monitoring/edge-monitoring-server/templates/prometheusrule.yaml

# Copyright 2020-present Open Networking Foundation
# SPDX-License-Identifier: LicenseRef-ONF-Member-Only-1.0

{{- if .Values.alerts.enabled }}
apiVersion: monitoring.coreos.com/v1
kind: PrometheusRule
metadata:
  name: {{ include "edge-monitoring-server.fullname" . }}
  labels:
    {{- include "edge-monitoring-server.labels" . | nindent 4 }}
spec:
  groups:
    - name: ace-e2e-tests-v2.rules
      rules:
      - alert: ScheduledDowntime
        annotations:
          message: The cluster {{`{{ .Labels.name }}`}} is undergoing scheduled maintenance.
        expr: aetheredge_in_maintenance_window{endpoint="metrics80"} > 0
        for: 1m
        labels:
          severity: info
      - alert: SingleEdgeTestNotReporting
        annotations:
          message: |
            The E2E test on cluster {{`{{ .Labels.name }}`}} has not reported results for at least 5 minutes.
        expr: (time() - aetheredge_last_update{endpoint="metrics80"}) > 300
        for: 1m
        labels:
          severity: critical
      - alert: SingleEdgeConnectTestFailing
        annotations:
          message: |
            The E2E test on cluster {{`{{ .Labels.name }}`}} is reporting UE connect failure for at least 10 minutes.
        expr: aetheredge_connect_test_ok{endpoint="metrics80"} < 1
        for: 10m
        labels:
          severity: critical
      - alert: SingleEdgePingTestFailing
        annotations:
          message: |
            The E2E test on cluster {{`{{ .Labels.name }}`}} is reporting that UE cannot ping the Internet for at least 10 minutes.
        expr: aetheredge_ping_test_ok{endpoint="metrics80"} < 1
        for: 11m
        labels:
          severity: critical
{{- if .Values.alerts.manyEdgeConnectTestsFailing }}
      - alert: ManyEdgeConnectTestsFailing
        annotations:
          message: |
            Over half of the clusters are reporting UE connect failures.
        expr: avg(clamp_max(aetheredge_connect_test_ok{endpoint="metrics80"} + aetheredge_in_maintenance_window{endpoint="metrics80"}, 1)) < 0.5
        for: 5m
        labels:
          severity: critical
{{- end }}
{{- end }}