ts-azure-services / local-mlflow-aml-endpoint

A repo to showcase a local dev workflow to build an MLFlow model and then load it for inferencing into Azure ML.

local-mlflow-aml-endpoint

This repo demonstrates a workflow to develop a local ML model using mlflow and then leverage Azure ML to setup a real-time inference endpoint. To use this workflow, ensure that you have the latest Azure ML CLI v2 installed (refer this link).

Steps

• Create a local virtual environment and install needed libraries using the make install command.
• Create the Azure ML environment using the make infra command:
  • Ensure you have a sub.env file with your subscription listed as a single line command as SUB_ID=<your subscription>.
• Then, start a local mlflow backend server, and in a separate terminal, trigger a local run, using make local_run.
• Once you have the local model artifacts, run aml_deploy. This will do a few things:
  • Consolidate all the model artifacts in a model directory.
  • Overwrite the conda.yaml file with a more extensive definition.
  • Trigger a number of Azure CLI commands to register the model in AML and trigger a managed real-time deployment. (This can also be manually done through the Portal as detailed here.
• Once the endpoint is up, the last series of commands - make get_prediction - can help generate some test data from the original dataset which can be manually tested through the Portal 'Test' feature for the managed online endpoint.