Deploying a Machine Learning Model with Flask, Gunicorn, and Kubernetes on AWS EKS

This project demonstrates how to deploy a machine learning model using Flask and Gunicorn as a web service endpoint, and then serve it using Kubernetes specifically in AWS EKS. The production-ready model is fetched from Neptune.AI model registry and served via Flask application (app.py). The endpoint is tested using the requests library from Python.

Prerequisites

• Python 3.x installed on your machine
• Docker installed on your machine
• Access to AWS EKS (Amazon Elastic Kubernetes Service) cluster
• Neptune.AI account with the model tagged as 'PRODUCTION' in the model registry

Setup

1. Clone this repository:

git clone git@github.com:HichamBenbriqa/ML_Kubernetes_EKS.git
cd your-project

Install the required Python packages:

poetry install

Set up your Neptune.AI credentials by exporting them as environment variables:

export NEPTUNE_API_TOKEN='your-neptune-api-token'
export NEPTUNE_PROJECT_NAME='your-neptune-project-name'

Run the Flask application to serve the model:

python app.py

Testing

Ensure that the Flask application is running.
Run the provided test script to test the endpoint:

python test_endpoint.py

Containerization and Deployment

Dockerize the Flask application:

docker build -t your-docker-image-name .
docker run -p 8001:8001 -it image-id

Push the Docker image to Docker Hub or AWS ECR:

# For Docker Hub
docker tag your-docker-image-name your-dockerhub-username/your-docker-image-name
docker push your-dockerhub-username/your-docker-image-name

# For AWS ECR
aws ecr get-login-password --region your-region | docker login --username AWS --password-stdin your-aws-account-id.dkr.ecr.your-region.amazonaws.com
docker tag your-docker-image-name your-aws-account-id.dkr.ecr.your-region.amazonaws.com/your-docker-image-name
docker push your-aws-account-id.dkr.ecr.your-region.amazonaws.com/your-docker-image-name

Deploy the Docker image to AWS EKS:

# Create Kubernetes deployment
kubectl apply -f deployment.yaml

# Expose deployment as a service
kubectl expose deployment your-deployment-name --type=LoadBalancer --port=8000 --target-port=8000

# Get the external IP address
kubectl get svc