Drop more heavy metrics (#455)

After these, we're down to useful metrics like node_cpu_seconds_total or
grpc_io_client_roundtrip_latency_bucket which we wouldn't want to drop
but perhaps relabel (for example to aggregate CPU usage across all CPUs
rather than generating 8 metrics for all 30 or 100 vCPUs on every node,
or to drop pod IDs from grpc client metrics so they're aggregated across
different instances of the same client) but I've never tried this.

Tested as for #453.

src/app_charts/prometheus/prometheus-robot.values.yaml

@@ -124,7 +124,7 @@ kubeApiServer:
     metricRelabelings:
     # Drop high cardinality apiserver metrics.
     - action: drop
-      regex: "apiserver_request.*|etcd_request.*|apiserver_watch.*|code_*"
+      regex: "apiserver_(request|response|watch|admission).*|etcd_request.*|code_*"
       sourceLabels: [__name__]
     relabelings:
     - sourceLabels: [__meta_kubernetes_pod_node_name]
@@ -148,7 +148,7 @@ kubelet:
       regex: 'container_fs_(io_current|io_time_seconds_total|io_time_weighted_seconds_total|reads_merged_total|sector_reads_total|sector_writes_total|writes_merged_total)'
     - sourceLabels: [__name__]
       action: drop
-      regex: 'container_memory_(mapped_file|swap)'
+      regex: 'container_memory_(mapped_file|swap|failures_total)'
     - sourceLabels: [__name__]
       action: drop
       regex: 'container_(file_descriptors|tasks_state|threads_max)'