Machine Learning Service for Real-Time Predictions

This is an example of Machine Learning REST API application for online prediction. It covers topics such as:

• How to cache model, and when it might be useful to use asynchronous code;
• How to organize a pipeline updates without redeployment;
• How to improve the health check endpoint for a Machine Learning service;
• How to split and organize production and local development;

You can find the full article on Medium.

Installation

• Run make init;
• Edit the .env file and set the database credentials;
• Run make up;
• Open in browser 127.0.0.1/health. Server status and database connection should be active.
• You can find the API documentation by link 127.0.0.1/docs.

The project structure:

|--docker
|--src
    |--ml_pipelines
    |--models
    |--scripts
        |--train.py
    |--sql
        |--migrations
    |--tests
        |--unit_tests
        |--integration_tests
    |--utils
|--server.py

• The server.py file is entry point for our REST API service.
• We will save the ML artifacts in the ml_pipelines folder.
• The models folder is for storage Pydantic schemas. It defines the properties and types to validate some data.
• The script scripts/train.py trains the ML model.

How to use

• make restart will stop running containers, remove them and start application.
• make bash will create a new bash session in the container.

Model training

Before you can start making any predictions, you need to train the Machine Learning model.

• Run make up and start application;
• In a new tab in the project folder run: make bash;
• Now you are inside the Docker container and can trigger the script to train the new model: python scripts/training.py;
• The ML pipeline is ready. Open the documentation 127.0.0.1/docs and try to make a prediction.

Database

• You can find training history of the model in the ml_pipeline table;
• All predictions are stored in the predictions table;