Additional Amazon SageMaker AI features — Studio, MLflow, MLOps, Feature Store, and more.
SageMaker Studio
Web-based IDE for end-to-end ML development:
| Feature | Details |
|---|
| Unified interface | All ML tools in one place |
| JupyterLab | Familiar notebook environment |
| Code Editor | Full IDE based on Code-OSS (VS Code) |
| Customizable | Bring your own images and kernels |
| Collaboration | Shared spaces for teams |
Studio Components
| Tool | Purpose |
|---|
| Notebooks | Interactive development |
| Experiments | Track and compare runs |
| Pipelines | Visual workflow builder |
| Model Registry | Version and manage models |
| Debugger | Debug training jobs |
| Profiler | Identify bottlenecks |
MLflow Integration
Experiment tracking and model management:
| Feature | Details |
|---|
| Experiment tracking | Log parameters, metrics, artifacts |
| Run comparison | Compare experiments side-by-side |
| Model registry | Version control for models |
| Managed service | No infrastructure to manage |
| Open source | Compatible with standard MLflow API |
What You Can Track
- Training parameters
- Model metrics (accuracy, loss, etc.)
- Model artifacts
- Code versions
- Environment configurations
MLflow is integrated into SageMaker Studio — no separate setup needed.
MLOps (Machine Learning Operations)
End-to-end ML lifecycle management:
SageMaker Pipelines
| Feature | Details |
|---|
| Visual builder | Drag-and-drop workflow design |
| CI/CD for ML | Automated training and deployment |
| Step types | Processing, training, tuning, inference |
| Orchestration | Manage complex dependencies |
| Versioning | Track pipeline versions |
Model Registry
| Feature | Details |
|---|
| Central catalog | All models in one place |
| Version control | Track model iterations |
| Approval workflows | Staged deployments |
| Metadata | Store training info, metrics |
| Lineage | Track model provenance |
Model Monitor
| Feature | Details |
|---|
| Data drift | Detect input distribution changes |
| Model quality | Monitor prediction accuracy |
| Bias drift | Track fairness metrics over time |
| Feature attribution | Explain predictions |
| Alerts | CloudWatch integration |
Feature Store
Centralized repository for ML features:
| Feature | Details |
|---|
| Single source of truth | Consistent features across training and inference |
| Offline store | S3-based for training |
| Online store | Low-latency for inference |
| Feature groups | Organized collections |
| Time travel | Historical feature values |
| Sharing | Cross-team feature reuse |
Benefits
- Reduce duplicate feature engineering
- Ensure train-serve consistency
- Enable feature discovery and reuse
- Simplify compliance and auditability
Clarify (Explainability & Bias)
ML explainability and fairness:
Key Characteristics
| Feature | Details |
|---|
| Model-agnostic | Works with any ML model, not framework-specific |
| Pre and post-deployment | Explain model behavior before AND after deployment |
| Per-instance explanations | Explain individual predictions during inference |
Bias Detection
| Stage | What It Checks |
|---|
| Pre-training | Bias in training data |
| Post-training | Bias in model predictions |
| Continuous | Bias drift over time |
Explainability Methods
| Method | Use Case |
|---|
| SHAP | Feature importance — based on Shapley values (game theory) |
| Partial Dependence Plots (PDPs) | Show marginal effect of features on predictions |
| Model cards | Document model behavior and characteristics |
Important Point: Know the difference:
- Clarify = Explainability — why did the model predict this? (SHAP, feature attribution)
- Model Monitor = Drift detection — is the model still accurate? (data drift, quality monitoring)
Debugger & Profiler
SageMaker Debugger
| Feature | Details |
|---|
| Real-time monitoring | Watch training as it happens |
| Built-in rules | Detect common issues (vanishing gradients, etc.) |
| Custom rules | Define your own debugging logic |
| Automatic actions | Stop training on issues |
SageMaker Profiler
| Feature | Details |
|---|
| Resource utilization | CPU, GPU, memory, I/O |
| Bottleneck detection | Find training slowdowns |
| Recommendations | Optimization suggestions |
| Timeline view | Visual profiling |
Edge & IoT Deployment
SageMaker Edge
| Feature | Details |
|---|
| Edge Manager | Deploy and manage edge models |
| Neo | Optimize models for edge hardware |
| Model packaging | Compile for specific devices |
| OTA updates | Update deployed models |
Supported Devices
- NVIDIA Jetson
- Intel OpenVINO
- ARM processors
- Custom hardware
Autopilot (AutoML)
Automated machine learning:
| Feature | Details |
|---|
| Auto feature engineering | Automatic data prep |
| Algorithm selection | Tests multiple algorithms |
| Hyperparameter tuning | Automatic optimization |
| Explainability | Understand auto-generated models |
| Notebooks | See what Autopilot did |
Use Cases
- Quick baseline models
- Non-ML experts building models
- Rapid prototyping
- Feature engineering ideas
TL;DR
- Studio = Web IDE with notebooks, code editor, visual tools
- MLflow = Managed experiment tracking (no infrastructure)
- Pipelines = CI/CD for ML workflows
- Model Registry = Version control and approval workflows
- Model Monitor = Drift detection, quality monitoring
- Feature Store = Centralized features for train/serve consistency
- Clarify = Bias detection and explainability (SHAP)
- Debugger/Profiler = Real-time training insights
- Edge = Deploy to IoT and edge devices with Neo
- Autopilot = AutoML for quick baseline models
Resources
SageMaker Studio 🔴
Web-based ML development environment.
SageMaker MLOps 🔴
End-to-end ML lifecycle management.
SageMaker Feature Store 🔴
Centralized feature repository.
SageMaker Clarify 🔴
Explainability and bias detection.