Skip to content
Docs home

SELF-HOSTING · UPDATED AUGUST 2026

Observability

The wired stack: Prometheus metrics, Grafana dashboards, Loki + Promtail logs, Sentry error context, and the Alertmanager sink.

The stack

A self-hosted QUE fleet runs Prometheus for metrics, Grafana for dashboards, Loki with Promtail for log aggregation, Sentry for error context, and Alertmanager for alert routing. Grafana's datasources are provisioned from JSON configuration declaring the three datasources; dashboards themselves are provisioned from JSON rather than hand-built per environment.

Metrics

The application exposes a Prometheus-format metrics endpoint authenticated by the metrics bearer token. Counters cover webhook deliveries and rejections — including the alertmanager webhook sink's accepted and refused payloads — so alerting can fire on delivery failure rather than on silence. The health endpoint supplies the database latency signal that dashboards display.

Errors and alerts

Server errors report to the fleet's own self-hosted Sentry through the envelope protocol — the application never imports the Sentry SDK on the server; it speaks the protocol directly, with linear-backoff retries, coalescing of identical in-flight payloads, and the same forbidden-key redaction as everywhere else. Alertmanager's webhook receiver posts to the application's alert sink, which is fail-closed: unauthenticated alerts are refused and counted, accepted alerts are classified firing/resolved and counted. The sink never dispatches — Alertmanager remains the only notifier in the fleet.

A dashboard without data flow is a visual

Documented here is the actual wiring: which endpoint exposes which counter, which receiver accepts which header, which datasource Grafana declares. A dashboard is operational evidence only when its metrics path is known end to end.

STILL READING

Something here still feels worth checking together.

This documentation describes the product exactly as it is implemented. If a behavior you expected is not here — or is here and does not match your deployment — tell the team. Implementation questions start on the contact page.