MLflow Deployment

07/26/2024 Contributors

This section describes the tasks that you must complete to deploy MLflow in your Kubernetes cluster.

It is possible to deploy MLflow on platforms other than Kubernetes. Deploying MLflow on platforms other than Kubernetes is outside of the scope of this solution.

Prerequisites

Before you perform the deployment exercise that is outlined in this section, we assume that you have already performed the following tasks:

You already have a working Kubernetes cluster.
You have already installed and configured NetApp Astra Trident in your Kubernetes cluster. For more details on Astra Trident, refer to the Astra Trident documentation.

Install Helm

MLflow is deployed using Helm, a popular package manager for Kubernetes. Before you deploy MLflow, you must install Helm on your Kubernetes control node. To install Helm, follow the installation instructions in the official Helm documentation.

Set Default Kubernetes StorageClass

Before you deploy MLflow, you must designate a default StorageClass within your Kubernetes cluster. To designate a default StorageClass within your cluster, follow the instructions outlined in the Kubeflow Deployment section. If you have already designated a default StorageClass within your cluster, then you can skip this step.

Deploy MLflow

Once the pre-requisites have been met, you can start with MLflow deployment using the helm chart.

Configure MLflow Helm Chart Deployment.

Before we deploy MLflow using the Helm chart, we can configure the deployment to use NetApp Trident Storage Class and change other parameters to suit our needs using a config.yaml file. An example of config.yaml file can be found at: https://github.com/bitnami/charts/blob/main/bitnami/mlflow/values.yaml

You can set the Trident storageClass under the global.defaultStorageClass parameter in the config.yaml file (e.g. storageClass: "ontap-flexvol").

Installing the Helm Chart

The Helm chart can be installed with the custom config.yaml file for MLflow using the following command:

helm install oci://registry-1.docker.io/bitnamicharts/mlflow -f config.yaml --generate-name --namespace jupyterhub

The command deploys MLflow on the Kubernetes cluster in the custom configuration via the provided config.yaml file. MLflow is deployed in the given namespace and a random release name is given via kubernetes for the release.

Check Deployment

After the Helm chart is done deploying, you can check if the service is accessible using:

kubectl get service -n jupyterhub

Replace jupyterhub with the namespace you used during deployment.

You should see the following services:

NAME                              TYPE        CLUSTER-IP      EXTERNAL-IP   PORT(S)           AGE
mlflow-1719843029-minio           ClusterIP   10.233.22.4     <none>        80/TCP,9001/TCP   25d
mlflow-1719843029-postgresql      ClusterIP   10.233.5.141    <none>        5432/TCP          25d
mlflow-1719843029-postgresql-hl   ClusterIP   None            <none>        5432/TCP          25d
mlflow-1719843029-tracking        NodePort    10.233.2.158    <none>        30002:30002/TCP   25d

We edited the config.yaml file to use NodePort service to access MLflow on port 30002.

Access MLflow

Once all the services related to MLflow are up and running you can access it using the given NodePort or LoadBalancer IP address (e.g. http://10.61.181.109:30002)