fix: make metrics endpoint responsive by reducing the chatter (#15055)

peerOnlineCounter was making NxN calls to many peers, this
can be really long and tedious if there are random servers
that are going down.

Instead we should calculate online peers from the point of
view of "self" and return those online and offline appropriately
by performing a healthcheck.

cmd/metrics-v2.go

@@ -1347,7 +1347,7 @@ func getNodeHealthMetrics() *MetricsGroup {
 			return
 		}
 		metrics = make([]Metric, 0, 16)
-		nodesUp, nodesDown := GetPeerOnlineCount()
+		nodesUp, nodesDown := globalNotificationSys.GetPeerOnlineCount()
 		metrics = append(metrics, Metric{
 			Description: getNodeOnlineTotalMD(),
 			Value:       float64(nodesUp),
@@ -1932,11 +1932,9 @@ func (c *minioClusterCollector) Collect(out chan<- prometheus.Metric) {
 	}
 
 	// Call peer api to fetch metrics
-	peerCh := globalNotificationSys.GetClusterMetrics(GlobalContext)
-	selfCh := ReportMetrics(GlobalContext, c.metricsGroups)
 	wg.Add(2)
-	go publish(peerCh)
-	go publish(selfCh)
+	go publish(ReportMetrics(GlobalContext, c.metricsGroups))
+	go publish(globalNotificationSys.GetClusterMetrics(GlobalContext))
 	wg.Wait()
 }