cgn-metrics — Telemetry and power aggregator

Prometheus scrape fan-in plus chassis / GPU power probes. Exposes a single /metrics endpoint (default :9092) and optional /federate for upstream Prometheus; feeds the router’s power scoring term.

Overview

cgn-metrics collects:

Scraped targets (router, agent, cgn-kvcached /metrics endpoints) declared in [metrics].scrape_targets
Redfish / IPMI style power inputs when configured
NVML / DCGM signals from agent scrapes

It merges everything with labels so dashboards stay consistent. Details: Observability.

Features

Configurable scrape_interval (default 10s in the example config)
scrape_targets list with per-target name + url; lines are tagged cgn_target=... at /federate
listen_admin binds the HTTP server (default 0.0.0.0:9092)
Redfish credentials via [metrics].redfish_* fields

Architecture

Downstream /metrics → cgn-metrics → :9092/metrics (+ /federate) → Prometheus. The router consumes normalized power gauges when scoring.

Configuration (highlights)

Key	Type	Default	Description
`[metrics].listen_admin`	string	`0.0.0.0:9092`	HTTP listen
`[metrics].scrape_interval`	duration	`10s`	Poll cadence
`[metrics].scrape_targets`	array	`[]`	`{ name, url }` entries
`[metrics].redfish_url`	string	`""`	Optional BMC REST

Example

[cluster]
name = "prod"
etcd = ["http://etcd:2379"]

[metrics]
listen_admin    = "0.0.0.0:9092"
scrape_interval = "10s"

scrape_targets = [
  { name = "router", url = "http://127.0.0.1:9091/metrics" },
  { name = "agent",  url = "http://127.0.0.1:9191/metrics" },
]

Dependencies

Scrape targets reachable from the metrics pod/host
Prometheus (or compatible) for long-term storage — optional but typical