milvus/internal/storage/field_stats.go
wei liu c6a1c49e02
enhance: Use Blocked Bloom Filter instead of basic bloom fitler impl. (#33405)
issue: #32995
To speed up the construction and querying of Bloom filters, we chose a
blocked Bloom filter instead of a basic Bloom filter implementation.

WARN: This PR is compatible with old version bf impl, but if fall back
to old milvus version, it may causes bloom filter deserialize failed.

In single Bloom filter test cases with a capacity of 1,000,000 and a
false positive rate (FPR) of 0.001, the blocked Bloom filter is 5 times
faster than the basic Bloom filter in both querying and construction, at
the cost of a 30% increase in memory usage.

- Block BF construct time	{"time": "54.128131ms"}
- Block BF size	                {"size": 3021578}
- Block BF Test cost	        {"time": "55.407352ms"}
- Basic BF construct time	{"time": "210.262183ms"}
- Basic BF size	                {"size": 2396308}
- Basic BF Test cost	        {"time": "192.596229ms"}

In multi Bloom filter test cases with a capacity of 100,000, an FPR of
0.001, and 100 Bloom filters, we reuse the primary key locations for all
Bloom filters to avoid repeated hash computations. As a result, the
blocked Bloom filter is also 5 times faster than the basic Bloom filter
in querying.

- Block BF TestLocation cost    {"time": "529.97183ms"}
- Basic BF TestLocation cost	{"time": "3.197430181s"}

