ML Problem Steps

1. Design
   • We gather the requirements to see if the ML is a suitable solution for the problem.
2. Train
   • We train the ML model, and part of this process involves experiment tracking to store and, of course, track the possible metrics.
   • We have a training pipeline that produces an ML model.
3. Operate
   • We deploy the ML model:
     1. Batch/Offline
        • Run regularly, as we don’t need the predictions immediately.
     2. Online
        • It’s up and running all the time, as we need the predictions immediately.
        • It has 2 types:
          1. Web Service
          2. Streaming

Deployment

Batch

Overview

• We run the model at regular intervals (Hourly, Daily, Monthly)
• We have our data in the database, and a scoring job runs regularly to apply the ML model to the data.
• We mostly use the batch mode, for example, the marketing tasks (Predict the churn).

Web Service

Overview

• The user is using a specific service/website/app (backend) that needs to generate predictions immediately, and the ML model is always up and running.
• Such as the Ride Duration Service to predict the duration before starting the ride/trip.
• It presents a 1:1 relationship between the client (Backend) and the server (ML Model).

Deploying

1. Install scikit-learn that is compatible with the ML model itself (such as the .bin file created in the first module)

   pipenv install scikit-learn==1.0.2 scipy==1.7.3 numpy==1.21.6 flask --python=3.9
   

2. Start the venv

   pipenv shell
   

3. Create the predict.py and test.py files to deploy the ML model on a Flask web service.

4. Install the gunicorn package for production, as Flask is for developing.

   pipenv install gunicorn