X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/96e2ac0bfb678221faaa0feba7039deeb853ddf0..50aea11082eaa90bd37593b9c987d38a082f4e1b:/cookbooks/prometheus/templates/default/alert_rules.yml.erb?ds=sidebyside

diff --git a/cookbooks/prometheus/templates/default/alert_rules.yml.erb b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
index fdb90765b..6751fe744 100644
--- 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"
+  - 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
@@ -488,6 +497,11 @@ groups:
           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
@@ -655,6 +669,13 @@ groups:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           new_errors: "{{ $value }}"
+  - name: resolved
+    rules:
+      - alert: dnssec validation failures
+        expr: rate(resolved_dnssec_verdicts_total{result="bogus"}[1m]) > 1
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
   - name: smart
     rules:
       - alert: smart failure
@@ -793,7 +814,7 @@ groups:
   - 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