Files
milvus/internal/querynodev2/segments/result.go
T
89f107fa00 enhance: optimize RPC protobuf decoding with fastpb codec (#50743)
issue: #50742

## What this PR does

Optimizes protobuf decoding on Milvus RPC read/write hot paths with a
hand-written decoder in `pkg/util/fastpb`. The final change contains no
gRPC connection-pool or connection-count changes.

## Fast decoder coverage

The gRPC `releaseCodec.Unmarshal` path dispatches these top-level
messages to fastpb:

- `internalpb.RetrieveResults`
- `milvuspb.InsertRequest`
- `milvuspb.UpsertRequest`

`schemapb.SearchResultData` is not a top-level codec-dispatched message.
It is decoded directly from `SlicedBlob` at the search/reduce hot paths
in:

- `internal/proxy/search_reduce_util.go`
- `internal/querynodev2/segments/result.go`
- `internal/querynodev2/local_worker.go`
- `internal/querynodev2/tasks/search_task_go_reduce.go`

Nested hot messages such as `FieldData`, scalar/vector arrays, and IDs
are decoded by the same package.

Main optimizations:

- varchar string arrays use one backing byte arena, reducing per-string
allocations
- packed float vectors use a single memcpy into self-owned memory
- cold, complex, unknown, and mismatched fields delegate or merge
through the official protobuf codec
- no new external dependency or source-buffer aliasing

## Compatibility boundary

The compatibility contract depends on the data path:

- **Untrusted client ingress (`InsertRequest` and `UpsertRequest`)**
validates proto3 strings as UTF-8 and is tested differentially against
`proto.Unmarshal`, including malformed input behavior.
- **Trusted internal result paths (`SearchResultData`,
`RetrieveResults`, and direct `FieldData` decoding)** assume canonical
Milvus/segcore protobuf output and intentionally skip UTF-8 validation
for performance. Therefore these paths do not claim equivalence for
arbitrary adversarial wire containing invalid UTF-8.

Within that boundary, the decoder preserves the protobuf behaviors
exercised by the differential tests:

- proto2 groups fall back to the official codec
- known fields with mismatched wire types fall back to the official
codec
- unknown and future fields are preserved through an official merge pass
- repeated singular messages merge from the raw wire, including
explicitly encoded proto3 default values
- repeated message-valued oneof members merge correctly, while different
variants remain last-wins
- mixed packed and unpacked scalar encodings preserve wire order
- descriptor field-set tripwires force decoder review when the protobuf
schema changes

The unsafe packed-float copy assumes the currently supported
little-endian Milvus targets such as x86-64 and arm64.

## Benchmarks

fastpb vs official `proto.Unmarshal`, using 1000-row payloads:

| path | payload | speedup | allocations/op |
|---|---|---:|---:|
| search / query | varchar | about 2x | about 1000-2035 to about 10-17 |
| search / query | float vector, dim 768 | about 6x | unchanged |
| insert | varchar with UTF-8 validation | about 1.8x | 1019 to about 10
|

## Verification

- unit and differential tests cover field descriptors, UTF-8 contracts,
wire-type mismatches, proto2 groups, unknown fields, packed/unpacked
ordering, oneof behavior, and repeated-message merge semantics
- randomized canonical-message equivalence tests cover search, retrieve,
insert, and upsert payloads
- `go test ./util/fastpb ./util/resource` passes
- `go vet ./util/fastpb ./util/resource` passes
- current full CI and end-to-end status is tracked by the PR checks

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Signed-off-by: xiaofanluan <xiaofan.luan@zilliz.com>
Signed-off-by: xiaofanluan <xf@hjjaq.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: xiaofanluan <xf@hjjaq.com>
2026-07-16 19:56:38 +08:00

