X-Git-Url: https://git.openstreetmap.org/chef.git/blobdiff_plain/2f2297f6e834712d6dd28e00ff1d9bc6a6450009..2c1479b9b228b234ea3514060ac2071e4caa1e4a:/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 417641cf0..3dd1af873 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: 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 @@ -216,13 +246,18 @@ 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: jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!="running"} > 0 + expr: jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!~"running.*"} > 0 for: 5m labels: alertgroup: "{{ $labels.site }}" - alert: juniper power alarm - expr: jnxOperatingState{jnxOperatingContentsIndex="2",jnxOperatingState!="running"} > 0 + expr: jnxOperatingState{jnxOperatingContentsIndex="2",jnxOperatingState!~"running.*"} > 0 for: 5m labels: alertgroup: "{{ $labels.site }}" @@ -353,8 +388,8 @@ 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: @@ -366,6 +401,23 @@ 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