milvus/internal/indexnode/task.go
zhenshan.cao 1287ca699a
Refine usage of TimeRecorder.Record (#23142)
Signed-off-by: zhenshan.cao <zhenshan.cao@zilliz.com>
2023-03-30 18:56:22 +08:00

559 lines
19 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 indexnode
import (
"context"
"encoding/json"
"fmt"
"runtime"
"runtime/debug"
"strconv"
"strings"
"time"
"github.com/cockroachdb/errors"
"go.uber.org/zap"
"github.com/milvus-io/milvus-proto/go-api/commonpb"
"github.com/milvus-io/milvus-proto/go-api/schemapb"
"github.com/milvus-io/milvus/internal/log"
"github.com/milvus-io/milvus/internal/metrics"
"github.com/milvus-io/milvus/internal/proto/indexpb"
"github.com/milvus-io/milvus/internal/storage"
"github.com/milvus-io/milvus/internal/util/funcutil"
"github.com/milvus-io/milvus/internal/util/indexcgowrapper"
"github.com/milvus-io/milvus/internal/util/indexparamcheck"
"github.com/milvus-io/milvus/internal/util/indexparams"
"github.com/milvus-io/milvus/internal/util/metautil"
"github.com/milvus-io/milvus/internal/util/paramtable"
"github.com/milvus-io/milvus/internal/util/retry"
"github.com/milvus-io/milvus/internal/util/timerecord"
)
var (
errCancel = fmt.Errorf("canceled")
diskUsageRatio = 4.0
)
type Blob = storage.Blob
type taskInfo struct {
cancel context.CancelFunc
state commonpb.IndexState
fileKeys []string
serializedSize uint64
failReason string
// task statistics
statistic *indexpb.JobInfo
}
type task interface {
Ctx() context.Context
Name() string
Prepare(context.Context) error
LoadData(context.Context) error
BuildIndex(context.Context) error
SaveIndexFiles(context.Context) error
OnEnqueue(context.Context) error
SetState(state commonpb.IndexState, failReason string)
GetState() commonpb.IndexState
Reset()
}
// IndexBuildTask is used to record the information of the index tasks.
type indexBuildTask struct {
ident string
cancel context.CancelFunc
ctx context.Context
cm storage.ChunkManager
index indexcgowrapper.CodecIndex
savePaths []string
req *indexpb.CreateJobRequest
BuildID UniqueID
nodeID UniqueID
ClusterID string
collectionID UniqueID
partitionID UniqueID
segmentID UniqueID
fieldID UniqueID
fieldData storage.FieldData
indexBlobs []*storage.Blob
newTypeParams map[string]string
newIndexParams map[string]string
serializedSize uint64
tr *timerecord.TimeRecorder
statistic indexpb.JobInfo
node *IndexNode
}
func (it *indexBuildTask) Reset() {
it.ident = ""
it.cancel = nil
it.ctx = nil
it.cm = nil
it.index = nil
it.savePaths = nil
it.req = nil
it.fieldData = nil
it.indexBlobs = nil
it.newTypeParams = nil
it.newIndexParams = nil
it.tr = nil
it.node = nil
}
// Ctx is the context of index tasks.
func (it *indexBuildTask) Ctx() context.Context {
return it.ctx
}
// Name is the name of task to build index.
func (it *indexBuildTask) Name() string {
return it.ident
}
func (it *indexBuildTask) SetState(state commonpb.IndexState, failReason string) {
it.node.storeTaskState(it.ClusterID, it.BuildID, state, failReason)
}
func (it *indexBuildTask) GetState() commonpb.IndexState {
return it.node.loadTaskState(it.ClusterID, it.BuildID)
}
// OnEnqueue enqueues indexing tasks.
