fix(infra-monitoring): add alerts for blackbox-exporter

infra-monitoring/templates/exporter/blackbox/prometheus-rules.yaml

@@ -0,0 +1,119 @@
+{{- if and .Values.prometheus.exporter.blackbox.enabled (.Capabilities.APIVersions.Has "monitoring.coreos.com/v1/PrometheusRule") }}
+{{- $filter := `job=~"probe/.*"` }}
+---
+apiVersion: "monitoring.coreos.com/v1"
+kind: "PrometheusRule"
+metadata:
+  name: prometheus-blackbox-exporter
+  labels:
+    {{- toYaml .Values.commons.prometheus.rules.labels | nindent 4 }}
+spec:
+  groups:
+    - name: "Exporter It-Self"
+      rules:
+        - alert: "Blackbox-Exporter down"
+          expr: 'up{ {{ $filter }} } == 0'
+          for: "5m"
+          labels:
+            severity: "critical"
+            detectedBy: "Blackbox"
+          annotations:
+            {{`
+            summary: "Blackbox-Exporter to probe {{  $labels.instance | stripPort }} is not reachable for more than 5 minutes."
+            `}}
+
+        - alert: "BlackboxProbeFailed"
+          expr: 'probe_success{ {{ $filter }}, instancetype!="server" } == 0'
+          for: "5m"
+          labels:
+            severity: "critical"
+            detectedBy: "Blackbox"
+          annotations:
+            {{`
+            summary: "Probe of the target {{  $labels.instance | stripPort }} failed."
+            `}}
+
+    - name: "ICMP"
+      rules:
+        - alert: "BlackboxProbeIcmpFailed"
+          expr: 'probe_success{ {{ $filter }}, instancetype="server" } == 0'
+          for: "5m"
+          labels:
+            severity: "critical"
+            detectedBy: "Blackbox"
+          annotations:
+            {{`
+            summary: "Icmp request of the target {{  $labels.instance | stripPort }} failed for more than 5 minutes."
+            `}}
+
+        - alert: "HostnameCouldNotBeResolvedByDNS"
+          expr: 'probe_icmp_duration_seconds{ {{ $filter }}, phase="resolve"} == 0'
+          for: "5m"
+          labels:
+            severity: "critical"
+            detectedBy: "Blackbox"
+          annotations:
+            {{`
+            summary: "Instance {{  $labels.instance | stripPort }} could not be resolved by blackbox-exporter while ICMP requests."
+            description: "Could not resolve hostname for more than 5 minutes."
+            `}}
+
+        - alert: "SlowPing"
+          expr: 'avg_over_time(probe_icmp_duration_seconds{ {{ $filter }} }[1m]) > 1'
+          for: "5m"
+          labels:
+            severity: "warning"
+            detectedBy: "Blackbox"
+          annotations:
+            {{`
+            summary: "ping duration is {{ $value }} s for {{  $labels.instance | stripPort }}."
+            `}}
+
+    - name: "SSL Certificates"
+      rules:
+        - alert: "CertificateAboutToExpire"
+          expr: 'probe_ssl_earliest_cert_expiry{ {{ $filter }} } - time() <= 86400'
+          for: "1m"
+          labels:
+            severity: "critical"
+            detectedBy: "Blackbox"
+          annotations:
+            {{`
+            summary: "SSL certificate for domain {{  $labels.instance }} has almost expired"
+            `}}
+
+        - alert: "CertificateAboutToExpire"
+          expr: 'probe_ssl_earliest_cert_expiry{ {{ $filter }} } - time() < 86400 * 6'
+          for: "1m"
+          labels:
+            severity: "warning"
+            detectedBy: "Blackbox"
+          annotations:
+            {{`
+            summary: "SSL certificate for domain {{  $labels.instance }} will expire in {{ $value | humanizeDuration }}."
+            `}}
+
+    - name: "HTTP Probes"
+      rules:
+        - alert: "HttpStatusCode"
+          expr: 'probe_http_status_code{ {{ $filter }}, module != "http_404" } <= 199 OR probe_http_status_code{ {{ $filter }}, module != "http_404" } >= 400'
+          for: "5m"
+          labels:
+            severity: "critical"
+            detectedBy: "Blackbox"
+          annotations:
+            {{`
+            summary: "HTTP status code for domain {{  $labels.instance | stripPort }} is {{ $value }}"
+            `}}
+
+        - alert: "HttpSlowRequests"
+          expr: 'avg_over_time(probe_http_duration_seconds{ {{ $filter }} }[1m]) > 1'
+          for: "5m"
+          labels:
+            severity: "warning"
+            detectedBy: "Blackbox"
+          annotations:
+            {{`
+            summary: 'HTTP responce time for domain {{  $labels.instance | stripPort }} is {{ printf "%.0f" $value }} s.'
+            `}}
+{{- end }}