Track Azure Databricks machine learning experiments with MLflow and Azure Machine Learning

MLflow is an open-source library for managing the life cycle of your machine learning experiments. You can use MLflow to integrate Azure Databricks with Azure Machine Learning to ensure you get the best from both of the products.

In this article, you learn:

Prerequisites

• The azureml-mlflow package, which handles the connectivity with Azure Machine Learning, including authentication.
• An Azure Databricks workspace and cluster.
• An Azure Machine Learning Workspace.

See which access permissions you need to perform your MLflow operations with your workspace.

Example notebooks

The Training models in Azure Databricks and deploying them on Azure Machine Learning repository demonstrates how to train models in Azure Databricks and deploy them in Azure Machine Learning. It also describes how to track the experiments and models with the MLflow instance in Azure Databricks. It describes how to use Azure Machine Learning for deployment.

Install libraries

To install libraries on your cluster:

1. Navigate to the Libraries tab and select Install New.

   

2. In the Package field, type azureml-mlflow and then select Install. Repeat this step as necessary to install other packages to your cluster for your experiment.

   

Track Azure Databricks runs with MLflow

You can configure Azure Databricks to track experiments using MLflow in two ways:

• Track in both Azure Databricks workspace and Azure Machine Learning workspace (dual-tracking)
• Track exclusively on Azure Machine Learning

By default, when you link your Azure Databricks workspace, dual-tracking is configured for you.

Dual-track on Azure Databricks and Azure Machine Learning

Linking your Azure Databricks workspace to your Azure Machine Learning workspace enables you to track your experiment data in the Azure Machine Learning workspace and Azure Databricks workspace at the same time. This configuration is called Dual-tracking.

Dual-tracking in a private link enabled Azure Machine Learning workspace isn't currently supported. Configure exclusive tracking with your Azure Machine Learning workspace instead.

Dual-tracking isn't currently supported in Azure operated by 21Vianet. Configure exclusive tracking with your Azure Machine Learning workspace instead.

To link your Azure Databricks workspace to a new or existing Azure Machine Learning workspace:

1. Sign in to the Azure portal.

2. Navigate to your Azure Databricks workspace Overview page.

3. Select Link Azure Machine Learning workspace.

   

After you link your Azure Databricks workspace with your Azure Machine Learning workspace, MLflow tracking is automatically tracked in the following places:

• The linked Azure Machine Learning workspace.
• Your original Azure Databricks workspace.

You can use then MLflow in Azure Databricks in the same way that you're used to. The following example sets the experiment name as usual in Azure Databricks and start logging some parameters.

import mlflow 

experimentName = "/Users/{user_name}/{experiment_folder}/{experiment_name}" 
mlflow.set_experiment(experimentName) 

with mlflow.start_run():
   mlflow.log_param('epochs', 20)
   pass

Track exclusively on Azure Machine Learning workspace

If you prefer to manage your tracked experiments in a centralized location, you can set MLflow tracking to only track in your Azure Machine Learning workspace. This configuration has the advantage of enabling easier path to deployment using Azure Machine Learning deployment options.

Configure the MLflow tracking URI to point exclusively to Azure Machine Learning, as shown in the following example:

Configure tracking URI

1. Get the tracking URI for your workspace.

   • Azure CLI
   • Python
   • Studio
   • Manually

   APPLIES TO: Azure CLI ml extension v2 (current)

   1. Sign in and configure your workspace.

      az account set --subscription <subscription>
      az configure --defaults workspace=<workspace> group=<resource-group> location=<location> 
      

   2. You can get the tracking URI using the az ml workspace command.

      az ml workspace show --query mlflow_tracking_uri
      

2. Configure the tracking URI.

   • Use MLflow SDK
   • Use environment variables

   The method set_tracking_uri() points the MLflow tracking URI to that URI.

   import mlflow
   
   mlflow.set_tracking_uri(mlflow_tracking_uri)
   

Configure authentication

After you configure tracking, configure how to authenticate to the associated workspace. By default, the Azure Machine Learning plugin for MLflow opens a browser to interactively prompt for credentials. For other ways to configure authentication for MLflow in Azure Machine Learning workspaces, see Configure MLflow for Azure Machine Learning: Configure authentication.

For interactive jobs where there's a user connected to the session, you can rely on Interactive Authentication and hence no further action is required.

For those scenarios where unattended execution is required, you have to configure a service principal to communicate with Azure Machine Learning.

import os

os.environ["AZURE_TENANT_ID"] = "<AZURE_TENANT_ID>"
os.environ["AZURE_CLIENT_ID"] = "<AZURE_CLIENT_ID>"
os.environ["AZURE_CLIENT_SECRET"] = "<AZURE_CLIENT_SECRET>"

export AZURE_TENANT_ID="<AZURE_TENANT_ID>"
export AZURE_CLIENT_ID="<AZURE_CLIENT_ID>"
export AZURE_CLIENT_SECRET="<AZURE_CLIENT_SECRET>"

Name experiment in Azure Machine Learning

When you configure MLflow to exclusively track experiments in Azure Machine Learning workspace, the experiment naming convention has to follow the one used by Azure Machine Learning. In Azure Databricks, experiments are named with the path to where the experiment is saved, for instance /Users/alice@contoso.com/iris-classifier. However, in Azure Machine Learning, you provide the experiment name directly. The same experiment would be named iris-classifier directly.

mlflow.set_experiment(experiment_name="experiment-name")

Track parameters, metrics and artifacts

After this configuration, you can use MLflow in Azure Databricks in the same way as you're used to. For more information, see Log & view metrics and log files.

