X-Git-Url: https://git.openstreetmap.org/chef.git/blobdiff_plain/c4f0a0e16117988b7f9488712b101bb08f8bc151..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 5b2ec56ac..3dd1af873 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: multiple fastly healthchecks failing - expr: count(fastly_healthcheck_status == 0) > 4 + expr: count(fastly_healthcheck_status == 0) by (service) > 4 for: 5m labels: alertgroup: fastly @@ -246,13 +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: 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 }}" @@ -396,6 +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