MLflowModel format: MLflow-compatible application deployments

Application deployments (see Deploying Snorkel-built models) can be packaged in the MLflowModel format. The MLflowModel package format is compatible with the standard model format defined by the MLflow project. Deployments packaged in this format will be compatible with all systems that use MLflow models.

Deployment structure

Deployments of the MLflowModel format will be of the pyfunc model flavor (see MLflowPyfunc). As of v0.84.0, these deployments will have the following directory structure:

• workflow_<ID>

  • MLmodel: The model metadata yaml file.
  • conda.yml: The conda environment file that describes the packages that are required to run the model.
  • data/: The directory that contains a description of the application DAG and any required model data (e.g. weights).
  • wheels/snorkelflowmlflow-XXX-py3-none-any.whl: A wheel file that contains the Snorkel proprietary source code that is needed to run the model.

Python environment and dependencies

MLflowModel deployments are only compatible with Python 3.8.

In order to run an application deployment, the packages listed in conda.yml need to be installed. Depending on the application template, operator configuration, model architecture, etc., additional packages are also needed. Please contact your Snorkel representative for more details.