mleap

Deploying a Spark model in SQL Server big data cluster using MLeap bundles

Contents

About this sample
Before you begin
Run this sample
Sample details
Related links

About this sample

This sample assumes you are familiar with MLeap. MLeap provides simple interfaces to execute entire ML pipelines. For additional information on MLeap see the documentation. MLeap allows you to take a trained Spark model and use it outside of Spark, for example for scoring. After you have created and trained your model in Spark, you can serialize the model as an MLeap bundle for use outside of Spark. This allows model portability and using the model for scoring outside of Spark. This example demonstrates how to serialize a trained model as MLeap bundle and how to deploy it a RESTful web service with a single line of code in SQL Server big data cluster. Also, this example shows how to use an sample input in MLeap frame format to test it.

Inputs

Parameter	Description
bundleFileName	The file name of the MLeap model to run.

Outputs

There are no outputs defined as it is dependent on what the model returns.

Before you begin

To run this sample, you need the following prerequisites.

Software prerequisites:

1. SQL Server big data cluster CTP 2.3 or later.
2. azdata. Refer to installing azdata document on setting up the azdata and connecting to a SQL Server 2019 big data cluster.

Run this sample

1. Clone or download this sample on your computer.

2. Log in to the SQL Server big data cluster using the command below using the IP address of the controller-svc-external in your cluster. If you are not familiar with azdata you can refer to the documentation and then return to this sample.

   azdata login -e https://<ip-address-of-controller-svc-external>:30080 -u <user-name>

3. This example uses a TensorFlow Machine Learning Model that uses public US Census data predict income. More details and information on the example are here. The application you will be deploying as part of this sample is a Random Forest Model that was built in Spark and has been serialized as an MLeap bundle.

   Deploy the app using the create command and pass the location of the spec file. In the example below, the spec file is expected to be in the mleap folder:

   azdata app create --spec ./mleap/

4. Check the deployment by running the following command:

   azdata app list -n mleap-census -v [version]

   Once the app is listed as Ready you can continue to the next step.

5. Now that the app has been deployed you can test if the app works correctly by passing in a sample input that is available in the mleap folder. The deployed app is a RESTful webservice that is Swagger compliant. For this sample we will show you how you can test this using the CLI.

   To test the app, run the command below. The input parameter is a MLeapFrame, a json file that describes the parameters and the values provided to the model for predicting income. Note that the input parameter has a special character '@' to indicate that a json file is being passed. This command needs to be run within the mleap folder.

   azdata app run --name mleap-census --version [version] --input schema=@census_frame.json

   The result will be a json output that includes the predicted income along with additional data.

   RESTful web service. Note that any app you create is also accessible using a RESTful web service that is Swagger compliant. See step 6 in the Addpy sample for detailed instructions on how to call the web service.

6. You can clean up the sample by running the following commands:

   # delete app
   azdata app delete --name mleap-census --version [version]

Sample details

Spec file

The spec file serves as the data required for deploying the app and it contains basic information about the app such as the name, version, inputs, outputs, replicas and poolsize you need for this app. As you can see this example uses the Mleap runtime and specifies the bundleFileName to point to the Spark model serialized as MLeap bundle.

Here is the spec file for this application:

name: mleap-census
version: v1
runtime: Mleap
bundleFileName: census_bundle.zip
replicas: 2
poolsize: 2

Next Steps

Please refer to Train and Create machine learning models with Spark on how to train your model in Spark within SQL Server big data clusters and export it to MLeap.