This short post follows-on from Moez Ali's tutorial on how to train a model with PyCaret and serve predictions using FastAPI. We're going to take this one step further and add some MLOps magic by deploying the FastAPI prediction service to a Kubernetes cluster using Bodywork.
Bodywork is a Python package that exposes a CLI for configuring Kubernetes to orchestrate your ML pipelines and deploy your prediction services, without having to build Docker images or learn how to configure Kubernetes.
In the original post Moez Ali guides us through the process of training and saving a model using PyCaret and then demonstrates how to use FastAPI to develop a Python module that serves predictions from the model. If you work through the post you should end-up with a project directory that looks something like this,
|-- pycaret-mlops/
|-- diamond-pipeline.pkl
|-- serve_predictions.py
Create a GitHub repo and commit all of these file to it. We've worked through this for you and you can find our project repo at https://github.com/AlexIoannides/pycaret-mlops, where we've reproduced all of the training code in the train_model.ipynb notebook, to make life simpler.
Bodywork is a Python package that exposes a CLI for configuring Kubernetes to orchestrate your ML pipelines and deploy your prediction services. Install it using Pip,
$ pip install bodywork
Create a file called bodywork.yaml and add the following,
version: "1.1"
pipeline:
name: pycaret-diamond-prices
docker_image: bodyworkml/bodywork-core:3.0
DAG: serve-predictions
stages:
serve-predictions:
executable_module_path: serve_predictions.py
requirements:
- fastapi==0.68.1
- uvicorn==0.15.0
- pycaret==2.3.3
cpu_request: 1.0
memory_request_mb: 500
service:
max_startup_time_seconds: 600
replicas: 1
port: 8000
ingress: true
logging:
log_level: INFOThis requests that two containers (or replicas) running the FastAPI prediction service are created, with automatic load-balancing between them - nice! For more info on what this all these configuration parameters mean, head to the Bodywork docs.
Commit this file to your Git repo and push to GitHub,
$ git add bodywork.yaml
$ git commit -m "Add bodywork deployment config"
$ git push origin master
If you don't have a Kubernetes cluster handy, then download Minikube so you can test locally. Then, start and configure a cluster as follows,
$ minikube start --kubernetes-version=v1.22.6 --addons=ingress --cpus=2 --memory=4g
$ bodywork create deployment https://github.com/AlexIoannides/pycaret-mlops.git
This will stream the logs to your terminal so you can keep track of progress.
First create a route to your cluster,
$ minikube kubectl -- -n ingress-nginx port-forward service/ingress-nginx-controller 8080:80
And then use the test_prediction_service.ipynb notebook to test the service using Python.
