Commonly Confused Services on MLS-C01

Deciding signal

A SageMaker Training Job is a single managed compute run that takes data from S3, runs a training algorithm on ML instances, and writes model artifacts back to S3. It answers one question: produce a trained model artifact. SageMaker Pipelines is a CI/CD orchestration service for ML workflows — it chains preprocessing, training, evaluation, approval gates, and registration into a repeatable, version-controlled DAG. Pipelines is the right answer when the scenario describes automating the end-to-end ML lifecycle, running the same workflow on new data batches, or enforcing quality gates before model registration.

Deciding signal

Real-Time Endpoints serve synchronous inference requests — the caller waits for the response, latency matters, and the endpoint must stay running. Best for user-facing applications. Batch Transform runs inference on large datasets stored in S3 without a persistent endpoint — it spins up instances, scores all records, and shuts down. Best for offline scoring jobs where latency is irrelevant and cost efficiency matters. Async Inference accepts requests with large payloads or long processing times and returns results via S3 notification — it queues requests and scales the endpoint down to zero when idle. Best for large document or video processing where synchronous timeout limits would be exceeded.

Deciding signal

Data Wrangler is a visual data preparation tool inside SageMaker Studio — it imports data from S3, Redshift, or other sources, applies transformations with a point-and-click UI, and exports the prepared dataset. It is a data preparation tool, not a feature store. Feature Store is a repository for ML features — engineered values that are stored with consistent naming, versioned, and reused across training and inference. It has an online store (low-latency lookup) and offline store (S3-backed for training). Glue is a serverless ETL service for transforming and loading data into S3, databases, or data warehouses — it is more general-purpose and not ML-specific. When the scenario involves preparing a dataset for a specific training run, Data Wrangler. When it involves storing and reusing computed features across multiple models, Feature Store.

Deciding signal

Ground Truth manages the labeling of training data — images, text, video — using a workforce (Amazon Mechanical Turk, private team, or third party). The output is a labeled dataset for training a model. A2I (Augmented AI) inserts human review into a production ML prediction workflow when a model returns low-confidence predictions. It integrates with Textract, Rekognition, and custom models. The output is a reviewed prediction with a human override where needed. The timing is the distinction: Ground Truth is pre-training (producing training data); A2I is post-deployment (reviewing production predictions).

Deciding signal

EMR provisions a managed cluster of EC2 instances running Hadoop, Spark, Hive, Presto, or other distributed frameworks. You control the cluster size, instance types, and framework configuration — it is the right answer when you need full control of the Spark environment, custom libraries, or processing that exceeds Glue limits. Glue is serverless ETL: you define jobs in PySpark or Python and Glue manages the compute — no cluster management required. Athena is a serverless SQL query engine that reads data directly from S3 — there is no data movement or ETL step. The distinction is the operation: ad-hoc SQL queries on S3 data (Athena), managed ETL jobs without cluster management (Glue), or custom Spark processing with full cluster control (EMR).

Deciding signal

Canvas is a visual, no-code tool for business analysts. It imports data, automatically trains a model, and produces predictions — no ML expertise required. It is designed for non-technical users. Autopilot is AutoML for ML practitioners: it trains and evaluates many candidate model/algorithm combinations, produces leaderboard results, and generates explainability reports (SageMaker Clarify integration). Autopilot exposes the full training configuration and notebooks so practitioners can inspect and reproduce what was done. The user is the key: business analysts using a point-and-click UI (Canvas) versus ML practitioners who want automated training with visibility into the process (Autopilot).

Deciding signal

Built-in algorithms (XGBoost, Linear Learner, DeepAR, BlazingText, etc.) are optimized AWS-managed training algorithm containers — you provide training data and hyperparameters, and SageMaker handles the rest. They are the right answer for standard ML tasks where an AWS-optimized implementation of a known algorithm is available. BYOC lets you package a custom training or inference script in a Docker container and run it on SageMaker infrastructure — required when you need a custom framework, library, or algorithm not available as a built-in. JumpStart provides a hub of pre-trained foundation models and ML models that can be deployed or fine-tuned — you are not training from scratch. JumpStart is the right answer when a pre-trained model is sufficient and you do not want to start training from scratch.

Deciding signal

SageMaker Hyperparameter Tuning automatically searches a defined hyperparameter space — learning rate, batch size, regularization — across multiple training jobs to find the combination that maximizes your objective metric. It manages the search strategy (Bayesian, random, grid) and selects the best model. SageMaker Experiments provides a framework for organizing, tracking, and comparing training runs — you log metrics, parameters, and artifacts from any training job and visualize them in a comparison view. HPO automates the search; Experiments provides observability and comparison across any training runs, including manual ones. When the scenario asks how to automatically find the best hyperparameters, HPO. When it asks how to track and compare runs across an ML project, Experiments.

Deciding signal

Clarify analyzes models for bias in training data or predictions and provides feature attribution (SHAP values) to explain individual predictions. It runs as a processing job during training and evaluation to assess fairness and explainability. Model Monitor continuously monitors deployed models in production for data quality drift (input feature distribution changes), model quality drift (accuracy degradation), bias drift, and feature attribution drift. It generates alerts when metrics exceed thresholds. Clarify is used at training time to understand the model; Model Monitor is used in production to detect when the model or data has changed. Both can work together — Clarify can baseline the bias metrics that Model Monitor then tracks.

Deciding signal

Kinesis Data Streams requires you to build consumer applications that read records in real time. It supports multiple consumers, configurable retention, and replay — giving full control over how streaming data is processed before it reaches a model or feature store. Kinesis Data Firehose automatically delivers data to a destination (S3, Redshift, OpenSearch) with optional Lambda transformation — no consumer code required. For ML pipelines, Data Streams is the right answer when the scenario involves real-time feature computation before writing to a Feature Store. Firehose is the right answer when streaming data should land in S3 for batch training with minimal code.