Prometheus

2 min read

MonitoringConfig → ServiceMonitor + PrometheusRule

02-prometheus wraps Prometheus Operator. One MonitoringConfig CR creates two resources: a ServiceMonitor that tells Prometheus what to scrape, and a PrometheusRule with an alert that fires when the error rate exceeds the threshold. The developer declares a threshold — Orkestra writes the PromQL.

ork init --pack ecosystem-composition
cd ecosystem-composition/02-prometheus

One CR, two resources

This is the first example where one internal CR maps to multiple ecosystem resources. The pattern is the same — declare the internal CRD in your vocabulary, implement the mapping in the Katalog:

spec:
  targetDeployment: my-webapp
  port: "8080"
  team: platform
  errorRateThreshold: "0.05"

What Orkestra creates:

ServiceMonitor — scrape configuration:

apiVersion: monitoring.coreos.com/v1
kind: ServiceMonitor
spec:
  selector:
    matchLabels:
      app: my-webapp
  endpoints:
    - port: "8080"
      interval: "30s"
      path: /metrics

PrometheusRule — alert expression:

apiVersion: monitoring.coreos.com/v1
kind: PrometheusRule
spec:
  groups:
    - rules:
        - alert: HighErrorRate
          expr: |
            rate(http_requests_total{job="my-webapp",status=~"5.."}[5m])
            /
            rate(http_requests_total{job="my-webapp"}[5m])
            > 0.05

The developer wrote errorRateThreshold: "0.05". They did not write the PromQL expression, the rate window, or the metric label selectors. The platform team wrote those once in the Katalog.

Why this matters

PromQL is expressive. It is also opaque to developers who are not SREs. Common failure modes:

  • Wrong label selectors — alert fires for everything or nothing
  • Wrong rate window — too short produces noise, too long misses incidents
  • Inconsistent metric naming — alert does not fire because the team named their counter differently

With the mapping:

  • The label selector is derived from spec.targetDeployment — always matches the right pods
  • The rate window is fixed at 5m by the platform
  • The metric names are enforced by convention

The SRE team writes the PromQL template once. Every team that creates a MonitoringConfig gets a working, correctly-labelled alert without writing a single PromQL expression.

Try it

ork init --pack ecosystem-composition
cd ecosystem-composition/02-prometheus
# Follow steps in README

→ 03 — Crossplane — infrastructure provisioning with an approval gate.