milvus/internal/storage/utils.go

// Licensed to the LF AI & Data foundation under one
// or more contributor license agreements. See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership. The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License. You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

package storage

import (
	"bytes"
	"encoding/binary"
	"errors"
	"fmt"
	"io"
	"sort"
	"strconv"

	"github.com/golang/protobuf/proto"

	"github.com/milvus-io/milvus/internal/mq/msgstream"

	"github.com/milvus-io/milvus/internal/util/typeutil"

	"github.com/milvus-io/milvus/internal/log"
	"github.com/milvus-io/milvus/internal/proto/schemapb"
	"go.uber.org/zap"

	"github.com/milvus-io/milvus/internal/common"

	"github.com/milvus-io/milvus/internal/proto/commonpb"

	"github.com/milvus-io/milvus/internal/kv"
)

// GetBinlogSize get size of a binlog file.
//		normal binlog file, error = nil;
//		key not exist, size = 0, error = nil;
//		key not in binlog format, size = (a not accurate number), error != nil;
//		failed to read event reader, size = (a not accurate number), error != nil;
func GetBinlogSize(kv kv.DataKV, key string) (int64, error) {

	return kv.GetSize(key)
}

// EstimateMemorySize get approximate memory size of a binlog file.
//		1, key not exist, size = 0, error != nil;
//		2, failed to read event header, size = 0, error != nil;
//		3, invalid event length, size = 0, error != nil;
//		4, failed to read descriptor event, size = 0, error != nil;
//		5, original_size not in extra, size = 0, error != nil;
//		6, original_size not in int format, size = 0, error != nil;
//		7, normal binlog with original_size, return original_size, error = nil;
func EstimateMemorySize(kv kv.DataKV, key string) (int64, error) {
	total := int64(0)

	header := &eventHeader{}
	headerSize := binary.Size(header)

	startPos := binary.Size(MagicNumber)
	endPos := startPos + headerSize

	// get header
	headerContent, err := kv.LoadPartial(key, int64(startPos), int64(endPos))
	if err != nil {
		return total, err
	}

	buffer := bytes.NewBuffer(headerContent)

	header, err = readEventHeader(buffer)
	if err != nil {
		return total, err
	}

	if header.EventLength <= 0 {
		return total, fmt.Errorf("key %v not in binlog format", key)
	}

	var desc *descriptorEvent
	endPos = startPos + int(header.EventLength)
	descContent, err := kv.LoadPartial(key, int64(startPos), int64(endPos))
	if err != nil {
		return total, err
	}

	buffer = bytes.NewBuffer(descContent)

	desc, err = ReadDescriptorEvent(buffer)
	if err != nil {
		return total, err
	}

	sizeStr, ok := desc.Extras[originalSizeKey]
	if !ok {
		return total, fmt.Errorf("key %v not in extra information", originalSizeKey)
	}

	size, err := strconv.Atoi(fmt.Sprintf("%v", sizeStr))
	if err != nil {
		return total, fmt.Errorf("%v not in valid format, value: %v", originalSizeKey, sizeStr)
	}

	total = int64(size)

	return total, nil
}

//////////////////////////////////////////////////////////////////////////////////////////////////

func checkTsField(data *InsertData) bool {
	tsData, ok := data.Data[common.TimeStampField]
	if !ok {
		return false
	}

	_, ok = tsData.(*Int64FieldData)
	return ok
}

func checkRowIDField(data *InsertData) bool {
	rowIDData, ok := data.Data[common.RowIDField]
	if !ok {
		return false
	}

