you enter the same deployment name as you used when creating the OAuth client ID.
  4. Click Create Deployment.

Here’s a partial screenshot of the deployment UI, showing all the fields in the form:

Kubeflow will be available at the following URI:

https://<deployment_name>.endpoints.<project>.cloud.goog/

It can take 10-15 minutes for the URI to become available. You can watch for updates in the information box on the deployment UI. If the deployment takes longer than expected, try accessing the above URI anyway.

Deleting your Kubeflow Deployment

To delete your Kubeflow deployment and reclaim all the resources from the UI:

  1. Open the Deployment Manager tab in the cloud console for your project through the browser. This should list all the available deployments in your project. Please, make sure that the selected project is the same as the one you used for your Kubeflow deployment.

  2. Select your Kubeflow deployment with the deployment name you used at the time of creation and click Delete button at the top.

This should delete any running nodes in your deployments, delete service accounts that were created and reclaim all resources.

Deploy Kubeflow on GKE using the command line

If you prefer to have more control over the deployment process and configuration, you can use the command line instead of the UI to deploy Kubeflow.

Follow these steps to deploy Kubeflow:

  1. Create environment variables from the OAuth client ID and secret that you obtained earlier:

    export CLIENT_ID=<CLIENT_ID from OAuth page>
    export CLIENT_SECRET=<CLIENT_SECRET from OAuth page>
    
  2. Run the following script to download kfctl.sh:

    mkdir ${KUBEFLOW_SRC}
    cd ${KUBEFLOW_SRC}
    export KUBEFLOW_TAG=<a href="https://github.com/kubeflow/kubeflow/releases/tag/v0.4.1">v0.4.1</a>
    
    curl https://raw.githubusercontent.com/kubeflow/kubeflow/${KUBEFLOW_TAG}/scripts/download.sh | bash
    
    • KUBEFLOW_SRC - a directory where you want to download the source to.
    • KUBEFLOW_TAG - a tag corresponding to the Kubeflow version to check out, such as master for the latest code.
    • Note: You can also just clone the Kubeflow repository using git.
  3. Run the following scripts to set up and deploy Kubeflow:

    ${KUBEFLOW_SRC}/scripts/kfctl.sh init ${KFAPP} --platform gcp --project ${PROJECT}
    cd ${KFAPP}
    ${KUBEFLOW_SRC}/scripts/kfctl.sh generate platform
    ${KUBEFLOW_SRC}/scripts/kfctl.sh apply platform
    ${KUBEFLOW_SRC}/scripts/kfctl.sh generate k8s
    ${KUBEFLOW_SRC}/scripts/kfctl.sh apply k8s
    
    • ${KFAPP} - the name of a directory where you want kubeflow configurations to be stored. This directory is created when you run init. The contents of this directory are described in the next section.
  4. Check the resources deployed in namespace kubeflow:

    kubectl -n kubeflow get  all
    
  5. Kubeflow will be available at the following URI:

    https://<deployment_name>.endpoints.<project>.cloud.goog/
    
    • It can take 10-15 minutes for the URI to become available. Kubeflow needs to provision a signed SSL certificate and register a DNS name.
    • If you own/manage the domain or a subdomain with Cloud DNS then you can configure this process to be much faster. See kubeflow/kubeflow#731.
    • While you wait you can access Kubeflow services by using kubectl proxy and kubectl port-forward to connect to services in the cluster.
  6. We recommend that you check in the contents of your ${KFAPP} directory into source control.

  7. To delete your deployment and reclaim all resources:

    cd ${KFAPP}
    ${KUBEFLOW_SRC}/scripts/kfctl.sh delete all
    

Understanding the deployment process

The deployment process is controlled by 4 different commands:

  • init - one time setup.
  • generate - creates config files defining the various resources.
  • apply - creates or updates the resources.
  • delete - deletes the resources.

With the exception of init, all commands take an argument which describes the set of resources to apply the command to; this argument can be one of the following:

  • platform - all GCP resources; that is, anything that doesn’t run on Kubernetes.
  • k8s - all resources that run on Kubernetes.
  • all - GCP and Kubernetes resources.

App layout

Your Kubeflow app directory contains the following files and directories:

  • env.sh defines several environment variables related to your Kubeflow deployment.

    • The values are set when you run init.
    • The values are snapshotted inside env.sh to make your app self contained.
  • ${KFAPP}/gcp_config is a directory that contains Deployment Manager configuration files defining your GCP infrastructure.

    • The directory is created when you run kfctl.sh generate platform.
    • You can modify these configurations to customize your GCP infrastructure.
  • ${KFAPP}/k8s_specs is a directory that contains YAML specifications for some daemons deployed on your Kubernetes Engine cluster.

  • ${KFAPP}/ks_app is a directory that contains the ksonnet application for Kubeflow.

    • The directory is created when you run kfctl.sh generate k8s.
    • You can use ksonnet to customize Kubeflow.

GCP service accounts

Creating a deployment using kfctl.sh creates three service accounts in your GCP project. These service accounts are created using the principle of least privilege. The three service accounts are:

  • ${KFAPP}-admin is used for some admin tasks like configuring the load balancers. The principle is that this account is needed to deploy Kubeflow but not needed to actually run jobs.
  • ${KFAPP}-user is intended to be used by training jobs and models to access GCP resources (Cloud Storate, BigQuery, etc.). This account has a much smaller set of privileges compared to admin.
  • ${KFAPP}-vm is used only for the virtual machine (VM) service account. This account has minimal permissions, needed to send metrics and logs to Stackdriver.

Next steps