Vultr DocsLatest Articles

How to Deploy MLFlow on the Vultr Kubernetes Engine

MLFlowMachine LearningKubernetes
DL
Updated on July 25, 2024
How to Deploy MLFlow on the Vultr Kubernetes Engine header image

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 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.

MLflow UI showing the diabetes linear regression experiment

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:

Next Steps

To implement more solutions, visit the following resources: