A simple docker-compose app to easily deploy an mlflow server, together with an nginx reverse proxy for authentication.
Step-by-step instructions on how to deploy can be found at the bottom of this README.
If you are deploying to AWS, check out the AWSREADME file.
The file docker-compose.yml tells docker what to do.
If you open it, you will see two "services" specified: mlflowui and nginxauth.
The docker-compose.yml is moreover telling docker to look into the mlflowui and nginxauth subdirectories (of this repo) to see the precise instructions to follow (which are contained in a file called Dockerfile).
Here are a few more details about these (for people familiar with docker).
The first service is mlflowui, the mlflow server.
The corresponding dockerfile is pretty straightforward: use a python3 base image and install the necessary packages.
In docker-compose.yml you'll find the entrypoint under command.
This can be customized depending on your needs.
The simplest way to use this is by
command: mlflow server --host 0.0.0.0
this will keep all artifacts and parameters (aka the mlflow backend) inside the docker container.
If you need these to persist in case the containers go down, you should use a volume.
A more robust way to deal with artifacts and the backend is to have them stored elsewhere (for example in an S3 bucket and a SQL server, respectively).
To do this, you need to populate the .mlfenv file accordingly (see below).
The second service is nginxauth, an nginx reverse proxy.
This is how we force users to provide credentials.
To use this service, we first need to create the .htpasswd file (see instructions below).
To deploy locally, follow the steps below.
Pick a username, here we use incredibleuser, and run:
htpasswd -c nginxauth/.htpasswd incredibleuser
this will prompt you to enter a password.
If you don't care about storing artifacts and parameters elsewhere, then swap the line beginning with command in the docker-compose.yml with command: mlflow server --host 0.0.0.0, and ignore the rest of this section.
If you want to store experiment parameters (the mlflow backend) in a separate SQL server, first populate the .mlfenv file.
Experiment metadata, such as parameters and metrics, are recorded in the BACKEND_STORE_URI.
This can either be a local path, or a sql backend.
A database URI looks like this: dialect://USER:PASS@HOST:PORT/NAME.
When using a custom backend, mlflow requires a location for artifacts.
Populate the .mlfenv file with the address.
The default provided will just store artifacts inside the mlflow container (in this case restarting the container will delete the artifacts).
Once you've populated the .mlfenv file with backend and artifacts, you're ready to go!
docker-compose up --build -d
To kill
docker-compose down --remove-orphans