	_, ok = rowIDData.(*Int64FieldData)
	return ok
}

func checkNumRows(fieldDatas ...FieldData) bool {
	if len(fieldDatas) <= 0 {
		return true
	}

	numRows := fieldDatas[0].RowNum()
	for i := 1; i < len(fieldDatas); i++ {
		if numRows != fieldDatas[i].RowNum() {
			return false
		}
	}

	return true
}

type fieldDataList struct {
	IDs   []FieldID
	datas []FieldData
}

func (ls fieldDataList) Len() int {
	return len(ls.IDs)
}

func (ls fieldDataList) Less(i, j int) bool {
	return ls.IDs[i] < ls.IDs[j]
}

func (ls fieldDataList) Swap(i, j int) {
	ls.IDs[i], ls.IDs[j] = ls.IDs[j], ls.IDs[i]
	ls.datas[i], ls.datas[j] = ls.datas[j], ls.datas[i]
}

func sortFieldDataList(ls fieldDataList) {
	sort.Sort(ls)
}

// TransferColumnBasedInsertDataToRowBased transfer column-based insert data to row-based rows.
// Note:
//	- ts column must exist in insert data;
//	- row id column must exist in insert data;
//	- the row num of all column must be equal;
//	- num_rows = len(RowData), a row will be assembled into the value of blob with field id order;
func TransferColumnBasedInsertDataToRowBased(data *InsertData) (
	Timestamps []uint64,
	RowIDs []int64,
	RowData []*commonpb.Blob,
	err error,
) {
	if !checkTsField(data) {
		return nil, nil, nil,
			errors.New("cannot get timestamps from insert data")
	}

	if !checkRowIDField(data) {
		return nil, nil, nil,
			errors.New("cannot get row ids from insert data")
	}

	tss := data.Data[common.TimeStampField].(*Int64FieldData)
	rowIds := data.Data[common.RowIDField].(*Int64FieldData)

	ls := fieldDataList{}
	for fieldID := range data.Data {
		if fieldID == common.TimeStampField || fieldID == common.RowIDField {
			continue
		}

		ls.IDs = append(ls.IDs, fieldID)
		ls.datas = append(ls.datas, data.Data[fieldID])
	}

	// checkNumRows(tss, rowIds, ls.datas...) // don't work
	all := []FieldData{tss, rowIds}
	all = append(all, ls.datas...)
	if !checkNumRows(all...) {
		return nil, nil, nil,
			errors.New("columns of insert data have different length")
	}

	sortFieldDataList(ls)

	numRows := tss.RowNum()
	rows := make([]*commonpb.Blob, numRows)
	for i := 0; i < numRows; i++ {
		blob := &commonpb.Blob{}
		var buffer bytes.Buffer

		for j := 0; j < ls.Len(); j++ {
			d := ls.datas[j].GetRow(i)
			err := binary.Write(&buffer, common.Endian, d)
			if err != nil {
				return nil, nil, nil,
					fmt.Errorf("failed to get binary row, err: %v", err)
			}
		}

		blob.Value = buffer.Bytes()
		rows[i] = blob
	}

	utss := make([]uint64, tss.RowNum())
	for i := 0; i < tss.RowNum(); i++ {
		utss[i] = uint64(tss.Data[i])
	}

	return utss, rowIds.Data, rows, nil
}

///////////////////////////////////////////////////////////////////////////////////////////

// TODO: remove these functions to proper file.

// GetDimFromParams get dim from params.
func GetDimFromParams(params []*commonpb.KeyValuePair) (int, error) {
	var dim int
	var err error
	for _, t := range params {
		if t.Key == "dim" {
			dim, err = strconv.Atoi(t.Value)
			if err != nil {
				return -1, err
			}
			return dim, nil
		}
	}
	return -1, errors.New("dim not found in params")
}

// ReadBinary read data in bytes and write it into receiver.
//  The receiver can be any type in int8, int16, int32, int64, float32, float64 and bool
//  ReadBinary uses LittleEndian ByteOrder.
func ReadBinary(reader io.Reader, receiver interface{}, dataType schemapb.DataType) {
	err := binary.Read(reader, common.Endian, receiver)
	if err != nil {
		log.Error("binary.Read failed", zap.Any("data type", dataType), zap.Error(err))
	}
}

// It will save my life if golang support generic programming.

