X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/0743983d24e536a0eb55f05feaf394e936868440..644f12460af63db5fa70baa490def90dd7374ad7:/cookbooks/prometheus/templates/default/alert_rules.yml.erb?ds=sidebyside

diff --git a/cookbooks/prometheus/templates/default/alert_rules.yml.erb b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
index 6dc966050..1fac68869 100644
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -105,13 +105,20 @@ groups:
           pressure: "{{ $value | humanizePercentage }}"
   - name: database
     rules:
-      - alert: postgres replication delay
-        expr: pg_replication_lag_seconds > 30
-        for: 15m
+      - alert: active rails queries
+        expr: sum(pg_stat_activity_count{datname="openstreetmap",usename="rails",state="active"}) by (instance) > 50 and on (instance) chef_role{name="db-master"}
+        for: 5m
         labels:
           alertgroup: database
         annotations:
-          delay: "{{ $value | humanizeDuration }}"
+          queries: "{{ $value }}"
+      - alert: active cgimap queries
+        expr: sum(pg_stat_activity_count{datname="openstreetmap",usename="cgimap",state="active"}) by (instance) > 30 and on (instance) chef_role{name="db-master"}
+        for: 5m
+        labels:
+          alertgroup: database
+        annotations:
+          delay: "{{ $value }}"
   - name: discourse
     rules:
       - alert: discourse job failure rate
@@ -416,6 +423,13 @@ groups:
           connections_used: "{{ $value | humanizePercentage }}"
   - name: network
     rules:
+      - alert: interface redundancy lost
+        expr: node_bonding_active < 2 and on (instance, master) label_replace(chef_network_interface{bond_mode="802.3ad"}, "master", "$1", "name", "(.*)")
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          link_count: "{{ $value }}"
       - alert: interface transmit rate
         expr: rate(node_network_transmit_bytes_total[1m]) / node_network_speed_bytes > 0.99
         for: 5m
@@ -572,13 +586,6 @@ groups:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           new_deadlocks: "{{ $value }}"
-      - alert: postgresql slow queries
-        expr: pg_slow_queries > 0
-        for: 5m
-        labels:
-          alertgroup: "{{ $labels.instance }}"
-        annotations:
-          queries: "{{ $value }}"
       - alert: postgresql idle transactions
         expr: sum(pg_process_idle_seconds_count{state="idle in transaction"}) by (instance, server) > sum(pg_process_idle_seconds_bucket{state="idle in transaction",le="300"}) by (instance, server)
         for: 5m