This is an alpine base mlflow image.

Why

• The official mlflow image is debian based and huge (>2GB)
• The alpine based is smaller (~800MB)
• default entrypoint starts model serving

TODO

[ ] pin mlflow version [ ] use matching tag version [X] default entrypoint should start tracking server

Usage

There are two operation modes of this image:

1. MLflow tracking: Start a tracking server. Note that at this moment this is a sandbox tracking server and it doesn't preserve it's state beyond the lifetime of the container.
2. MLflow model serving: Start a model serving REST API

Starting a tracking server

A tracking server is started by default, simply starting a container: docker run -d -p 5000:5000 ghcr.io/ezmldev/mlflow

Starting a model serving

You want to pass the MLFLOW_SERVICE_TYPE=serving environment variable to start model serving. The MLflow model can be provided from a named volume, like:

docker run -d -e MLFLOW_SERVICE_TYPE=serving -p 5000:5000 -v model:/model ghcr.io/ezmldev/mlflow

Or even from a local dir

docker run -d MLFLOW_SERVICE_TYPE=serving -p 5000:5000 -v $PWD/my-model:/model ghcr.io/ezmldev/mlflow

Testing model serving

curl \
  -H "content-type: application/json" \
  localhost:5000/invocations \
  -d '{"inputs": [["feature1", "feature2"]]}'

You can also specify an alternative model location by the env var MLFLOW_MODEL_URI.