Automatic blue/green deployment using Docker Compose and NGINX.

Introduction

Simulation of the following mechanisms:

1. Event-triggered training, once there are enough data added, automatically start the training job
2. Once the trianing is done, the newest model will be deployed to the resting host
3. After deployment, traffic is directed to the host with newest model and the old one is now resting

Usages

Run system

Run docker-compose up --build -d

Install dependencies

Run pip3 install -r requirements.txt

Correctness test

Run python3 test/test.py [local_model_path] to sending post request to the host and print the response.
The local_model_path is an optional argument. If used, the program will also print inference result of given model.

Load stress test

1. Run locust -f test/stress.py -H http://localhost:8080/
2. Go to http://localhost:8089/ and you will see the GUI of locust load stress test
3. Input your desired number of users and spawn frequency and start the test

Adding data

Run python3 addData.py [number_of_data] to add data (data is defined in config.py) to csv file.
The number_of_data is a required argument, it should be either 1 or 2 representing how many data should be added.
If number of new data exceed BATCHSIZE (defined in config.py), this code will automatically retrain and deploy.

Batch size

In config.py, I have set a variable BATCHSIZE equals to 2, which means when adding two or more data to the csv, the system will retrain the model. It could be set to any positive integer.

System spec

1. Docker 20.10.7, docker-compose 1.29.2
2. Python 3.8.5
3. zsh 5.8

Other details

1. Model is for predicting house price in Boston. Check this competition on Kaggle.
2. For simplicity, basic linear regression model is used.
3. Currently this project has no frontend, use test.py to check result instead.