diff --git a/pkg/lightning/backend/kv/sql2kv.go b/pkg/lightning/backend/kv/sql2kv.go
index 2f66834c2..6466149d3 100644
--- a/pkg/lightning/backend/kv/sql2kv.go
+++ b/pkg/lightning/backend/kv/sql2kv.go
@@ -406,6 +406,14 @@ func (kvcodec *tableKVEncoder) Encode(
 	return kvPairs(pairs), nil
 }
 
+func (kvs kvPairs) Size() uint64 {
+	size := uint64(0)
+	for _, kv := range kvs {
+		size += uint64(len(kv.Key) + len(kv.Val))
+	}
+	return size
+}
+
 func (kvs kvPairs) ClassifyAndAppend(
 	data *Rows,
 	dataChecksum *verification.KVChecksum,
diff --git a/pkg/lightning/backend/kv/types.go b/pkg/lightning/backend/kv/types.go
index 299f4a8cb..4ebf65f90 100644
--- a/pkg/lightning/backend/kv/types.go
+++ b/pkg/lightning/backend/kv/types.go
@@ -35,6 +35,9 @@ type Row interface {
 		indices *Rows,
 		indexChecksum *verification.KVChecksum,
 	)
+
+	// Size represents the total kv size of this Row.
+	Size() uint64
 }
 
 // Rows represents a collection of encoded rows.
diff --git a/pkg/lightning/backend/noop/noop.go b/pkg/lightning/backend/noop/noop.go
index 00e344b84..b026f4212 100644
--- a/pkg/lightning/backend/noop/noop.go
+++ b/pkg/lightning/backend/noop/noop.go
@@ -154,6 +154,10 @@ func (e noopEncoder) Encode(log.Logger, []types.Datum, int64, []int) (kv.Row, er
 
 type noopRow struct{}
 
+func (r noopRow) Size() uint64 {
+	return 0
+}
+
 func (r noopRow) ClassifyAndAppend(*kv.Rows, *verification.KVChecksum, *kv.Rows, *verification.KVChecksum) {
 }
 
diff --git a/pkg/lightning/backend/tidb/tidb.go b/pkg/lightning/backend/tidb/tidb.go
index 91ff5e00f..86ad569c5 100644
--- a/pkg/lightning/backend/tidb/tidb.go
+++ b/pkg/lightning/backend/tidb/tidb.go
@@ -93,6 +93,10 @@ func NewTiDBBackend(db *sql.DB, onDuplicate string) backend.Backend {
 	return backend.MakeBackend(&tidbBackend{db: db, onDuplicate: onDuplicate})
 }
 
+func (row tidbRow) Size() uint64 {
+	return uint64(len(row))
+}
+
 func (row tidbRow) ClassifyAndAppend(data *kv.Rows, checksum *verification.KVChecksum, _ *kv.Rows, _ *verification.KVChecksum) {
 	rows := (*data).(tidbRows)
 	// Cannot do `rows := data.(*tidbRows); *rows = append(*rows, row)`.
diff --git a/pkg/lightning/restore/restore.go b/pkg/lightning/restore/restore.go
index 93a244c48..f18dd851f 100644
--- a/pkg/lightning/restore/restore.go
+++ b/pkg/lightning/restore/restore.go
@@ -2519,6 +2519,7 @@ func (cr *chunkRestore) encodeLoop(
 		canDeliver := false
 		kvPacket := make([]deliveredKVs, 0, maxKvPairsCnt)
 		var newOffset, rowID int64
+		var kvSize uint64
 	outLoop:
 		for !canDeliver {
 			readDurStart := time.Now()
@@ -2554,8 +2555,16 @@ func (cr *chunkRestore) encodeLoop(
 				return
 			}
 			kvPacket = append(kvPacket, deliveredKVs{kvs: kvs, columns: columnNames, offset: newOffset, rowID: rowID})
-			if len(kvPacket) >= maxKvPairsCnt || newOffset == cr.chunk.Chunk.EndOffset {
+			kvSize += kvs.Size()
+			failpoint.Inject("mock-kv-size", func(val failpoint.Value) {
+				kvSize += uint64(val.(int))
+			})
+			// pebble cannot allow > 4.0G kv in one batch.
+			// we will meet pebble panic when import sql file and each kv has the size larger than 4G / maxKvPairsCnt.
+			// so add this check.
+			if kvSize >= minDeliverBytes || len(kvPacket) >= maxKvPairsCnt || newOffset == cr.chunk.Chunk.EndOffset {
 				canDeliver = true
+				kvSize = 0
 			}
 		}
 		encodeTotalDur += encodeDur
diff --git a/pkg/lightning/restore/restore_test.go b/pkg/lightning/restore/restore_test.go
index 61bce987b..c5249a9f1 100644
--- a/pkg/lightning/restore/restore_test.go
+++ b/pkg/lightning/restore/restore_test.go
@@ -1150,6 +1150,57 @@ func (s *chunkRestoreSuite) TestEncodeLoopForcedError(c *C) {
 	c.Assert(kvsCh, HasLen, 0)
 }
 
