Amazon SageMaker Cheat Sheet

Updated 2026-05-21

Next Topic: Anomaly Detection in Machine Learning Cheat Sheet

🧠Study flashcards on this topic122 cards · spaced repetition→

Amazon SageMaker is AWS's fully managed end-to-end machine learning platform, covering every stage from data preparation and feature engineering through model training, tuning, deployment, and production monitoring. It eliminates the heavy lifting of infrastructure provisioning so practitioners can focus on model quality. The platform spans a wide spectrum — from no-code tools like Canvas and Autopilot, to low-level Python SDK primitives like Estimator and HyperparameterTuner — making it relevant whether you're a data scientist, ML engineer, or business analyst. The critical mental model is that SageMaker is not a single service but a suite of tightly integrated capabilities: data flows from processing → training → registry → endpoint, and MLOps tooling (Pipelines, Projects, Model Monitor) wraps that lifecycle for automation and governance.

What This Cheat Sheet Covers

This topic spans 19 focused tables and 141 indexed concepts, 122 flashcards. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: SageMaker Studio and Development EnvironmentsTable 2: Training Jobs and the Estimator APITable 3: Distributed TrainingTable 4: Hyperparameter TuningTable 5: SageMaker Processing JobsTable 6: SageMaker Pipelines and MLOps WorkflowTable 7: Model RegistryTable 8: Inference Endpoint TypesTable 9: Advanced Endpoint PatternsTable 10: SageMaker Feature StoreTable 11: SageMaker Clarify — Bias and ExplainabilityTable 12: SageMaker Model MonitorTable 13: SageMaker Experiments and LineageTable 14: SageMaker JumpStartTable 15: SageMaker Autopilot and Canvas (AutoML)Table 16: SageMaker Data WranglerTable 17: SageMaker Ground Truth — Data LabelingTable 18: SageMaker MLOps and ProjectsTable 19: SageMaker Built-in Algorithms

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: SageMaker Studio and Development Environments

SageMaker Studio is the primary integrated development environment for the platform, providing a browser-based workspace where data scientists and engineers access notebooks, experiments, pipelines, and model governance in one place. Understanding the Studio architecture — domains, user profiles, and application types — is foundational for configuring team access and customizing compute environments.

Feature	Example	Description
SageMaker Unified Studio	Single browser IDE for SQL, notebooks, pipelines, MLflow	Unified IDE that consolidates data engineering, ML development, and generative AI workflows into one experience with access to all AWS analytics and AI services.
Domain	`CreateDomain` with VPC, auth mode, default settings	• Top-level organizational unit that scopes a team's Studio environment • contains user profiles, apps, and shared resources within a VPC
UserProfile	`CreateUserProfile` with per-user execution role	Per-user configuration within a domain that can override domain defaults for instance types, images, and execution roles.
JupyterLab Space	Persistent workspace with kernel gateway app	• Private or shared compute environment running JupyterLab • persists notebooks and installed packages between sessions
Code Editor (VS Code)	Connect local VS Code / Cursor to SageMaker compute	• Browser or local IDE experience powered by Code-OSS • supports local IDE connections via AWS Toolkit so you use your own editor against SageMaker compute
KernelGateway App	Kernel running `conda_pytorch_p310` environment	• Application type that manages compute kernels for notebooks • each kernel runs in its own managed environment
Lifecycle Configuration (LCC)	Shell script that runs `pip install` on kernel startup	• Shell script attached to Studio apps or notebook instances that runs on create or start • used to pre-install packages or customize the environment
Custom Images	ECR image → `AppImageConfig` → attach to Domain	• Bring-your-own Docker images for JupyterLab or CodeEditor apps • registered via `AppImageConfig` and made available to users in a domain
SageMaker Notebook Instances (classic)	Standalone `ml.t3.medium` notebook with git integration	• Standalone managed EC2 notebook instance (pre-Studio) • still used for certain legacy workloads • lifecycle configs run on create or start

Table 2: Training Jobs and the Estimator API

Training jobs are the core compute primitive: SageMaker provisions instances, copies data from S3, runs your script, saves model artifacts back to S3, and terminates the instance. The Estimator class in the Python SDK is the primary interface, with framework-specific subclasses (PyTorch, TensorFlow, SKLearn, etc.) that handle image selection and serialization automatically.

