Machine Learning Implementation and Operations
Machine Learning Implementation and Operations
This domain focuses on deploying, monitoring, and securing ML models on AWS. Know SageMaker Endpoints for real-time inference, Batch Transform for offline, and SageMaker Pipelines for orchestration. Understand deployment strategies: blue/green, canary, and linear. For monitoring, use SageMaker Model Monitor to detect data drift and quality issues. Security: IAM roles, VPC configuration, encryption at rest (S3, EBS) and in transit (TLS). Also know SageMaker Debugger for training debugging and SageMaker Model Registry for versioning. For cost optimization, use SageMaker Savings Plans and auto-scaling endpoints.
Traps: Model Monitor requires a baseline (e.g., from training data) and a schedule. Endpoints can be auto-scaled based on invocations. Do not confuse Model Monitor with Debugger: Monitor is for post-deployment drift; Debugger is for training. Another trap: Batch Transform is for large batches and does not require a persistent endpoint. Also, for A/B testing, use production variants on an endpoint. Ensure least privilege IAM policies; do not use root.
Worked example: You deploy a model to a real-time endpoint and want to detect feature drift. Create a baseline from training data, enable Model Monitor with a hourly schedule, and set alerts via CloudWatch. For cost, enable auto-scaling with target tracking on invocations per instance. Rule of thumb: Always monitor endpoints for drift; use Batch Transform for infrequent large jobs; use Pipelines to automate retraining.