X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/7dade8a248c510d4ee5ed24b505fe8ed1e299d41..28326173066466bb20f54bad702b2885b64f925c:/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 6a9ad6120..8c1aeee48 100644
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -105,6 +105,36 @@ groups:
           alertgroup: database
         annotations:
           delay: "{{ $value | humanizeDuration }}"
+  - name: dublin
+    rules:
+      - alert: pdu current draw
+        expr: rPDU2PhaseStatusCurrent{site="dublin",rPDU2PhaseStatusIndex="1"} / 10 > 13
+        for: 5m
+        labels:
+          alertgroup: "dublin"
+        annotations:
+          current: "{{ $value | humanize }}A"
+      - alert: site current draw
+        expr: sum(rPDU2PhaseStatusCurrent{site="dublin",rPDU2PhaseStatusIndex="1"} / 10) > 17
+        for: 5m
+        labels:
+          alertgroup: "dublin"
+        annotations:
+          current: "{{ $value | humanize }}A"
+      - alert: site temperature
+        expr: min(rPDU2SensorTempHumidityStatusTempC{site="dublin"}) / 10 < 18 or min(rPDU2SensorTempHumidityStatusTempC{site="dublin"}) / 10 > 25
+        for: 5m
+        labels:
+          alertgroup: "dublin"
+        annotations:
+          temperature: "{{ $value | humanize }}C"
+      - alert: site humidity
+        expr: max(rPDU2SensorTempHumidityStatusRelativeHumidity{site="dublin"}) / 100 < 0.25 or max(rPDU2SensorTempHumidityStatusRelativeHumidity{site="dublin"}) / 100 > 0.65
+        for: 5m
+        labels:
+          alertgroup: "dublin"
+        annotations:
+          humidity: "{{ $value | humanizePercentage }}"
   - name: fastly
     rules:
       - alert: fastly error rate
@@ -115,12 +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: fastly healthcheck 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
@@ -217,12 +247,12 @@ groups:
   - name: juniper
     rules:
       - alert: juniper fan alarm
-        expr: jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!="running"} > 0
+        expr: sum_over_time(jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!~"running.*"}[6m]) > 0
         for: 5m
         labels:
           alertgroup: "{{ $labels.site }}"
       - alert: juniper power alarm
-        expr: jnxOperatingState{jnxOperatingContentsIndex="2",jnxOperatingState!="running"} > 0
+        expr: sum_over_time(jnxOperatingState{jnxOperatingContentsIndex="2",jnxOperatingState!~"running.*"}[6m]) > 0
         for: 5m
         labels:
           alertgroup: "{{ $labels.site }}"
@@ -320,7 +350,7 @@ groups:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           error_rate: "{{ $value | humanizePercentage }}"
-      - alert: interface transmit errors
+      - alert: wireguard interface transmit errors
         expr: rate(node_network_transmit_errs_total{device=~"wg.*"}[1m]) / rate(node_network_transmit_packets_total{device=~"wg.*"}[1m]) > 0.05
         for: 1h
         labels:
@@ -353,8 +383,8 @@ groups:
   - name: overpass
     rules:
       - alert: overpass osm database age
-        expr: overpass_database_age_seconds{database="osm"} > 300
-        for: 5m
+        expr: overpass_database_age_seconds{database="osm"} > 3600
+        for: 1h
         labels:
           alertgroup: overpass
         annotations:
@@ -537,7 +567,7 @@ groups:
         for: 5m
         labels:
           alertgroup: "{{ $labels.instance }}"
-      - alert: systemd failed service
+      - alert: systemd failed chef client service
         expr: node_systemd_unit_state{state="failed",name="chef-client.service"} == 1
         for: 6h
         labels: