reduce Prometheus cardinality round 3: drop 44k more series
- cadvisor: drop unused network error/dropped counters, unused cpu
metrics (load_avg, system, user), unused memory metrics (cache,
failcnt, kernel, mapped_file, max_usage, rss, swap, active/inactive)
- kubelet: drop all unused histogram buckets (storage_operation, csi,
volume_operation, image_pull, http_requests, rest_client, pod_worker,
volume_metric, cgroup_manager) + kubernetes_feature_enabled
- apiserver: drop flowcontrol/rest_client histograms, longrunning_requests
- traefik: drop all router-level metrics (keep service + entrypoint)
- service-endpoints: drop coredns histograms, node_filesystem_*
Post-relabel: 332k → 99k (-70%), ingestion: 5,480 → 1,659 samples/sec (-70%)
06490b0634