Amazon SageMaker Sub-Services

SageMaker sub-services — Canvas (no-code ML), Autopilot (AutoML), Ground Truth (data labelling), Data Wrangler (feature prep), Feature Store, Pipelines (MLOps), Model Monitor (drift), Clarify (bias/explainability), JumpStart (foundation models). AIF-C01 Domain 3 core.

Updated Invalid Date·

aws sagemaker sagemaker-studio canvas autopilot ground-truth data-wrangler feature-store pipelines model-monitor clarify jumpstart aif-c01 saa-c03

SageMaker is a platform, not a single service. AIF-C01 Domain 3 (Foundations of Generative AI Applications, 28%) and Domain 1 (AI/ML Fundamentals, 20%) test scenario-to-sub-service matching. The exam gives a job role or use case and expects you to identify which SageMaker capability solves it.

Sub-Services at a Glance

Sub-Service	Who uses it	What it does
SageMaker Studio	All ML practitioners	Unified IDE — notebook, experiments, pipelines
Canvas	Business analysts, non-coders	No-code ML — point-and-click model building
Autopilot	Data scientists wanting AutoML	AutoML — auto feature engineering, algorithm selection, tuning
Ground Truth	Data engineers	Data labelling — human + automated labelling workflows
Data Wrangler	Data scientists	Visual feature engineering and data preparation
Feature Store	ML engineers	Centralised, shared feature storage (online + offline)
Pipelines	MLOps engineers	CI/CD for ML — DAG-based training pipelines
Experiments	Data scientists	Track and compare training runs
Model Monitor	MLOps engineers	Detect data/model drift in production
Clarify	ML engineers, compliance	Bias detection and model explainability
JumpStart	All ML practitioners	Model hub — pre-trained models, fine-tuning, 1-click deploy
Model Registry	MLOps engineers	Versioned model catalogue with approval workflow
A2I (Augmented AI)	Product engineers	Human review routing for low-confidence predictions

Sub-Service Deep Dives

SageMaker Studio

The unified web-based IDE for all SageMaker services. Replaces the need to navigate separate AWS console pages.

Jupyter notebooks, terminals, file browser
Access to all sub-services from a single interface
Supports multiple ML practitioners on the same domain simultaneously
SageMaker Studio Classic (old) → SageMaker Studio (new, JupyterLab-based, 2023+)

Exam trigger: "unified ML environment", "central IDE for the ML team", "notebook environment with SageMaker integration"

Amazon SageMaker Canvas

No-code ML for business analysts and non-technical users.

Capabilities:

Upload data (CSV, S3) via point-and-click — no SQL, no Python
Canvas automatically detects column types and suggests the prediction target
Runs AutoML under the hood (Autopilot) to build and evaluate models
Time series forecasting, binary/multi-class classification, regression
Generates predictions and confidence scores; exports to S3 or connects to QuickSight
Ready Signal integration: can trigger retraining when data updates

Who uses it: Business analysts, domain experts who own the data but lack ML expertise.

Exam trigger: "non-technical users build ML models", "business analysts predict outcomes without coding", "no-code forecasting"

Amazon SageMaker Autopilot

AutoML for data scientists — automatically explores feature engineering, algorithm selection, and hyperparameter tuning.

Capabilities:

Input: tabular data (CSV/Parquet in S3) + target column
Output: ranked model leaderboard with accuracy metrics; deployable best model
Generates explainability reports (via Clarify integration)
Modes: AutoPilot (full auto), HPO (custom algorithm + tune), Ensemble (combine multiple models)
Supports: binary/multi-class classification, regression, time series forecasting (2023+)
Produces editable notebooks showing what it tried — "glass box" AutoML

vs Canvas: Autopilot requires a data scientist; Canvas requires no ML knowledge. Autopilot gives code artefacts; Canvas gives a GUI and predictions.

Exam trigger: "AutoML", "automatically select the best algorithm", "auto-tune hyperparameters", "data scientist wants fast baseline"

Amazon SageMaker Ground Truth

Managed data labelling service.

Capabilities:

Send unlabelled data to human labellers (internal workforce, Amazon Mechanical Turk, or private vendor)
Active learning: model labels high-confidence items automatically, humans handle low-confidence only
Label types: image classification, bounding boxes, named entity recognition, text classification, semantic segmentation, video object tracking, 3D point clouds
Produces labelled datasets ready for training
Ground Truth Plus: fully managed labelling service (AWS handles the workforce)

Exam trigger: "label training data", "human review of images", "build training dataset", "annotate data", "bounding box labelling"

Amazon SageMaker Data Wrangler

Visual data preparation and feature engineering — no code required.

