Serverless Machine Learning API

The goal of the project is to familiarize myself with deploying statistical models into the cloud. The goal is not to build/create the most interesting statistical model that solves an interesting problem.

Architecture

Steps to build & test "ML Prediction Server" (locally)

1. Run train_model.py - $ python train_model.py.

• Trains & saves a model.
• Saves a test data set.

2. Run $ docker build --tag ml-server:latest ..

• Build a docker image using the model build in #1 and app.py.

3. Run $ docker run --publish <host_port>:<container_port> ml-server, which in this case, is: $ docker run --publish 80:80 ml-server

• Runs the container, mapping port 80 of the container to port 80 of the host.

4. (In a separate shell) Run $ python test_api.py 127.0.0.1 from the command line.

• Loads the test data set from #1, and sends a sample to the Docker Flask API. Should output the class predictions for the requested instances.

Steps to move "ML Prediction Server" to AWS Cloud

1. Complete all steps in "Steps to recreate/test (locally)" to ensure everything works properly.
2. Change directories to aws_setup $ cd aws_setup.
3. Run setup_aws_resources.sh $ bash setup_aws_resources.sh.

Steps to test if the service is working

1. Run $ python test_api.py <instance_public_ip_address>. The accuracy of the test set will be printed.

TODO

• Add a NAT Gateway so the ALB can the be only component in a public subnet (with the Fargate task in private subnets)