diff --git a/mvcc/backend/backend.go b/mvcc/backend/backend.go
index cea328ca3..9a6d2ac9a 100644
--- a/mvcc/backend/backend.go
+++ b/mvcc/backend/backend.go
@@ -229,6 +229,8 @@ func (b *backend) Defrag() error {
 }
 
 func (b *backend) defrag() error {
+	now := time.Now()
+
 	// TODO: make this non-blocking?
 	// lock batchTx to ensure nobody is using previous tx, and then
 	// close previous ongoing tx.
@@ -285,6 +287,9 @@ func (b *backend) defrag() error {
 	atomic.StoreInt64(&b.size, size)
 	atomic.StoreInt64(&b.sizeInUse, size-(int64(db.Stats().FreePageN)*int64(db.Info().PageSize)))
 
+	took := time.Since(now)
+	defragDurations.Observe(took.Seconds())
+
 	return nil
 }
 
diff --git a/mvcc/backend/metrics.go b/mvcc/backend/metrics.go
index 76fcb9cf2..b266d1e0f 100644
--- a/mvcc/backend/metrics.go
+++ b/mvcc/backend/metrics.go
@@ -27,8 +27,21 @@ var (
 		// highest bucket start of 0.001 sec * 2^13 == 8.192 sec
 		Buckets: prometheus.ExponentialBuckets(0.001, 2, 14),
 	})
+
+	defragDurations = prometheus.NewHistogram(prometheus.HistogramOpts{
+		Namespace: "etcd",
+		Subsystem: "disk",
+		Name:      "backend_defrag_duration_seconds",
+		Help:      "The latency distribution of backend defragmentation.",
+
+		// 100 MB usually takes 1 sec, so start with 10 MB of 100 ms
+		// lowest bucket start of upper bound 0.1 sec (100 ms) with factor 2
+		// highest bucket start of 0.1 sec * 2^12 == 409.6 sec
+		Buckets: prometheus.ExponentialBuckets(.01, 2, 13),
+	})
 )
 
 func init() {
 	prometheus.MustRegister(commitDurations)
+	prometheus.MustRegister(defragDurations)
 }