ML Implementation & Operations

CloudTrail logs API calls/management events across the account, providing an audit trail of who did what (e.g., who created or invoked a SageMaker endpoint) and when.

CloudWatch collects metrics, logs, and supports alarms/dashboards - e.g., endpoint invocation counts, latency, CPU/GPU/memory utilization, and custom metrics.

CloudTrail answers 'who made which API call' (audit), while CloudWatch answers 'how is the system performing' (metrics/logs/alarms).

High Invocation4XXErrors/5XXErrors, rising ModelLatency and OverheadLatency, and high CPUUtilization/GPUUtilization or MemoryUtilization on the variant.

Emit logs/metrics to CloudWatch, create CloudWatch Alarms on error rates or failed pipeline steps, and trigger SNS notifications or Lambda remediation; use EventBridge for state-change events.