MLflow

In the Ultra Transcenders books

AI-300

Each book explains MLflow in context, with comparison tables and the common traps.

Terms in this definition

• Azure Machine Learning

  Azure platform where you train, deploy, monitor and retrain models of your own, practising MLOps through tools such as pipelines, online endpoints, model monitoring and prompt flow.

• API

  Short for application programming interface: a contract that client code calls programmatically, for example a web API secured with tokens or the Files, Images or Responses APIs.

• Workspace

  Teams in Power BI and Microsoft Fabric collaborate in this folder-style container, which groups items such as reports, semantic models and lakehouses, controls who can access them and is assigned a capacity.

Related terms

• Azure Machine Learning SDK v2

  Python SDK currently used with Azure Machine Learning, centred on the MLClient class. It has no logging API, so MLflow is used to log metrics.

• MLflow model

  When a model is stored in MLflow format, with an MLmodel file and conda requirements, you can deploy it to batch or online endpoints with no scoring script or environment of your own.

• MLflow tracking URI

  The location MLflow sends logged data to. Azure Machine Learning jobs configure it for you; outside them you supply it with mlflow.set_tracking_uri.

• mlflow.sklearn.log_model

  Function that saves a scikit-learn model in MLflow format at whatever artifact path you supply, such as "model". That path is used literally and has nothing to do with the experiment's name.

• ONNX

  An open model format designed so a model built with one machine learning framework can run in another. Fabric's batch scoring function supports it as an MLflow flavour.

• PREDICT

  A Fabric function for scoring MLflow models in batches from a notebook, provided the models have signatures; you can call it via MLFlowTransformer, Spark SQL or a PySpark UDF. Fabric Warehouse does not support the T-SQL PREDICT statement.

• runs:/ URI

  MLflow URI of the form runs:/<run-id>/<path> that points to a model stored in a run's default artifact location, allowing it to be registered with lineage. The equivalent Azure ML form is azureml://jobs/<job>/outputs/artifacts/paths/<path>.

• Scoring script

  Script for a deployment, usually score.py: init() loads the model as the container starts and run() handles scoring for each request. MLflow models do not require one.