Add IPv6 addresses to dulcy

[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 7afa799e8ad2b234f9537b2c01ef80ff4be8aa86..6751fe74423f2b386e78dd82176cc0c400c19d2a 100644 (file)
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -25,14 +25,14 @@ groups:
         annotations:
           current: "{{ $value | humanize }}kVA"
       - alert: site temperature
         annotations:
           current: "{{ $value | humanize }}kVA"
       - alert: site temperature

-        expr: min(rPDU2SensorTempHumidityStatusTempC{site="amsterdam"}) / 10 < 18 or min(rPDU2SensorTempHumidityStatusTempC{site="amsterdam"}) / 10 > 26
+        expr: min(rPDU2SensorTempHumidityStatusTempC{site="amsterdam"}) / 10 < 15 or min(rPDU2SensorTempHumidityStatusTempC{site="amsterdam"}) / 10 > 32

         for: 6m
         labels:
           alertgroup: "amsterdam"
         annotations:
           temperature: "{{ $value | humanize }}C"
       - alert: site humidity
         for: 6m
         labels:
           alertgroup: "amsterdam"
         annotations:
           temperature: "{{ $value | humanize }}C"
       - alert: site humidity

-        expr: max(rPDU2SensorTempHumidityStatusRelativeHumidity{site="amsterdam"}) / 100 < 0.25 or max(rPDU2SensorTempHumidityStatusRelativeHumidity{site="amsterdam"}) / 100 > 0.65
+        expr: max(rPDU2SensorTempHumidityStatusRelativeHumidity{site="amsterdam"}) / 100 < 0.08 or max(rPDU2SensorTempHumidityStatusRelativeHumidity{site="amsterdam"}) / 100 > 0.8

         for: 6m
         labels:
           alertgroup: "amsterdam"
         for: 6m
         labels:
           alertgroup: "amsterdam"


@@ -118,7 +118,7 @@ groups:
         labels:
           alertgroup: database
         annotations:
         labels:
           alertgroup: database
         annotations:

-          delay: "{{ $value }}"
+          queries: "{{ $value }}"

   - name: discourse
     rules:
       - alert: discourse job failure rate
   - name: discourse
     rules:
       - alert: discourse job failure rate


@@ -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


@@ -655,6 +669,13 @@ groups:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           new_errors: "{{ $value }}"
           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
   - name: smart
     rules:
       - alert: smart failure


@@ -793,7 +814,7 @@ 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