Halves sample volume on all default-scrape jobs (cAdvisor, node-exporter,
service-endpoints, etc.). Memory id 559's earlier scrape-2m tuning was
applied live but not codified — this restores the Helm template.
Companion changes to keep alerting fidelity:
- evaluation_interval kept at 1m (alerts evaluate every minute)
- snmp-ups job pinned to scrape_interval=30s so PowerOutage /
LowUPSBattery detect within ~30s instead of 2m
- 3 alerts bumped from for:1m to for:3m (HighGPUTemp, LowUPSBattery,
PowerOutage) for stability above the new 2m global cadence
Other jobs that already had per-job overrides (snmp-idrac 1m,
redfish-idrac 3m, kubernetes-pods 5m, kubernetes-services 5m) unaffected.
Expected: 50-150m sustained CPU saving on Prometheus + apiserver.
Verification ongoing — apiserver settles ~minutes after Prometheus
config reload due to initial-target-scrape burst.
b92e1166a8