How to Deploy MLFlow on the Vultr Kubernetes Engine

Introduction

MLflow is a versatile open-source Machine Learning (ML) platform for model lifecycle management. It supports experimentation, duplicability, deployment, and a central model registry. MLflow is used to track machine learning workload experiments, packaging and deployment of code, and collaborative management. Frequently used by data scientists and by MLOps professionals.

This article explains how to deploy MLflow on a Vultr Kubernetes Engine (VKE) cluster with a working environment and demo application.

Prerequisites

Before you begin:

• Deploy a Vultr Kubernetes Engine(VKE) cluster

• Deploy a Vultr Ubuntu instance to use as your management server

• Using SSH, access the server

• Install and configure Kubectl to access the cluster

• Install the Helm package manager

    $ sudo snap install helm --classic

Deploy Persistent Volume and Persistent Volume Claims (PVCs)

MLflow requires persistent storage to store artifacts and experiment data. In this section, deploy a PV and PVC to the cluster as described in the steps below.

1. Create a new file mlflow-pv-pvc.yaml

    $ touch mlflow-pv-pvc.yaml
   

2. Using a text editor such as nano, edit the file

    $ nano mlflow-pv-pvc.yaml
   

3. Add the following contents to the file

    apiVersion: v1
    kind: PersistentVolume
    metadata:
      name: mlflow-pv
      labels:
        type: local
    spec:
      storageClassName: manual
      capacity:
        storage: 10Gi
      accessModes:
        - ReadWriteOnce
      hostPath:
        path: "/mnt/data"
    ---
    apiVersion: v1
    kind: PersistentVolumeClaim
    metadata:
      name: mlflow-pvc
    spec:
      storageClassName: manual
      accessModes:
        - ReadWriteOnce
      resources:
        requests:
          storage: 10Gi

   Save the file

4. Apply the configuration to your cluster

    $ kubectl apply -f mlflow-pv-pvc.yaml
   

5. Verify the deployed Cluster PVs

    $ kubectl get pv

   Output:

    NAME        CAPACITY   ACCESS MODES   RECLAIM POLICY   STATUS   CLAIM                STORAGECLASS   REASON   AGE
    mlflow-pv   10Gi       RWO            Retain           Bound    default/mlflow-pvc   manual                
   

6. Verify the deployed Cluster PVCs

    $ kubectl get pvc

   Output:

    NAME         STATUS   VOLUME      CAPACITY   ACCESS MODES   STORAGECLASS   AGE
    mlflow-pvc   Bound    mlflow-pv   10Gi       RWO            manual         24s

Install MLflow

1. Using Helm, add community-charts to your repositories

    $ helm repo add community-charts https://community-charts.github.io/helm-charts
   

2. Update the Helm repository index

    $ helm repo update
   

3. Install MLflow. Replace vultr with your desired MLflow label

    $ helm install vultr community-charts/mlflow
   

4. View the cluster deployment to verify that MLflow is ready and available

    $ kubectl get deployments

   Your output should look like the one below:

    NAME           READY   UP-TO-DATE   AVAILABLE   AGE
    vultr-mlflow    1/1     1            1           19h

Expose the MLflow Forwarding Service For External Access

To access your MLflow deployment over the Internet, set up a forwarding service to expose the application for external access as described in the steps below.

1. Create a new file mlflow-service.yaml

    $ touch mlflow-service.yaml
   

2. Edit the file

    $ nano mlflow-service.yaml
   

3. Add the following contents to the file

    apiVersion: v1
    kind: Service
    metadata:
      name: mlflow-service
    spec:
      selector:
        app: mlflow
      ports:
        - protocol: TCP
          port: 80
          targetPort: 5000
      type: LoadBalancer
   

4. Apply the service to your cluster

    $ kubectl apply -f mlflow-service.yaml
   

5. Wait for at least 3 minutes, view the cluster services, and verify the MLflow External IP value

     $ kubectl get services

   Your output should look like the one below:

    NAME             TYPE           CLUSTER-IP      EXTERNAL-IP   PORT(S)        AGE
    kubernetes       ClusterIP      10.96.0.1       <none>        443/TCP        40m
    mlflow-service   LoadBalancer   10.110.234.98   192.0.2.100   80:31545/TCP   64s
    vultr-mlflow     ClusterIP      10.102.93.29    <none>        5000/TCP       101s

Create a sample MLflow Experiment

To test the MLflow functionalities, create a sample linear regression experiment as described in the steps below.

