Prerequisites

Before running the tutorial, configure a Docker secret and an AWS credentials secret. AWS credentials are needed in both the source cluster (the one training the model) and the target cluster (the one to which you'll deploy the model).

ENV setup

Set the environment variable NAMESPACE to the namespace in which you will be training and/or deploying your model.

export NAMESPACE=user1

Create a Docker secret

Create a docker configuration file, docker-config.json, with base64 encoded credentials.

1. Encode your credentials with the following command:

   CODE

   echo -n "<username>:<password>" | base64

2. Create the docker-config.json file with the following format:

   CODE

   {
     "auths": {
       "https://index.docker.io/v1/": {
               "auth": "<username:password in base64>"
       }
     }
   }

3. Create a secret from your docker-config.json file:

   CODE

   kubectl --kubeconfig target/kube.config create secret generic docker-secret -n ${NAMESPACE} --from-file=docker-config.json

Create an AWS secret

1. Set your AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY to environment variables:

   CODE

   AWS_ACCESS_KEY_ID=...
   AWS_SECRET_ACCESS_KEY=...

2. Create a secret with the following command:

   CODE

   kubectl --kubeconfig target/kube.config create secret -n ${NAMESPACE} generic aws-credentials \
       --from-literal=AWS_ACCESS_KEY_ID="${AWS_ACCESS_KEY_ID}" \
       --from-literal=AWS_SECRET_ACCESS_KEY="${AWS_SECRET_ACCESS_KEY}" \
       --from-literal=AWS_REGION=us-west-2 \
       --from-literal=S3_ENDPOINT=https://s3.us-west-2.amazonaws.com

Create a PodDefault which References the Docker and AWS Secrets:

cat << EOF | kubectl apply -f -
apiVersion: "kubeflow.org/v1alpha1"
kind: PodDefault
metadata:
  name: inject-credentials
  namespace: ${NAMESPACE}
spec:
  selector:
    matchLabels:
      inject-credentials: "true"
  desc: "AWS and Docker credentials"
  volumeMounts:
    - name: docker-secret-volume
      mountPath: /home/kubeflow/.docker/
  volumes:
    - name: docker-secret-volume
      secret:
        secretName: docker-secret
  envFrom:
  - secretRef:
      name: aws-credentials
EOF

Launch your Notebook Server

1. Create a notebook server based on the mesosphere/kubeflow:...-jupyter-spark-3.3.0-tensorflow-2.x.x image

2. Select AWS and Docker credentials under the Configuration section

3. Select Launch > Connect.

Upload the Tutorials and Run the Corresponding Notebook

Upload this tutorial to the notebook, and run the corresponding step. In the source cluster, run through Build-Notebook.ipynb. In the target deployment cluster, run through Deploy-Notebook.ipynb.