nova/meta/prometheus.yml

{% from "nova/map.jinja" import controller, compute, monitoring with context %}

{%- set is_controller = controller.get('enabled', False) %}
{%- set is_compute = compute.get('enabled', False) %}

{%- if is_controller or is_compute %}
{%- if is_compute and
       exporters is defined %}
{%- set packages = exporters.get('libvirt', {}).get('packages', ('libvirt-exporter', )) %}
  {%- load_yaml as new_exporters_cfg %}
exporters:
  libvirt:
    enabled: true
{%- if packages is defined %}
    packages:
    {% for pkg in packages %}
    - {{ pkg }}
    {% endfor %}
{%- endif %}
    services:
      qemu:
        enabled: true
        bind:
          address: 0.0.0.0
          port: 9177
  {%- endload %}
{{ new_exporters_cfg|yaml(False) }}
{%- endif %}

server:
  alert:
{%- if is_controller %}
{%- set minor_threshold = monitoring.services_failed_warning_threshold_percent|float %}
{%- set major_threshold = monitoring.services_failed_critical_threshold_percent|float %}
{%- set minor_compute_threshold = monitoring.computes_failed_warning_threshold_percent|float %}
{%- set major_compute_threshold = monitoring.computes_failed_critical_threshold_percent|float %}
{% raw %}
    NovaAPIOutage:
      if: >-
        max(openstack_api_check_status{name=~"nova.*|placement"}) == 0
      labels:
        severity: critical
        service: nova
      annotations:
        summary: "Nova API outage"
        description: >-
          Nova API is not accessible for all available Nova endpoints in the OpenStack service catalog.
    NovaAPIDown:
      if: >-
        openstack_api_check_status{name=~"nova.*|placement"} == 0
      labels:
        severity: major
        service: nova
      annotations:
        summary: "{{ $labels.name }} endpoint is not accessible"
        description: >-
          Nova API is not accessible for the {{ $labels.name }} endpoint.
    NovaAPIServiceDown:
      if: >-
        http_response_status{name=~"nova-api"} == 0
      for: 2m
      labels:
        severity: minor
        service: nova
      annotations:
        summary: "Host nova-api endpoint is not accessible"
        description: >-
          The host nova-api endpoint on the {{ $labels.host }} node is not accessible for at least 2 minutes.
    NovaServiceDown:
      if: >-
        openstack_nova_service_state == 0
      labels:
        severity: minor
        service: nova
      annotations:
        summary: "{{ $labels.binary }} service is down"
        description: >-
          The {{ $labels.binary }} service on the {{ $labels.hostname }} node is down.
{%- endraw %}
    NovaServicesDownMinor:
      if: >-
        count(openstack_nova_service_state{binary!~"nova-compute"} == 0) by (binary) >= on (binary) count(openstack_nova_service_state{binary!~"nova-compute"}) by (binary) * {{minor_threshold}} and count(openstack_nova_service_state{binary!~"nova-compute"} == 0) by (binary) < on (binary) count(openstack_nova_service_state{binary!~"nova-compute"}) by (binary) * {{major_threshold}}
      labels:
        severity: minor
        service: nova
      annotations:
        summary: "{{minor_threshold * 100}}%{%- raw %} of {{ $labels.binary }} services are down"
        description: >-
          {{ $value }} {{ $labels.binary }} services are down {%- endraw %}(at least {{minor_threshold * 100}}%).
    NovaComputeServicesDownMinor:
      if: >-
        count(openstack_nova_service_state{binary="nova-compute"} == 0) >= count(openstack_nova_service_state{binary="nova-compute"}) * {{minor_compute_threshold}} and count(openstack_nova_service_state{binary="nova-compute"} == 0) < count(openstack_nova_service_state{binary="nova-compute"}) * {{major_compute_threshold}}
      labels:
        severity: minor
        service: nova
      annotations:
        summary: "{{minor_compute_threshold * 100}}%{%- raw %} of nova-compute services are down"
        description: >-
          {{ $value }} nova-compute services are down {%- endraw %}(at least {{minor_compute_threshold * 100}}%).
    NovaServicesDownMajor:
      if: >-
        count(openstack_nova_service_state{binary!~"nova-compute"} == 0) by (binary) >= on (binary) count(openstack_nova_service_state{binary!~"nova-compute"}) by (binary) * {{major_threshold}}
      labels:
        severity: major
        service: nova
      annotations:
        summary: "{{major_threshold * 100}}%{%- raw %} of {{ $labels.binary }} services are down"
        description: >-
          {{ $value }} {{ $labels.binary }} services are down {%- endraw %}(at least {{major_threshold * 100}}%).
    NovaComputeServicesDownMajor:
      if: >-
        count(openstack_nova_service_state{binary="nova-compute"} == 0) >= count(openstack_nova_service_state{binary="nova-compute"}) * {{major_compute_threshold}}
      labels:
        severity: major
        service: nova
      annotations:
        summary: "{{major_compute_threshold * 100}}%{%- raw %} of nova-compute services are down"
        description: >-
          {{ $value }} nova-compute services are down {%- endraw %}(at least {{major_compute_threshold * 100}}%).{%- raw %}
