Mlflow MLserver Docker

• MLflow
• mlserver

This is a utility package and CLI for building mlserver docker images based on MLflow models on a remote tracking server.

Currently, the S3 backend is the only officially supported backend, but any backend should work just fine as long as you have all the required dependencies installed. If you're interested in other backends let me know.

Installation

I recommend using pipx, but you can use any method you like such as poetry or plain pip.

pipx install mlflow-mlserver-docker

Usage

Configure access to the mlflow tracking server and the S3 backend using environment variables, following their respective documentation.

mlflow-mlserver-docker build runs:/fg8934ug54eg9hrdegu904/model --tag myimage:mytag

Any mlflow artifact URL should work fine, as long as the model uses the MLflow packaging format and contains conda.yaml. I have only tested it with scikit-learn models logged with mlflow.autolog().

Development

poetry install

Make sure docker is running.

Run image locally

docker run -it -p 8080:8080 myimage:mytag