Invoke Model Endpoint From External Clients

Integration Overview (Lesson 96)

SageMaker SDK: Specifically designed for SageMaker, offering high-level abstractions.
Boto3 SDK: General AWS SDK for Python, providing more control but requires detailed configuration.

Serverless compute service ideal for hosting endpoint integration logic.
Simplifies the architecture by abstracting server management and scaling.
Lambda functions can preprocess input data, invoke the SageMaker endpoint, and post-process the results.

Combining these services allows you to create RESTful APIs for your endpoints.
Platform-agnostic: Facilitates endpoint invocation regardless of the client's programming language.
Scalable and maintainable: Manages traffic, authorizes API calls, and monitors performance.

Consistent data preprocessing is crucial both during training and prediction.
The same transformations (e.g., one-hot encoding, scaling) applied during model training should be replicated when preparing data for prediction.
This can be embedded in Lambda functions or within client applications.

In this lab, we will invoke a SageMaker endpoint using the SageMaker SDK from local environment.
We will use the XGBoost bike rental model, so chek Lesson 78 if you didn't create it yet.
You can create the endpoint using ml_admin user.
Use invoke_using_sagemaker_sdk.ipynb notebook locally.
This notebook uses the ml_user_predict credentials, so make sure you have configured it beforehand.
Also check if your ~/.aws/credentials and ~/.aws/config files contains this user.

Use invoke_using_boto3.ipynb notebook locally.
Refer to the previous lab for the setup.
Boto3 also requires some work in the client side. Next lab implements this logic as a micro-service using lambda serverless infrastructure.

When lambda functions run our code, it needs permission to access services that are referenced.
We can grant permission to a lambda function using an IAM role.
Use your my_admin account.
Select Roles and create a new role.
Role Configuration:
Select AWS service as Lambda.
Assign SageMakerInvokeEndpoint policy for SageMaker endpoint invocation and AWSLambdaBasicExecutionRole for Lambda function logging.
Give name : lambda_sagemaker_invoke_endpoint

Under Services, select Lambda. Then Create Function.
We can Author from Scratch:
Name: bikerental_prediction
Runtime: Python 3.6 or later.
Permissions: Use the role lambda_sagemaker_invoke_endpoint.
Implementing the Function:
Use the code from invoke_using_boto3_lambda.py in the IntegrationExamples folder.
Define the entry point as lambda_handler.
Deploy changes.
Environment Variable:
Set ENDPOINT_NAME to your SageMaker endpoint name.
Save and Test the Function:
Create a test event using LambdaTestEvent.txt file.
First create an event for a single observation (Bike Single Observation)
Name the event as BikeSingleObservation
Click Test and check the results.

All the messages that you print in the lambda function will be logged to CloudWatch.
View logs under the Monitoring tab or directly in the CloudWatch service.

At this point we can expose the AWS Lambda function as is, or as a RESTful API using API Gateway.
API Gateway approach provides a clean interface for backend services, security management, and scalability.
So client will have access to the API Gateway, which will forward the request to the Lambda function, which will invoke the SageMaker endpoint.

Login to AWS Management Console with my_admin account, navigate to API Gateway.
Create a new REST API named BikeRentalPrediction.
Select Create Method ans select POST method.
Select Integration Type as Lambda Function.

Copy and paste the same single observation (Bike Single Observation) in Lesson 101.