275 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 segments
import (
"context"
"math"
"github.com/samber/lo"
"go.opentelemetry.io/otel"
"google.golang.org/protobuf/proto"
"github.com/milvus-io/milvus-proto/go-api/v3/schemapb"
"github.com/milvus-io/milvus/internal/util/reduce"
"github.com/milvus-io/milvus/internal/util/segcore"
"github.com/milvus-io/milvus/pkg/v3/mlog"
"github.com/milvus-io/milvus/pkg/v3/proto/internalpb"
"github.com/milvus-io/milvus/pkg/v3/proto/segcorepb"
"github.com/milvus-io/milvus/pkg/v3/util/fastpb"
"github.com/milvus-io/milvus/pkg/v3/util/merr"
"github.com/milvus-io/milvus/pkg/v3/util/paramtable"
"github.com/milvus-io/milvus/pkg/v3/util/typeutil"
)
var _ typeutil.ResultWithID = &internalpb.RetrieveResults{}
var _ typeutil.ResultWithID = &segcorepb.RetrieveResults{}
func ReduceSearchOnQueryNode(ctx context.Context, results []*internalpb.SearchResults, info *reduce.ResultInfo) (*internalpb.SearchResults, error) {
if info.GetIsAdvance() {
return ReduceAdvancedSearchResults(ctx, results)
}
return ReduceSearchResults(ctx, results, info)
}
func ReduceSearchResults(ctx context.Context, results []*internalpb.SearchResults, info *reduce.ResultInfo) (*internalpb.SearchResults, error) {
results = lo.Filter(results, func(result *internalpb.SearchResults, _ int) bool {
return result != nil && (result.GetSlicedBlob() != nil || result.GetResultData() != nil)
})
if len(results) == 1 {
mlog.Debug(ctx, "Shortcut return ReduceSearchResults", mlog.Any("result info", info))
return results[0], nil
}
ctx, sp := otel.Tracer(typeutil.QueryNodeRole).Start(ctx, "ReduceSearchResults")
defer sp.End()
channelsMvcc := make(map[string]uint64)
isTopkReduce := false
isRecallEvaluation := false
for _, r := range results {
for ch, ts := range r.GetChannelsMvcc() {
channelsMvcc[ch] = ts
}
if r.GetIsTopkReduce() {
isTopkReduce = true
}
if r.GetIsRecallEvaluation() {
isRecallEvaluation = true
}
// shouldn't let new SearchResults.MetricType to be empty, though the req.MetricType is empty
if info.GetMetricType() == "" {
info.SetMetricType(r.MetricType)
}
}
searchResultData, err := DecodeSearchResults(ctx, results)
if err != nil {
mlog.Warn(ctx, "shard leader decode search results errors", mlog.Err(err))
return nil, err
}
mlog.Debug(ctx, "shard leader get valid search results", mlog.Int("numbers", len(searchResultData)))
for i, sData := range searchResultData {
mlog.Debug(ctx, "reduceSearchResultData",
mlog.Int("result No.", i),
mlog.Int64("nq", sData.NumQueries),
mlog.Int64("topk", sData.TopK),
mlog.Int("ids.len", typeutil.GetSizeOfIDs(sData.Ids)),
mlog.Int("fieldsData.len", len(sData.FieldsData)))
}
searchReduce := InitSearchReducer(info)
reducedResultData, err := searchReduce.ReduceSearchResultData(ctx, searchResultData, info)
if err != nil {
mlog.Warn(ctx, "shard leader reduce errors", mlog.Err(err))
return nil, err
}
searchResults, err := EncodeSearchResultData(ctx, reducedResultData, info.GetNq(), info.GetTopK(), info.GetMetricType())
if err != nil {
mlog.Warn(ctx, "shard leader encode search result errors", mlog.Err(err))
return nil, err
}
requestCosts := lo.FilterMap(results, func(result *internalpb.SearchResults, _ int) (*internalpb.CostAggregation, bool) {
// delegator node won't be used to load sealed segment if stream node is enabled
// and if growing segment doesn't exists, delegator won't produce any cost metrics
// so we deprecate the EnableWorkerSQCostMetrics param
return result.GetCostAggregation(), true
})
searchResults.CostAggregation = mergeRequestCost(requestCosts)
if searchResults.CostAggregation == nil {
searchResults.CostAggregation = &internalpb.CostAggregation{}
}
relatedDataSize := lo.Reduce(results, func(acc int64, result *internalpb.SearchResults, _ int) int64 {
return acc + result.GetCostAggregation().GetTotalRelatedDataSize()
}, 0)
storageCost := lo.Reduce(results, func(acc segcore.StorageCost, result *internalpb.SearchResults, _ int) segcore.StorageCost {
acc.ScannedRemoteBytes += result.GetScannedRemoteBytes()
acc.ScannedTotalBytes += result.GetScannedTotalBytes()
return acc
}, segcore.StorageCost{})
searchResults.CostAggregation.TotalRelatedDataSize = relatedDataSize
searchResults.ChannelsMvcc = channelsMvcc
searchResults.IsTopkReduce = isTopkReduce
searchResults.IsRecallEvaluation = isRecallEvaluation
searchResults.ScannedRemoteBytes = storageCost.ScannedRemoteBytes
searchResults.ScannedTotalBytes = storageCost.ScannedTotalBytes
return searchResults, nil
}
func ReduceAdvancedSearchResults(ctx context.Context, results []*internalpb.SearchResults) (*internalpb.SearchResults, error) {
_, sp := otel.Tracer(typeutil.QueryNodeRole).Start(ctx, "ReduceAdvancedSearchResults")