---------

Signed-off-by: Wei Liu <wei.liu@zilliz.com>
2024-05-31 17:49:45 +08:00

470 lines
13 KiB
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 (
"encoding/json"
"fmt"
"go.uber.org/zap"
"github.com/milvus-io/milvus-proto/go-api/v2/schemapb"
"github.com/milvus-io/milvus/internal/util/bloomfilter"
"github.com/milvus-io/milvus/pkg/common"
"github.com/milvus-io/milvus/pkg/log"
"github.com/milvus-io/milvus/pkg/util/merr"
"github.com/milvus-io/milvus/pkg/util/paramtable"
)
// FieldStats contains statistics data for any column
// todo: compatible to PrimaryKeyStats
type FieldStats struct {
FieldID int64 `json:"fieldID"`
Type schemapb.DataType `json:"type"`
Max ScalarFieldValue `json:"max"` // for scalar field
Min ScalarFieldValue `json:"min"` // for scalar field
BFType bloomfilter.BFType `json:"bfType"` // for scalar field
BF bloomfilter.BloomFilterInterface `json:"bf"` // for scalar field
Centroids []VectorFieldValue `json:"centroids"` // for vector field
}
// UnmarshalJSON unmarshal bytes to FieldStats
func (stats *FieldStats) UnmarshalJSON(data []byte) error {
var messageMap map[string]*json.RawMessage
err := json.Unmarshal(data, &messageMap)
if err != nil {
return err
}
if value, ok := messageMap["fieldID"]; ok && value != nil {
err = json.Unmarshal(*messageMap["fieldID"], &stats.FieldID)
if err != nil {
return err
}
} else {
return fmt.Errorf("invalid fieldStats, no fieldID")
}
stats.Type = schemapb.DataType_Int64
value, ok := messageMap["type"]
if !ok {
value, ok = messageMap["pkType"]
}
if ok && value != nil {
var typeValue int32
err = json.Unmarshal(*value, &typeValue)
if err != nil {
return err
}
if typeValue > 0 {
stats.Type = schemapb.DataType(typeValue)
}
}
isScalarField := false
switch stats.Type {
case schemapb.DataType_Int8:
stats.Max = &Int8FieldValue{}
stats.Min = &Int8FieldValue{}
isScalarField = true
case schemapb.DataType_Int16:
stats.Max = &Int16FieldValue{}
stats.Min = &Int16FieldValue{}
isScalarField = true
case schemapb.DataType_Int32:
stats.Max = &Int32FieldValue{}
stats.Min = &Int32FieldValue{}
isScalarField = true
case schemapb.DataType_Int64:
stats.Max = &Int64FieldValue{}
stats.Min = &Int64FieldValue{}
isScalarField = true
case schemapb.DataType_Float:
stats.Max = &FloatFieldValue{}
stats.Min = &FloatFieldValue{}
isScalarField = true
case schemapb.DataType_Double:
stats.Max = &DoubleFieldValue{}
stats.Min = &DoubleFieldValue{}
isScalarField = true
case schemapb.DataType_String:
stats.Max = &StringFieldValue{}
stats.Min = &StringFieldValue{}
isScalarField = true
case schemapb.DataType_VarChar:
stats.Max = &VarCharFieldValue{}
stats.Min = &VarCharFieldValue{}
isScalarField = true
case schemapb.DataType_FloatVector:
stats.Centroids = []VectorFieldValue{}
isScalarField = false
default:
// unsupported data type
}
if isScalarField {
if value, ok := messageMap["max"]; ok && value != nil {
err = json.Unmarshal(*messageMap["max"], &stats.Max)
if err != nil {
return err
}
}
if value, ok := messageMap["min"]; ok && value != nil {
err = json.Unmarshal(*messageMap["min"], &stats.Min)
if err != nil {
return err
}
}
// compatible with primaryKeyStats
if maxPkMessage, ok := messageMap["maxPk"]; ok && maxPkMessage != nil {
err = json.Unmarshal(*maxPkMessage, stats.Max)
if err != nil {
return err
}
}
if minPkMessage, ok := messageMap["minPk"]; ok && minPkMessage != nil {
err = json.Unmarshal(*minPkMessage, stats.Min)
if err != nil {
return err
}
}
bfType := bloomfilter.BasicBF
if bfTypeMessage, ok := messageMap["bfType"]; ok && bfTypeMessage != nil {
err := json.Unmarshal(*bfTypeMessage, &bfType)
if err != nil {
return err
}
stats.BFType = bfType
}
if bfMessage, ok := messageMap["bf"]; ok && bfMessage != nil {
bf, err := bloomfilter.UnmarshalJSON(*bfMessage, bfType)
if err != nil {