+func (s *chunkRestoreSuite) TestEncodeLoopDeliverLimit(c *C) {
+	ctx := context.Background()
+	kvsCh := make(chan []deliveredKVs, 4)
+	deliverCompleteCh := make(chan deliverResult)
+	kvEncoder, err := kv.NewTableKVEncoder(s.tr.encTable, &kv.SessionOptions{
+		SQLMode:   s.cfg.TiDB.SQLMode,
+		Timestamp: 1234567898,
+	})
+	c.Assert(err, IsNil)
+
+	dir := c.MkDir()
+	fileName := "db.limit.000.csv"
+	err = ioutil.WriteFile(filepath.Join(dir, fileName), []byte("1,2,3\r\n4,5,6\r\n7,8,9\r"), 0o644)
+	c.Assert(err, IsNil)
+
+	store, err := storage.NewLocalStorage(dir)
+	c.Assert(err, IsNil)
+	cfg := config.NewConfig()
+
+	reader, err := store.Open(ctx, fileName)
+	c.Assert(err, IsNil)
+	w := worker.NewPool(ctx, 1, "io")
+	p := mydump.NewCSVParser(&cfg.Mydumper.CSV, reader, 111, w, false)
+	s.cr.parser = p
+
+	rc := &Controller{pauser: DeliverPauser, cfg: cfg}
+	c.Assert(failpoint.Enable(
+		"github.com/pingcap/br/pkg/lightning/restore/mock-kv-size", "return(110000000)"), IsNil)
+	_, _, err = s.cr.encodeLoop(ctx, kvsCh, s.tr, s.tr.logger, kvEncoder, deliverCompleteCh, rc)
+
+	// we have 3 kvs total. after the failpoint injected.
+	// we will send one kv each time.
+	count := 0
+	for {
+		kvs, ok := <-kvsCh
+		if !ok {
+			break
+		}
+		count += 1
+		if count <= 3 {
+			c.Assert(kvs, HasLen, 1)
+		}
+		if count == 4 {
+			// we will send empty kvs before encodeLoop exists
+			// so, we can receive 4 batch and 1 is empty
+			c.Assert(kvs, HasLen, 0)
+			break
+		}
+	}
+}
+
 func (s *chunkRestoreSuite) TestEncodeLoopDeliverErrored(c *C) {
 	ctx := context.Background()
 	kvsCh := make(chan []deliveredKVs)
diff --git a/tests/lightning_checkpoint_parquet/run.sh b/tests/lightning_checkpoint_parquet/run.sh
index 9f1004095..31666bd3b 100755
--- a/tests/lightning_checkpoint_parquet/run.sh
+++ b/tests/lightning_checkpoint_parquet/run.sh
@@ -41,9 +41,9 @@ set +e
 run_lightning -d "$DBPATH" --backend tidb --enable-checkpoint=1 2> /dev/null
 set -e
 run_sql 'SELECT count(*), sum(iVal) FROM `cppq_tsr`.tbl'
-check_contains "count(*): 32"
-# sum(0..31)
-check_contains "sum(iVal): 496"
+check_contains "count(*): 1"
+# sum(0)
+check_contains "sum(iVal): 0"
 
 # check chunk offset and update checkpoint current row id to a higher value so that
 # if parse read from start, the generated rows will be different