SageMaker

🧠 What is Amazon SageMaker?

Amazon SageMaker is a fully managed service that enables data scientists and developers to build, train, and deploy machine learning models quickly at scale.

✅ It supports the full ML lifecycle: data labeling → training → tuning → hosting → monitoring — all in one platform.

---

🧰 SageMaker: Key Features and Modules

Module	Purpose
Studio	Web-based IDE for ML development (JupyterLab-like)
Data Wrangler	Prepare and visualize data without writing code
Feature Store	Store and reuse features across models
Ground Truth	Data labeling with human annotators + ML assistance
Training Jobs	Train ML models at scale using built-in or custom containers
Hyperparameter Tuning	Automatically tune model parameters
Inference Endpoints	Deploy models via REST APIs (real-time or batch)
Model Monitor	Detect drift in production
Pipelines	Automate ML workflows (CI/CD for ML)

---

🎯 Use Cases

Industry	Example Use Case
eCommerce	Product recommendation, customer churn
Finance	Fraud detection, credit scoring
Healthcare	Medical image analysis, disease prediction
Manufacturing	Predictive maintenance
Retail	Demand forecasting
NLP / Vision	Sentiment analysis, object detection

---

🧪 Supported ML Frameworks

• ✅ Built-in: XGBoost, Linear Learner, KNN, etc.

• ✅ Frameworks: TensorFlow, PyTorch, MXNet, Scikit-Learn, HuggingFace

• ✅ Bring Your Own Container (BYOC): Custom Docker for any toolset

---

🧑‍💻 Example: Train a Model with Boto3 (Python SDK)

Step 1: Upload data to S3

import boto3

s3 = boto3.client('s3')
s3.upload_file('train.csv', 'my-sagemaker-bucket', 'train/train.csv')

---

Step 2: Train a built-in XGBoost model

import sagemaker
from sagemaker import image_uris

session = sagemaker.Session()
role = 'arn:aws:iam::123456789012:role/sagemaker-role'

xgboost_image = image_uris.retrieve("xgboost", region='us-east-1', version='1.3-1')

estimator = sagemaker.estimator.Estimator(
    image_uri=xgboost_image,
    role=role,
    instance_count=1,
    instance_type='ml.m5.large',
    output_path='s3://my-sagemaker-bucket/output',
    sagemaker_session=session
)

estimator.set_hyperparameters(objective='reg:squarederror', num_round=100)

estimator.fit({'train': 's3://my-sagemaker-bucket/train/train.csv'})

---

Step 3: Deploy model as a REST API

predictor = estimator.deploy(
    initial_instance_count=1,
    instance_type='ml.m5.large'
)

response = predictor.predict([1.2, 3.4, 5.6](1.2,%203.4,%205.6.md))
print(response)

---

🚀 SageMaker Deployment Options

Type	Use Case
Real-time Endpoint	For low-latency inference
Batch Transform	For offline, large datasets
Asynchronous	For long-duration inference
Edge Deployment	Deploy to IoT devices using SageMaker Neo

---

🧪 Example: SageMaker Pipelines (ML CI/CD)

from sagemaker.workflow.pipeline import Pipeline
from sagemaker.workflow.steps import ProcessingStep, TrainingStep
from sagemaker.workflow.parameters import ParameterString

# Define steps using Step Functions syntax
pipeline = Pipeline(
    name="my-ml-pipeline",
    steps=[data_processing_step, model_training_step, model_registration_step]
)
pipeline.upsert(role_arn=role)
pipeline.start()

---

💰 Pricing Overview

Component	Pricing Model
Studio Notebook	Pay per compute (CPU/GPU) instance-hour
Training Jobs	Per instance type/hour + optional S3 storage cost
Inference Endpoint	Per instance/hour + data transfer
Ground Truth Labeling	Per object labeled
Pipelines	Pay only for compute used in each step

✅ Free Tier: 250 hours/month of ml.t2.medium for notebooks for 2 months.

---

🛡️ Security

Feature	Support
IAM roles and policies	✅ Granular access control
VPC support	✅ Yes (training + inference)
Encryption	✅ S3 + EBS with KMS
PrivateLink	✅ SageMaker via VPC endpoint
Audit logging	✅ CloudTrail + Model Monitor

---

🧱 Terraform Support

Amazon SageMaker has rich Terraform support. Example for deploying a model:

1. IAM Role

resource "aws_iam_role" "sagemaker_execution" {
  name = "sagemaker-execution-role"

  assume_role_policy = jsonencode({
    Version = "2012-10-17",
    Statement = [{
      Effect = "Allow",
      Principal = { Service = "sagemaker.amazonaws.com" },
      Action    = "sts:AssumeRole"
    }]
  })
}

2. Model Deployment

resource "aws_sagemaker_model" "example" {
  name               = "my-model"
  execution_role_arn = aws_iam_role.sagemaker_execution.arn
  primary_container {
    image           = "382416733822.dkr.ecr.us-east-1.amazonaws.com/xgboost:latest"
    model_data_url  = "s3://my-sagemaker-bucket/output/model.tar.gz"
  }
}

---

🧠 Comparison with Other Services

Service	Use Case
Comprehend	Pre-trained NLP tasks
Rekognition	Pre-trained computer vision
SageMaker	Custom ML models (NLP, CV, etc.)
Bedrock	Foundation models (LLMs) via API

---

✅ TL;DR Summary

Feature	SageMaker
Full ML lifecycle support	✅ Yes
Auto-scaling training	✅ Yes
Built-in algorithms	✅ 15+ included
Custom model/container	✅ Yes
Framework support	TensorFlow, PyTorch, SKLearn
Deployment options	Real-time, batch, async, edge
CI/CD automation	✅ With SageMaker Pipelines
Free Tier	✅ 250 hours/month (2 months)
Terraform support	✅ Yes (IAM, model, endpoint)

---