log.Warn("Failed to unmarshal bloom filter, use AlwaysTrueBloomFilter instead of return err", zap.Error(err))
bf = bloomfilter.AlwaysTrueBloomFilter
}
stats.BF = bf
}
} else {
stats.initCentroids(data, stats.Type)
err = json.Unmarshal(*messageMap["centroids"], &stats.Centroids)
if err != nil {
return err
}
}
return nil
}
func (stats *FieldStats) initCentroids(data []byte, dataType schemapb.DataType) {
type FieldStatsAux struct {
FieldID int64 `json:"fieldID"`
Type schemapb.DataType `json:"type"`
Max json.RawMessage `json:"max"`
Min json.RawMessage `json:"min"`
BF bloomfilter.BloomFilterInterface `json:"bf"`
Centroids []json.RawMessage `json:"centroids"`
}
// Unmarshal JSON into the auxiliary struct
var aux FieldStatsAux
if err := json.Unmarshal(data, &aux); err != nil {
return
}
for i := 0; i < len(aux.Centroids); i++ {
switch dataType {
case schemapb.DataType_FloatVector:
stats.Centroids = append(stats.Centroids, &FloatVectorFieldValue{})
default:
// other vector datatype
}
}
}
func (stats *FieldStats) UpdateByMsgs(msgs FieldData) {
switch stats.Type {
case schemapb.DataType_Int8:
data := msgs.(*Int8FieldData).Data
// return error: msgs must has one element at least
if len(data) < 1 {
return
}
b := make([]byte, 8)
for _, int8Value := range data {
pk := NewInt8FieldValue(int8Value)
stats.UpdateMinMax(pk)
common.Endian.PutUint64(b, uint64(int8Value))
stats.BF.Add(b)
}
case schemapb.DataType_Int16:
data := msgs.(*Int16FieldData).Data
// return error: msgs must has one element at least
if len(data) < 1 {
return
}
b := make([]byte, 8)
for _, int16Value := range data {
pk := NewInt16FieldValue(int16Value)
stats.UpdateMinMax(pk)
common.Endian.PutUint64(b, uint64(int16Value))
stats.BF.Add(b)
}
case schemapb.DataType_Int32:
data := msgs.(*Int32FieldData).Data
// return error: msgs must has one element at least
if len(data) < 1 {
return
}
b := make([]byte, 8)
for _, int32Value := range data {
pk := NewInt32FieldValue(int32Value)
stats.UpdateMinMax(pk)
common.Endian.PutUint64(b, uint64(int32Value))
stats.BF.Add(b)
}
case schemapb.DataType_Int64:
data := msgs.(*Int64FieldData).Data
// return error: msgs must has one element at least
if len(data) < 1 {
return
}
b := make([]byte, 8)
for _, int64Value := range data {
pk := NewInt64FieldValue(int64Value)
stats.UpdateMinMax(pk)
common.Endian.PutUint64(b, uint64(int64Value))
stats.BF.Add(b)
}
case schemapb.DataType_Float:
data := msgs.(*FloatFieldData).Data
// return error: msgs must has one element at least
if len(data) < 1 {
return
}
b := make([]byte, 8)
for _, floatValue := range data {
pk := NewFloatFieldValue(floatValue)
stats.UpdateMinMax(pk)
common.Endian.PutUint64(b, uint64(floatValue))
stats.BF.Add(b)
}
case schemapb.DataType_Double:
data := msgs.(*DoubleFieldData).Data
// return error: msgs must has one element at least
if len(data) < 1 {
return
}
b := make([]byte, 8)
for _, doubleValue := range data {
pk := NewDoubleFieldValue(doubleValue)
stats.UpdateMinMax(pk)
common.Endian.PutUint64(b, uint64(doubleValue))
stats.BF.Add(b)
}
case schemapb.DataType_String:
data := msgs.(*StringFieldData).Data
// return error: msgs must has one element at least
if len(data) < 1 {
return
}
for _, str := range data {
pk := NewStringFieldValue(str)
stats.UpdateMinMax(pk)
stats.BF.AddString(str)
}
case schemapb.DataType_VarChar:
data := msgs.(*StringFieldData).Data
// return error: msgs must has one element at least
if len(data) < 1 {
return
}
for _, str := range data {
pk := NewVarCharFieldValue(str)
stats.UpdateMinMax(pk)
stats.BF.AddString(str)
}
default:
// TODO::
}
}
func (stats *FieldStats) Update(pk ScalarFieldValue) {
stats.UpdateMinMax(pk)
switch stats.Type {
case schemapb.DataType_Int8:
data := pk.GetValue().(int8)
b := make([]byte, 8)
common.Endian.PutUint64(b, uint64(data))
stats.BF.Add(b)
