MLOps Meetup Demo

This repo demonstrates how to take the basic model training code in src/mlops_demo/train_basic.py and add MLOps tools to it to track experiments and package the resulting model into a deployable docker container. This is done using MLFlow and BentoML.

Open this demo in your browser

1. Open https://github.com/radix-ai/mlops-meetup-demo in your browser.
2. Click on Code and select Create codespace to start a Dev Container with GitHub Codespaces.

Tasks

Start a MLFlow Tracking server

To track runs with MLFlow we need to run an MLFLow tracking server. The following command runs a local server.

mlflow server \
    --backend-store-uri sqlite:///mlflow.db \
    --default-artifact-root file:///app/mlflow/ \
    --host 127.0.0.1

Train a new model

With the MLFlow server running run the following.

poetry run python src/mlops_demo/train.py

Register latest model in BentoML

Register th latest model from MLFlow in BentoML so that it can be used by the BentoML service.

python /app/src/mlops_demo/register_model.py

Run the model in a local BentoML Service

We can run the service we defined in src/mlops_demo/service.py and the iris_model we registered in train.py as follows.

bentoml serve src.mlops_demo.service:svc --reload --port=8000

Build a deployment ready docker image with BentoML

This uses the service defined in src/mlops_demo/service.py and the iris_model we registered in train.py to build a deployable docker image. Additional config is defined in bentofile.yaml.

1. Build a bento.

bentoml build

2. Turn it into docker container.

bentoml containerize iris_service:latest