milvus/internal/util/importutilv2/parquet/field_reader.go

616 lines
17 KiB
Go
Raw Normal View History

// 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 parquet
import (
"context"
"encoding/json"
"fmt"
"github.com/apache/arrow/go/v12/arrow"
"github.com/apache/arrow/go/v12/arrow/array"
"github.com/apache/arrow/go/v12/parquet/pqarrow"
"github.com/samber/lo"
"golang.org/x/exp/constraints"
"github.com/milvus-io/milvus-proto/go-api/v2/schemapb"
"github.com/milvus-io/milvus/internal/storage"
"github.com/milvus-io/milvus/pkg/util/merr"
"github.com/milvus-io/milvus/pkg/util/typeutil"
)
type FieldReader struct {
columnIndex int
columnReader *pqarrow.ColumnReader
dim int
field *schemapb.FieldSchema
}
func NewFieldReader(ctx context.Context, reader *pqarrow.FileReader, columnIndex int, field *schemapb.FieldSchema) (*FieldReader, error) {
columnReader, err := reader.GetColumn(ctx, columnIndex)
if err != nil {
return nil, err
}
var dim int64 = 1
if typeutil.IsVectorType(field.GetDataType()) && !typeutil.IsSparseFloatVectorType(field.GetDataType()) {
dim, err = typeutil.GetDim(field)
if err != nil {
return nil, err
}
}
cr := &FieldReader{
columnIndex: columnIndex,
columnReader: columnReader,
dim: int(dim),
field: field,
}
return cr, nil
}
func (c *FieldReader) Next(count int64) (any, error) {
switch c.field.GetDataType() {
case schemapb.DataType_Bool:
return ReadBoolData(c, count)
case schemapb.DataType_Int8:
return ReadIntegerOrFloatData[int8](c, count)
case schemapb.DataType_Int16:
return ReadIntegerOrFloatData[int16](c, count)
case schemapb.DataType_Int32:
return ReadIntegerOrFloatData[int32](c, count)
case schemapb.DataType_Int64:
return ReadIntegerOrFloatData[int64](c, count)
case schemapb.DataType_Float:
data, err := ReadIntegerOrFloatData[float32](c, count)
if err != nil {
return nil, err
}
if data == nil {
return nil, nil
}
return data, typeutil.VerifyFloats32(data.([]float32))
case schemapb.DataType_Double:
data, err := ReadIntegerOrFloatData[float64](c, count)
if err != nil {
return nil, err
}
if data == nil {
return nil, nil
}
return data, typeutil.VerifyFloats64(data.([]float64))
case schemapb.DataType_VarChar, schemapb.DataType_String:
return ReadStringData(c, count)
case schemapb.DataType_JSON:
return ReadJSONData(c, count)
case schemapb.DataType_BinaryVector, schemapb.DataType_Float16Vector, schemapb.DataType_BFloat16Vector:
return ReadBinaryData(c, count)
case schemapb.DataType_FloatVector:
arrayData, err := ReadIntegerOrFloatArrayData[float32](c, count)
if err != nil {
return nil, err
}
if arrayData == nil {
return nil, nil
}
vectors := lo.Flatten(arrayData.([][]float32))
return vectors, nil
case schemapb.DataType_SparseFloatVector:
return ReadSparseFloatVectorData(c, count)
case schemapb.DataType_Array:
return ReadArrayData(c, count)
default:
return nil, merr.WrapErrImportFailed(fmt.Sprintf("unsupported data type '%s' for field '%s'",
c.field.GetDataType().String(), c.field.GetName()))
}
}
func (c *FieldReader) Close() {}
func ReadBoolData(pcr *FieldReader, count int64) (any, error) {
chunked, err := pcr.columnReader.NextBatch(count)
if err != nil {
return nil, err
}
data := make([]bool, 0, count)
for _, chunk := range chunked.Chunks() {
dataNums := chunk.Data().Len()
boolReader, ok := chunk.(*array.Boolean)
if !ok {
return nil, WrapTypeErr("bool", chunk.DataType().Name(), pcr.field)
}
