DriftWatch: A Microservice-based Drift Detection Framework

DriftWatch is a cloud-native microservices framework designed to explore, implement, and test data and concept drift detection in machine learning systems. It provides a scalable, modular architecture to integrate drift detection techniques into production pipelines, enabling real-time observability and adaptive decision-making.

About DriftWatch

This framework focuses on:

• Data Drift Detection: Monitoring feature distribution changes that could degrade model performance.
• Concept Drift Detection: Identifying shifts in the underlying target variable or relationships.
• Integration with MLOps Pipelines: Seamlessly integrating drift detection in CI/CD workflows.
• Adaptability: Supporting retraining, re-tuning, or model switching strategies.

Architecture

DriftWatch employs a microservices architecture with multiple modular components for real-time data transformation, drift monitoring, and visualization. Each service is designed to work independently, communicating over gRPC for high performance and scalability.

Microservices Overview

Service	Language	Description
frontend	Go	Exposes an HTTP server to serve the website. Does not require signup/login and generates session IDs for all users automatically.
cartservice	C#	Stores the items in the user's shopping cart in Redis and retrieves it.
productcatalogservice	Go	Provides the list of products from a JSON file and ability to search products and get individual products.
currencyservice	Node.js	Converts one money amount to another currency. Uses real values fetched from European Central Bank. It's the highest QPS service.
paymentservice	Node.js	Charges the given credit card info (mock) with the given amount and returns a transaction ID.
shippingservice	Go	Gives shipping cost estimates based on the shopping cart. Ships items to the given address (mock)
emailservice	Python	Sends users an order confirmation email (mock).
checkoutservice	Go	Retrieves user cart, prepares order and orchestrates the payment, shipping and the email notification.
recommendationservice	Python	Recommends other products based on what's given in the cart.
adservice	Java	Provides text ads based on given context words.
loadgenerator	Python/Locust	Continuously sends requests imitating realistic user shopping flows to the frontend.

Key Features

1. Pluggable Drift Detection Algorithms:

   • Statistical Process Control (SPC)
   • Ensemble-based methods
   • Adaptive sliding window techniques
   • Concept-specific approaches

2. Visualization and Reporting:

   • Drift heatmaps
   • Real-time dashboards

3. Extensible Framework:

   • Add custom algorithms with minimal configuration.
   • Integrates with CI/CD pipelines via Kubernetes or Terraform.

---

Deployment Options

DriftWatch supports both local and cloud deployment to provide flexibility for development and production environments.

Quickstart with Skaffold for Local Deployment

DriftWatch supports Skaffold to simplify the build and deployment process, reducing dependencies like GDrive setup.

Prerequisites

• Kubernetes cluster (e.g., Minikube, Kind)
• kubectl
• Skaffold installed locally

Build and Deploy Using Skaffold

1. Clone the repository:

   git clone https://github.com/your-org/driftwatch.git
   cd driftwatch

2. Ensure Skaffold is installed and configured:

   sudo snap install skaffold
   skaffold version

3. Build and deploy DriftWatch:

skaffold build

skaffold dev

• The skaffold.yaml configuration ensures that all services are built and deployed in the correct order.
• By default, Skaffold uses your local Docker environment to build images and deploys them to your Kubernetes cluster.

4. Access the frontend service:

   Use kubectl to fetch the external IP of the frontend:

   kubectl get service frontend-external | awk '{print $4}'

   Visit http://<EXTERNAL_IP> in your browser.

Documentation

• Development Guide: Run and develop locally.
• Integration Guide: Add DriftWatch to existing pipelines.

Contributing

We welcome contributions! Please see our CONTRIBUTING.md for guidelines on how to get involved.

Acknowledgment

This project utilizes components and microservices derived from the GoogleCloudPlatform/microservices-demo repository. These resources are used exclusively for academic and experimental purposes to facilitate research and exploration in drift detection methodologies.