// TODO: string type.

func readFloatVectors(blobReaders []io.Reader, dim int) []float32 {
	ret := make([]float32, 0)
	for _, r := range blobReaders {
		var v = make([]float32, dim)
		ReadBinary(r, &v, schemapb.DataType_FloatVector)
		ret = append(ret, v...)
	}
	return ret
}

func readBinaryVectors(blobReaders []io.Reader, dim int) []byte {
	ret := make([]byte, 0)
	for _, r := range blobReaders {
		var v = make([]byte, dim/8)
		ReadBinary(r, &v, schemapb.DataType_BinaryVector)
		ret = append(ret, v...)
	}
	return ret
}

func readBoolArray(blobReaders []io.Reader) []bool {
	ret := make([]bool, 0)
	for _, r := range blobReaders {
		var v bool
		ReadBinary(r, &v, schemapb.DataType_Bool)
		ret = append(ret, v)
	}
	return ret
}

func readInt8Array(blobReaders []io.Reader) []int8 {
	ret := make([]int8, 0)
	for _, r := range blobReaders {
		var v int8
		ReadBinary(r, &v, schemapb.DataType_Int8)
		ret = append(ret, v)
	}
	return ret
}

func readInt16Array(blobReaders []io.Reader) []int16 {
	ret := make([]int16, 0)
	for _, r := range blobReaders {
		var v int16
		ReadBinary(r, &v, schemapb.DataType_Int16)
		ret = append(ret, v)
	}
	return ret
}

func readInt32Array(blobReaders []io.Reader) []int32 {
	ret := make([]int32, 0)
	for _, r := range blobReaders {
		var v int32
		ReadBinary(r, &v, schemapb.DataType_Int32)
		ret = append(ret, v)
	}
	return ret
}

func readInt64Array(blobReaders []io.Reader) []int64 {
	ret := make([]int64, 0)
	for _, r := range blobReaders {
		var v int64
		ReadBinary(r, &v, schemapb.DataType_Int64)
		ret = append(ret, v)
	}
	return ret
}

func readFloatArray(blobReaders []io.Reader) []float32 {
	ret := make([]float32, 0)
	for _, r := range blobReaders {
		var v float32
		ReadBinary(r, &v, schemapb.DataType_Float)
		ret = append(ret, v)
	}
	return ret
}

func readDoubleArray(blobReaders []io.Reader) []float64 {
	ret := make([]float64, 0)
	for _, r := range blobReaders {
		var v float64
		ReadBinary(r, &v, schemapb.DataType_Double)
		ret = append(ret, v)
	}
	return ret
}

func RowBasedInsertMsgToInsertData(msg *msgstream.InsertMsg, collSchema *schemapb.CollectionSchema) (idata *InsertData, err error) {
	blobReaders := make([]io.Reader, 0)
	for _, blob := range msg.RowData {
		blobReaders = append(blobReaders, bytes.NewReader(blob.GetValue()))
	}

	idata = &InsertData{
		Data: make(map[FieldID]FieldData),
		// TODO: handle Infos.
		Infos: nil,
	}

	for _, field := range collSchema.Fields {
		switch field.DataType {
		case schemapb.DataType_FloatVector:
			dim, err := GetDimFromParams(field.TypeParams)
			if err != nil {
				log.Error("failed to get dim", zap.Error(err))
				return nil, err
			}

			vecs := readFloatVectors(blobReaders, dim)
			idata.Data[field.FieldID] = &FloatVectorFieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    vecs,
				Dim:     dim,
			}

		case schemapb.DataType_BinaryVector:
			var dim int
			dim, err := GetDimFromParams(field.TypeParams)
			if err != nil {
				log.Error("failed to get dim", zap.Error(err))
				return nil, err
			}

			vecs := readBinaryVectors(blobReaders, dim)
			idata.Data[field.FieldID] = &BinaryVectorFieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    vecs,
				Dim:     dim,
			}

		case schemapb.DataType_Bool:
			idata.Data[field.FieldID] = &BoolFieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    readBoolArray(blobReaders),
			}

		case schemapb.DataType_Int8:
			idata.Data[field.FieldID] = &Int8FieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    readInt8Array(blobReaders),
			}

		case schemapb.DataType_Int16:
			idata.Data[field.FieldID] = &Int16FieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    readInt16Array(blobReaders),
			}

		case schemapb.DataType_Int32:
			idata.Data[field.FieldID] = &Int32FieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    readInt32Array(blobReaders),
			}

