mlflow-docker-compose

Deploy mlflow with docker-compose

Deploy

1. Login Google Cloud Platform

In this script, mlflow stores artifacts on Google Cloud Storage.
It means you must set up GCP Credentials.
If you already have application_default_credentials.json, go next chapter.

$ gcloud auth application-default login

application_default_credentials.json will be saved ${HOME}/.config/gcloud/

2. Create .env file

In docker-composa.yaml, some parameters is loaded from .env file.
Set following parameters in .env.

• HOST: host name(If you don't use domain, any name is accepted. If use, speciy it)
• POSTGRES_USER: postgresql db user
• POSTGRES_PASSWORD: postgresql db user password
• GCP_STORAGE_BUCKET: Google Cloud Storage bucket name mlflow will store artifact
• CREDENTIALS_PATH: Path to application_default_credentials.json
• GCLOUD_PROJECT: GCP Project name you use

HOST=mlflow.dev
POSTGRES_USER=demo-user
POSTGRES_PASSWORD=demo-password
GCP_STORAGE_BUCKET=demo-bucket
CREDENTIALS_PATH=~/.config/gcloud/application_default_credentials.json
GCLOUD_PROJECT=demo-project

3. Set up NGINX Basic Authentication

Because mlflow doesn't provide authentication, use NGINX proxy for basic authentication system.

$ sudo echo "{USER_NAME}:$(openssl passwd -apr1 {PASSWORD})" >> ${HOST}

${HOST} is host name you set in chapter 2.

4. Build and deploy

Build mlflow Dockerfilw, and then deploy applications.

$ sudo docker-compose build
$ sudo docker-compose up -d

Client

To use Basic authentication, mlflow use following parameters passing HTTP authentication.
Set following environment parameters in local, same as 3. Set up NGINX Basic Authentication

• MLFLOW_TRACKING_USERNAME
• MLFLOW_TRACKING_PASSWORD

See also https://www.mlflow.org/docs/latest/tracking.html#logging-to-a-tracking-server

Update MLflow version

If you want update MLflow, stop container and remove images, and then rebuild MLflow container.

$ sudo docker-compose stop mlflow && \
  sudo docker-compose rm mlflow && \
  docker images mlflow-docker-compose_mlflow --format '{{.ID}}'|xargs docker rmi && \
  sudo docker-compose build && \
  sudo docker-compose up -d

On-Premises version

In case you'd like to run MLflow on your on-premises server, on_premises folder is for you. Batteries (sample setting files) included.