fix: correct memory size estimation on arrays (#40312)

See: #40342

---------

Signed-off-by: Ted Xu <ted.xu@zilliz.com>

internal/flushcommon/syncmgr/pack_writer_test.go

@@ -150,14 +150,14 @@ func TestBulkPackWriter_Write(t *testing.T) {
 					{
 						EntriesNum: 10,
 						LogPath:    "files/delta_log/123/456/789/10000",
-						LogSize:    594,
+						LogSize:    592,
 						MemorySize: 283,
 					},
 				},
 			},
 			wantStats:     map[int64]*datapb.FieldBinlog{},
 			wantBm25Stats: map[int64]*datapb.FieldBinlog{},
-			wantSize:      594,
+			wantSize:      592,
 			wantErr:       nil,
 		},
 	}

internal/storage/serde.go

@@ -20,6 +20,7 @@ import (
 	"encoding/binary"
 	"io"
 	"math"
+	"strconv"
 	"sync"
 
 	"github.com/apache/arrow/go/v17/arrow"
@@ -32,6 +33,7 @@ import (
 
 	"github.com/milvus-io/milvus-proto/go-api/v2/schemapb"
 	"github.com/milvus-io/milvus/pkg/v2/common"
+	"github.com/milvus-io/milvus/pkg/v2/util/typeutil"
 )
 
 type Record interface {
@@ -713,8 +715,9 @@ type singleFieldRecordWriter struct {
 	schema      *arrow.Schema
 	writerProps *parquet.WriterProperties
 
-	numRows             int
+	numRows              int
-	writtenUncompressed uint64
+	writtenUncompressed  uint64
+	memoryExpansionRatio int
 }
 
 func (sfw *singleFieldRecordWriter) Write(r Record) error {
@@ -728,21 +731,45 @@ func (sfw *singleFieldRecordWriter) Write(r Record) error {
 }
 
 func (sfw *singleFieldRecordWriter) GetWrittenUncompressed() uint64 {
-	return sfw.writtenUncompressed
+	return sfw.writtenUncompressed * uint64(sfw.memoryExpansionRatio)
 }
 
 func (sfw *singleFieldRecordWriter) Close() error {
 	return sfw.fw.Close()
 }
 
-func newSingleFieldRecordWriter(fieldId FieldID, field arrow.Field, writer io.Writer, opts ...RecordWriterOptions) (*singleFieldRecordWriter, error) {
+func newSingleFieldRecordWriter(field *schemapb.FieldSchema, writer io.Writer, opts ...RecordWriterOptions) (*singleFieldRecordWriter, error) {
+	// calculate memory expansion ratio
+	// arrays are serialized by protobuf, where int values may be compacted, see https://protobuf.dev/reference/go/size
+	// to correct the actual size, we need to multiply the memory expansion ratio accordingly.
+	determineMemoryExpansionRatio := func(field *schemapb.FieldSchema) int {
+		if field.DataType == schemapb.DataType_Array {
+			switch field.GetElementType() {
+			case schemapb.DataType_Int16:
+				return 2
+			case schemapb.DataType_Int32:
+				return 4
+			case schemapb.DataType_Int64:
+				return 8
+			}
+		}
+		return 1
+	}
+	dim, _ := typeutil.GetDim(field)
 	w := &singleFieldRecordWriter{
-		fieldId: fieldId,
+		fieldId: field.FieldID,
-		schema:  arrow.NewSchema([]arrow.Field{field}, nil),
+		schema: arrow.NewSchema([]arrow.Field{
+			{
+				Name:     strconv.Itoa(int(field.FieldID)),
+				Type:     serdeMap[field.DataType].arrowType(int(dim)),
+				Nullable: true, // No nullable check here.
+			},
+		}, nil),
 		writerProps: parquet.NewWriterProperties(
 			parquet.WithMaxRowGroupLength(math.MaxInt64), // No additional grouping for now.
 			parquet.WithCompression(compress.Codecs.Zstd),
 			parquet.WithCompressionLevel(3)),
+		memoryExpansionRatio: determineMemoryExpansionRatio(field),
 	}
 	for _, o := range opts {
 		o(w)

internal/storage/serde_events.go

@@ -322,13 +322,7 @@ func (bsw *BinlogStreamWriter) GetRecordWriter() (RecordWriter, error) {
 		return bsw.rw, nil
 	}
 