    NovaServiceOutage:
      if: >-
        count(openstack_nova_service_state == 0) by (binary) == on (binary) count(openstack_nova_service_state) by (binary)
      labels:
        severity: critical
        service: nova
      annotations:
        summary: "{{ $labels.binary }} service outage"
        description: >-
          All {{ $labels.binary }} services are down.
    NovaTotalFreeVCPUsLow:
      if: >-
        (100.0 * openstack_nova_total_free_vcpus) / (openstack_nova_total_free_vcpus + openstack_nova_total_used_vcpus) < 10.0
      for: 1m
      labels:
        severity: warning
        service: nova
      annotations:
        summary: "VCPU low limit for new instances"
        description: >-
            VPCU low limit for 1 minutes
    NovaTotalFreeMemoryLow:
      if: >-
        (100.0 * openstack_nova_total_free_ram) / (openstack_nova_total_free_ram + openstack_nova_total_used_ram) < 10.0
      for: 1m
      labels:
        severity: warning
        service: nova
      annotations:
        summary: "Memory low limit for new instances"
        description: >-
            Memory low limit for 1 minutes
    NovaTotalFreeVCPUsShortage:
      if: >-
        (100.0 * openstack_nova_total_free_vcpus) / (openstack_nova_total_free_vcpus + openstack_nova_total_used_vcpus) < 2.0
      for: 1m
      labels:
        severity: critical
        service: nova
      annotations:
        summary: "VCPU shortage for new instances"
        description: >-
            VPCU shortage for 1 minutes
    NovaTotalFreeMemoryShortage:
      if: >-
        (100.0 * openstack_nova_total_free_ram) / (openstack_nova_total_free_ram + openstack_nova_total_used_ram) < 2.0
      for: 1m
      labels:
        severity: critical
        service: nova
      annotations:
        summary: "Memory shortage for new instances"
        description: >-
            Memory shortage for 1 minutes
    NovaAggregatesFreeVCPUsLow:
      if: >-
        (100.0 * openstack_nova_aggregate_free_vcpus) / (openstack_nova_aggregate_free_vcpus + openstack_nova_aggregate_used_vcpus) < 10.0
      for: 1m
      labels:
        severity: warning
        service: nova
        aggregate: "{{ $labels.aggregate }}"
      annotations:
        summary: "VCPU low limit for new instances on aggregate {{ $labels.aggregate }}"
        description: >-
            VPCU low limit for 1 minutes on aggregate {{ $labels.aggregate }}
    NovaAggregatesFreeMemoryLow:
      if: >-
        (100.0 * openstack_nova_aggregate_free_ram) / (openstack_nova_aggregate_free_ram + openstack_nova_aggregate_used_ram) < 10.0
      for: 1m
      labels:
        severity: warning
        service: nova
        aggregate: "{{ $labels.aggregate }}"
      annotations:
        summary: "Memory low limit for new instances on aggregate {{ $labels.aggregate }}"
        description: >-
            Memory low limit for 1 minutes on aggregate {{ $labels.aggregate }}
    NovaAggregatesFreeVCPUsShortage:
      if: >-
        (100.0 * openstack_nova_aggregate_free_vcpus) / (openstack_nova_aggregate_free_vcpus + openstack_nova_aggregate_used_vcpus) < 2.0
      for: 1m
      labels:
        severity: critical
        service: nova
        aggregate: "{{ $labels.aggregate }}"
      annotations:
        summary: "VCPU shortage for new instances on aggregate {{ $labels.aggregate }}"
        description: >-
            VPCU shortage for 1 minutes on aggregate {{ $labels.aggregate }}
    NovaAggregatesFreeMemoryShortage:
      if: >-
        (100.0 * openstack_nova_aggregate_free_ram) / (openstack_nova_aggregate_free_ram + openstack_nova_aggregate_used_ram) < 2.0
      for: 1m
      labels:
        severity: critical
        service: nova
        aggregate: "{{ $labels.aggregate }}"
      annotations:
        summary: "Memory shortage for new instances on aggregate {{ $labels.aggregate }}"
        description: >-
            Memory shortage for 1 minutes on aggregate {{ $labels.aggregate }}
{%- endraw %}
{%- endif %}
    NovaErrorLogsTooHigh:
      {%- set log_threshold = monitoring.error_log_rate.warn|float %}
      if: >-
        sum(rate(log_messages{service="nova",level=~"(?i:(error|emergency|fatal))"}[5m])) without (level) > {{ log_threshold }}
{%- raw %}
      labels:
        severity: warning
        service: nova
      annotations:
        summary: "High number of errors in Nova logs"
        description: "The rate of errors in Nova logs over the last 5 minutes is too high on the {{ $labels.host }} node (current value={{ $value }}, threshold={%- endraw %}{{ log_threshold }})."
{%- if is_compute and exporters is defined %}
{%- raw %}
    LibvirtDown:
      if: >-
        libvirt_up == 0
      for: 2m
      labels:
        severity: critical
        service: libvirt
      annotations:
        summary: "Failure to gather Libvirt metrics"
        description: "The Libvirt metric exporter fails to gather metrics on the {{ $labels.host }} node for at least 2 minutes."
{%- endraw %}
{%- include "prometheus/_exporters_config.sls" %}
{%- endif %}
{%- endif %}