func (it *indexBuildTask) OnEnqueue(ctx context.Context) error {
it.statistic.StartTime = time.Now().UnixMicro()
it.statistic.PodID = it.node.GetNodeID()
log.Ctx(ctx).Info("IndexNode IndexBuilderTask Enqueue", zap.Int64("buildID", it.BuildID), zap.Int64("segID", it.segmentID))
return nil
}
func (it *indexBuildTask) Prepare(ctx context.Context) error {
log.Ctx(ctx).Info("Begin to prepare indexBuildTask", zap.Int64("buildID", it.BuildID),
zap.Int64("Collection", it.collectionID), zap.Int64("SegmentID", it.segmentID))
typeParams := make(map[string]string)
indexParams := make(map[string]string)
// type params can be removed
for _, kvPair := range it.req.GetTypeParams() {
key, value := kvPair.GetKey(), kvPair.GetValue()
typeParams[key] = value
indexParams[key] = value
}
for _, kvPair := range it.req.GetIndexParams() {
key, value := kvPair.GetKey(), kvPair.GetValue()
indexParams[key] = value
}
it.newTypeParams = typeParams
it.newIndexParams = indexParams
it.statistic.IndexParams = it.req.GetIndexParams()
// ugly codes to get dimension
if dimStr, ok := typeParams["dim"]; ok {
var err error
it.statistic.Dim, err = strconv.ParseInt(dimStr, 10, 64)
if err != nil {
log.Ctx(ctx).Error("parse dimesion failed", zap.Error(err))
// ignore error
}
}
log.Ctx(ctx).Info("Successfully prepare indexBuildTask", zap.Int64("buildID", it.BuildID),
zap.Int64("Collection", it.collectionID), zap.Int64("SegmentIf", it.segmentID))
return nil
}
func (it *indexBuildTask) LoadData(ctx context.Context) error {
getValueByPath := func(path string) ([]byte, error) {
data, err := it.cm.Read(ctx, path)
if err != nil {
if errors.Is(err, ErrNoSuchKey) {
return nil, ErrNoSuchKey
}
return nil, err
}
return data, nil
}
getBlobByPath := func(path string) (*Blob, error) {
value, err := getValueByPath(path)
if err != nil {
return nil, err
}
return &Blob{
Key: path,
Value: value,
}, nil
}
toLoadDataPaths := it.req.GetDataPaths()
keys := make([]string, len(toLoadDataPaths))
blobs := make([]*Blob, len(toLoadDataPaths))
loadKey := func(idx int) error {
keys[idx] = toLoadDataPaths[idx]
blob, err := getBlobByPath(toLoadDataPaths[idx])
if err != nil {
return err
}
blobs[idx] = blob
return nil
}
// Use runtime.GOMAXPROCS(0) instead of runtime.NumCPU()
// to respect CPU quota of container/pod
// gomaxproc will be set by `automaxproc`, passing 0 will just retrieve the value
err := funcutil.ProcessFuncParallel(len(toLoadDataPaths), runtime.GOMAXPROCS(0), loadKey, "loadKey")
if err != nil {
log.Ctx(ctx).Warn("loadKey failed", zap.Error(err))
return err
}
loadFieldDataLatency := it.tr.CtxRecord(ctx, "load field data done")
metrics.IndexNodeLoadFieldLatency.WithLabelValues(strconv.FormatInt(paramtable.GetNodeID(), 10)).Observe(float64(loadFieldDataLatency.Milliseconds()))
err = it.decodeBlobs(ctx, blobs)
if err != nil {
log.Ctx(ctx).Info("failed to decode blobs", zap.Int64("buildID", it.BuildID),
zap.Int64("Collection", it.collectionID), zap.Int64("SegmentIf", it.segmentID), zap.Error(err))
} else {
log.Ctx(ctx).Info("Successfully load data", zap.Int64("buildID", it.BuildID),
zap.Int64("Collection", it.collectionID), zap.Int64("SegmentIf", it.segmentID))
}
blobs = nil
debug.FreeOSMemory()
return err
}
func (it *indexBuildTask) BuildIndex(ctx context.Context) error {
// support build diskann index
indexType := it.newIndexParams["index_type"]
if indexType == indexparamcheck.IndexDISKANN {
return it.BuildDiskAnnIndex(ctx)
}
dataset := indexcgowrapper.GenDataset(it.fieldData)
dType := dataset.DType
var err error
if dType != schemapb.DataType_None {
it.index, err = indexcgowrapper.NewCgoIndex(dType, it.newTypeParams, it.newIndexParams, it.req.GetStorageConfig())
if err == nil {
err = it.index.Build(dataset)
}
if err != nil {
log.Ctx(ctx).Error("failed to build index", zap.Error(err))
return err
}
}
buildIndexLatency := it.tr.RecordSpan()
metrics.IndexNodeKnowhereBuildIndexLatency.WithLabelValues(strconv.FormatInt(paramtable.GetNodeID(), 10)).Observe(float64(buildIndexLatency.Milliseconds()))
indexBlobs, err := it.index.Serialize()
if err != nil {
log.Ctx(ctx).Error("IndexNode index Serialize failed", zap.Error(err))
return err
}
log.Ctx(ctx).Info("index serialize done", zap.Int64("buildID", it.BuildID),
zap.Duration("duration", it.tr.RecordSpan()))
// use serialized size before encoding
it.serializedSize = 0
for _, blob := range indexBlobs {
it.serializedSize += uint64(len(blob.Value))
}
// early release index for gc, and we can ensure that Delete is idempotent.