1. Create a new directory Models

    $ mkdir Models
   

2. Switch to the directory

    $ cd Models
   

3. Using pip, install the necessary experiment dependencies

    $ pip install mlflow scikit-learn shap matplotlib

   When using Conda, install the dependencies in your environment:

    $ conda install mlflow scikit-learn shap matplotlib
   

4. Create a new environment variable with your desired experiment name

    $ export MLFLOW_EXPERIMENT_NAME='my-sample-experiment'
   

5. Export a new mlflow_TRACKING_URI with your MLflow service external IP HTTP URL as the value. For example 192.0.2.100

    $ export MLFLOW_TRACKING_URI='http://192.0.2.100'
   

6. Create a new file main.py

    $ touch main.py
   

7. Edit the file

    $ nano main.py
   

8. Add the following contents to the file

    # Import Libraries
   
    import os
   
    import numpy as np
    import shap
    from sklearn.datasets import load_diabetes
    from sklearn.linear_model import LinearRegression
   
    import mlflow
    from mlflow.artifacts import download_artifacts
    from mlflow.tracking import MlflowClient
   
    # Prepare the Training Data
   
    X, y = load_diabetes(return_X_y=True, as_frame=True)
    X = X.iloc[:50, :4]
    y = y.iloc[:50]
   
    # Train a model
   
    model = LinearRegression()
    model.fit(X, y)
   
    # Log an explanation
   
    with mlflow.start_run() as run:
        mlflow.shap.log_explanation(model.predict, X)
   
    # List Artifacts
   
    client = MlflowClient()
    artifact_path = "model_explanations_shap"
    artifacts = [x.path for x in client.list_artifacts(run.info.run_id, artifact_path)]
    print("# artifacts:")
    print(artifacts)
   
    # Load the logged explanation
   
    dst_path = download_artifacts(run_id=run.info.run_id, artifact_path=artifact_path)
    base_values = np.load(os.path.join(dst_path, "base_values.npy"))
    shap_values = np.load(os.path.join(dst_path, "shap_values.npy"))
   
    # Show a Force Plot
   
    shap.force_plot(float(base_values), shap_values[0, :], X.iloc[0, :], matplotlib=True)

   Save and close the file.

The above Python application imports all necessary libraries and uses the diabetes dataset for training data. The dataset contains a total of 442 samples. Then, the model is trained and an explanation is logged to MLflow using SHAP (Shapley Additive Explanations) that plots the output for a visual representation of the data.

Test the MLflow Experiement

1. Run the Python application

    $ python3 main.py
   

2. Using a web browser such as Chrome, visit your MLflow external IP address

    http://192.0.2.100

Verify that the application is included in your MLflow experiments. It uses a linear regression statistical approach to model a relationship between a scalar response and one or more variables.

Conclusion

You have deployed MLflow on a Vultr Kubernetes Engine (VKE) cluster. To test the service operations, implement more examples from the MLflow repository. For more information about MLflow, visit the following documentation resources:

• MLflow documentation
• MLflow Recipes
• MLflow Projects

Next Steps

To implement more solutions, visit the following resources:

• Deploy a Machine Learning Model to Production with TensorFlow Serving
• How to Build a Machine Learning Classifier in Python with Scikit Learn
• Introduction to Vultr Cloud GPUs
• Fine Tune a Hugging Face Diffuser Model on Vultr Cloud GPU