pipelines/manifests/gcp_marketplace

Kubeflow Pipelines for GKE Marketplace

Kubeflow Pipelines can be installed using either of the following approaches:

• Using the Google Cloud Platform Console

• Using the command line

Using the Google Cloud Platform Marketplace

Get up and running with a few clicks! Install this Kubeflow Pipelines app to a Google Kubernetes Engine cluster using Google Cloud Marketplace. Follow the on-screen instructions.

Using the command line

You can use Google Cloud Shell or a local workstation to complete these steps.

Prerequisites

Set up command-line tools

You'll need the following tools in your development environment. If you are using Cloud Shell, these tools are installed in your environment by default.

• gcloud
• kubectl
• docker
• git

Configure gcloud as a Docker credential helper:

gcloud auth configure-docker

You can install Kubeflow Pipelines in an existing GKE cluster or create a new GKE cluster.

• If you want to create a new Google GKE cluster, follow the instructions from the section Create a GKE cluster onwards.

• If you have an existing GKE cluster, ensure that the cluster nodes have a minimum 3 node cluster with each node having a minimum of 2 vCPU and running k8s version 1.9 and follow the instructions from section Install the application resource definition onwards.

Create a GKE cluster

Kubeflow Pipelines requires a minimum 3 node cluster with each node having a minimum of 2 vCPU and k8s version 1.9. Available machine types can be seen here.

Create a new cluster from the command line:

export CLUSTER=kubeflow-pipelines-cluster
export ZONE=us-west1-a
export MACHINE_TYPE=n1-standard-2

gcloud container clusters create "$CLUSTER" --zone "$ZONE" --machine-type "$MACHINE_TYPE"

Configure kubectl to connect to the new cluster:

gcloud container clusters get-credentials "$CLUSTER" --zone "$ZONE"

Install the application resource definition

An application resource is a collection of individual Kubernetes components, such as services, stateful sets, deployments, and so on, that you can manage as a group.

To set up your cluster to understand application resources, run the following command:

kubectl apply -f "https://raw.githubusercontent.com/GoogleCloudPlatform/marketplace-k8s-app-tools/master/crd/app-crd.yaml"

You need to run this command once.

The application resource is defined by the Kubernetes SIG-apps community. The source code can be found on github.com/kubernetes-sigs/application.

Prerequisites for using Role-Based Access Control

You must grant your user the ability to create roles in Kubernetes by running the following command.

kubectl create clusterrolebinding cluster-admin-binding \
  --clusterrole cluster-admin \
  --user $(gcloud config get-value account)

You need to run this command once.

Install the Application

Run installer script

Set your application instance name and the Kubernetes namespace to deploy:

# set the application instance name
export APP_INSTANCE_NAME=kubeflow-pipelines-app

# set the Kubernetes namespace the application was originally installed
export NAMESPACE=<namespace>

Creat the namespace

kubectl create namespace $NAMESPACE

Download token for your service account which you want to use for calling GCP APIs from the pipelines.

gcloud iam service-accounts keys create application_default_credentials.json --iam-account [your-service-account]
export SERVICE_ACCOUNT_TOKEN="$(cat application_default_credentials.json | base64 -w 0)"

Follow the instruction and install mpdev TODO: The official mpdev won't work because it doesn't have permission to deploy CRD. The latest unofficial build will have right permission. Remove following instruction when change is in prod.

BIN_FILE="$HOME/bin/mpdev"
docker run gcr.io/cloud-marketplace-staging/marketplace-k8s-app-tools/k8s/dev:unreleased-pr396 cat /scripts/dev > "$BIN_FILE"
chmod +x "$BIN_FILE"
export MARKETPLACE_TOOLS_TAG=unreleased-pr396
export MARKETPLACE_TOOLS_IMAGE=gcr.io/cloud-marketplace-staging/marketplace-k8s-app-tools/k8s/dev

Run the install script

mpdev scripts/install  --deployer=gcr.io/ml-pipeline/google/pipelines/deployer:0.2   --parameters='{"name": "'$APP_INSTANCE_NAME'", "namespace": "'$NAMESPACE'", "serviceAccountCredential": "'$SERVICE_ACCOUNT_TOKEN'"}'

Or if using CloudSQL and GCS,

mpdev scripts/install  --deployer=gcr.io/ml-pipeline/google/pipelines/deployer:0.2   --parameters='{"name": "'$APP_INSTANCE_NAME'", "namespace": "'$NAMESPACE'", "serviceAccountCredential": "'$SERVICE_ACCOUNT_TOKEN'", "managedstorage.enabled": true, "managedstorage.cloudsqlInstanceConnectionName": "[your-name]", "managedstorage.dbPassword": "[your-pwd]"}'

Watch the deployment come up with

kubectl get pods -n $NAMESPACE --watch

Get public endpoint

kubectl describe configmap inverse-proxy-config -n $NAMESPACE | grep googleusercontent.com
 

Delete the Application

kubectl delete applications -n $NAMESPACE $APP_INSTANCE_NAME