terraform-google-mlflow

A terraform module to deploy a mlflow server with nginx basic authentication on Google Cloud. This module install a development MLflow for quick experiment.

• It uses a local sqlite for the database.
• There is only one hardcoded user account, accessed through basic authentication.
• There is no SSL.

Prerequisites

To create ressources on Google Cloud, terraform needs:

• A Google Cloud Platform project
• A service accound with the permissions needed to create ressources in the project and its json key file
• A working terraform >= 0.13

Deploying MLflow with terraform on Google Cloud

1. In an empty directory, copy the service account json key file.

2. Create a file with a .tf extention and add the module configuration with your own values:

   module "mlflow" {
     source                = "github.com/benqua/terraform-google-mlflow"
     project               = "project-name"
     cred_file             = "project-name-service-accound-key-file-XXXNNNXXXNNX.json"
     service_account_email = "[email protected]"
     ssh_user              = "your ssh username"
     ssh_pub_key_file      = "~/.ssh/id_rsa.pub"
     zone                  = "europe-west6-b"
     machine_type          = "n1-standard-1"
     mlflow_password       = "PasswordT0AccessMLflow"
   }
   
   output "mlflow_ip" {
     value       = module.mlflow.cip
     description = "IP to access mlflow server (HTTP or SSH)"
   }
   

   Check other variables and their default values.

3. Run terraform init

4. Run terraform plan and check the output

5. Run terraform apply and answer yes

6. Wait. Once the ressources are created, terraform will output mlflow_ip, the IP to access the MLFlow server.

Note that:

• Once the ressources are created, it takes a few minutes before the MLflow server is available. This time is needed to configure the instance, download and run MLflow and the Nginx proxy (for basic auth).
• The MLflow server is accessible through HTTP (and not HTTPS).

Warning

A terraform destroy will destroy the mlflow server and the disck containing the metrics and the S3 bucket containing the models.