references/aks-anomaly.md
Azure Kubernetes Service (AKS) Cost and Utilization Anomalies
Detect unexplained AKS cost changes or investigate a supplied cluster or namespace window. Keep billed-cost evidence separate from operational signals.
Choose Mode
| Request | Mode | |---|---| | Did anything unusual happen? | Detect across about 60 complete billing days | | User supplies a date or window | Investigate that window plus 28-56 preceding days |
Confirm the cluster or namespace, timezone, and exact interval. Use only complete billing days and report each currency separately.
Collect Cost Evidence
- Follow AKS cost analysis and the
cost query workflow.
query_aks_costsaccepts only/subscriptions/{subscriptionId}and supports daily or monthly granularity, not hourly cost attribution. - Use
query_aks_costsfor cluster or namespace allocation, keeping the same returned cost column (CostorCostUSD) and currency for the baseline and anomaly. Preserve the column, period, currency, and coverage.query_aks_costscannot report Reservation or Savings Plan charges or benefits separately by namespace; route that analysis tocost-optimization. - Query cluster allocation first, then the affected cluster and namespaces.
Separate real namespaces from
#idle charges#,#service charges#,#system charges#, and#unallocated charges#. - For resource-level contributors such as VM scale sets, disks, networking,
or monitoring, use the narrowest supported
query_costscall or approved fallback process. Use the same metric, scope, filters, and currency for both periods. Do not assign a shared workspace's full cost to one cluster.
Missing add-on data, inaccessible scopes, incomplete days, and absent rows are unavailable evidence, not zero cost.
Detect Candidate Windows
Prefer native Cost Management anomaly insights when the available tool exposes them. Otherwise compare complete daily values with comparable preceding weekdays. Derive expected and observed cost from the same operation, cost column or metric, scope, filters, and currency. Report absolute and percentage delta.
Use the anomaly score returned by native Cost Management anomaly insights when available, preserving its reported meaning and threshold. Otherwise, report the absolute and percentage difference from comparable preceding weekdays and describe the result as a candidate anomaly. For insufficient or strongly seasonal history, report the limitation instead of forcing a baseline.
Explain Contributors
Query only candidate and comparable baseline windows:
- Compare namespace allocation and idle/system/service/unallocated categories.
- Identify resource-level cost deltas when the available dimensions support it.
- Compare workload creation and schedules, resource-request or limit changes, autoscaler decisions, disruption controls, node-local blockers, and scale-down eligibility before and during the candidate window.
- Correlate scaling, upgrades, SKU, node-pool, storage, monitoring, and network changes through Resource Graph change history. Validate generated queries before execution; empty history is inconclusive.
- Use historical Azure Monitor or Container Insights node count, workload, autoscaler, storage, and ingestion evidence when available.
- Treat current Kubernetes metrics only as evidence that a condition persists, not as proof of a historical cause.
Distinguish verified from likely contributors, quantify attributable portions when supported, and report any material unexplained remainder.
Route Follow-up
| Finding | Handoff |
|---|---|
| Pod rightsizing, VPA, Cluster Autoscaler, or Spot configuration | azure-kubernetes when available |
| NAP readiness or node-pool fragmentation | aks-cost-optimization when available |
| PVC/PV-to-Azure storage optimization | aks-cost-optimization when available |
| Reservation or Savings Plan effect | cost-optimization |
| Namespace allocation unavailable | AKS enablement |
| Budget or alert creation | cost-governance |
Do not resize pools, change autoscaling, enable add-ons, create alerts, or modify workloads without the owning workflow's required confirmation.
Output
For each anomaly, report the period and timezone, baseline method and coverage, cost column or metric and currency, expected and observed cost, absolute and percentage delta, affected cluster/namespaces/resources, verified and likely contributors, unexplained remainder, evidence gaps, confidence, and next handoff.