1.5. Reflection Checkpoint: Foundational ML & AWS ML Landscape

Key Takeaways

• The ML workflow lifecycle (problem definition → data ingestion → preparation → EDA → feature engineering → training → evaluation → deployment → monitoring → MLOps) is the map you use to place any AWS service into context — most exam questions live in data preparation, training, evaluation, and deployment.
• Data quality and bias management aren't separate concerns — poor data quality and bias (selection, measurement, historical, algorithmic) both produce inaccurate or unfair models regardless of how sophisticated the algorithm is; SageMaker Clarify detects bias both pre- and post-training.
• Algorithm choice follows from problem type (regression/classification/clustering) and data characteristics, while evaluation metric choice follows from the business cost of false positives vs. false negatives — picking the wrong metric can make a bad model look good.
• Scalability and performance decisions split into three tracks: data (S3 + distributed processing), training (distributed training, Spot Instances, GPU instance selection), and inference (real-time for low latency, batch for high-throughput non-urgent workloads, async for large payloads).
• ML security and governance layers exactly like general AWS security (encryption, IAM least privilege, VPC isolation, CloudTrail/Config auditing) but adds ML-specific governance: Model Registry for version/lineage tracking and SageMaker Clarify for responsible AI.
• MLOps applies DevOps principles (automation, reproducibility, CI/CD, version control) to the ML lifecycle specifically so models can be retrained and redeployed automatically when drift is detected.
• The AWS ML stack has three layers matched to three skill levels: AI Services (pre-trained APIs — no ML expertise needed), SageMaker (managed custom model building), and foundational compute/storage/networking (EC2, S3, VPC) that everything else runs on.

Connecting Forward

Phase 1 built your map of the ML lifecycle and the AWS services available at each stage. Phase 2 zooms into the first real stage of that lifecycle in depth: data engineering — how you actually ingest, store, transform, and govern the data that everything downstream depends on.

Self-Check Questions

• A loan-approval model was trained on historical data that reflects past lending decisions, which themselves were affected by discriminatory practices. Which type of bias is this, and which AWS tool would you use to detect it before training?
• Your team needs to score 10 million records overnight for a report due the next morning — no user is waiting on an individual prediction. Which SageMaker inference option fits, and why would a real-time endpoint be the wrong (and more expensive) choice here?