Train and Deploy Machine Learning Models on Kubernetes with Kubeflow and Seldon-Core

Using:

• kubeflow
• seldon-core

The example will be the MNIST handwritten digit classification task. We will train 3 different models to solve this task:

• A TensorFlow neural network model.
• A scikit-learn random forest model.
• An R least squares model.

We will then show various rolling deployments

1. Deploy the single Tensorflow model.
2. Do a rolling update to an AB test of the Tensorflow model and the sklearn model.
3. Do a rolling update to a Multi-armed Bandit over all 3 models to direct traffic in real time to the best model.

In the follow we will:

1. Install kubeflow and seldon-core on a kubernetes cluster
2. Train the models
3. Serve the models

Requirements

• gcloud
• kubectl
• ksonnet
• argo

Setup

There is a consolidated script to create the demo which can be found here. For a step by step guide do the following:

1. Install kubeflow on GKE. This should create kubeflow in a namespace kubeflow. We suggest you use the command line install so you can easily modify your Ksonnet installation. Ensure you have the environment variables KUBEFLOW_SRC and KFAPP set. OAUTH is preferred as with basic auth port-forwarding to ambassador is insufficient

2. Install seldon. Go to your Ksonnet application folder setup in the previous step and run

   cd ${KUBEFLOW_SRC}/${KFAPP}/ks_app
   
   ks pkg install kubeflow/seldon
   ks generate seldon seldon
   ks apply default -c seldon
   

3. Install Helm

   kubectl -n kube-system create sa tiller
   kubectl create clusterrolebinding tiller --clusterrole cluster-admin --serviceaccount=kube-system:tiller
   helm init --service-account tiller
   kubectl rollout status deploy/tiller-deploy -n kube-system
   

4. Create an NFS disk and persistent volume claim called nfs-1. You can follow one guide on create an NFS volume using Google Filestore here. A consolidated set of steps is shown here

5. Add Cluster Roles so Argo can start jobs successfully

   kubectl create clusterrolebinding my-cluster-admin-binding --clusterrole=cluster-admin --user=$(gcloud info --format="value(config.account)")
   kubectl create clusterrolebinding default-admin2 --clusterrole=cluster-admin --serviceaccount=kubeflow:default
   

6. Install Seldon Analytics Dashboard

   helm install seldon-core-analytics --name seldon-core-analytics --set grafana_prom_admin_password=password --set persistence.enabled=false --repo https://storage.googleapis.com/seldon-charts --namespace kubeflow
   

7. Port forward the dashboard when running

   kubectl port-forward $(kubectl get pods -n kubeflow -l app=grafana-prom-server -o jsonpath='{.items[0].metadata.name}') -n kubeflow 3000:3000
   

8. Visit http://localhost:3000/dashboard/db/prediction-analytics?refresh=5s&orgId=1 and login using "admin" and the password you set above when launching with helm.

MNIST models

Tensorflow Model

• Python training code
• Python runtime prediction code
• Dockerfile to wrap runtime prediction code to run under seldon-Core.

SKLearn Model

• Python training code
• Python runtime prediction code
• Dockerfile to wrap runtime prediction code to run under seldon-Core.

R Model

• R training code
• R runtime prediction code
• Dockerfile to wrap runtime prediction code to run under seldon-Core.

Train the Models

Follow the steps in ./notebooks/training.ipynb to:

• Run Argo Jobs for each model to:
  • Creating training images and push to repo
  • Run training
  • Create runtime prediction images and push to repo
  • Deploy individual runtime model

To push to your own repo the Docker images you will need to setup your docker credentials as a Kubernetes secret containing a config.json. To do this you can find your docker home (typically ~/.docker) and run kubectl create secret generic docker-config --from-file=config.json=${DOCKERHOME}/config.json --type=kubernetes.io/config to create a secret.

Serve the Models

Follow the steps in ./notebooks/serving.ipynb to:

1. Deploy the single Tensorflow model.
2. Do a rolling update to an AB test of the Tensorflow model and the sklearn model.
3. Do a rolling update to a Multi-armed Bandit over all 3 models to direct traffic in real time to the best model.

To ensure the notebook can run successfully install the python dependencies:

pip install -r notebooks/requirements.txt

If you have installed the Seldon-Core analytics you can view them on the grafana dashboard: