diff --git a/pkg/scheduler/metrics/metrics.go b/pkg/scheduler/metrics/metrics.go
index 2d709cb27c8..f8a0ed39989 100644
--- a/pkg/scheduler/metrics/metrics.go
+++ b/pkg/scheduler/metrics/metrics.go
@@ -195,11 +195,13 @@ var (
 			StabilityLevel: metrics.ALPHA,
 		},
 	)
-	PreemptionVictims = metrics.NewGauge(
-		&metrics.GaugeOpts{
-			Subsystem:      SchedulerSubsystem,
-			Name:           "pod_preemption_victims",
-			Help:           "Number of selected preemption victims",
+	PreemptionVictims = metrics.NewHistogram(
+		&metrics.HistogramOpts{
+			Subsystem: SchedulerSubsystem,
+			Name:      "pod_preemption_victims",
+			Help:      "Number of selected preemption victims",
+			// we think #victims>50 is pretty rare, therefore [50, +Inf) is considered a single bucket.
+			Buckets:        metrics.LinearBuckets(5, 5, 10),
 			StabilityLevel: metrics.ALPHA,
 		})
 	PreemptionAttempts = metrics.NewCounter(
diff --git a/pkg/scheduler/scheduler.go b/pkg/scheduler/scheduler.go
index 209bd2518e9..df8fa1dcf89 100644
--- a/pkg/scheduler/scheduler.go
+++ b/pkg/scheduler/scheduler.go
@@ -473,7 +473,7 @@ func (sched *Scheduler) preempt(state *framework.CycleState, fwk framework.Frame
 			sched.Recorder.Eventf(victim, preemptor, v1.EventTypeNormal, "Preempted", "Preempting", "Preempted by %v/%v on node %v", preemptor.Namespace, preemptor.Name, nodeName)
 
 		}
-		metrics.PreemptionVictims.Set(float64(len(victims)))
+		metrics.PreemptionVictims.Observe(float64(len(victims)))
 	}
 	// Clearing nominated pods should happen outside of "if node != nil". Node could
 	// be nil when a pod with nominated node name is eligible to preempt again,