X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/67c0ac7256b64455f3f95cc727485bbe05a10515..cc57b9e9a6f06eb2fdce5ed83aa5f9ed6cd90cad:/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 1736da60c..239731607 100644
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -105,6 +105,15 @@ groups:
           alertgroup: database
         annotations:
           delay: "{{ $value | humanizeDuration }}"
+  - name: discourse
+    rules:
+      - alert: discourse job failure rate
+        expr: rate(discourse_job_failures[5m]) > 0
+        for: 5m
+        labels:
+          alertgroup: discourse
+        annotations:
+          failure_rate: "{{ $value }} jobs/s"
   - name: dublin
     rules:
       - alert: pdu current draw
@@ -530,6 +539,11 @@ groups:
           alertgroup: "prometheus"
   - name: raid
     rules:
+      - alert: raid controller battery failed
+        expr: ohai_controller_info{battery_status="failed"} > 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
       - alert: raid array degraded
         expr: ohai_array_info{status="degraded"} > 0
         for: 5m
@@ -606,7 +620,7 @@ groups:
   - name: statuscake
     rules:
       - alert: statuscake uptime check failing
-        expr: statuscake_uptime{status="down",paused="false"} > 0
+        expr: statuscake_paused == 0 and statuscake_up == 0
         for: 10m
         labels:
           alertgroup: statuscake
@@ -618,8 +632,8 @@ groups:
         labels:
           alertgroup: "{{ $labels.instance }}"
       - alert: systemd failed chef client service
-        expr: node_systemd_unit_state{state="failed",name="chef-client.service"} == 1
-        for: 6h
+        expr: sum_over_time(node_systemd_unit_state{state="inactive",name="chef-client.service"}[6h]) == 0
+        for: 0m
         labels:
           alertgroup: "{{ $labels.instance }}"
   - name: tile