milvus/internal/util/indexparamcheck/conf_adapter.go

365 lines
10 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 indexparamcheck
import (
"strconv"
"github.com/milvus-io/milvus-proto/go-api/schemapb"
"github.com/milvus-io/milvus/internal/util/funcutil"
)
const (
// L2 represents Euclidean distance
L2 = "L2"
// IP represents inner product distance
IP = "IP"
// HAMMING represents hamming distance
HAMMING = "HAMMING"
// JACCARD represents jaccard distance
JACCARD = "JACCARD"
// TANIMOTO represents tanimoto distance
TANIMOTO = "TANIMOTO"
// SUBSTRUCTURE represents substructure distance
SUBSTRUCTURE = "SUBSTRUCTURE"
// SUPERSTRUCTURE represents superstructure distance
SUPERSTRUCTURE = "SUPERSTRUCTURE"
MinNBits = 1
MaxNBits = 16
DefaultNBits = 8
// MinNList is the lower limit of nlist that used in Index IVFxxx
MinNList = 1
// MaxNList is the upper limit of nlist that used in Index IVFxxx
MaxNList = 65536
// DefaultMinDim is the smallest dimension supported in Milvus
DefaultMinDim = 1
// DefaultMaxDim is the largest dimension supported in Milvus
DefaultMaxDim = 32768
// If Dim = 32 and raw vector data = 2G, query node need 24G disk space When loading the vectors' disk index
// If Dim = 2, and raw vector data = 2G, query node need 240G disk space When loading the vectors' disk index
// So DiskAnnMinDim should be greater than or equal to 32 to avoid running out of disk space
DiskAnnMinDim = 32
DiskAnnMaxDim = 1024
HNSWMinEfConstruction = 8
HNSWMaxEfConstruction = 512
HNSWMinM = 4
HNSWMaxM = 64
MinNTrees = 1
// too large of n_trees takes much time, if there is real requirement, change this threshold.
MaxNTrees = 1024
// DIM is a constant used to represent dimension
DIM = "dim"
// Metric is a constant used to metric type
Metric = "metric_type"
// NLIST is a constant used to nlist in Index IVFxxx
NLIST = "nlist"
NBITS = "nbits"
IVFM = "m"
EFConstruction = "efConstruction"
HNSWM = "M"
PQM = "PQM"
NTREES = "n_trees"
IndexMode = "index_mode"
CPUMode = "CPU"
GPUMode = "GPU"
)
// METRICS is a set of all metrics types supported for float vector.
var METRICS = []string{L2, IP} // const
// BinIDMapMetrics is a set of all metric types supported for binary vector.
var BinIDMapMetrics = []string{HAMMING, JACCARD, TANIMOTO, SUBSTRUCTURE, SUPERSTRUCTURE} // const
var BinIvfMetrics = []string{HAMMING, JACCARD, TANIMOTO} // const
var supportDimPerSubQuantizer = []int{32, 28, 24, 20, 16, 12, 10, 8, 6, 4, 3, 2, 1} // const
var supportSubQuantizer = []int{96, 64, 56, 48, 40, 32, 28, 24, 20, 16, 12, 8, 4, 3, 2, 1} // const
type ConfAdapter interface {
// CheckTrain returns true if the index can be built with the specific index parameters.
CheckTrain(map[string]string) bool
CheckValidDataType(dType schemapb.DataType) bool
}
// BaseConfAdapter checks if a `FLAT` index can be built.
type BaseConfAdapter struct {
}
// CheckTrain check whether the params contains supported metrics types
func (adapter *BaseConfAdapter) CheckTrain(params map[string]string) bool {
if !CheckIntByRange(params, DIM, DefaultMinDim, DefaultMaxDim) {
return false
}
return CheckStrByValues(params, Metric, METRICS)
}
// CheckValidDataType check whether the field data type is supported for the index type
func (adapter *BaseConfAdapter) CheckValidDataType(dType schemapb.DataType) bool {
return true
}
func newBaseConfAdapter() *BaseConfAdapter {
return &BaseConfAdapter{}
}
// IVFConfAdapter checks if a IVF index can be built.
type IVFConfAdapter struct {
BaseConfAdapter
}
// CheckTrain returns true if the index can be built with the specific index parameters.
func (adapter *IVFConfAdapter) CheckTrain(params map[string]string) bool {
if !CheckIntByRange(params, NLIST, MinNList, MaxNList) {
return false
}
// skip check number of rows
return adapter.BaseConfAdapter.CheckTrain(params)
}
func newIVFConfAdapter() *IVFConfAdapter {
return &IVFConfAdapter{}
}
// IVFPQConfAdapter checks if a IVF_PQ index can be built.
type IVFPQConfAdapter struct {
IVFConfAdapter
}
// CheckTrain checks if ivf-pq index can be built with the specific index parameters.