Feature	Example	Description
Estimator	`est = Estimator(image_uri, role, instance_type="ml.p3.2xlarge", instance_count=1)`	• Base class for launching training jobs • wraps Docker image, IAM role, instance config, and hyperparameters into a single `fit()` call

The rest of this section is part of the full cheat sheet

Table 3: Distributed Training

Distributed training becomes necessary when a model or dataset doesn't fit on a single GPU or when you want to cut training wall-clock time by parallelizing across many accelerators. SageMaker provides AWS-optimized libraries for both data parallelism and model parallelism, and the HyperPod service manages the persistent cluster infrastructure for the largest workloads.

Feature	Example	Description
SageMaker Distributed Data Parallel (SMDDP)	`distribution={"smdistributed": {"dataparallel": {"enabled": True}}}`	• AWS-optimized AllReduce/AllGather collective operations for data parallelism • achieves near-linear scaling efficiency across multiple instances by optimizing for AWS networking fabric

The rest of this section is part of the full cheat sheet

Table 4: Hyperparameter Tuning

Hyperparameter tuning (Automatic Model Tuning, AMT) runs many training jobs over a defined parameter search space to find the configuration that optimizes an objective metric. It wraps any Estimator — including custom containers — making it algorithm-agnostic.

Feature	Example	Description
HyperparameterTuner	`tuner = HyperparameterTuner(estimator, "validation:rmse", {"lr": ContinuousParameter(0.001, 0.1)}, max_jobs=20, max_parallel_jobs=4)`	• Main SDK class for AMT • wraps an estimator with a search space, objective metric, and job budget • calling `tuner.fit()` launches the tuning job

The rest of this section is part of the full cheat sheet

Table 1: SageMaker Studio and Development Environments

Feature	Example	Description
SageMaker Unified Studio	Single browser IDE for SQL, notebooks, pipelines, MLflow	Unified IDE that consolidates data engineering, ML development, and generative AI workflows into one experience with access to all AWS analytics and AI services.
Domain	`CreateDomain` with VPC, auth mode, default settings	• Top-level organizational unit that scopes a team's Studio environment • contains user profiles, apps, and shared resources within a VPC
UserProfile	`CreateUserProfile` with per-user execution role	Per-user configuration within a domain that can override domain defaults for instance types, images, and execution roles.
JupyterLab Space	Persistent workspace with kernel gateway app	• Private or shared compute environment running JupyterLab • persists notebooks and installed packages between sessions
Code Editor (VS Code)	Connect local VS Code / Cursor to SageMaker compute	• Browser or local IDE experience powered by Code-OSS • supports local IDE connections via AWS Toolkit so you use your own editor against SageMaker compute
KernelGateway App	Kernel running `conda_pytorch_p310` environment	• Application type that manages compute kernels for notebooks • each kernel runs in its own managed environment
Lifecycle Configuration (LCC)	Shell script that runs `pip install` on kernel startup	• Shell script attached to Studio apps or notebook instances that runs on create or start • used to pre-install packages or customize the environment
Custom Images	ECR image → `AppImageConfig` → attach to Domain	• Bring-your-own Docker images for JupyterLab or CodeEditor apps • registered via `AppImageConfig` and made available to users in a domain
SageMaker Notebook Instances (classic)	Standalone `ml.t3.medium` notebook with git integration	• Standalone managed EC2 notebook instance (pre-Studio) • still used for certain legacy workloads • lifecycle configs run on create or start

Table 2: Training Jobs and the Estimator API

Feature	Example	Description
Estimator	`est = Estimator(image_uri, role, instance_type="ml.p3.2xlarge", instance_count=1)`	• Base class for launching training jobs • wraps Docker image, IAM role, instance config, and hyperparameters into a single `fit()` call

Table 3: Distributed Training

Feature	Example	Description
SageMaker Distributed Data Parallel (SMDDP)	`distribution={"smdistributed": {"dataparallel": {"enabled": True}}}`	• AWS-optimized AllReduce/AllGather collective operations for data parallelism • achieves near-linear scaling efficiency across multiple instances by optimizing for AWS networking fabric

Table 4: Hyperparameter Tuning

Feature	Example	Description
HyperparameterTuner	`tuner = HyperparameterTuner(estimator, "validation:rmse", {"lr": ContinuousParameter(0.001, 0.1)}, max_jobs=20, max_parallel_jobs=4)`	• Main SDK class for AMT • wraps an estimator with a search space, objective metric, and job budget • calling `tuner.fit()` launches the tuning job