Log models with MLflow

After your model is trained, you can log it to the tracking server with the mlflow.<model_flavor>.log_model() method. <model_flavor> refers to the framework associated with the model. Learn what model flavors are supported.

In the following example, a model created with the Spark library MLLib is being registered.

mlflow.spark.log_model(model, artifact_path = "model")

The flavor spark doesn't correspond to the fact that you're training a model in a Spark cluster. Instead, it follows from the training framework used. You can train a model using TensorFlow with Spark. The flavor to use would be tensorflow.

Models are logged inside of the run being tracked. That fact means that models are available in either both Azure Databricks and Azure Machine Learning (default) or exclusively in Azure Machine Learning if you configured the tracking URI to point to it.

Register models in the registry with MLflow

As opposed to tracking, model registries can't operate at the same time in Azure Databricks and Azure Machine Learning. They have to use either one or the other. By default, model registries use the Azure Databricks workspace. If you choose to set MLflow tracking to only track in your Azure Machine Learning workspace, the model registry is the Azure Machine Learning workspace.

If you use the default configuration, the following code logs a model inside the corresponding runs of both Azure Databricks and Azure Machine Learning, but it registers it only on Azure Databricks.

mlflow.spark.log_model(model, artifact_path = "model", 
                       registered_model_name = 'model_name')  

• If a registered model with the name doesn’t exist, the method registers a new model, creates version 1, and returns a ModelVersion MLflow object.
• If a registered model with the name already exists, the method creates a new model version and returns the version object.

Use Azure Machine Learning registry with MLflow

If you want to use Azure Machine Learning Model Registry instead of Azure Databricks, we recommend that you set MLflow tracking to only track in your Azure Machine Learning workspace. This approach removes the ambiguity of where models are being registered and simplifies the configuration.

If you want to continue using the dual-tracking capabilities but register models in Azure Machine Learning, you can instruct MLflow to use Azure Machine Learning for model registries by configuring the MLflow Model Registry URI. This URI has the same format and value that the MLflow that tracks URI.

mlflow.set_registry_uri(azureml_mlflow_uri)

For a complete example of this scenario, see Training models in Azure Databricks and deploying them on Azure Machine Learning.

Deploy and consume models registered in Azure Machine Learning

Models registered in Azure Machine Learning Service using MLflow can be consumed as:

• An Azure Machine Learning endpoint (real-time and batch). This deployment allows you to use Azure Machine Learning deployment capabilities for both real-time and batch inference in Azure Container Instances, Azure Kubernetes, or Managed Inference Endpoints.
• MLFlow model objects or Pandas user-defined functions (UDFs), which can be used in Azure Databricks notebooks in streaming or batch pipelines.

Deploy models to Azure Machine Learning endpoints

You can use the azureml-mlflow plugin to deploy a model to your Azure Machine Learning workspace. For more information about how to deploy models to the different targets How to deploy MLflow models.

Deploy models to Azure Databricks for batch scoring using UDFs

You can choose Azure Databricks clusters for batch scoring. By using Mlflow, you can resolve any model from the registry you're connected to. You usually use one of the following methods:

• If your model was trained and built with Spark libraries like MLLib, use mlflow.pyfunc.spark_udf to load a model and used it as a Spark Pandas UDF to score new data.
• If your model wasn't trained or built with Spark libraries, either use mlflow.pyfunc.load_model or mlflow.<flavor>.load_model to load the model in the cluster driver. You need to orchestrate any parallelization or work distribution you want to happen in the cluster. MLflow doesn't install any library your model requires to run. Those libraries need to be installed in the cluster before running it.

The following example shows how to load a model from the registry named uci-heart-classifier and used it as a Spark Pandas UDF to score new data.

from pyspark.sql.types import ArrayType, FloatType 

model_name = "uci-heart-classifier"
model_uri = "models:/"+model_name+"/latest"

#Create a Spark UDF for the MLFlow model 
pyfunc_udf = mlflow.pyfunc.spark_udf(spark, model_uri) 

For more ways to reference models from the registry, see Loading models from registry.

After the model is loaded, you can use this command to score new data.

#Load Scoring Data into Spark Dataframe 
scoreDf = spark.table({table_name}).where({required_conditions}) 

#Make Prediction 
preds = (scoreDf 
           .withColumn('target_column_name', pyfunc_udf('Input_column1', 'Input_column2', ' Input_column3', …)) 
        ) 

display(preds) 

Clean up resources

If you want to keep your Azure Databricks workspace, but no longer need the Azure Machine Learning workspace, you can delete the Azure Machine Learning workspace. This action results in unlinking your Azure Databricks workspace and the Azure Machine Learning workspace.

If you don't plan to use the logged metrics and artifacts in your workspace, delete the resource group that contains the storage account and workspace.

1. In the Azure portal, search for Resource groups. Under services, select Resource groups.
2. In the Resource groups list, find and select the resource group that you created to open it.
3. In the Overview page, select Delete resource group.
4. To verify deletion, enter the resource group's name.

Related content

• Deploy MLflow models as an Azure web service
• Manage your models
• Track experiment jobs with MLflow and Azure Machine Learning
• Azure Databricks and MLflow