fix: Fix a bug that bulkimport cannot handle empty struct list (#45693)

issue: https://github.com/milvus-io/milvus/issues/42148

Signed-off-by: yhmo <yihua.mo@zilliz.com>

internal/util/importutilv2/csv/row_parser.go

@@ -189,6 +189,15 @@ func (r *rowParser) reconstructArrayForStructArray(structName string, subFieldsM
 			flatStructs[fieldName] = append(flatStructs[fieldName], value)
 		}
 	}
+
+	// struct list can be empty, len(structs) can be zero
+	// fill flatStructs with empty list for each sub field if len(structs) is zero
+	if len(structs) == 0 {
+		for subFieldName := range subFieldsMap {
+			flatStructs[subFieldName] = make([]any, 0)
+		}
+	}
+
 	return flatStructs, nil
 }
 

internal/util/importutilv2/json/row_parser.go

@@ -45,7 +45,7 @@ type rowParser struct {
 	dynamicField         *schemapb.FieldSchema
 	functionOutputFields map[string]int64
 
-	structArrays      map[string]interface{}
+	structArrays      map[string][]string
 	allowInsertAutoID bool
 
 	timezone string
@@ -89,10 +89,14 @@ func NewRowParser(schema *schemapb.CollectionSchema) (RowParser, error) {
 	)
 	allowInsertAutoID, _ := pkgcommon.IsAllowInsertAutoID(schema.GetProperties()...)
 
-	sturctArrays := lo.SliceToMap(
+	structArrays := lo.SliceToMap(
 		schema.GetStructArrayFields(),
-		func(sa *schemapb.StructArrayFieldSchema) (string, interface{}) {
+		func(sa *schemapb.StructArrayFieldSchema) (string, []string) {
-			return sa.GetName(), nil
+			subFieldNames := lo.Map(sa.Fields, func(field *schemapb.FieldSchema, _ int) string {
+				fieldName, _ := typeutil.ExtractStructFieldName(field.GetName())
+				return fieldName
+			})
+			return sa.GetName(), subFieldNames
 		},
 	)
 
@@ -103,7 +107,7 @@ func NewRowParser(schema *schemapb.CollectionSchema) (RowParser, error) {
 		pkField:              pkField,
 		dynamicField:         dynamicField,
 		functionOutputFields: functionOutputFields,
-		structArrays:         sturctArrays,
+		structArrays:         structArrays,
 		allowInsertAutoID:    allowInsertAutoID,
 		timezone:             common.GetSchemaTimezone(schema),
 	}, nil
@@ -221,12 +225,20 @@ func (r *rowParser) Parse(raw any) (Row, error) {
 			return nil, merr.WrapErrImportFailed(fmt.Sprintf("the field '%s' is output by function, no need to provide", key))
 		}
 
-		if _, ok := r.structArrays[key]; ok {
+		if subFieldNames, ok := r.structArrays[key]; ok {
 			values, err := reconstructArrayForStructArray(value)
 			if err != nil {
 				return nil, err
 			}
 
+			// a struct list can be empty, the values could be an empty map
+			// make an empty list for each sub field
+			if len(values) == 0 {
+				for i := 0; i < len(subFieldNames); i++ {
+					values[subFieldNames[i]] = make([]any, 0)
+				}
+			}
+
 			for subKey, subValue := range values {
 				// Pass struct name for sub-fields
 				if err := handleField(key, subKey, subValue); err != nil {

internal/util/importutilv2/parquet/struct_field_reader.go

@@ -103,10 +103,7 @@ func (r *StructFieldReader) Next(count int64) (any, any, error) {
 }
 
 func (r *StructFieldReader) toScalarField(data []interface{}) (*schemapb.ScalarField, error) {
-	if len(data) == 0 {
+	// struct list can be empty, len(data) can be zero, build an empty ScalarField if len(data) is zero
-		return nil, nil
-	}
-
 	switch r.field.GetElementType() {
 	case schemapb.DataType_Bool:
 		boolData := make([]bool, len(data))
@@ -320,7 +317,9 @@ func (r *StructFieldReader) readArrayOfVectorField(chunked *arrow.Chunked) (any,
 						}
 					}
 				}
-				if len(allVectors) > 0 {
+				// struct list could be empty, len(allVectors) can be zero
+				// build an empty VectorField if len(allVectors) is zero
+				if len(allVectors) >= 0 {
 					vectorField := &schemapb.VectorField{
 						Dim: int64(r.dim),
 						Data: &schemapb.VectorField_FloatVector{