SageMaker Sub-Features

Beyond the core training and inference components, SageMaker offers a rich set of sub-features that cover every aspect of the ML lifecycle — from data preparation and labeling to model explainability, monitoring, and CI/CD for ML. Understanding which feature to use for each task is essential for building production ML systems.

Data Preparation and Labeling

Feature	What It Does	When to Use
Data Wrangler	Visual data preparation inside Studio. Transforms, analyzes, and visualizes data. Connects to S3, Redshift, Athena, Snowflake	Data prep, feature engineering, EDA before modeling. Does NOT accept direct uploads — must import from a data source
Feature Store	Centralized store for ML features with two modes: Online (low-latency) and Offline (S3-backed)	Share and reuse features across teams. Online mode serves features at inference time; Offline mode provides historical data for training
Ground Truth	Data labeling service supporting image, text, video, and 3D labeling	Building labeled training datasets
Canvas	No-code ML for business analysts — point-and-click model building	When users with no coding knowledge need to build ML predictions

Data Wrangler Key Capabilities

• Quick Model: Generates feature importance scores to identify which features matter most
• Multicollinearity Detection: Uses PCA/VIF to find correlated features
• Bias Report: Assesses fairness in your data before training

Ground Truth Workforce Options

Workforce	When to Use
Private workforce	Sensitive or confidential data
Amazon Mechanical Turk	Non-sensitive data at scale
Active learning	Automatically labels high-confidence samples; humans label the rest — reduces labeling cost

Model Development and Tuning

Feature	What It Does	When to Use
Autopilot	AutoML — automatically explores data, selects algorithms, trains and tunes models	When you want a quick baseline model or full automated ML pipeline
Automatic Model Tuning (HPO)	Hyperparameter optimization with Bayesian, Random, Grid, and Hyperband strategies	Finding optimal hyperparameters for your model
Experiments	Track, compare, and evaluate ML experiments — logs parameters, metrics, artifacts	Comparing different preprocessing or training approaches
Training Compiler	Optimizes deep learning training by compiling computation graphs	Speed up PyTorch or TensorFlow training (up to 50% faster) without code changes

HPO Strategy Guide

Strategy	Characteristics	Best For
Bayesian	Uses results of previous evaluations to choose next set	Most efficient with fewer jobs
Random	Samples randomly from the search space	Good baseline, easily parallelized
Grid	Tests every combination	Small, discrete search spaces
Hyperband	Early-stops poorly performing configurations	Fastest overall, large search spaces
Warm start	Reuses results from previous tuning jobs	Iterating on prior tuning work

Model Explainability and Monitoring

Feature	What It Does	When to Use
Clarify	Bias detection and model explainability using SHAP values	Explain individual predictions, detect bias pre-training and post-training
Model Monitor	Monitors deployed models for data drift, model quality, bias drift, and feature attribution drift	Production monitoring — detect when model performance degrades
Debugger	Monitors training jobs in real time — captures tensors, gradients, and weights	Training-time debugging: detect vanishing gradients, overfitting, etc. Can trigger CloudWatch alarms and auto-stop training

Clarify vs. Model Monitor vs. Debugger

	When It Runs	What It Detects
Clarify	Pre-training or post-training (on demand)	Bias in data/model, feature contributions (SHAP)
Model Monitor	Continuously on deployed endpoints	Data drift, model quality degradation, bias drift over time
Debugger	During training jobs	Training issues — vanishing gradients, overfitting, poor convergence

MLOps and Deployment

Feature	What It Does	When to Use
Pipelines	ML workflow orchestration — CI/CD for ML	Automated workflows: data prep, train, evaluate, deploy
Model Registry	Central catalog for model versioning, approval status, and lineage	Tracking model versions across dev/staging/production
Neo	Compiles models for optimized inference on edge devices (ARM, Intel, NVIDIA)	Deploying models to edge hardware
Inference Recommender	Benchmarks model across instance types for best price-performance	Choosing the right endpoint instance type
Elastic Inference	Attaches fractional GPU acceleration to CPU instances	When a full GPU is underutilized — reduce inference cost

The MLOps Pattern

A common production pattern chains these features together:

Pipelines (orchestrate workflow) → Model Registry (version and approve) → Model Monitor (detect drift) → EventBridge (trigger retraining)

When to Use

Use SageMaker sub-features when you need integrated, purpose-built tools for specific ML lifecycle tasks. The key is matching each task to the right feature rather than building custom solutions.

Flashcards

1 / 10

Question

What is the difference between SageMaker Ground Truth and Amazon A2I?

Click to reveal

Answer

Ground Truth is for labeling training data before model building. A2I (Augmented AI) is for human review of model predictions in production. Don't confuse the two.

Key Insight

Data Wrangler does NOT accept direct file uploads — you must import data from S3, Redshift, Athena, or another supported data source. If you need feature importance scores quickly, use Data Wrangler's Quick Model rather than training a full Autopilot model.