		case schemapb.DataType_Int64:
			idata.Data[field.FieldID] = &Int64FieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    nil,
			}

			fieldData := idata.Data[field.FieldID].(*Int64FieldData)
			switch field.FieldID {
			case 0: // rowIDs
				fieldData.Data = append(fieldData.Data, msg.RowIDs...)
			case 1: // Timestamps
				for _, ts := range msg.Timestamps {
					fieldData.Data = append(fieldData.Data, int64(ts))
				}
			default:
				fieldData.Data = readInt64Array(blobReaders)
			}

		case schemapb.DataType_Float:
			idata.Data[field.FieldID] = &FloatFieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    readFloatArray(blobReaders),
			}

		case schemapb.DataType_Double:
			idata.Data[field.FieldID] = &DoubleFieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    readDoubleArray(blobReaders),
			}
		}
	}

	return idata, nil
}

func ColumnBasedInsertMsgToInsertData(msg *msgstream.InsertMsg, collSchema *schemapb.CollectionSchema) (idata *InsertData, err error) {
	srcFields := make(map[FieldID]*schemapb.FieldData)
	for _, field := range msg.FieldsData {
		srcFields[field.FieldId] = field
	}

	idata = &InsertData{
		Data: make(map[FieldID]FieldData),
		// TODO: handle Infos.
		Infos: nil,
	}

	for _, field := range collSchema.Fields {
		switch field.DataType {
		case schemapb.DataType_FloatVector:
			dim, err := GetDimFromParams(field.TypeParams)
			if err != nil {
				log.Error("failed to get dim", zap.Error(err))
				return nil, err
			}

			srcData := srcFields[field.FieldID].GetVectors().GetFloatVector().GetData()

			fieldData := &FloatVectorFieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    make([]float32, 0, len(srcData)),
				Dim:     dim,
			}
			fieldData.Data = append(fieldData.Data, srcData...)

			idata.Data[field.FieldID] = fieldData

		case schemapb.DataType_BinaryVector:
			dim, err := GetDimFromParams(field.TypeParams)
			if err != nil {
				log.Error("failed to get dim", zap.Error(err))
				return nil, err
			}

			srcData := srcFields[field.FieldID].GetVectors().GetBinaryVector()

			fieldData := &BinaryVectorFieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    make([]byte, 0, len(srcData)),
				Dim:     dim,
			}
			fieldData.Data = append(fieldData.Data, srcData...)

			idata.Data[field.FieldID] = fieldData

		case schemapb.DataType_Bool:
			srcData := srcFields[field.FieldID].GetScalars().GetBoolData().GetData()

			fieldData := &BoolFieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    make([]bool, 0, len(srcData)),
			}
			fieldData.Data = append(fieldData.Data, srcData...)

			idata.Data[field.FieldID] = fieldData

		case schemapb.DataType_Int8:
			srcData := srcFields[field.FieldID].GetScalars().GetIntData().GetData()

			fieldData := &Int8FieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    make([]int8, 0, len(srcData)),
			}
			int8SrcData := make([]int8, len(srcData))
			for i := 0; i < len(srcData); i++ {
				int8SrcData[i] = int8(srcData[i])
			}
			fieldData.Data = append(fieldData.Data, int8SrcData...)

			idata.Data[field.FieldID] = fieldData

		case schemapb.DataType_Int16:
			srcData := srcFields[field.FieldID].GetScalars().GetIntData().GetData()

			fieldData := &Int16FieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    make([]int16, 0, len(srcData)),
			}
			int16SrcData := make([]int16, len(srcData))
			for i := 0; i < len(srcData); i++ {
				int16SrcData[i] = int16(srcData[i])
			}
			fieldData.Data = append(fieldData.Data, int16SrcData...)

			idata.Data[field.FieldID] = fieldData

		case schemapb.DataType_Int32:
			srcData := srcFields[field.FieldID].GetScalars().GetIntData().GetData()

			fieldData := &Int32FieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    make([]int32, 0, len(srcData)),
			}
			fieldData.Data = append(fieldData.Data, srcData...)

