Every declared service has its requested replicas and its update is completed.
remigirard.dev · monitoring handbook
Monitor without
guessing.
Start with the symptom, gather read-only evidence, and use the interface that answers the next question. Grafana shows the overview and logs, Prometheus provides exact metrics, Alertmanager records action, and Traefik explains the live request path.
The 60-second health check
Know what “healthy” means before debugging.
A green dashboard alone is not proof. These four signals cover the declared service, its public path, current alerts, and the freshness of the evidence.
The external probe reaches the real hostname with a valid certificate and expected response.
Alertmanager has no firing alert without a known owner, cause, or response.
Prometheus targets are up and timestamps follow the current server time.
Open the right interface
Each tool answers a different question.
Grafana
Start here for service health, capacity, request rates, errors, containers, and hosts. Use Explore with the Loki data source for centralized service logs.
02 · EVIDENCEPrometheus
Ask precise metric questions, inspect targets, validate timestamps, and confirm what Grafana panels are querying.
03 · ACTIONAlertmanager
See firing and silenced alerts, their labels, ownership, and routing. An alert is a prompt to diagnose—not the diagnosis itself.
04 · REQUEST PATHTraefik
Inspect active routers, middleware, services, TLS entry points, and the backend selected for each public hostname.
When something looks wrong
Follow one short evidence path.
Use the real HTTPS address and note time, status, and affected route.
Compare desired replicas, running tasks, update state, and endpoint checks.
Open Grafana Explore, select Loki, and find the first service error around the observed timestamp.
Check whether the failure is application, capacity, host, or network related.
Use Traefik when the task is healthy but the hostname or certificate path is not.
Same evidence from the repository
The dashboards do not replace declared operations.
Read-only first
Run diagnosis before changing anything. TOAD reports the application contract, Swarm convergence, failed tasks, and public checks without exposing credentials.
Use the full verifier for platform-wide DNS, TLS, mTLS, monitoring, quorum, image, and replica evidence.
cd openstackV3
pnpm run app -- diagnose NAME --json
pnpm run app -- logs NAME
pnpm run verify -- STACK --json
pnpm run maintenance -- status --json
Administrative boundary
Public explanation, private control plane.
Browser access
Import the operator PKCS #12 identity into your browser or operating-system certificate store. Select it when opening an administrative hostname. Issue separate identities for separate people or devices.
Expected rejection
A browser without the client certificate should be rejected before Grafana, Prometheus, Alertmanager, or Traefik can respond. There is no shared dashboard password.
Open the public routing proof →