diff --git a/server/storage/backend/batch_tx.go b/server/storage/backend/batch_tx.go
index 60be4ce6d74..a436a72a3a6 100644
--- a/server/storage/backend/batch_tx.go
+++ b/server/storage/backend/batch_tx.go
@@ -73,7 +73,8 @@ type batchTx struct {
 	tx      *bolt.Tx
 	backend *backend
 
-	pending int
+	pending                 int
+	pendingDeleteOperations int
 }
 
 // Lock is supposed to be called only by the unit test.
@@ -105,7 +106,23 @@ func (t *batchTx) LockOutsideApply() {
 }
 
 func (t *batchTx) Unlock() {
-	if t.pending >= t.backend.batchLimit {
+	// We need to commit the transaction if there is any pending deleting
+	// operation, otherwise etcd might run into a situation that it haven't
+	// finished committing the data into backend storage (note: etcd
+	// periodically commits the bbolt transactions instead of on each
+	// request) when it applies next request. Accordingly, etcd may still
+	// read the stale data from bbolt when processing next request. So it
+	// breaks the linearizability.
+	//
+	// Note we don't need to commit the transaction for put requests if
+	// it doesn't exceed the batch limit, because there is a buffer on top
+	// of the bbolt. Each time when etcd reads data from backend storage,
+	// it will read data from both bbolt and the buffer. But there is no
+	// such a buffer for delete requests.
+	//
+	// Please also refer to
+	// https://github.com/etcd-io/etcd/pull/17119#issuecomment-1857547158
+	if t.pending >= t.backend.batchLimit || t.pendingDeleteOperations > 0 {
 		t.commit(false)
 	}
 	t.Mutex.Unlock()
@@ -133,6 +150,7 @@ func (t *batchTx) UnsafeDeleteBucket(bucket Bucket) {
 		)
 	}
 	t.pending++
+	t.pendingDeleteOperations++
 }
 
 // UnsafePut must be called holding the lock on the tx.
@@ -223,6 +241,7 @@ func (t *batchTx) UnsafeDelete(bucketType Bucket, key []byte) {
 		)
 	}
 	t.pending++
+	t.pendingDeleteOperations++
 }
 
 // UnsafeForEach must be called holding the lock on the tx.
@@ -277,6 +296,7 @@ func (t *batchTx) commit(stop bool) {
 		atomic.AddInt64(&t.backend.commits, 1)
 
 		t.pending = 0
+		t.pendingDeleteOperations = 0
 		if err != nil {
 			t.backend.lg.Fatal("failed to commit tx", zap.Error(err))
 		}