			idata.Data[field.FieldID] = fieldData

		case schemapb.DataType_Int64:
			fieldData := &Int64FieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    make([]int64, 0),
			}

			switch field.FieldID {
			case 0: // rowIDs
				fieldData.Data = make([]int64, 0, len(msg.RowIDs))
				fieldData.Data = append(fieldData.Data, msg.RowIDs...)
			case 1: // Timestamps
				fieldData.Data = make([]int64, 0, len(msg.Timestamps))
				for _, ts := range msg.Timestamps {
					fieldData.Data = append(fieldData.Data, int64(ts))
				}
			default:
				srcData := srcFields[field.FieldID].GetScalars().GetLongData().GetData()
				fieldData.Data = make([]int64, 0, len(srcData))
				fieldData.Data = append(fieldData.Data, srcData...)
			}

			idata.Data[field.FieldID] = fieldData

		case schemapb.DataType_Float:
			srcData := srcFields[field.FieldID].GetScalars().GetFloatData().GetData()

			fieldData := &FloatFieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    make([]float32, 0, len(srcData)),
			}
			fieldData.Data = append(fieldData.Data, srcData...)

			idata.Data[field.FieldID] = fieldData

		case schemapb.DataType_Double:
			srcData := srcFields[field.FieldID].GetScalars().GetDoubleData().GetData()

			fieldData := &DoubleFieldData{
				NumRows: []int64{int64(msg.NRows())},
				Data:    make([]float64, 0, len(srcData)),
			}
			fieldData.Data = append(fieldData.Data, srcData...)

			idata.Data[field.FieldID] = fieldData
		case schemapb.DataType_String, schemapb.DataType_VarChar:
			srcData := srcFields[field.FieldID].GetScalars().GetStringData().GetData()

			fieldData := &StringFieldData{
				NumRows: []int64{int64(msg.NumRows)},
				Data:    make([]string, 0, len(srcData)),
			}

			fieldData.Data = append(fieldData.Data, srcData...)
			idata.Data[field.FieldID] = fieldData
		}
	}

	return idata, nil
}

func InsertMsgToInsertData(msg *msgstream.InsertMsg, schema *schemapb.CollectionSchema) (idata *InsertData, err error) {
	if msg.IsRowBased() {
		return RowBasedInsertMsgToInsertData(msg, schema)
	}
	return ColumnBasedInsertMsgToInsertData(msg, schema)
}

func mergeBoolField(data *InsertData, fid FieldID, field *BoolFieldData) {
	if _, ok := data.Data[fid]; !ok {
		fieldData := &BoolFieldData{
			NumRows: []int64{0},
			Data:    nil,
		}
		data.Data[fid] = fieldData
	}
	fieldData := data.Data[fid].(*BoolFieldData)
	fieldData.Data = append(fieldData.Data, field.Data...)
	fieldData.NumRows[0] += int64(field.RowNum())
}

func mergeInt8Field(data *InsertData, fid FieldID, field *Int8FieldData) {
	if _, ok := data.Data[fid]; !ok {
		fieldData := &Int8FieldData{
			NumRows: []int64{0},
			Data:    nil,
		}
		data.Data[fid] = fieldData
	}
	fieldData := data.Data[fid].(*Int8FieldData)
	fieldData.Data = append(fieldData.Data, field.Data...)
	fieldData.NumRows[0] += int64(field.RowNum())
}

func mergeInt16Field(data *InsertData, fid FieldID, field *Int16FieldData) {
	if _, ok := data.Data[fid]; !ok {
		fieldData := &Int16FieldData{
			NumRows: []int64{0},
			Data:    nil,
		}
		data.Data[fid] = fieldData
	}
	fieldData := data.Data[fid].(*Int16FieldData)
	fieldData.Data = append(fieldData.Data, field.Data...)
	fieldData.NumRows[0] += int64(field.RowNum())
}