defer sp.End()
channelsMvcc := make(map[string]uint64)
relatedDataSize := int64(0)
isTopkReduce := false
searchResults := &internalpb.SearchResults{
IsAdvanced: true,
}
storageCost := segcore.StorageCost{}
for index, result := range results {
if result.GetIsTopkReduce() {
isTopkReduce = true
}
relatedDataSize += result.GetCostAggregation().GetTotalRelatedDataSize()
storageCost.ScannedRemoteBytes += result.GetScannedRemoteBytes()
storageCost.ScannedTotalBytes += result.GetScannedTotalBytes()
for ch, ts := range result.GetChannelsMvcc() {
channelsMvcc[ch] = ts
}
searchResults.NumQueries = result.GetNumQueries()
// we just append here, no need to split subResult and reduce
// defer this reduction to proxy
subResult := &internalpb.SubSearchResults{
MetricType: result.GetMetricType(),
NumQueries: result.GetNumQueries(),
TopK: result.GetTopK(),
SlicedBlob: result.GetSlicedBlob(),
ResultData: result.GetResultData(),
SlicedNumCount: result.GetSlicedNumCount(),
SlicedOffset: result.GetSlicedOffset(),
ReqIndex: int64(index),
}
searchResults.SubResults = append(searchResults.SubResults, subResult)
}
searchResults.ChannelsMvcc = channelsMvcc
requestCosts := lo.FilterMap(results, func(result *internalpb.SearchResults, _ int) (*internalpb.CostAggregation, bool) {
// delegator node won't be used to load sealed segment if stream node is enabled
// and if growing segment doesn't exists, delegator won't produce any cost metrics
// so we deprecate the EnableWorkerSQCostMetrics param
return result.GetCostAggregation(), true
})
searchResults.CostAggregation = mergeRequestCost(requestCosts)
if searchResults.CostAggregation == nil {
searchResults.CostAggregation = &internalpb.CostAggregation{}
}
searchResults.CostAggregation.TotalRelatedDataSize = relatedDataSize
searchResults.IsTopkReduce = isTopkReduce
searchResults.ScannedRemoteBytes = storageCost.ScannedRemoteBytes
searchResults.ScannedTotalBytes = storageCost.ScannedTotalBytes
return searchResults, nil
}
func SelectSearchResultData(dataArray []*schemapb.SearchResultData, resultOffsets [][]int64, offsets []int64, qi int64) int {
var (
sel = -1
maxDistance = -float32(math.MaxFloat32)
resultDataIdx int64 = -1
)
for i, offset := range offsets { // query num, the number of ways to merge
if offset >= dataArray[i].Topks[qi] {
continue
}
idx := resultOffsets[i][qi] + offset
distance := dataArray[i].Scores[idx]
if distance > maxDistance {
sel = i
maxDistance = distance
resultDataIdx = idx
} else if distance == maxDistance {
if sel == -1 {
// A bad case happens where knowhere returns distance == +/-maxFloat32
// by mistake.
mlog.Warn(context.TODO(), "a bad distance is found, something is wrong here!", mlog.Float32("score", distance))
} else if typeutil.ComparePK(
typeutil.GetPK(dataArray[i].GetIds(), idx),
typeutil.GetPK(dataArray[sel].GetIds(), resultDataIdx)) {
sel = i
maxDistance = distance
resultDataIdx = idx
}
}
}
return sel
}
func DecodeSearchResults(ctx context.Context, searchResults []*internalpb.SearchResults) ([]*schemapb.SearchResultData, error) {
_, sp := otel.Tracer(typeutil.QueryNodeRole).Start(ctx, "DecodeSearchResults")
defer sp.End()
results := make([]*schemapb.SearchResultData, 0)
for _, partialSearchResult := range searchResults {
if partialSearchResult.ResultData != nil {
// Pre-decoded by delegator — use directly, no unmarshal needed.
results = append(results, partialSearchResult.ResultData)
} else if partialSearchResult.SlicedBlob != nil {
var partialResultData schemapb.SearchResultData
// fastpb: hand-written decoder for the search reduce hot path
// (wire-equivalent to proto.Unmarshal, ~2x varchar / ~6x vector).
err := fastpb.UnmarshalSearchResultData(partialSearchResult.SlicedBlob, &partialResultData)
if err != nil {
return nil, err
}
results = append(results, &partialResultData)
}
}
return results, nil
}
func EncodeSearchResultData(ctx context.Context, searchResultData *schemapb.SearchResultData,
nq int64, topk int64, metricType string,
) (searchResults *internalpb.SearchResults, err error) {
_, sp := otel.Tracer(typeutil.QueryNodeRole).Start(ctx, "EncodeSearchResultData")
defer sp.End()
searchResults = &internalpb.SearchResults{
Status: merr.Success(),
NumQueries: nq,
TopK: topk,
MetricType: metricType,
}
if searchResultData != nil && paramtable.Get().QueryNodeCfg.EnableResultZeroCopy.GetAsBool() {
// New path: embed struct directly, skip marshal
searchResults.ResultData = searchResultData
} else if searchResultData != nil {
// Legacy path: marshal to SlicedBlob
slicedBlob, err := proto.Marshal(searchResultData)
if err != nil {
return nil, err
}
searchResults.SlicedBlob = slicedBlob
}
return searchResults, err
}