X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/e85f7cbf201686dac598d8b285585d9af468490d..44b52281d387a2a26df9cd9ac5e228c0976089fd:/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 30a14a471..23b94727c 100644
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -115,7 +115,12 @@ groups:
         annotations:
           error_rate: "{{ $value | humanizePercentage }}"
       - alert: fastly healthcheck failing
-        expr: fastly_healthcheck_status == 0
+        expr: count(fastly_healthcheck_status == 0) > 0
+        for: 15m
+        labels:
+          alertgroup: fastly
+      - alert: fastly healthcheck failing
+        expr: count(fastly_healthcheck_status == 0) > 4
         for: 5m
         labels:
           alertgroup: fastly
@@ -402,7 +407,7 @@ groups:
         annotations:
           connections_used: "{{ $value | humanizePercentage }}"
       - alert: postgresql deadlocks
-        expr: increase(pg_stat_database_deadlocks[1m]) > 5
+        expr: increase(pg_stat_database_deadlocks{datname!="nominatim"}[1m]) > 5
         for: 0m
         labels:
           alertgroup: "{{ $labels.instance }}"
@@ -439,6 +444,22 @@ groups:
         for: 5m
         labels:
           alertgroup: "{{ $labels.instance }}"
+  - name: rasdaemon
+    rules:
+      - alert: memory controller errors
+        expr: increase(rasdaemon_mc_events_total[1m]) > 0
+        for: 0m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          new_errors: "{{ $value }}"
+      - alert: pcie aer errors
+        expr: increase(rasdaemon_aer_events_total[1m]) > 0
+        for: 0m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          new_ercrors: "{{ $value }}"
   - name: smart
     rules:
       - alert: smart failure
@@ -447,7 +468,7 @@ groups:
         labels:
           alertgroup: "{{ $labels.instance }}"
       - alert: smart ssd wearout approaching
-        expr: smart_percentage_used >= 90
+        expr: smart_percentage_used >= 80
         for: 60m
         labels:
           alertgroup: "{{ $labels.instance }}"