MLops Project: Smoke Detection Model

This project is created to learn best practices of MLops and submitted as my project in DataTalksClub's MLops Zoomcamp.

https://github.com/DataTalksClub/mlops-zoomcamp

Project Background and Overview

The goal of this project is to build an end-to-end MLops to train, store, register, deploy, serve, and monitor ML model. The problem that this project aiming to solve is to create a pipeline of smoke detection model (binary classification), using public data shared on Kaggle.

The solution overview:

• Cloud resource provision using IaC tool (Terraform)
• ML experiment tracking & model registry using MLflow
• Orchestration of ML pipeline using Prefect
• ML deployment using Flask (real-time inference)
• ML monitoring (real-time) using Evidently, Grafana, and Prometheus

Getting Started

Clone this repository and follow the guide below.

S3 Provision with Terrform

---

First, create a new S3 bucket using Terraform (IaC tool). Open 01_terraform/main.tf and change the bucket name:

resource "aws_s3_bucket" "b" {
  bucket = "{YOUR_BUCKET_NAME}"

  tags = {
    Name = "MLops bucket"
  }
}

Create the bucket by executing

> cd 01_terraform
> terraform init
> terraform plan
> terraform apply
> cd ..

Run MLflow tracking server

---

First, run the MLflow webserver and database by running

docker-compose --profile mlflow up

MLflow webserver is ran inside a container that exposes port 5000 with the host machine. To access the web UI, open web browser and navigate to http://localhost:5000.

Run Prefect (Orchestrator)

---

To run Prefect, first run the webserver (Prefect Orion) and Database by running

docker-compose --profile orion up

The server can be accessed on http://localhost:4200

You can run a flow from another container that is connected to the Orion API by running Prefect CLI container

docker-compose run prefect_cli

To run the Prefect Flow, run python model_training.py from inside the prefect_cli container. The command will execute the Flow to train several models (RandomForest, LightGBM, XGBoost, and CatBoost), and do hyperparameter tuning for XGBoost. The model artifacts are stored in S3 bucket.

Finally, run Prefect Agent to execute tasks in queue that already being scheduled in Prefect deployments by running

docker-compose --profile agent up

Model Registry

---

Open MLflow server and choose the model to deploy to production.

Select the best performing model from MLflow Tracking Server UI, and register the model to MLflow Model Registry.

After registering the model, select one best performing model and change the state to production.

Model Deployment & Monitoring

---

Deploy the production-ready model with real-time inference by running it inside a Flask application, that is ran inside a Docker container. At the same time, run monitoring services as well (MongoDB, Prometheus, and Grafana).

docker-compose --profile monitoring up

Test the service by executing prediction_test.py

To monitor the data, access it on Grafana http://localhost:3000

Best Practices

• To run each container as background service, add -d flag behind the command to daemonize it. This will keep your terminal clean and keep you uninterrupted.
• Install pre-commit to automatically format (with black) and do some checks when committing.
• CI test during push to remote with GitHub workflow

References

• Reference for this docker-compose architecture & code (MLflow part): https://towardsdatascience.com/deploy-mlflow-with-docker-compose-8059f16b6039
• Reference for the prefect docker-compose: https://github.com/rpeden/prefect-docker-compose
• Reference for the Prefect orchestrator: https://github.com/hakymulla/Diabetes-ML-ps
• Reference for Monitoring architecture: https://github.com/AntonisCSt/Mlops_project_semicon/blob/main/evidently_service/Dockerfile