Amazon SageMaker Deep Dive
Amazon SageMaker is AWS's fully managed ML platform covering the entire lifecycle from data preparation to model monitoring. The AIF-C01 exam tests your ability to map each SageMaker feature to the right scenario.
---
What is Amazon SageMaker?
A fully managed service that lets data scientists and developers build, train, and deploy ML models without managing servers. AWS handles the infrastructure automatically.
SageMaker covers: data collection and preparation, model build and training, hyperparameter tuning, model deployment, and performance monitoring.
---
Built-in ML Algorithms
Ready-to-use algorithms without writing code from scratch: Supervised learning: linear regression, classification, KNN Unsupervised learning: PCA, K-means, anomaly detection Specialized: NLP text summarization, image classification and detection
---
Automatic Model Tuning (AMT)
AMT automates hyperparameter optimization. Define a target metric, and AMT searches the hyperparameter space to find the best combination. Use this whenever the exam mentions automated hyperparameter optimization.
---
4 Deployment Types Compared
| Type | Latency | Max Payload | Best For | |------|---------|-------------|----------| | Real-Time | ms–seconds | 6 MB | Fraud detection, live recommendations | | Serverless | low (cold start) | 4 MB | Intermittent chatbots | | Asynchronous | minutes–hours | 1 GB | Medical imaging, large file processing | | Batch Transform | slowest | unlimited | Whole-dataset predictions |
---
Key Features Summary
Studio: unified IDE for all SageMaker features Data Wrangler: no-code data preparation and feature engineering Feature Store: centralized feature storage and sharing across teams Clarify: bias detection and model explainability Ground Truth: human labeling of training data (RLHF) Model Monitor: production drift detection Model Registry: version control and team sharing Pipelines: ML CI/CD automation JumpStart: pre-trained model hub (more models than Bedrock) Canvas: no-code ML for non-technical users
---
Exam Quick Reference
4 deployment types and when to use each Ground Truth = label training data; A2I = human review of low-confidence predictions Clarify = bias + explainability; Model Monitor = production drift JumpStart = developer hub; Canvas = no-code Fine-tuning cost order: Prompt Engineering < RAG < Instruction < Domain Adaptation