for i := 0; i < dataNums; i++ {
data = append(data, boolReader.Value(i))
}
}
if len(data) == 0 {
return nil, nil
}
return data, nil
}
func ReadIntegerOrFloatData[T constraints.Integer | constraints.Float](pcr *FieldReader, count int64) (any, error) {
chunked, err := pcr.columnReader.NextBatch(count)
if err != nil {
return nil, err
}
data := make([]T, 0, count)
for _, chunk := range chunked.Chunks() {
dataNums := chunk.Data().Len()
switch chunk.DataType().ID() {
case arrow.INT8:
int8Reader := chunk.(*array.Int8)
for i := 0; i < dataNums; i++ {
data = append(data, T(int8Reader.Value(i)))
}
case arrow.INT16:
int16Reader := chunk.(*array.Int16)
for i := 0; i < dataNums; i++ {
data = append(data, T(int16Reader.Value(i)))
}
case arrow.INT32:
int32Reader := chunk.(*array.Int32)
for i := 0; i < dataNums; i++ {
data = append(data, T(int32Reader.Value(i)))
}
case arrow.INT64:
int64Reader := chunk.(*array.Int64)
for i := 0; i < dataNums; i++ {
data = append(data, T(int64Reader.Value(i)))
}
case arrow.FLOAT32:
float32Reader := chunk.(*array.Float32)
for i := 0; i < dataNums; i++ {
data = append(data, T(float32Reader.Value(i)))
}
case arrow.FLOAT64:
float64Reader := chunk.(*array.Float64)
for i := 0; i < dataNums; i++ {
data = append(data, T(float64Reader.Value(i)))
}
default:
return nil, WrapTypeErr("integer|float", chunk.DataType().Name(), pcr.field)
}
}
if len(data) == 0 {
return nil, nil
}
return data, nil
}
func ReadStringData(pcr *FieldReader, count int64) (any, error) {
chunked, err := pcr.columnReader.NextBatch(count)
if err != nil {
return nil, err
}
data := make([]string, 0, count)
for _, chunk := range chunked.Chunks() {
dataNums := chunk.Data().Len()
stringReader, ok := chunk.(*array.String)
if !ok {
return nil, WrapTypeErr("string", chunk.DataType().Name(), pcr.field)
}
for i := 0; i < dataNums; i++ {
data = append(data, stringReader.Value(i))
}
}
if len(data) == 0 {
return nil, nil
}
return data, nil
}
func ReadJSONData(pcr *FieldReader, count int64) (any, error) {
// JSON field read data from string array Parquet
data, err := ReadStringData(pcr, count)
if err != nil {
return nil, err
}
if data == nil {
return nil, nil
}
byteArr := make([][]byte, 0)
for _, str := range data.([]string) {
var dummy interface{}
err = json.Unmarshal([]byte(str), &dummy)
if err != nil {
return nil, err
}
if pcr.field.GetIsDynamic() {
var dummy2 map[string]interface{}
err = json.Unmarshal([]byte(str), &dummy2)
if err != nil {
return nil, err
}
}
byteArr = append(byteArr, []byte(str))
}
return byteArr, nil
}
func ReadBinaryData(pcr *FieldReader, count int64) (any, error) {
dataType := pcr.field.GetDataType()
chunked, err := pcr.columnReader.NextBatch(count)
if err != nil {
return nil, err
}
data := make([]byte, 0, count)
for _, chunk := range chunked.Chunks() {
rows := chunk.Data().Len()
switch chunk.DataType().ID() {
case arrow.BINARY:
binaryReader := chunk.(*array.Binary)
for i := 0; i < rows; i++ {
data = append(data, binaryReader.Value(i)...)
}
case arrow.LIST:
listReader := chunk.(*array.List)
if !isVectorAligned(listReader.Offsets(), pcr.dim, dataType) {
return nil, merr.WrapErrImportFailed("%s not aligned", dataType.String())
}
uint8Reader, ok := listReader.ListValues().(*array.Uint8)
if !ok {
return nil, WrapTypeErr("binary", listReader.ListValues().DataType().Name(), pcr.field)
}
data = append(data, uint8Reader.Uint8Values()...)
