4.3.3 · Lesson 13 of 16
Alert on Cost, Latency, and Availability Regressions
Create CloudWatch alarms and SNS notifications when token spend, latency, throttling, or interventions regress.
WHAT
Alert automatically when GenAI cost, latency, throttling, or quality signals regress.
HOW
- Metric filters on token counts and latency from invocation logs.
- Alarms on throttling and intervention metrics.
- SNS to operations/finops when thresholds breach.
InputTokenCount spike
↓
CloudWatch Alarm
↓
SNS → FinOps / OpsKey points
- CloudWatch Metric Filters
- SNS alerts
- Token budget thresholds
- Throttling alarms
- Intervention spikes
Exam focus
Pair invocation metrics with SNS alarms — e.g., alert when InputTokenCount exceeds a budget window.