This is a project to deploy MLFlow on AWS in a microservices architecture using ECS FARGATE and AURORA serverless DB. If you want to deploy MLFLow to a EC2 instance, check out the ec2 branch.

• AWS IAM User for the infrastructure deployment, with admin permissions
• awscli, intall running pip install awscli
• terraform >= 0.12
• setup your IAM User credentials inside ~/.aws/credentials

This is what you are going to build.

First, you have to rename config.tf.example to config.tf and fulfill it with your real variables' values.

Use terraform to build the infrastructure:

cd infrastructure
terraform init
terraform plan
terraform apply

By default, the infrastructure is deployed in us-east-1.

When the infrastructure is created, you will have build the docker image for MLFlow and push it to the created ECR repository. First you have to set up these environment variables:

export IMAGE=mlflow-dev
export AWS_ACCOUNT=<YOUR_ACCOUNT_NUMBER>
export AWS_ACCESS=<YOUR_AWS_ACCESS_KEY_ID>
export AWS_SECRET=<YOUR_AWS_SECRET_ACCESS_KEY>
export AWS_DEFAULT_REGION=us-east-1
export DB_PASS=<YOUR_MASTER_DB_PASSWORD>
export BUCKET_NAME=<YOUR_BUCKET_NAME>
export DB_ENDPOINT=<YOUR_DB_ENDPOINT>:3306

or you can set them inside .bashrc or in any manner you like.

Then, build the image and push it to ECR.

sh build_and_push.sh

After that, you can check your App Load Balancer DNS to access MLFlow's UI. Go to your browser and type http://<YOUR_ALB_PUBLIC_DNS>.

To set mlflow experiments and runs to this tracking server, in your python code do

mlflow.set_tracking_uri("http://<YOUR_ALB_PUBLIC_DNS>")

To start a new experiment or retrieve an old one, do

if mlflow.get_experiment_by_name("EXPERIMENT_NAME") is None:
    mlflow.create_experiment("EXPERIMENT_NAME")
mlflow.set_experiment("EXPERIMENT_NAME")

That's it! Now you have an MLFlow tracking server running on AWS. If you need more computing power, you can easily configure a auto-scaling group within the service.

If you want to clean up everything you created, just type:

terraform destroy

and terraform will do it's work!

This is a work in progress so there are a few things we would still like to improve:

• Put Aurora in a private subnet and configure routing.