default:
return nil, WrapTypeErr("binary", chunk.DataType().Name(), pcr.field)
}
}
if len(data) == 0 {
return nil, nil
}
return data, nil
}
func ReadSparseFloatVectorData(pcr *FieldReader, count int64) (any, error) {
data, err := ReadStringData(pcr, count)
if err != nil {
return nil, err
}
if data == nil {
return nil, nil
}
byteArr := make([][]byte, 0, count)
maxDim := uint32(0)
for _, str := range data.([]string) {
rowVec, err := typeutil.CreateSparseFloatRowFromJSON([]byte(str))
if err != nil {
return nil, merr.WrapErrImportFailed(fmt.Sprintf("Invalid JSON string for SparseFloatVector: '%s', err = %v", str, err))
}
byteArr = append(byteArr, rowVec)
elemCount := len(rowVec) / 8
maxIdx := typeutil.SparseFloatRowIndexAt(rowVec, elemCount-1)
if maxIdx+1 > maxDim {
maxDim = maxIdx + 1
}
}
return &storage.SparseFloatVectorFieldData{
SparseFloatArray: schemapb.SparseFloatArray{
Dim: int64(maxDim),
Contents: byteArr,
},
}, nil
}
func checkVectorAlignWithDim(offsets []int32, dim int32) bool {
for i := 1; i < len(offsets); i++ {
if offsets[i]-offsets[i-1] != dim {
return false
}
}
return true
}
func isVectorAligned(offsets []int32, dim int, dataType schemapb.DataType) bool {
if len(offsets) < 1 {
return false
}
switch dataType {
case schemapb.DataType_BinaryVector:
return checkVectorAlignWithDim(offsets, int32(dim/8))
case schemapb.DataType_FloatVector:
return checkVectorAlignWithDim(offsets, int32(dim))
case schemapb.DataType_Float16Vector, schemapb.DataType_BFloat16Vector:
return checkVectorAlignWithDim(offsets, int32(dim*2))
case schemapb.DataType_SparseFloatVector:
// JSON format, skip alignment check
return true
default:
return false
}
}
func ReadBoolArrayData(pcr *FieldReader, count int64) (any, error) {
chunked, err := pcr.columnReader.NextBatch(count)
if err != nil {
return nil, err
}
data := make([][]bool, 0, count)
for _, chunk := range chunked.Chunks() {
listReader, ok := chunk.(*array.List)
if !ok {
return nil, WrapTypeErr("list", chunk.DataType().Name(), pcr.field)
}
boolReader, ok := listReader.ListValues().(*array.Boolean)
if !ok {
return nil, WrapTypeErr("boolArray", chunk.DataType().Name(), pcr.field)
}
offsets := listReader.Offsets()
for i := 1; i < len(offsets); i++ {
start, end := offsets[i-1], offsets[i]
elementData := make([]bool, 0, end-start)
for j := start; j < end; j++ {
elementData = append(elementData, boolReader.Value(int(j)))
}
data = append(data, elementData)
}
}
if len(data) == 0 {
return nil, nil
}
return data, nil
}
func ReadIntegerOrFloatArrayData[T constraints.Integer | constraints.Float](pcr *FieldReader, count int64) (any, error) {
chunked, err := pcr.columnReader.NextBatch(count)
if err != nil {
return nil, err
}
data := make([][]T, 0, count)
getDataFunc := func(offsets []int32, getValue func(int) T) {
for i := 1; i < len(offsets); i++ {
start, end := offsets[i-1], offsets[i]
elementData := make([]T, 0, end-start)
for j := start; j < end; j++ {
elementData = append(elementData, getValue(int(j)))
}
data = append(data, elementData)
}
}
for _, chunk := range chunked.Chunks() {
listReader, ok := chunk.(*array.List)
if !ok {
return nil, WrapTypeErr("list", chunk.DataType().Name(), pcr.field)
}
offsets := listReader.Offsets()
dataType := pcr.field.GetDataType()
if typeutil.IsVectorType(dataType) && !isVectorAligned(offsets, pcr.dim, dataType) {
return nil, merr.WrapErrImportFailed("%s not aligned", dataType.String())
}
valueReader := listReader.ListValues()
switch valueReader.DataType().ID() {
case arrow.INT8:
int8Reader := valueReader.(*array.Int8)
getDataFunc(offsets, func(i int) T {
return T(int8Reader.Value(i))
})
case arrow.INT16:
int16Reader := valueReader.(*array.Int16)
getDataFunc(offsets, func(i int) T {
return T(int16Reader.Value(i))
})
case arrow.INT32:
int32Reader := valueReader.(*array.Int32)
getDataFunc(offsets, func(i int) T {
return T(int32Reader.Value(i))
})
case arrow.INT64:
int64Reader := valueReader.(*array.Int64)
getDataFunc(offsets, func(i int) T {
return T(int64Reader.Value(i))
})
case arrow.FLOAT32:
float32Reader := valueReader.(*array.Float32)
getDataFunc(offsets, func(i int) T {
return T(float32Reader.Value(i))
})
case arrow.FLOAT64:
float64Reader := valueReader.(*array.Float64)
getDataFunc(offsets, func(i int) T {
return T(float64Reader.Value(i))
})
default:
