A tool to deploy a mostly serverless MLflow on a GCP project with one command

• 1. one-click-mlflow
  • 1.1. How to use
    • 1.1.1. Pre-requesites
    • 1.1.2. Deploying
    • 1.1.3. What it does
    • 1.1.4. Other available make commands
    • 1.1.5. Pushing logs and artifacts

• A GCP project on which you are owner
• Terraform, make, and jq installed
• Initialized gcloud SDK with your owner account

Clone the repo

Run make one-click-mlflow and let the wizard guide you.

If you want to see the innards, you can run it in debug mode: DEBUG=true make one-click-mlflow

• Enables the necessary services
• Builds and deploys the MLFlow docker image
• Creates a private IP CloudSQL (MySQL) database for the tracking server
• Creates an AppEngine Flex on the default service for the web UI, secured by IAP
• Manages all the network magic
• Creates the mlflow-log-pusher service account

• make deploy: builds and pushes the application image and (re)deploys the infrastructure
• make docker: builds and pushes the application image
• make apply: (re)deploys the infrastructure
• make destroy: destroys the infrastructure. Will not delete the OAuth consent screen, and the app engine application.

Once the deployment successful, you can start pushing to your MLFlow instance.

cd examples
python3 -m venv venv 
source venv/bin/activate
pip install -r requirements.txt
python track_experiment.py

You can than adapt examples/track_experiment.py and examples/mlflow_config.py to suit your application's needs.