This repository contains the necessary files and deployments scripts to deploy REMLA-app in a Kubernetes/Minikube setting and perform automated blue-green deployment.

REMLA-app is a machine learning application that predicts tags for titles of StackOverflow posts using multilabel classification. The application provides a simple web interface where users can provide a title and get the model's predicted tags back. Users can then provide feedback on the quality of the predicted tags.

The application is based on the jupyter notebook provided by luiscruz/remla-baseline-project.

The REMLA-app can be deployed in two ways: singular deployment or blue-green deployment.

• Helm v3.9
• Minikube v1.25

To deploy only one instance ("blue") of REMLA-app using Minikube:

• Run ./deploy-simple.sh
• Access the web interface at <ingress_ip>/

To deploy two instances ("blue" and "green") of REMLA-app using Minikube:

• First perform singular deployment.
• Run ./deploy-bg.sh green <version>
• Install the alert_actor, e.g. run pip install -e alert_actor from the project root
• Run alert_actor from the project root
• Access the web interface at <ingress_ip>/

The automated blue-green deployment will switch traffic to the green deployment. Once one of the metrics used in k8s/alert_rules.yml drops below their specified thresholds, a Prometheus alert will fire which will trigger the alert_actor to switch traffic back to the old version.

Descriptions of the scripts' usage is also found by running ./<script_name> --help.

Deploys a singular deployment. This script will set up a Minikube cluster and deploy only the blue version of REMLA-app. The image version used by blue is kept up-to-date in this repository with the latest released image version.

Note: When running the script for the first time, you will be prompted to set a password for the Grafana dashboard.

Deploys the given color deployment with the given version. The script will wait until the deployment is successful, then switch traffic to the newly deployed version.

Switches traffic to the given color deployment. This script is used by the alert_actor.

Undeploys/deletes the deployment of the given deployment; used to free up resources when you no longer need a stand-by version.

Listens for specific alerts on localhost:8081/webhook. When it receives an alert labeled rollback="true", it switches traffic to the color deployment that is on stand-by (i.e. not currently running in production). The alert_actor needs access to the Kubernetes manifest containing the service used to access REMLA-app. The given default path is k8s/services.yml but a different file path can also be given by running alert_actor <path_to_service_yml>.

• On MacOS: run minikube tunnel. The <ingress_ip> is then localhost.
• Otherwise: run minikube service list. The <ingress_ip> can be found in the “URL” column of ingress-nginx-controller.

You will need the following:

• Scripts: deploy-bg.sh and switch-traffic.sh. Modify the variables at the top of the scripts to match your environment. The other scripts may also be used for convenience.
• The alert_actor/ directory to enable automatic alert-based switching between blue-green.
• The k8s/ directory. Modify the manifests to point to your own application images.
• Implement your own Prometheus alert(s). Modify k8s/alert_rules.yml to match your desired alert(s) configuration and add rollback: "true" to the labels of all alerts that should initiate rollbacks.

The Grafana dashboard is accessible on <ingress_ip>/dashboard.

We provide a custom dashboard that displays our application-specific metrics (user_satisfaction, num_pred) which is stored in k8s/grafana-dashboards/metrics_dashboard.json.

Log in to the Grafana dashboard with username admin and the password you set when running deploy-simple.sh for the first time.

To import the dashboard go to Create -> Import and copy-paste the json in k8s/grafana-dashboards/metrics_dashboard.json.

To prepare your local environment for development, install all requirements using pip install -r requirements.txt. We make use of the conventional commits specification for our commit messages, checkout Conventional Commits for format. We've got checks on CI that enforce this specification. To help you, we've defined a git hook that runs these checks after every commit. You can configure it by running pre-commit install.