Files
HiveGrid-Wireframes/03 Quellen und Referenzen/hivegrid-cloud/infra/prometheus/README.md
T

59 lines
2.1 KiB
Markdown

# Prometheus — HiveGrid
Scrapes the HiveGrid runtime metrics every 15s. The web core/admin process and
the separately deployed portal process each have their own process-local
registry and therefore their own scrape target.
## Files
- `prometheus.yml` — scrape jobs (api / worker / web-core / web-portal)
- `docker-compose.yml` — Coolify-friendly compose stack
- `secrets/metrics_token` — bearer token (chmod 600, do NOT commit)
## First-time deploy on Hetzner staging
```sh
# 1. Create the secret file with the same value as METRICS_TOKEN in Coolify
mkdir -p secrets
openssl rand -hex 32 > secrets/metrics_token
chmod 600 secrets/metrics_token
# 2. Create the shared docker network (also used by Grafana)
docker network create hivegrid-observability
# 3. Start the stack
docker compose up -d
```
Prometheus is deliberately reachable only as `prometheus:9090` on the
`hivegrid-observability` Docker network. It has no host-port binding and no
lifecycle HTTP endpoint. Use Grafana for dashboards; temporary operator access
must use an authenticated tunnel rather than public ingress.
## Verifying scrape targets
```sh
curl -s -H "Authorization: Bearer $(cat secrets/metrics_token)" \
https://api.hivegrid.de/metrics | head -20
curl -s -H "Authorization: Bearer $(cat secrets/metrics_token)" \
https://app.hivegrid.de/api/metrics | head -20
curl -s -H "Authorization: Bearer $(cat secrets/metrics_token)" \
https://portal.hivegrid.de/api/metrics | head -20
```
Through an authenticated operator tunnel, the Prometheus `/targets` view should
list all four jobs as **UP**. Do not also scrape `admin.hivegrid.de`: admin and
core share the same process-local registry, so doing so would double-count.
## Cardinality budget
- `ai_tokens_consumed_total` — capped at ~100 tenant buckets + 1 `other`
- `subscription_active` — capped by the plan-id table (single digits today)
- All other metrics have bounded label sets (HTTP routes, queue names)
If a panel suddenly shows millions of series, the most likely culprit
is a new histogram with an unbounded label — search the codebase for
`labels({` and verify.