func (adapter *IVFPQConfAdapter) CheckTrain(params map[string]string) bool {
if !adapter.IVFConfAdapter.CheckTrain(params) {
return false
}
return adapter.checkPQParams(params)
}
func (adapter *IVFPQConfAdapter) checkPQParams(params map[string]string) bool {
dimStr, dimensionExist := params[DIM]
if !dimensionExist {
return false
}
dimension, err := strconv.Atoi(dimStr)
if err != nil { // invalid dimension
return false
}
// nbits can be set to default: 8
nbitsStr, nbitsExist := params[NBITS]
var nbits int
if !nbitsExist {
nbits = 8
} else {
nbits, err = strconv.Atoi(nbitsStr)
if err != nil { // invalid nbits
return false
}
}
mStr, ok := params[IVFM]
if !ok {
return false
}
m, err := strconv.Atoi(mStr)
if err != nil || m == 0 { // invalid m
return false
}
mode, ok := params[IndexMode]
if !ok {
mode = CPUMode
}
if mode == GPUMode && !adapter.checkGPUPQParams(dimension, m, nbits) {
return false
}
return adapter.checkCPUPQParams(dimension, m)
}
func (adapter *IVFPQConfAdapter) checkGPUPQParams(dimension, m, nbits int) bool {
/*
* Faiss 1.6
* Only 1, 2, 3, 4, 6, 8, 10, 12, 16, 20, 24, 28, 32 dims per sub-quantizer are currently supported with
* no precomputed codes. Precomputed codes supports any number of dimensions, but will involve memory overheads.
*/
subDim := dimension / m
return funcutil.SliceContain(supportSubQuantizer, m) && funcutil.SliceContain(supportDimPerSubQuantizer, subDim) && nbits == 8
}
func (adapter *IVFPQConfAdapter) checkCPUPQParams(dimension, m int) bool {
return (dimension % m) == 0
}
func newIVFPQConfAdapter() *IVFPQConfAdapter {
return &IVFPQConfAdapter{}
}
// IVFSQConfAdapter checks if a IVF_SQ index can be built.
type IVFSQConfAdapter struct {
IVFConfAdapter
}
func (adapter *IVFSQConfAdapter) checkNBits(params map[string]string) bool {
// cgo will set this key to DefaultNBits (8), which is the only value Milvus supports.
_, exist := params[NBITS]
if exist {
// 8 is the only supported nbits.
return CheckIntByRange(params, NBITS, DefaultNBits, DefaultNBits)
}
return true
}
// CheckTrain returns true if the index can be built with the specific index parameters.
func (adapter *IVFSQConfAdapter) CheckTrain(params map[string]string) bool {
if !adapter.checkNBits(params) {
return false
}
return adapter.IVFConfAdapter.CheckTrain(params)
}
func newIVFSQConfAdapter() *IVFSQConfAdapter {
return &IVFSQConfAdapter{}
}
type BinIDMAPConfAdapter struct {
BaseConfAdapter
}
// CheckTrain checks if a binary flat index can be built with the specific parameters.
func (adapter *BinIDMAPConfAdapter) CheckTrain(params map[string]string) bool {
if !CheckIntByRange(params, DIM, DefaultMinDim, DefaultMaxDim) {
return false
}
return CheckStrByValues(params, Metric, BinIDMapMetrics)
}
func newBinIDMAPConfAdapter() *BinIDMAPConfAdapter {
return &BinIDMAPConfAdapter{}
}
// BinIVFConfAdapter checks if a bin IFV index can be built.
type BinIVFConfAdapter struct {
BaseConfAdapter
}
// CheckTrain checks if a binary ivf index can be built with specific parameters.
func (adapter *BinIVFConfAdapter) CheckTrain(params map[string]string) bool {
if !CheckIntByRange(params, DIM, DefaultMinDim, DefaultMaxDim) {
return false
}
if !CheckIntByRange(params, NLIST, MinNList, MaxNList) {
return false
}
if !CheckStrByValues(params, Metric, BinIvfMetrics) {
return false
}
// skip checking the number of rows
return true
}
func newBinIVFConfAdapter() *BinIVFConfAdapter {
return &BinIVFConfAdapter{}
}
// HNSWConfAdapter checks if a hnsw index can be built.
type HNSWConfAdapter struct {
BaseConfAdapter
}
// CheckTrain checks if a hnsw index can be built with specific parameters.
func (adapter *HNSWConfAdapter) CheckTrain(params map[string]string) bool {
if !CheckIntByRange(params, EFConstruction, HNSWMinEfConstruction, HNSWMaxEfConstruction) {
return false
}
if !CheckIntByRange(params, HNSWM, HNSWMinM, HNSWMaxM) {
return false
}
return adapter.BaseConfAdapter.CheckTrain(params)
}
func newHNSWConfAdapter() *HNSWConfAdapter {
return &HNSWConfAdapter{}
}
// ANNOYConfAdapter checks if an ANNOY index can be built.
type ANNOYConfAdapter struct {
BaseConfAdapter
}
// CheckTrain checks if an annoy index can be built with specific parameters.
func (adapter *ANNOYConfAdapter) CheckTrain(params map[string]string) bool {
if !CheckIntByRange(params, NTREES, MinNTrees, MaxNTrees) {
return false
}
return adapter.BaseConfAdapter.CheckTrain(params)
}
func newANNOYConfAdapter() *ANNOYConfAdapter {
return &ANNOYConfAdapter{}
}
type DISKANNConfAdapter struct {
BaseConfAdapter
}
func (adapter *DISKANNConfAdapter) CheckTrain(params map[string]string) bool {
if !CheckIntByRange(params, DIM, DiskAnnMinDim, DiskAnnMaxDim) {
return false
}
return adapter.BaseConfAdapter.CheckTrain(params)
}
// CheckValidDataType check whether the field data type is supported for the index type
func (adapter *DISKANNConfAdapter) CheckValidDataType(dType schemapb.DataType) bool {
vecDataTypes := []schemapb.DataType{
schemapb.DataType_FloatVector,
}
return funcutil.SliceContain(vecDataTypes, dType)
}
func newDISKANNConfAdapter() *DISKANNConfAdapter {
return &DISKANNConfAdapter{}
}