X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/c51d171917d4bbba2b6221128ca1cebb863870cc..06f75772a25febcfd00bfb8607ce3931520adb22:/cookbooks/prometheus/templates/default/alert_rules.yml.erb?ds=inline

diff --git a/cookbooks/prometheus/templates/default/alert_rules.yml.erb b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
index 396de8de4..d717e4f0a 100644
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -4,7 +4,7 @@ groups:
   - name: amsterdam
     rules:
       - alert: uplink
-        expr: ifOperStatus{site="amsterdam",ifName=~"ge-[01]/2/2"} != 1
+        expr: junos_interface_up{site="amsterdam",name=~"ge-[01]/2/2"} != 1
         for: 6m
         labels:
           alertgroup: "amsterdam"
@@ -117,7 +117,7 @@ groups:
   - name: dublin
     rules:
       - alert: uplink
-        expr: ifOperStatus{site="dublin",ifName=~"ge-[01]/2/2"} != 1
+        expr: junos_interface_up{site="dublin",name=~"ge-[01]/2/2"} != 1
         for: 6m
         labels:
           alertgroup: "dublin"
@@ -160,13 +160,23 @@ groups:
           alertgroup: fastly
         annotations:
           error_rate: "{{ $value | humanizePercentage }}"
-      - alert: fastly healthcheck failing
-        expr: count(fastly_healthcheck_status == 0) by (service) > 0
+      - alert: fastly frontend healthcheck warning
+        expr: count(fastly_healthcheck_status == 0) by (service, datacenter) > 2
         for: 15m
         labels:
           alertgroup: fastly
-      - alert: multiple fastly healthchecks failing
-        expr: count(fastly_healthcheck_status == 0) by (service) > 4
+      - alert: fastly frontend healthcheck critical
+        expr: count(fastly_healthcheck_status == 0) by (service, datacenter) == count(fastly_healthcheck_status) by (service, datacenter)
+        for: 5m
+        labels:
+          alertgroup: fastly
+      - alert: fastly backend healthcheck warning
+        expr: count(fastly_healthcheck_status == 0) by (service, backend) > 10
+        for: 15m
+        labels:
+          alertgroup: fastly
+      - alert: fastly backend healthcheck critical
+        expr: count(fastly_healthcheck_status == 0) by (service, backend) == count(fastly_healthcheck_status) by (service, backend)
         for: 5m
         labels:
           alertgroup: fastly
@@ -262,20 +272,34 @@ groups:
           alertgroup: "{{ $labels.instance }}"
   - name: juniper
     rules:
+      - alert: juniper red alarms
+        expr: juniper_alarms_red_count > 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.site }}"
+        annotations:
+          alarm_count: "{{ $value }} alarms"
+      - alert: juniper yellow alarms
+        expr: juniper_alarms_yellow_count > 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.site }}"
+        annotations:
+          alarm_count: "{{ $value }} alarms"
       - alert: juniper cpu alarm
-        expr: jnxOperating5MinLoadAvg{jnxOperatingContentsIndex="9"} / 200 > 0.5
+        expr: junos_route_engine_load_average_five / 2 > 0.5
         for: 5m
         labels:
           alertgroup: "{{ $labels.site }}"
         annotations:
           load_average: "{{ $value | humanizePercentage }}"
       - alert: juniper fan alarm
-        expr: jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!~"running.*"} > 0
+        expr: junos_environment_fan_up != 1
         for: 5m
         labels:
           alertgroup: "{{ $labels.site }}"
       - alert: juniper power alarm
-        expr: jnxOperatingState{jnxOperatingContentsIndex="2",jnxOperatingState!~"running.*"} > 0
+        expr: junos_environment_power_up != 1
         for: 5m
         labels:
           alertgroup: "{{ $labels.site }}"