return nil, WrapTypeErr("integerArray|floatArray", chunk.DataType().Name(), pcr.field)
}
}
if len(data) == 0 {
return nil, nil
}
return data, nil
}
func ReadStringArrayData(pcr *FieldReader, count int64) (any, error) {
chunked, err := pcr.columnReader.NextBatch(count)
if err != nil {
return nil, err
}
data := make([][]string, 0, count)
for _, chunk := range chunked.Chunks() {
listReader, ok := chunk.(*array.List)
if !ok {
return nil, WrapTypeErr("list", chunk.DataType().Name(), pcr.field)
}
stringReader, ok := listReader.ListValues().(*array.String)
if !ok {
return nil, WrapTypeErr("stringArray", chunk.DataType().Name(), pcr.field)
}
offsets := listReader.Offsets()
for i := 1; i < len(offsets); i++ {
start, end := offsets[i-1], offsets[i]
elementData := make([]string, 0, end-start)
for j := start; j < end; j++ {
elementData = append(elementData, stringReader.Value(int(j)))
}
data = append(data, elementData)
}
}
if len(data) == 0 {
return nil, nil
}
return data, nil
}
func ReadArrayData(pcr *FieldReader, count int64) (any, error) {
data := make([]*schemapb.ScalarField, 0, count)
elementType := pcr.field.GetElementType()
switch elementType {
case schemapb.DataType_Bool:
boolArray, err := ReadBoolArrayData(pcr, count)
if err != nil {
return nil, err
}
if boolArray == nil {
return nil, nil
}
for _, elementArray := range boolArray.([][]bool) {
data = append(data, &schemapb.ScalarField{
Data: &schemapb.ScalarField_BoolData{
BoolData: &schemapb.BoolArray{
Data: elementArray,
},
},
})
}
case schemapb.DataType_Int8:
int8Array, err := ReadIntegerOrFloatArrayData[int32](pcr, count)
if err != nil {
return nil, err
}
if int8Array == nil {
return nil, nil
}
for _, elementArray := range int8Array.([][]int32) {
data = append(data, &schemapb.ScalarField{
Data: &schemapb.ScalarField_IntData{
IntData: &schemapb.IntArray{
Data: elementArray,
},
},
})
}
case schemapb.DataType_Int16:
int16Array, err := ReadIntegerOrFloatArrayData[int32](pcr, count)
if err != nil {
return nil, err
}
if int16Array == nil {
return nil, nil
}
for _, elementArray := range int16Array.([][]int32) {
data = append(data, &schemapb.ScalarField{
Data: &schemapb.ScalarField_IntData{
IntData: &schemapb.IntArray{
Data: elementArray,
},
},
})
}
case schemapb.DataType_Int32:
int32Array, err := ReadIntegerOrFloatArrayData[int32](pcr, count)
if err != nil {
return nil, err
}
if int32Array == nil {
return nil, nil
}
for _, elementArray := range int32Array.([][]int32) {
data = append(data, &schemapb.ScalarField{
Data: &schemapb.ScalarField_IntData{
IntData: &schemapb.IntArray{
Data: elementArray,
},
},
})
}
case schemapb.DataType_Int64:
int64Array, err := ReadIntegerOrFloatArrayData[int64](pcr, count)
if err != nil {
return nil, err
}
if int64Array == nil {
return nil, nil
}
for _, elementArray := range int64Array.([][]int64) {
data = append(data, &schemapb.ScalarField{
Data: &schemapb.ScalarField_LongData{
LongData: &schemapb.LongArray{
Data: elementArray,
},
},
})
}
case schemapb.DataType_Float:
float32Array, err := ReadIntegerOrFloatArrayData[float32](pcr, count)
if err != nil {
return nil, err
}
if float32Array == nil {
return nil, nil
}
for _, elementArray := range float32Array.([][]float32) {
data = append(data, &schemapb.ScalarField{
Data: &schemapb.ScalarField_FloatData{
FloatData: &schemapb.FloatArray{
Data: elementArray,
},
},
})
}
case schemapb.DataType_Double:
float64Array, err := ReadIntegerOrFloatArrayData[float64](pcr, count)
if err != nil {
return nil, err
}
if float64Array == nil {
return nil, nil
}
for _, elementArray := range float64Array.([][]float64) {
data = append(data, &schemapb.ScalarField{
Data: &schemapb.ScalarField_DoubleData{
DoubleData: &schemapb.DoubleArray{
Data: elementArray,
},
},
})
}
case schemapb.DataType_VarChar, schemapb.DataType_String:
stringArray, err := ReadStringArrayData(pcr, count)
if err != nil {
return nil, err
}
if stringArray == nil {
return nil, nil
}
for _, elementArray := range stringArray.([][]string) {
data = append(data, &schemapb.ScalarField{
Data: &schemapb.ScalarField_StringData{
StringData: &schemapb.StringArray{
Data: elementArray,
},
},
})
}
default:
return nil, merr.WrapErrImportFailed(fmt.Sprintf("unsupported data type '%s' for array field '%s'",
elementType.String(), pcr.field.GetName()))
}
return data, nil
}