Deploy SageMaker Endpoint for Inference

sagemaker-endpoint.png

In this section, we will deploy the trained model from step 5 to Amazon SageMaker Endpoint, allowing inference calls via API or Lambda function. This is an important step to turn your ML model into a service that can be used in the real world.

🎯 Objectives

  • Create an endpoint from the model trained and registered in the previous step.
  • Test the endpoint with sample data.
  • Prepare the endpoint for integration with Lambda in the next step.

🧠 6.1 – Create a SageMaker Model from an Artifact

After training and registering the model (step 5), we will create a SageMaker Model based on that output.

  1. Go to SageMaker Console → Inference → Models
  2. Click Create model
  3. Configure:
  • Model name: ml-blog-model
  • Execution role: Select the IAM role created in the previous step (SageMakerExecutionRole)
  • Container definition:
  • Image: 382416733822.dkr.ecr.ap-southeast-1.amazonaws.com/xgboost:latest (or the image you used when training)
  • Model artifact location:
s3://ml-pipeline-bucket/model/xgboost-model.tar.gz
  1. Click Create model to complete. create-endpoint.png

📌 Note: The container image and artifact path must match the previously created train job.

⚙️ 6.2 – Create Endpoint Configuration

  1. Navigate to Inference → Endpoint configurations
  2. Select Create endpoint configuration
  3. Configuration:
  • Name: ml-blog-endpoint-config
  • Model name: ml-blog-model
  • Instance type: ml.m5.large (or ml.t2.medium if you want to save costs)
  • Initial instance count: 1
  1. Click Create endpoint configuration create-endpoint.png

🌐 6.3 – Deploy SageMaker Endpoint

  1. Navigate to Inference → Endpoints
  2. Select Create endpoint
  3. Enter:
  • Endpoint name: ml-blog-endpoint
  • Endpoint configuration: Select ml-blog-endpoint-config
  1. Click Create endpoint. The creation process will take a few minutes ⏳.

📸 Example deployment interface:

create-endpoint.png

  • Make sure the IAM Role has access to S3 and SageMaker (AmazonSageMakerFullAccess, AmazonS3ReadOnlyAccess).
  • The endpoint must be in InService state before using inference.

🧪 6.4 – Testing Endpoint with boto3 (Python)

Once the endpoint is in InService state, check the inference with the following Python code:

import boto3
import json

runtime = boto3.client('sagemaker-runtime')

payload = {
"features": [0.56, 0.32, 0.78, 0.12] # example input data
}

response = runtime.invoke_endpoint(
EndpointName='ml-blog-endpoint',
ContentType='application/json',
Body=json.dumps(payload)
)

result = json.loads(response['Body'].read().decode())
print("📊 Predicted result:", result)

✅ The result will return the predicted value (e.g., 1 or 0 for a classification model).

.

📊 6.5 – Endpoint Monitoring with CloudWatch

Go to CloudWatch → Logs to view the inference log. Monitor metrics such as:

  • Invocations
  • Invocation4XXErrors
  • ModelLatency

This helps evaluate model performance in production environments.

📌 You can enable Auto Scaling for the endpoint by using Application Auto Scaling to automatically scale up/down the number of instances based on inference traffic.

✅ Done

You have successfully deployed a SageMaker Endpoint from the trained model.