X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/bbbe7474976e33c5471f7ecdabd4cf2a44c25ee2..9169fe98d01e044dd4289099dc8d665089da9e30:/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 b78bbaddd..305afbd90 100644
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -3,13 +3,20 @@
 groups:
   - name: amsterdam
     rules:
-      - alert: uplink
+      - alert: he uplink
         expr: junos_interface_up{site="amsterdam",name=~"ge-[01]/2/2"} != 1
         for: 6m
         labels:
           alertgroup: "amsterdam"
         annotations:
           status: "{{ $value }}"
+      - alert: equinix uplink
+        expr: junos_interface_up{site="amsterdam",name=~"xe-[01]/2/0"} != 1
+        for: 6m
+        labels:
+          alertgroup: "amsterdam"
+        annotations:
+          status: "{{ $value }}"
       - alert: pdu current draw
         expr: rPDU2PhaseStatusCurrent{site="amsterdam",rPDU2PhaseStatusIndex="1"} / 10 > 28
         for: 6m
@@ -130,13 +137,20 @@ groups:
           failure_rate: "{{ $value }} jobs/s"
   - name: dublin
     rules:
-      - alert: uplink
+      - alert: he uplink
         expr: junos_interface_up{site="dublin",name=~"ge-[01]/2/2"} != 1
         for: 6m
         labels:
           alertgroup: "dublin"
         annotations:
           status: "{{ $value }}"
+      - alert: equinix uplink
+        expr: junos_interface_up{site="dublin",name=~"xe-[01]/2/0"} != 1
+        for: 6m
+        labels:
+          alertgroup: "dublin"
+        annotations:
+          status: "{{ $value }}"
       - alert: pdu current draw
         expr: rPDU2PhaseStatusCurrent{site="dublin",rPDU2PhaseStatusIndex="1"} / 10 > 28
         for: 6m
@@ -331,6 +345,15 @@ 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
@@ -660,6 +683,13 @@ 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