Capabilities:

300+ built-in data transforms (impute missing values, encode categoricals, normalise, one-hot encode)
Data flow: visual DAG of transformation steps
Connects to S3, Athena, Redshift, EMR, Snowflake, and more
Generates PySpark/Pandas code for the transformation steps (export to notebook)
Data Quality and Insights Report: statistics, class imbalance, target leakage detection
Exports directly to SageMaker Feature Store, Pipelines, or S3

vs Glue DataBrew: Both are visual, no-code data prep tools. Data Wrangler is purpose-built for ML and integrates directly into SageMaker. Glue DataBrew is for general data preparation for any downstream use.

Exam trigger: "prepare features for training", "visual feature engineering", "no-code data transformation for ML", "impute missing values without coding"

Amazon SageMaker Feature Store

Centralised repository for ML features — enables reuse across models and teams.

Two stores:

Online store: low-latency (<10ms) retrieval for real-time inference (backed by key-value store)
Offline store: historical features in S3 for training and batch inference

Capabilities:

Features written once, used by multiple models — eliminates training/serving skew
Time-travel queries: retrieve features as they existed at a specific point in time (critical for correct backtesting)
Feature groups: named collections of related features with a defined schema
Integrated with Data Wrangler, Pipelines, and Autopilot

Training/serving skew: when the feature pipeline used at training differs from the one at inference, causing degraded production performance. Feature Store solves this with a single shared pipeline.

Exam trigger: "share features across teams", "consistent features between training and inference", "feature reuse", "training/serving skew", "time-travel feature queries"

Amazon SageMaker Pipelines

MLOps CI/CD for machine learning — automates the end-to-end ML workflow.

Capabilities:

DAG-based pipeline definition in Python SDK
Steps: Processing, Training, Evaluation, RegisterModel, Condition (branch on metric), Transform (batch), CreateModel
Integrated with SageMaker Experiments for run tracking
Trigger via: scheduled events, S3 uploads, EventBridge, or API call
Version-controlled pipeline definitions stored in SageMaker

Exam trigger: "automate training workflow", "retrain on new data automatically", "MLOps pipeline", "trigger model retraining on schedule", "CI/CD for ML"

Amazon SageMaker Experiments

Track, compare, and organise ML training runs.

Capabilities:

Log metrics, parameters, and artefacts for each training run
Group runs into experiments; compare across runs with visual charts
Integrated into Studio and Pipelines
Supports automatic tracking when using SageMaker Training jobs

Exam trigger: "compare model training runs", "track hyperparameter experiments", "experiment management"

Amazon SageMaker Model Monitor

Detect and alert on data and model drift in production endpoints.

Monitor types:

Monitor	What it detects
Data Quality Monitor	Input feature distribution drift (vs training baseline)
Model Quality Monitor	Prediction drift (accuracy, precision drop)
Bias Drift Monitor	Emerging bias in live predictions (via Clarify)
Feature Attribution Drift Monitor	SHAP value changes over time

How it works:

Captures a sample of inference requests and predictions automatically
Runs a scheduled processing job (hourly, daily, etc.) against a baseline from training
Publishes metrics to CloudWatch; alerts via SNS
Integrates with SageMaker Clarify for bias monitoring

Exam trigger: "detect data drift", "model degradation in production", "alert when model accuracy drops", "monitor inference for distribution shift"

Amazon SageMaker Clarify

Bias detection and model explainability.

Capabilities:

Pre-training bias detection: analyse the training dataset for class imbalance and demographic disparity before training
Post-training bias detection: analyse model predictions for fairness issues across groups
Model explainability (SHAP): compute feature importance scores — which features drive predictions most?
Explainability reports: auto-generated PDFs attached to model cards
Integrated with Model Monitor for continuous production bias/attribution drift

Bias metrics: Demographic parity difference, disparate impact, conditional demographic disparity.

Exam trigger: "detect bias in training data", "explain model predictions", "feature importance", "fairness audit", "regulatory compliance for ML model", "SHAP values"

Responsible AI angle: Clarify is SageMaker's answer to Responsible AI requirements — it surfaces both data bias and prediction bias, and generates documentation for model governance.

Amazon SageMaker JumpStart

Model hub and deployment accelerator.

