Amazon SageMaker

Overview

Fully managed service for building, training, and deploying ML models.

SageMaker Components

graph TD
    A[SageMaker Studio] --> B[Data Preparation]
    B --> C[Training]
    C --> D[Model Registry]
    D --> E[Deployment]
    E --> F[Monitoring]

    B --> B1[Data Wrangler]
    B --> B2[Feature Store]
    B --> B3[Processing]

    C --> C1[Training Jobs]
    C --> C2[Experiments]
    C --> C3[Debugger]
    C --> C4[Automatic Model Tuning]

    E --> E1[Real-time Endpoints]
    E --> E2[Serverless]
    E --> E3[Batch Transform]

    F --> F1[Model Monitor]
    F --> F2[Clarify]

Key Features

Feature	Description
Studio	Unified IDE for ML
Data Wrangler	Visual data preparation
Feature Store	Centralized feature repository
Training	Managed training infrastructure
Experiments	Track and compare runs
Debugger	Debug and profile training
Clarify	Bias and explainability
Model Registry	Version and manage models
Pipelines	MLOps workflows
Model Monitor	Production monitoring

Instance Types

Category	Types	Use Case
General	ml.m5.*	Balanced workloads
Compute	ml.c5.*	CPU-intensive
GPU	ml.p3., ml.p4d.	Deep learning
GPU (inference)	ml.g4dn.*	Inference
Memory	ml.r5.*	Large datasets

Data Channels

Channel	Purpose
train	Training data
validation	Validation data
test	Test data
model	Pre-trained model

Training Modes

Mode	Description	Best For
File	Download to instance	Small/medium data
Pipe	Stream from S3	Large data
FastFile	POSIX access	Random access

Exam Focus Areas

!!! warning "Key Topics" - Training job configuration and instance selection - Built-in algorithms and when to use each - Endpoint types and deployment strategies - Model Monitor for production monitoring - Pipelines for MLOps automation