diff --git a/server/etcdserver/api/etcdhttp/metrics.go b/server/etcdserver/api/etcdhttp/metrics.go
index 48799e18a301..b369b01fd81a 100644
--- a/server/etcdserver/api/etcdhttp/metrics.go
+++ b/server/etcdserver/api/etcdhttp/metrics.go
@@ -47,7 +47,9 @@ func HandleMetricsHealth(lg *zap.Logger, mux *http.ServeMux, srv etcdserver.Serv
 // and its corresponding timeout.
 func HandleMetricsHealthForV3(lg *zap.Logger, mux *http.ServeMux, srv *etcdserver.EtcdServer) {
 	mux.Handle(PathMetrics, promhttp.Handler())
-	mux.Handle(PathHealth, NewHealthHandler(lg, func(excludedAlarms AlarmSet, serializable bool) Health { return checkV3Health(lg, srv, excludedAlarms, serializable) }))
+	mux.Handle(PathHealth, NewHealthHandler(lg, func(excludedAlarms AlarmSet, serializable bool) Health {
+		return checkV3Health(lg, srv, excludedAlarms, serializable)
+	}))
 }
 
 // HandlePrometheus registers prometheus handler on '/metrics'.
@@ -65,8 +67,13 @@ func NewHealthHandler(lg *zap.Logger, hfunc func(excludedAlarms AlarmSet, Serial
 			return
 		}
 		excludedAlarms := getExcludedAlarms(r)
-		// when the query parameter "consistency=s" is provided, then etcdserver should only check local etcd member's health
-		// instead of the etcd cluster's health.
+		// Kubernetes Probes (i.e. livenessProbe) use "/health" endpoint to make a decision whether to restart a specific container.
+		// In this case, it should only check local etcd member's health instead of etcd cluster's health. When the cluster isn't healthy,
+		// such as no raft leader, restarting the local etcd member cannot help, and it could even make the situation worse. So the endpoint
+		// should provide an option to let users choose to check the etcd cluster's health or local etcd member's health.
+		// The default behaviour is to check the etcd cluster's health so as to be backward compatible.
+		// When the query parameter "serializable=true" is provided, then etcdserver should only do serializable read, which means
+		// it checks local etcd member's health instead of the etcd cluster's health.
 		serializableFlag := getSerializableFlag(r)
 		h := hfunc(excludedAlarms, serializableFlag)
 		defer func() {
@@ -132,9 +139,9 @@ func getExcludedAlarms(r *http.Request) (alarms AlarmSet) {
 }
 
 func getSerializableFlag(r *http.Request) bool {
-	getConsistency, found := r.URL.Query()["consistency"]
+	getConsistency, found := r.URL.Query()["serializable"]
 	if found {
-		if len(getConsistency) == 1 && getConsistency[0] == "s" {
+		if len(getConsistency) == 1 && getConsistency[0] == "true" {
 			return true
 		}
 	}