Capabilities:

Catalogue of 700+ pre-trained models: foundation models (Llama 3, Mistral, Falcon, AI21, Stability AI), domain-specific models, and SageMaker built-in algorithm solutions
1-click fine-tuning: select a base model, provide fine-tuning data, launch a training job
1-click deployment: deploy any model to a SageMaker endpoint with auto-scaling
Solution templates: end-to-end ML pipelines for common use cases (fraud detection, demand forecasting)
Integrates with SageMaker Studio for browsing and deployment

vs Bedrock: JumpStart hosts and fine-tunes open-source models on your own SageMaker infrastructure. Bedrock provides managed API access to foundation models from Anthropic, Amazon, Meta, etc., without managing infrastructure.

Exam trigger: "deploy a pre-trained model", "fine-tune Llama on SageMaker", "model hub", "1-click foundation model deployment", "open-source LLM on AWS"

Amazon SageMaker Model Registry

Versioned catalogue of trained models with approval workflow.

Capabilities:

Register model versions with metadata: training metrics, dataset version, lineage
Approval workflow: Pending → Approved → Rejected; gates production deployment
Integrates with Pipelines (auto-register on passing evaluation) and CI/CD systems
Cross-account model sharing

Exam trigger: "model version control", "approval gate before production", "model governance", "model lineage"

Amazon Augmented AI (A2I)

Human review routing for low-confidence ML predictions.

Capabilities:

Works with Rekognition (content moderation), Textract (document extraction), or any custom ML model
Define confidence thresholds: predictions below threshold go to human review
Human task UI: customisable review interface
Integrates with Ground Truth workforce (Mechanical Turk, private, vendor)

Exam trigger: "human-in-the-loop", "review low-confidence predictions", "manual review of uncertain AI decisions", "escalate to human reviewer"

Decision Matrix: Which SageMaker Service?

Scenario	Service
Business analyst wants to predict customer churn without coding	Canvas
Data scientist wants best algorithm selected automatically	Autopilot
Team needs to label 10,000 images with bounding boxes	Ground Truth
Prepare and transform features visually before training	Data Wrangler
Share the same features between the training pipeline and real-time inference	Feature Store
Automate model retraining when new data arrives in S3	Pipelines
Compare 50 training runs to find the best hyperparameters	Experiments
Alert when production input data starts diverging from training distribution	Model Monitor
Audit a model for demographic bias before deployment	Clarify
Deploy Llama 3 on SageMaker with one click	JumpStart
Require manager approval before promoting a model to production	Model Registry
Route uncertain medical image classifications to a radiologist	A2I

Key Facts

Canvas: no-code ML for business analysts; runs Autopilot under the hood; no Python required
Autopilot: AutoML for data scientists; produces editable notebooks; auto-selects algorithm + tunes hyperparameters
Ground Truth: managed data labelling; active learning reduces human annotation cost; Ground Truth Plus is fully managed
Data Wrangler: visual feature engineering; 300+ transforms; exports to Feature Store, Pipelines, or S3
Feature Store: online store (<10ms latency for inference) + offline store (S3 for training); solves training/serving skew
Pipelines: DAG-based ML CI/CD; triggered by schedule, S3 events, or API; integrates with Experiments and Model Registry
Model Monitor: four monitor types (data quality, model quality, bias drift, attribution drift); publishes to CloudWatch
Clarify: pre- and post-training bias; SHAP feature importance; generates explainability reports; Responsible AI tool
JumpStart: 700+ models; 1-click fine-tune and deploy; open-source LLMs; vs Bedrock (managed API, no infrastructure)
A2I: human-in-the-loop for low-confidence predictions; works with Rekognition, Textract, or custom models

Connections

llms/ml-fundamentals — the ML concepts (supervised/unsupervised, evaluation metrics, lifecycle) that SageMaker operationalises
cloud/aws-ai-recognition-services — pre-built AI services for when custom SageMaker training is not needed
cloud/aws-analytics-services — Glue, Athena, and Redshift feed data into SageMaker training pipelines
apis/aws-bedrock — Bedrock for foundation model APIs without managing SageMaker infrastructure
landscape/aws-ai-practitioner — AIF-C01 study guide; Domains 1 and 3 cover this page directly
safety/responsible-ai — Clarify implements bias detection; connects to responsible AI principles
evals/methodology — eval framework selection and golden sets; Model Monitor and Experiments operationalise the same eval loop
observability/llmops — LLMOps maturity model; SageMaker Pipelines + Model Monitor implement the eval gate and drift detection patterns

Open Questions

When does SageMaker JumpStart become preferable to Bedrock for foundation model deployment — is it purely a cost/control decision?
Does SageMaker Canvas now support image/NLP tasks, or is it still restricted to tabular data?