🌿💻 The Packets Garden

Amazon SageMaker Overview

🌱 🪴 ⏱️ 6 min read

Main Concept

Amazon SageMaker is AWS’s fully managed end-to-end ML platform. It covers every stage of the ML lifecycle — from data preparation to model training, deployment, and monitoring — without managing underlying infrastructure.

SageMaker is the answer whenever a scenario requires building, training, or deploying a custom ML model, as opposed to consuming a pre-built managed service like Rekognition or Transcribe.

Key Idea

  • Managed AI services (Rekognition, Transcribe, Comprehend) → pre-built AI capabilities, no model building needed.

  • Amazon SageMaker → when you need to build, train, and deploy YOUR OWN custom ML model.

  • The exam distinguishes these two paths clearly — recognizing which one applies to a scenario is high-value.

SageMaker Studio

The central interface for everything in SageMaker — a unified, web-based environment where all ML development happens in one place.

Key Idea

  • Team collaboration → multiple team members work in the same environment.

  • End-to-end → data prep, training, tuning, debugging, deployment, and monitoring — all from one interface.

  • Automated workflows → orchestrate ML pipelines without switching tools.

Exam trigger: "end-to-end ML development in a single unified environment" → SageMaker Studio.

The SageMaker Features Map — By ML Pipeline Stage

Data Preparation

SageMaker Data Wrangler Visual tool for data exploration, transformation, and preparation — EDA without writing code.

Exam trigger: "visually prepare and explore data before training" → SageMaker Data Wrangler.

SageMaker Feature Store Centralized repository to store, share, and reuse ML features across teams and models.

Exam trigger: "store and reuse features across multiple ML models or teams" → SageMaker Feature Store.

SageMaker Ground Truth Managed data labeling service combining human labelers with automated labeling.

Exam trigger: "label training data at scale" → SageMaker Ground Truth.

Model Training & Tuning

Amazon SageMaker (core) Fully managed training environment — you bring your algorithm and data, SageMaker provisions compute and manages infrastructure.

Example from Maarek's lesson

Goal: predict a student’s AWS exam score. Input features: years of IT experience, years of AWS experience, hours spent on the course. Output: predicted exam score.

Historical data with known scores is used to train the model on SageMaker. Once trained, a new student inputs their profile and gets a predicted score — all without managing any servers.

SageMaker Automatic Model Tuning (AMT) Automates hyperparameter tuning. You define the objective metric — AMT handles the rest: choosing hyperparameter ranges, search strategy, run duration, and early stopping conditions.

Key Idea

  • You define → the objective metric (what to optimize for).

  • AMT handles → hyperparameter ranges, search strategy, early stopping.

  • Benefit → saves time and money by avoiding suboptimal configurations automatically.

Exam trigger: "automatically find the best hyperparameter values" → SageMaker AMT.

SageMaker JumpStart Pre-built ML solutions and foundation models ready to deploy with one click.

Exam trigger: "deploy a pre-built foundation model or ML solution quickly" → SageMaker JumpStart.

Model Deployment — The Four Types

This is a high-priority exam area. Maarek explicitly flags the keywords to watch for.

Key Idea: The four deployment types and their exam signal words

  • Real-time → low latency, small payload, one record, managed infrastructure.

  • Serverless → low latency, no infrastructure to manage, risk of cold start on first call.

  • Asynchronous → near-real time, large payload, one record at a time, long processing.

  • Batch Transform → high latency, entire dataset, multiple records processed concurrently.

Real-Time Inference

Latency:          Low
Payload size:     Up to 6 MB
Records:          One at a time
Processing time:  Up to 60 seconds
Infrastructure:   You configure CPU/GPU + auto-scaling

Exam signal words: "real-time", "immediate response", "small payload"

Serverless Inference

Latency:          Low (but cold start risk)
Payload size:     Up to 6 MB
Records:          One at a time
Processing time:  Up to 60 seconds
Infrastructure:   None to manage — auto-scaling built in

Key Idea: Cold Start

If the serverless endpoint has had no traffic for a period, the first request triggers infrastructure to boot up — adding latency to that first call. Subsequent calls are fast.

Exam signal words: "no infrastructure to manage", "serverless", "variable traffic"
Key differentiator from real-time: no infrastructure management + cold start risk.

Asynchronous Inference

Latency:          Near-real time (not immediate)
Payload size:     Up to 1 GB
Records:          One large record at a time
Processing time:  Up to 1 hour
Storage:          Requests and responses go through Amazon S3

Exam signal words: "near-real time", "large payload", "long processing time", "up to 1 GB"

Batch Transform

Latency:          High (minutes to hours)
Payload size:     100 MB per mini-batch (many mini-batches allowed)
Records:          Entire dataset — multiple records processed concurrently
Processing time:  Up to 1 hour
Storage:          Requests and responses go through Amazon S3

Exam signal words: "entire dataset", "multiple predictions", "batch", "concurrent processing"

Deployment Types — Quick Comparison Table

TypeLatencyPayloadRecordsKey Signal
Real-timeLow6 MB1immediate response
ServerlessLow + cold start6 MB1no infrastructure
AsynchronousNear-real time1 GB1large payload
Batch TransformHigh100 MB/mini-batchManyentire dataset

Monitoring & Responsible AI

SageMaker Model Monitor Continuously monitors deployed models for data quality, model drift, bias drift, and feature attribution drift.

Exam trigger: "detect model drift or performance degradation in production" → SageMaker Model Monitor.

SageMaker Clarify Detects bias in training data and model predictions. Provides explainability reports.

Exam trigger: "detect bias in ML model" or "explain model predictions" → SageMaker Clarify.

SageMaker Model Cards Standardized documentation capturing model purpose, training data, evaluation results, and limitations.

Exam trigger: "document model details for transparency or governance" → SageMaker Model Cards.

The Full SageMaker Feature Map

Studio               → unified end-to-end ML development environment
Data Wrangler        → prepare and explore data visually (EDA)
Feature Store        → store and reuse features across models
Ground Truth         → label training data at scale
Core SageMaker       → train custom ML models
AMT                  → automate hyperparameter tuning
JumpStart            → deploy pre-built models and solutions
Real-time endpoint   → low latency, one record, small payload
Serverless endpoint  → no infrastructure, cold start risk
Async endpoint       → near-real time, large payload, one record
Batch transform      → high latency, entire dataset, concurrent
Model Monitor        → detect drift and degradation in production
Clarify              → detect bias, explain predictions
Model Cards          → document models for transparency

Where SageMaker Appears in the Exam Domains

Domain 1, Task 1.2  → SageMaker as the managed platform for custom ML models
Domain 1, Task 1.3  → features mapped to each ML pipeline stage
                      Data Wrangler, Feature Store, Core, AMT, Model Monitor
Domain 2, Task 2.3  → SageMaker JumpStart for generative AI applications
Domain 4, Task 4.1  → Clarify (bias), Model Monitor (monitoring), A2I (human review)
Domain 4, Task 4.2  → Model Cards (transparency and explainability)
Domain 5, Task 5.1  → Model Cards (data lineage and governance)

What You Do NOT Need to Know

  • How to write SageMaker training scripts or code.
  • How to configure SageMaker endpoints technically.
  • The mathematics behind any SageMaker built-in algorithm.
  • How to implement bias detection technically.

You describe what each feature does and match it to the right scenario — nothing more.

Graph View