func mergeInt32Field(data *InsertData, fid FieldID, field *Int32FieldData) {
	if _, ok := data.Data[fid]; !ok {
		fieldData := &Int32FieldData{
			NumRows: []int64{0},
			Data:    nil,
		}
		data.Data[fid] = fieldData
	}
	fieldData := data.Data[fid].(*Int32FieldData)
	fieldData.Data = append(fieldData.Data, field.Data...)
	fieldData.NumRows[0] += int64(field.RowNum())
}

func mergeInt64Field(data *InsertData, fid FieldID, field *Int64FieldData) {
	if _, ok := data.Data[fid]; !ok {
		fieldData := &Int64FieldData{
			NumRows: []int64{0},
			Data:    nil,
		}
		data.Data[fid] = fieldData
	}
	fieldData := data.Data[fid].(*Int64FieldData)
	fieldData.Data = append(fieldData.Data, field.Data...)
	fieldData.NumRows[0] += int64(field.RowNum())
}

func mergeFloatField(data *InsertData, fid FieldID, field *FloatFieldData) {
	if _, ok := data.Data[fid]; !ok {
		fieldData := &FloatFieldData{
			NumRows: []int64{0},
			Data:    nil,
		}
		data.Data[fid] = fieldData
	}
	fieldData := data.Data[fid].(*FloatFieldData)
	fieldData.Data = append(fieldData.Data, field.Data...)
	fieldData.NumRows[0] += int64(field.RowNum())
}

func mergeDoubleField(data *InsertData, fid FieldID, field *DoubleFieldData) {
	if _, ok := data.Data[fid]; !ok {
		fieldData := &DoubleFieldData{
			NumRows: []int64{0},
			Data:    nil,
		}
		data.Data[fid] = fieldData
	}
	fieldData := data.Data[fid].(*DoubleFieldData)
	fieldData.Data = append(fieldData.Data, field.Data...)
	fieldData.NumRows[0] += int64(field.RowNum())
}

func mergeStringField(data *InsertData, fid FieldID, field *StringFieldData) {
	if _, ok := data.Data[fid]; !ok {
		fieldData := &StringFieldData{
			NumRows: []int64{0},
			Data:    nil,
		}
		data.Data[fid] = fieldData
	}
	fieldData := data.Data[fid].(*StringFieldData)
	fieldData.Data = append(fieldData.Data, field.Data...)
	fieldData.NumRows[0] += int64(field.RowNum())
}

func mergeBinaryVectorField(data *InsertData, fid FieldID, field *BinaryVectorFieldData) {
	if _, ok := data.Data[fid]; !ok {
		fieldData := &BinaryVectorFieldData{
			NumRows: []int64{0},
			Data:    nil,
			Dim:     field.Dim,
		}
		data.Data[fid] = fieldData
	}
	fieldData := data.Data[fid].(*BinaryVectorFieldData)
	fieldData.Data = append(fieldData.Data, field.Data...)
	fieldData.NumRows[0] += int64(field.RowNum())
}

func mergeFloatVectorField(data *InsertData, fid FieldID, field *FloatVectorFieldData) {
	if _, ok := data.Data[fid]; !ok {
		fieldData := &FloatVectorFieldData{
			NumRows: []int64{0},
			Data:    nil,
			Dim:     field.Dim,
		}
		data.Data[fid] = fieldData
	}
	fieldData := data.Data[fid].(*FloatVectorFieldData)
	fieldData.Data = append(fieldData.Data, field.Data...)
	fieldData.NumRows[0] += int64(field.RowNum())
}

// MergeFieldData merge field into data.
func MergeFieldData(data *InsertData, fid FieldID, field FieldData) {
	if field == nil {
		return
	}
	switch field := field.(type) {
	case *BoolFieldData:
		mergeBoolField(data, fid, field)
	case *Int8FieldData:
		mergeInt8Field(data, fid, field)
	case *Int16FieldData:
		mergeInt16Field(data, fid, field)
	case *Int32FieldData:
		mergeInt32Field(data, fid, field)
	case *Int64FieldData:
		mergeInt64Field(data, fid, field)
	case *FloatFieldData:
		mergeFloatField(data, fid, field)
	case *DoubleFieldData:
		mergeDoubleField(data, fid, field)
	case *StringFieldData:
		mergeStringField(data, fid, field)
	case *BinaryVectorFieldData:
		mergeBinaryVectorField(data, fid, field)
	case *FloatVectorFieldData:
		mergeFloatVectorField(data, fid, field)
	}
}

