Alerts Extension

ArgoCD’s “Healthy” means the pods are running, not that the service works. The alerts extension shows which Prometheus alerts fire for an application, inside ArgoCD: a full view per application and a tab per workload.

It is read-only and works with any server that speaks the Prometheus rules API: Prometheus, Thanos, Mimir and VictoriaMetrics (vmalert). Alertmanager is optional.

Registration

  • App view on any ArgoCD Application (Alerts)
  • Resource tabs on Pod, Deployment, StatefulSet, DaemonSet and Job

What it shows

Application view

  • Summary cards: firing, critical, warning, pending and silenced
  • Active alerts: name, summary, severity, state, how long it has been active, owner and links
  • Filters for owner, severity and free text across name, summary and labels
  • Rules: the alerting rules the application deploys itself, with their state, their health and the last evaluation error. A rule that fails to evaluate never fires, so a broken expression shows up here instead of staying silent.

Resource tabs

Active alerts about one workload or pod. An alert belongs to a workload when it carries kube-state-metrics’ label for that kind (deployment, statefulset, daemonset, job_name) or a pod label with a pod name that workload produces.

Which alerts an application sees

An alert is shown when its namespace label names one of the namespaces the application manages: its destination namespace plus every namespace in status.resources. This is the same authorization every ArgoPlane backend applies.

Alerts without a namespace label (node, cluster or control-plane alerts) are never shown, with one exception: an alert from a rule the application deploys itself appears in the app view even if the rule aggregates the namespace away. An alert from the application’s own rule that names a namespace the application does not manage is dropped, so a rule that queries another team’s namespace cannot surface their data.

App and platform alerts

Each alert is marked app or platform, without relying on any label convention of your install:

  • App: the rule comes from a PrometheusRule the ArgoCD Application deploys.
  • Platform: the rule comes from anywhere else, and the alert fires in one of the application’s namespaces.

prometheus-operator names each rule file after its PrometheusRule (<namespace>-<name>-<uid>.yaml). The backend reads the UID of every PrometheusRule in the application’s resource tree and matches the files exactly, so a similarly named rule elsewhere is never mistaken for the application’s own.

Without prometheus-operator (for example a standalone Thanos Ruler), every alert is shown as platform.

Severities are read from the severity label and mapped onto critical, warning and info. Common spellings such as page, error, high or p1 count as critical; anything unrecognised shows as unknown, with the original value next to it.

Links come from the runbook_url and dashboard_url annotations. With PROMETHEUS_EXTERNAL_URL set, every rule also links to its expression in Prometheus. Only http and https links are shown: annotations are written by whoever can create a PrometheusRule.

Backend

EndpointDescription
GET /api/v1/alertsActive alerts with a summary
GET /api/v1/rulesThe application’s own alerting rules

See the Alerts API reference.

Configuration

Environment VariableRequiredDefaultDescription
PORTNo8086HTTP server port
LOG_LEVELNoinfoLog level
PROMETHEUS_URLNokube-prometheus-stack serviceAny server with the Prometheus rules API
ALERTMANAGER_URLNoMarks silenced and inhibited alerts. Without it every active alert shows as firing or pending.
PROMETHEUS_EXTERNAL_URLNoBrowser-reachable Prometheus, for a link from every rule to its expression
NAMESPACE_LABELNonamespaceLabel that names an alert’s namespace
SEVERITY_LABELNoseverityLabel that holds an alert’s severity
LINK_ANNOTATIONSNorunbook_url,dashboard_urlAnnotations shown as links
CLUSTER_LABELNoShared Prometheus for several clusters: only alerts whose label matches the application’s destination cluster are shown
CLUSTER_MAPNoJSON map from ArgoCD cluster name or server URL to CLUSTER_LABEL value

CLUSTER_LABEL and CLUSTER_MAP work exactly as for the metrics and logs backends. See “Cluster filter” on the Metrics extension page.

Rules and Alertmanager data are fetched at most every 15 seconds and shared by all requests, because the rules API returns every rule on the server. If Alertmanager is unreachable, alerts are still returned with a warning that their silences are unknown. If the rules API is unreachable, the backend answers 502 rather than an empty list.

RBAC

rules:
  - apiGroups: ["monitoring.coreos.com"]
    resources: ["prometheusrules"]
    verbs: ["get"]
  - apiGroups: ["argoproj.io"]
    resources: ["applications"]
    verbs: ["get"]

Helm

extensions:
  alerts:
    enabled: true
    env:
      PROMETHEUS_URL: "http://kube-prometheus-kube-prome-prometheus.monitoring.svc:9090"
      ALERTMANAGER_URL: "http://kube-prometheus-kube-prome-alertmanager.monitoring.svc:9093"
Note
Prometheus only evaluates the PrometheusRules it selects. kube-prometheus-stack, by default, selects only rules carrying its release label. A rule your application deploys without it never fires and does not appear under Rules.

ArgoCD RBAC

p, role:developer, extensions, invoke, alerts, allow

Not yet supported

  • Silences: the extension does not create or expire them yet.
  • Grafana-managed alert rules.