Extremely lightweight development environment for a machine-learning prototype application.

The intend is to boost your data science dept. so that they can deliver an AI prototype in 3 weeks.

The app is organized into 4 micro-services:

• React Front-end
• Fast API Python / Fast API back-end
• PostgreSQL SQL database
• MLFLow Model engine

The front-end connects to the API by making HTTP requests for desired data. The API connects to the SQL database and the model engine. The services are containerized and managed with Docker Compose.

Download Docker desktop and follow its instructions to install it. This allows us to start using Docker containers.

Create a local copy of this repository and run:

docker-compose build

This spins up Compose and builds a local development environment according to our specifications in docker-compose.yml. Keep in mind that this file contains settings for development, and not production.

After the containers have been built (this may take a few minutes), run:

docker-compose up

Test the app by opening this from a browser:

http://localhost:3000/ 

The fruits listed are provided by the API at this endpoint:

http://localhost:4000/latest/fruits

See Open API documentation of the API here:

http://localhost:4000/latest/docs

In turn, the API issues SQL requests to PostgreSQL at:

http://localhost:5432

The "Predict" button will request a prediction from the API at this endpoint:

http://localhost:4000/latest/predict

In turn, the API will forward the request to the model engine at:

http://localhost:5000/invocations

A "Hello World model" has been added to models/model. It will serve all prediction requests by returning "Hello world".

To do something more interesting, download your own model files into models/model and rebuild.

To download a model served by Databricks, run the following function from a notebook:

%python
import mlflow
from mlflow.store.artifact.models_artifact_repo import ModelsArtifactRepository

def getModel(model_name, model_stage = "Staging", dest_path = "models"):
local_path = f'file:/databricks/driver/{dest_path}'
mlflow.set_tracking_uri("databricks")

dbutils.fs.mkdirs(local_path)
ModelsArtifactRepository(
    f'models:/{model_name}/{model_stage}').download_artifacts("", dst_path = dest_path)

dbutils.fs.cp("file:/databricks/driver/models/", "/FileStore/{local_path}/", True)
print(f'{model_stage} Model {model_name} can be downloaded like this:')
print(f'databricks fs cp dbfs:/FileStore/{dest_path}/ . --recursive --overwrite')
l = dbutils.fs.ls(f'/FileStore/{dest_path}/')
return l

You may need to install mlflow on the Databricks cluster:

%pip install mlflow

After that, you can get the model files by running this:

cd models/model
databricks fs cp dbfs:/FileStore/models/ . --recursive --overwrite

You may need to install the Databricks CLI.

models/model folder should contain files like these (depending on the model):

conda.yaml
MLmodel
python_model.pkl

Once you have replaced the model files, you may have to change the format of the input data.

• Create a Kubernetes deployment scripts to build and deploy to AWS AKS
• Expose client and api through HTTPS, set up routing and SSL certs in terraform

Feel free to use the code in this repository however you wish. Details are provided in LICENSE.md.