ml-ops

The project's goal is to show an User Interface (UI) that interactively displays the results of a machine learning (ML) model hosted on a server. The application allows users to analyze new data and provides easy access to the model's predictions.

Key functionalities of the server:

MLflow Integration: The server can load stored models from an MLflow server.

RabbitMQ: RabbitMQ is used as a message broker for asynchronous communication, enabling scalability and efficient background process management. The collaboration between the UI and the server supports both interactive visualization and real-time data processing.

This solution can be used as a base project for applications requiring effective data management, and real-time analytics.

File structure

ml-ops/
├── app/                      # Application layer, including UI and server logic
│   ├── __init__.py
│   ├── ui.py                 # User Interface logic
│   └── server.py             # Server-side logic for handling predictions and routing
│
├── models/                   # Model-related logic and operations
│   ├── load_model.py         # Directory for handling loading model from mlflow
│   └── model.py              # Training, saving, and processing models to mlflow
│
├── messaging/                # Messaging and communication logic (e.g., RabbitMQ-related code)
│   ├── __init__.py
│   └── commands.py           # Builds up RabbitMQ channels and connection
│
├── .gitignore                # Specifies files and directories to ignore in version control
├── README.md                 # Documentation about the project, including setup and usage instructions
└── requirements.txt          # Lists the Python dependencies required for the project

Usage

Requirements

Install requirement packages from the requirements.txt file

pip install -r requirements.txt

MLflow

Start a local server that tracks our model activity

mlflow server

The mlruns folder will appear in our directory, where the results of different runs will be stored. The User Interface for MLflow can typically be accessed via the following link:

http://127.0.0.1:5000.

Model

Create a model for Iris dataset and save it MLflow. Use the following command in the ml-ops/models folder:

python model.py

The code created a model and stored the model parameters in the artifacts folder of MLflow. You can get the run id in the Overview folder.

Server

In the ml-ops/app folder start the server with the help of the following command:

python server.py

Now the server is running, end point are ready for handling incoming messages.

The endpoint documentation can be reached at the following local link:

http://localhost:8000/docs

The server also initializes the rabbitMQ connections and adjust the channels for communications.

Get MLflow model endpoint: /model/{runId}

It loads the model data for the given run by the run ID.

Predict endpoint: /model/{queue}

It predicts the new results for the incoming new input and it uses the previously loaded model for that. It also calculates the metrics for the given input.

User Interface

I used the streamlite package for the creation of the user interface for handling new model input and showing results.

In the ml-ops/app folder run the following command.

streamlit run ui.py

1. First load the model by your run Id at the UI.

2. Load the iris_random_60_samples.csv file for test.

After the load succeeded, you can see the result metrics and other analytics.