1.3.2. SageMaker Components and When to Use Them
💡 First Principle: Each SageMaker component solves one specific problem in the ML lifecycle. The exam tests your ability to match the problem description to the correct component—not your knowledge of SageMaker's marketing materials.
| Component | Lifecycle Stage | Primary Use | Key Exam Signals |
|---|---|---|---|
| Data Wrangler | Data Prep | Visual data prep, feature engineering | "Minimal code," "visual interface," "data exploration" |
| Feature Store | Data Prep | Centralized feature repository | "Reuse features," "online/offline store," "feature consistency" |
| Ground Truth (closed to new customers) | Data Prep | Human data labeling workflows | "Labeling," "annotation," "human reviewers" |
| Training Jobs | Model Dev | Scalable model training | "Train at scale," "distributed training," "GPU instances" |
| Automatic Model Tuning (AMT) | Model Dev | Hyperparameter optimization | "Hyperparameter tuning," "Bayesian optimization," "best parameters" |
| JumpStart | Model Dev | Pre-trained models and solutions | "Pre-trained," "transfer learning," "quick start" |
| Canvas | Data Prep + Model Dev | No-code model building over your own data | "No ML team," "business analyst," "no code," "visual," new time-series forecasting |
| Autopilot | Model Dev | AutoML for tabular classification and regression | "AutoML," "automatically select the algorithm," "tabular," candidate leaderboard |
| Experiments | Model Dev | Track and compare training runs | "Compare experiments," "track metrics," "reproducibility" |
| Model Registry | Model Dev → Deploy | Model versioning and approval | "Model version," "approval workflow," "model governance" |
| Clarify (closed to new customers) | Data Prep + Model Dev + Monitor | Bias metrics and SHAP explanations | "Bias," "explainability," "SHAP values," "fairness" |
| Debugger (closed to new customers) | Model Dev | Training-time convergence rules and profiling | "Vanishing gradients," "exploding gradients," "overfitting during training" |
| TensorBoard in SageMaker AI | Model Dev | Inspect gradients, weights and activations across training steps | "Gradient histograms," "weight distributions," "vanishing or exploding gradients" |
| Endpoints | Deployment | Real-time inference serving | "Real-time predictions," "low latency," "endpoint" |
| Batch Transform | Deployment | Large-scale offline inference | "Batch predictions," "no persistent endpoint," "offline scoring" |
| Pipelines | Deployment | ML workflow orchestration | "Automate workflow," "DAG," "orchestrate steps" |
| Neo | Deployment | Model optimization for edge | "Edge deployment," "optimize model," "IoT" |
| Model Monitor (closed to new customers) | Monitoring | Scheduled data/model drift checks | "Drift," "data quality," "production monitoring," "baseline" |
| Inference Recommender | Monitoring | Endpoint sizing recommendations | "Rightsizing," "instance selection," "load testing" |
📌 Availability note (checked against AWS documentation, October 2026): Amazon SageMaker Clarify, Model Monitor, Ground Truth, Debugger and Augmented AI (A2I) have all been closed to new customers since 30 July 2026. Existing customers can continue to use them as normal, and AWS still makes security and availability improvements, but it plans no new features. The MLA-C02 exam guide no longer names Clarify, Model Monitor or Ground Truth; it phrases those skills as capabilities: "explain model outputs", "create model performance baselines and implement drift detection", and labeling and annotation. The exam guide likewise phrases the Debugger skill as "debug model convergence issues" and the A2I one as "human-in-the-loop workflows", without naming a tool. The concepts behind all five (bias metrics, SHAP, baselines, Data Capture, drift types, and gradient and convergence failure patterns) are still exam content. What changes is the tool a new team adopts:
| Capability | Closed tool | What AWS recommends for a new team |
|---|---|---|
| Pre- and post-training bias metrics (CI, DPL, DI, DPPL) | Clarify | Compute the published formulas with pandas and scikit-learn in a SageMaker Pipelines processing step, and log results to a SageMaker AI MLflow App |
| Explaining predictions (feature attribution) | Clarify | The open-source SHAP library, the same engine Clarify uses internally |
| Data quality and model quality drift | Model Monitor | AWS's open-source SageMaker AI monitoring solutions (Evidently AI and MLflow, run by EventBridge and Lambda or a scheduled SageMaker Pipeline, with SNS alerts), plus CloudWatch metrics and anomaly detection alarms |
| Bias drift and feature attribution drift | Model Monitor with Clarify | The same monitoring solutions, recomputing per-segment fairness metrics and SHAP importance on a schedule |
| Training convergence debugging and profiling (built-in rules, profiler) | Debugger | A SageMaker AI MLflow App for metric logging and run comparison, TensorBoard in SageMaker AI for gradient and weight histograms, and CloudWatch training-job metrics and alarms for utilization and automated alerts (see 3.3.3) |
| Foundation model evaluation | Clarify (FMEval) | The fmeval library or Amazon Bedrock Evaluations, with Amazon Bedrock Guardrails for runtime safety. Neither replaces tabular bias or explainability analysis |
| Human data labeling | Ground Truth | AWS documents no replacement. Existing customers continue with a private or vendor workforce |
| Human review of model predictions | A2I | AWS documents no direct replacement. To compare foundation-model outputs, Amazon Bedrock model evaluation jobs can use your own human workers (see 3.4.6) |
⚠️ Exam Trap: Bias analysis appears in two lifecycle stages: pre-training bias detection on the dataset (Stage 1) and bias drift on production predictions (Stage 4), whichever tool computes it. Similarly, Feature Store has an online store (low-latency inference lookups) and an offline store (batch training data). Questions test whether you understand these dual roles. When a scenario says the team is new to a service or is building in a new AWS account, rule out tools that are closed to new customers.
Reflection Question: A team notices their model's predictions disproportionately affect one demographic group. Which bias metrics and which AWS approach should they use, and does it matter whether the model is in development or already deployed?