// MergeInsertData merge insert datas. Maybe there are large write zoom if frequent inserts are met.
func MergeInsertData(datas ...*InsertData) *InsertData {
	ret := &InsertData{
		Data:  make(map[FieldID]FieldData),
		Infos: nil,
	}
	for _, data := range datas {
		if data != nil {
			for fid, field := range data.Data {
				MergeFieldData(ret, fid, field)
			}

			// TODO: handle storage.InsertData.Infos
			ret.Infos = append(ret.Infos, data.Infos...)
		}
	}
	return ret
}

// TODO: string type.
func GetPkFromInsertData(collSchema *schemapb.CollectionSchema, data *InsertData) (FieldData, error) {
	helper, err := typeutil.CreateSchemaHelper(collSchema)
	if err != nil {
		log.Error("failed to create schema helper", zap.Error(err))
		return nil, err
	}

	pf, err := helper.GetPrimaryKeyField()
	if err != nil {
		log.Warn("no primary field found", zap.Error(err))
		return nil, err
	}

	pfData, ok := data.Data[pf.FieldID]
	if !ok {
		log.Warn("no primary field found in insert msg", zap.Int64("fieldID", pf.FieldID))
		return nil, errors.New("no primary field found in insert msg")
	}

	var realPfData FieldData
	switch pf.DataType {
	case schemapb.DataType_Int64:
		realPfData, ok = pfData.(*Int64FieldData)
	case schemapb.DataType_VarChar:
		realPfData, ok = pfData.(*StringFieldData)
	default:
		//TODO
	}
	if !ok {
		log.Warn("primary field not in Int64 or VarChar format", zap.Int64("fieldID", pf.FieldID))
		return nil, errors.New("primary field not in Int64 or VarChar format")
	}

	return realPfData, nil
}

func boolFieldDataToPbBytes(field *BoolFieldData) ([]byte, error) {
	arr := &schemapb.BoolArray{Data: field.Data}
	return proto.Marshal(arr)
}

func stringFieldDataToPbBytes(field *StringFieldData) ([]byte, error) {
	arr := &schemapb.StringArray{Data: field.Data}
	return proto.Marshal(arr)
}

func binaryWrite(endian binary.ByteOrder, data interface{}) ([]byte, error) {
	buf := new(bytes.Buffer)
	err := binary.Write(buf, endian, data)
	if err != nil {
		return nil, err
	}
	return buf.Bytes(), nil
}

// FieldDataToBytes encode field data to byte slice.
// For some fixed-length data, such as int32, int64, float vector, use binary.Write directly.
// For binary vector, return it directly.
// For bool data, first transfer to schemapb.BoolArray and then marshal it. (TODO: handle bool like other scalar data.)
// For variable-length data, such as string, first transfer to schemapb.StringArray and then marshal it.
// TODO: find a proper way to store variable-length data. Or we should unify to use protobuf?
func FieldDataToBytes(endian binary.ByteOrder, fieldData FieldData) ([]byte, error) {
	switch field := fieldData.(type) {
	case *BoolFieldData:
		// return binaryWrite(endian, field.Data)
		return boolFieldDataToPbBytes(field)
	case *StringFieldData:
		return stringFieldDataToPbBytes(field)
	case *BinaryVectorFieldData:
		return field.Data, nil
	case *FloatVectorFieldData:
		return binaryWrite(endian, field.Data)
	case *Int8FieldData:
		return binaryWrite(endian, field.Data)
	case *Int16FieldData:
		return binaryWrite(endian, field.Data)
	case *Int32FieldData:
		return binaryWrite(endian, field.Data)
	case *Int64FieldData:
		return binaryWrite(endian, field.Data)
	case *FloatFieldData:
		return binaryWrite(endian, field.Data)
	case *DoubleFieldData:
		return binaryWrite(endian, field.Data)
	default:
		return nil, fmt.Errorf("unsupported field data: %s", field)
	}
}