-	fid := bsw.fieldSchema.FieldID
+	rw, err := newSingleFieldRecordWriter(bsw.fieldSchema, &bsw.buf, WithRecordWriterProps(getFieldWriterProps(bsw.fieldSchema)))
-	dim, _ := typeutil.GetDim(bsw.fieldSchema)
-	rw, err := newSingleFieldRecordWriter(fid, arrow.Field{
-		Name:     strconv.Itoa(int(fid)),
-		Type:     serdeMap[bsw.fieldSchema.DataType].arrowType(int(dim)),
-		Nullable: true, // No nullable check here.
-	}, &bsw.buf, WithRecordWriterProps(getFieldWriterProps(bsw.fieldSchema)))
 	if err != nil {
 		return nil, err
 	}
@@ -814,12 +808,7 @@ func (dsw *DeltalogStreamWriter) GetRecordWriter() (RecordWriter, error) {
 	if dsw.rw != nil {
 		return dsw.rw, nil
 	}
-	dim, _ := typeutil.GetDim(dsw.fieldSchema)
+	rw, err := newSingleFieldRecordWriter(dsw.fieldSchema, &dsw.buf, WithRecordWriterProps(getFieldWriterProps(dsw.fieldSchema)))
-	rw, err := newSingleFieldRecordWriter(dsw.fieldSchema.FieldID, arrow.Field{
-		Name:     dsw.fieldSchema.Name,
-		Type:     serdeMap[dsw.fieldSchema.DataType].arrowType(int(dim)),
-		Nullable: false,
-	}, &dsw.buf, WithRecordWriterProps(getFieldWriterProps(dsw.fieldSchema)))
 	if err != nil {
 		return nil, err
 	}

internal/storage/serde_events_test.go

@@ -97,9 +97,13 @@ func TestBinlogStreamWriter(t *testing.T) {
 	t.Run("test write", func(t *testing.T) {
 		size := 3
 
-		field := arrow.Field{Name: "bool", Type: arrow.FixedWidthTypes.Boolean}
+		field := &schemapb.FieldSchema{
+			FieldID:  1,
+			DataType: schemapb.DataType_Bool,
+		}
+
 		var w bytes.Buffer
-		rw, err := newSingleFieldRecordWriter(1, field, &w)
+		rw, err := newSingleFieldRecordWriter(field, &w)
 		assert.NoError(t, err)
 
 		builder := array.NewBooleanBuilder(memory.DefaultAllocator)
@@ -108,7 +112,7 @@ func TestBinlogStreamWriter(t *testing.T) {
 		defer arr.Release()
 		ar := array.NewRecord(
 			arrow.NewSchema(
-				[]arrow.Field{field},
+				[]arrow.Field{{Name: "bool", Type: arrow.FixedWidthTypes.Boolean}},
 				nil,
 			),
 			[]arrow.Array{arr},
@@ -213,6 +217,84 @@ func TestBinlogSerializeWriter(t *testing.T) {
 	})
 }
 
+func TestSize(t *testing.T) {
+	t.Run("test array of int", func(t *testing.T) {
+		size := 100
+		schema := &schemapb.CollectionSchema{Fields: []*schemapb.FieldSchema{
+			{
+				FieldID:     18,
+				Name:        "array",
+				DataType:    schemapb.DataType_Array,
+				ElementType: schemapb.DataType_Int32,
+			},
+		}}
+
+		writers := NewBinlogStreamWriters(0, 0, 0, schema.Fields)
+		writer, err := NewBinlogSerializeWriter(schema, 0, 0, writers, 7)
+		assert.NoError(t, err)
+
+		for i := 0; i < size; i++ {
+			e := int32(i)
+			d := []int32{e, e, e, e, e, e, e, e}
+			value := &Value{
+				Value: map[FieldID]any{
+					18: &schemapb.ScalarField{
+						Data: &schemapb.ScalarField_IntData{
+							IntData: &schemapb.IntArray{Data: d},
+						},
+					},
+				},
+			}
+			err := writer.Write(value)
+			assert.NoError(t, err)
+		}
+
+		err = writer.Close()
+		assert.NoError(t, err)
+		memSize := writer.WrittenMemorySize()
+		assert.Greater(t, memSize, uint64(8*4*size)) // written memory size should greater than data size
+		t.Log("writtern memory size", memSize)
+	})
+
+	t.Run("test array of varchar", func(t *testing.T) {
+		size := 100
+		schema := &schemapb.CollectionSchema{Fields: []*schemapb.FieldSchema{
+			{
+				FieldID:     18,
+				Name:        "array",
+				DataType:    schemapb.DataType_Array,
+				ElementType: schemapb.DataType_String,
+			},
+		}}
+
+		writers := NewBinlogStreamWriters(0, 0, 0, schema.Fields)
+		writer, err := NewBinlogSerializeWriter(schema, 0, 0, writers, 7)
+		assert.NoError(t, err)
+
+		for i := 0; i < size; i++ {
+			e := fmt.Sprintf("%4d", i)
+			d := []string{e, e, e, e, e, e, e, e}
+			value := &Value{
+				Value: map[FieldID]any{
+					18: &schemapb.ScalarField{
+						Data: &schemapb.ScalarField_StringData{
+							StringData: &schemapb.StringArray{Data: d},
+						},
+					},
+				},
+			}
+			err := writer.Write(value)
+			assert.NoError(t, err)
+		}
+
+		err = writer.Close()
+		assert.NoError(t, err)
+		memSize := writer.WrittenMemorySize()
+		assert.Greater(t, memSize, uint64(8*4*size)) // written memory size should greater than data size
+		t.Log("writtern memory size", memSize)
+	})
+}
+
 func BenchmarkSerializeWriter(b *testing.B) {
 	const (
 		dim     = 128