X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/daff29395a0b52fb994ff33bc2c7086f6a6436b6..e740c30b54cb87886c396d783432f12c6c9648a4:/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 ac45cc62c..be10977a4 100644
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -145,12 +145,12 @@ groups:
         annotations:
           error_rate: "{{ $value | humanizePercentage }}"
       - alert: fastly healthcheck failing
-        expr: count(fastly_healthcheck_status == 0) > 0
+        expr: count(fastly_healthcheck_status == 0) by (service) > 0
         for: 15m
         labels:
           alertgroup: fastly
-      - alert: multipe fastly healthchecks failing
-        expr: count(fastly_healthcheck_status == 0) > 4
+      - alert: multiple fastly healthchecks failing
+        expr: count(fastly_healthcheck_status == 0) by (service) > 4
         for: 5m
         labels:
           alertgroup: fastly
@@ -246,18 +246,28 @@ groups:
           alertgroup: "{{ $labels.instance }}"
   - name: juniper
     rules:
+      - alert: juniper cpu alarm
+        expr: jnxOperatingCPU{jnxOperatingContentsIndex="7"} > 30
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.site }}"
       - alert: juniper fan alarm
-        expr: sum_over_time(jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!="running"}[6m]) > 0
+        expr: jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!~"running.*"} > 0
         for: 5m
         labels:
           alertgroup: "{{ $labels.site }}"
       - alert: juniper power alarm
-        expr: sum_over_time(jnxOperatingState{jnxOperatingContentsIndex="2",jnxOperatingState!="running"}[6m]) > 0
+        expr: jnxOperatingState{jnxOperatingContentsIndex="2",jnxOperatingState!~"running.*"} > 0
         for: 5m
         labels:
           alertgroup: "{{ $labels.site }}"
   - name: mail
     rules:
+      - alert: exim down
+        expr: exim_up == 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
       - alert: exim queue length
         expr: exim_queue > exim_queue_limit
         for: 60m
@@ -383,8 +393,8 @@ groups:
   - name: overpass
     rules:
       - alert: overpass osm database age
-        expr: overpass_database_age_seconds{database="osm"} > 1800
-        for: 5m
+        expr: overpass_database_age_seconds{database="osm"} > 3600
+        for: 1h
         labels:
           alertgroup: overpass
         annotations:
@@ -396,6 +406,23 @@ groups:
           alertgroup: overpass
         annotations:
           age: "{{ $value | humanizeDuration }}"
+  - name: passenger
+    rules:
+      - alert: passenger down
+        expr: passenger_up == 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+      - alert: passenger queuing
+        expr: passenger_top_level_request_queue > 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+      - alert: passenger application queuing
+        expr: passenger_app_request_queue > 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
   - name: planet
     rules:
       - alert: planet dump overdue