X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/aad368fcb42c6955ccadbabe1f22aca5d31730dd..0d12afe2ac536a6e7696800798b2e553f36e1a44:/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 be6a89b43..0aa0b64cb 100644
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -22,10 +22,19 @@ groups:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           busy_workers: "{{ $value | humanizePercentage }}"
+  - name: chef
+    rules:
+      - alert: chef client not running
+        expr: time() - node_systemd_timer_last_trigger_seconds{name="chef-client.timer"} > 3600
+        for: 12h
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          down_time: "{{ $value | humanizeDuration }}"
   - name: cpu
     rules:
       - alert: cpu pressure
-        expr: rate(node_pressure_cpu_waiting_seconds_total[5m]) > 0.3
+        expr: rate(node_pressure_cpu_waiting_seconds_total[5m]) > 0.6
         for: 15m
         labels:
           alertgroup: "{{ $labels.instance }}"
@@ -367,10 +376,15 @@ groups:
   - name: systemd
     rules:
       - alert: systemd failed service
-        expr: node_systemd_unit_state{state="failed"} == 1
+        expr: node_systemd_unit_state{state="failed",name!="chef-client.service"} == 1
         for: 5m
         labels:
           alertgroup: "{{ $labels.instance }}"
+      - alert: systemd failed service
+        expr: node_systemd_unit_state{state="failed",name="chef-client.service"} == 1
+        for: 6h
+        labels:
+          alertgroup: "{{ $labels.instance }}"
   - name: tile
     rules:
       - alert: renderd replication delay