Exam Room · Advanced Generative AI Developer

Flash Card: AWS Cost Anomaly Detection

· 3 min read

Generative AI Development · part of The Exam Room

The three horizons sit under the guardrail layering: a CloudWatch alarm in minutes, an anomaly monitor in about a day, a budget across the month. This card is the middle one.

Flash card

AWS Cost Anomaly Detection: machine-learning detection of unusual spend against a baseline it learns for itself, delivered as part of AWS Cost Explorer at no charge.

  1. Monitors are scoped to an AWS service (Bedrock), to a cost allocation tag, to a cost category, or to a linked account, so a per-feature monitor becomes possible once inference profiles carry tags.
  2. It learns each monitor’s own pattern instead of comparing spend to a number somebody chose, so a seasonal or steadily growing workload does not alarm on its own growth.
  3. Alerts carry a dollar-impact threshold and go out by email or Amazon SNS, either individually or batched as a daily or weekly summary.
  4. Detection runs on billing data and lands roughly a day behind the spend, so cost anomaly detection is not the control that stops a runaway retry loop.
  5. It costs nothing to run, which makes a monitor per major workload cheap enough to be the default rather than a decision.

Pick it when

Pick it over a budget when what you want is “tell me when this looks unusual for us” rather than “tell me when this passes a number”, and over a CloudWatch alarm when the signal you want is dollars rather than tokens or invocations.

It's the wrong answer when

It is the wrong answer for a same-minute reaction to a retry storm, which wants a CloudWatch alarm on Invocations or token counts; for enforcing a hard ceiling, which wants a budget action, a service quota, or an application rate limiter; and for a workload with no spend history for the model to learn from.

These posts are LLM-aided. Backbone, original writing, and structure by Craig. Research and editing by Craig + LLM. Proof-reading by Craig.