workaround temporary statuscake IP list failure

[chef.git] / cookbooks / prometheus / templates / default / alert_rules.yml.erb

diff --git a/cookbooks/prometheus/templates/default/alert_rules.yml.erb b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
index fdb90765bf0618d8f704a00157e1499fce8d6b11..dfd885bb8055252fd72a99df4807eeba1f6c956c 100644 (file)
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -331,6 +331,15 @@ groups:
           alertgroup: "{{ $labels.site }}"
         annotations:
           power: "{{ $value }} dBm"
           alertgroup: "{{ $labels.site }}"
         annotations:
           power: "{{ $value }} dBm"

+  - name: load
+    rules:
+      - alert: load average
+        expr: sum(node_load5) by (instance) / count(node_cpu_frequency_max_hertz) by (instance) > 2
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          load: "{{ $value | humanizePercentage }}"

   - name: mail
     rules:
       - alert: exim down
   - name: mail
     rules:
       - alert: exim down


@@ -488,6 +497,11 @@ groups:
           alertgroup: nominatim
         annotations:
           delay: "{{ $value | humanizeDuration }}"
           alertgroup: nominatim
         annotations:
           delay: "{{ $value | humanizeDuration }}"

+      - alert: nominatim connections
+        expr: sum(nginx_connections_writing and on (instance) chef_role{name="nominatim"}) > 2500
+        for: 15m
+        labels:
+          alertgroup: nominatim

   - name: overpass
     rules:
       - alert: overpass osm database age
   - name: overpass
     rules:
       - alert: overpass osm database age


@@ -793,7 +807,7 @@ groups:
   - name: web
     rules:
       - alert: web error rate
   - name: web
     rules:
       - alert: web error rate

-        expr: sum(rate(api_call_count_total{status=~"50[0-8]|5[1-9][0-9]"}[5m])) by (instance) / sum(rate(api_call_count_total[5m])) by (instance) > 0.002
+        expr: sum(rate(api_call_count_total{status=~"50[0-8]|5[1-9][0-9]"}[5m])) by (instance) / sum(rate(api_call_count_total[5m])) by (instance) > 0.002 and sum(rate(api_call_count_total{status=~"50[0-8]|5[1-9][0-9]"}[5m])) by (instance) > 0.01

         for: 5m
         labels:
           alertgroup: web
         for: 5m
         labels:
           alertgroup: web