case schemapb.DataType_Int16:
data := pk.GetValue().(int16)
b := make([]byte, 8)
common.Endian.PutUint64(b, uint64(data))
stats.BF.Add(b)
case schemapb.DataType_Int32:
data := pk.GetValue().(int32)
b := make([]byte, 8)
common.Endian.PutUint64(b, uint64(data))
stats.BF.Add(b)
case schemapb.DataType_Int64:
data := pk.GetValue().(int64)
b := make([]byte, 8)
common.Endian.PutUint64(b, uint64(data))
stats.BF.Add(b)
case schemapb.DataType_Float:
data := pk.GetValue().(float32)
b := make([]byte, 8)
common.Endian.PutUint64(b, uint64(data))
stats.BF.Add(b)
case schemapb.DataType_Double:
data := pk.GetValue().(float64)
b := make([]byte, 8)
common.Endian.PutUint64(b, uint64(data))
stats.BF.Add(b)
case schemapb.DataType_String:
data := pk.GetValue().(string)
stats.BF.AddString(data)
case schemapb.DataType_VarChar:
data := pk.GetValue().(string)
stats.BF.AddString(data)
default:
// todo support vector field
}
}
// UpdateMinMax update min and max value
func (stats *FieldStats) UpdateMinMax(pk ScalarFieldValue) {
if stats.Min == nil {
stats.Min = pk
} else if stats.Min.GT(pk) {
stats.Min = pk
}
if stats.Max == nil {
stats.Max = pk
} else if stats.Max.LT(pk) {
stats.Max = pk
}
}
// SetVectorCentroids update centroids value
func (stats *FieldStats) SetVectorCentroids(centroids ...VectorFieldValue) {
stats.Centroids = centroids
}
func NewFieldStats(fieldID int64, pkType schemapb.DataType, rowNum int64) (*FieldStats, error) {
if pkType == schemapb.DataType_FloatVector {
return &FieldStats{
FieldID: fieldID,
Type: pkType,
}, nil
}
bfType := paramtable.Get().CommonCfg.BloomFilterType.GetValue()
return &FieldStats{
FieldID: fieldID,
Type: pkType,
BFType: bloomfilter.BFTypeFromString(bfType),
BF: bloomfilter.NewBloomFilterWithType(
uint(rowNum),
paramtable.Get().CommonCfg.MaxBloomFalsePositive.GetAsFloat(),
bfType),
}, nil
}
// FieldStatsWriter writes stats to buffer
type FieldStatsWriter struct {
buffer []byte
}
// GetBuffer returns buffer
func (sw *FieldStatsWriter) GetBuffer() []byte {
return sw.buffer
}
// GenerateList writes Stats slice to buffer
func (sw *FieldStatsWriter) GenerateList(stats []*FieldStats) error {
b, err := json.Marshal(stats)
if err != nil {
return err
}
sw.buffer = b
return nil
}
// GenerateByData writes data from @msgs with @fieldID to @buffer
func (sw *FieldStatsWriter) GenerateByData(fieldID int64, pkType schemapb.DataType, msgs ...FieldData) error {
statsList := make([]*FieldStats, 0)
bfType := paramtable.Get().CommonCfg.BloomFilterType.GetValue()
for _, msg := range msgs {
stats := &FieldStats{
FieldID: fieldID,
Type: pkType,
BFType: bloomfilter.BFTypeFromString(bfType),
BF: bloomfilter.NewBloomFilterWithType(
uint(msg.RowNum()),
paramtable.Get().CommonCfg.MaxBloomFalsePositive.GetAsFloat(),
bfType),
}
stats.UpdateByMsgs(msg)
statsList = append(statsList, stats)
}
return sw.GenerateList(statsList)
}
// FieldStatsReader reads stats
type FieldStatsReader struct {
buffer []byte
}
// SetBuffer sets buffer
func (sr *FieldStatsReader) SetBuffer(buffer []byte) {
sr.buffer = buffer
}
// GetFieldStatsList returns buffer as FieldStats
func (sr *FieldStatsReader) GetFieldStatsList() ([]*FieldStats, error) {
var statsList []*FieldStats
err := json.Unmarshal(sr.buffer, &statsList)
if err != nil {
// Compatible to PrimaryKey Stats
stats := &FieldStats{}
errNew := json.Unmarshal(sr.buffer, &stats)
if errNew != nil {
return nil, merr.WrapErrParameterInvalid("valid JSON", string(sr.buffer), err.Error())
}
return []*FieldStats{stats}, nil
}
return statsList, nil
}
func DeserializeFieldStats(blob *Blob) ([]*FieldStats, error) {
if len(blob.Value) == 0 {
return []*FieldStats{}, nil
}
sr := &FieldStatsReader{}
sr.SetBuffer(blob.Value)
stats, err := sr.GetFieldStatsList()
if err != nil {
return nil, err
}
return stats, nil
}