if err := it.index.Delete(); err != nil {
log.Ctx(ctx).Error("IndexNode indexBuildTask Execute CIndexDelete failed", zap.Error(err))
}
var serializedIndexBlobs []*storage.Blob
codec := storage.NewIndexFileBinlogCodec()
serializedIndexBlobs, err = codec.Serialize(
it.req.BuildID,
it.req.IndexVersion,
it.collectionID,
it.partitionID,
it.segmentID,
it.fieldID,
it.newIndexParams,
it.req.IndexName,
it.req.IndexID,
indexBlobs,
)
if err != nil {
log.Warn("failed to serialize index", zap.Error(err))
return err
}
encodeIndexFileDur := it.tr.RecordSpan()
metrics.IndexNodeEncodeIndexFileLatency.WithLabelValues(strconv.FormatInt(paramtable.GetNodeID(), 10)).Observe(float64(encodeIndexFileDur.Milliseconds()))
it.indexBlobs = serializedIndexBlobs
log.Ctx(ctx).Info("Successfully build index", zap.Int64("buildID", it.BuildID),
zap.Int64("Collection", it.collectionID), zap.Int64("SegmentID", it.segmentID))
return nil
}
func (it *indexBuildTask) BuildDiskAnnIndex(ctx context.Context) error {
// check index node support disk index
if !Params.IndexNodeCfg.EnableDisk.GetAsBool() {
log.Ctx(ctx).Error("IndexNode don't support build disk index",
zap.String("index type", it.newIndexParams["index_type"]),
zap.Bool("enable disk", Params.IndexNodeCfg.EnableDisk.GetAsBool()))
return errors.New("index node don't support build disk index")
}
// check load size and size of field data
localUsedSize, err := indexcgowrapper.GetLocalUsedSize()
if err != nil {
log.Ctx(ctx).Error("IndexNode get local used size failed")
return errors.New("index node get local used size failed")
}
usedLocalSizeWhenBuild := int64(float64(it.fieldData.GetMemorySize())*diskUsageRatio) + localUsedSize
maxUsedLocalSize := int64(Params.IndexNodeCfg.DiskCapacityLimit.GetAsFloat() * Params.IndexNodeCfg.MaxDiskUsagePercentage.GetAsFloat())
if usedLocalSizeWhenBuild > maxUsedLocalSize {
log.Ctx(ctx).Error("IndexNode don't has enough disk size to build disk ann index",
zap.Int64("usedLocalSizeWhenBuild", usedLocalSizeWhenBuild),
zap.Int64("maxUsedLocalSize", maxUsedLocalSize))
return errors.New("index node don't has enough disk size to build disk ann index")
}
dataset := indexcgowrapper.GenDataset(it.fieldData)
dType := dataset.DType
if dType != schemapb.DataType_None {
// TODO:: too ugly
it.newIndexParams["collection_id"] = strconv.FormatInt(it.collectionID, 10)
it.newIndexParams["partition_id"] = strconv.FormatInt(it.partitionID, 10)
it.newIndexParams["segment_id"] = strconv.FormatInt(it.segmentID, 10)
it.newIndexParams["field_id"] = strconv.FormatInt(it.fieldID, 10)
it.newIndexParams["index_build_id"] = strconv.FormatInt(it.req.GetBuildID(), 10)
it.newIndexParams["index_id"] = strconv.FormatInt(it.req.IndexID, 10)
it.newIndexParams["index_version"] = strconv.FormatInt(it.req.GetIndexVersion(), 10)
err = indexparams.SetDiskIndexBuildParams(it.newIndexParams, it.statistic.NumRows)
if err != nil {
log.Ctx(ctx).Error("failed to fill disk index params", zap.Error(err))
return err
}
jsonIndexParams, err := json.Marshal(it.newIndexParams)
if err != nil {
log.Ctx(ctx).Error("failed to json marshal index params", zap.Error(err))
return err
}
log.Ctx(ctx).Info("disk index params are ready",
zap.Int64("buildID", it.BuildID),
zap.String("index params", string(jsonIndexParams)))
it.index, err = indexcgowrapper.NewCgoIndex(dType, it.newTypeParams, it.newIndexParams, it.req.GetStorageConfig())
if err != nil {
log.Ctx(ctx).Error("failed to create index", zap.Error(err))
} else {
err = it.index.Build(dataset)
}
if err != nil {
if it.index != nil && it.index.CleanLocalData() != nil {
log.Ctx(ctx).Error("failed to clean cached data on disk after build index failed",
zap.Int64("buildID", it.BuildID),
zap.Int64("index version", it.req.GetIndexVersion()))
}
log.Ctx(ctx).Error("failed to build index", zap.Error(err))
return err
}
}
buildIndexLatency := it.tr.RecordSpan()
metrics.IndexNodeKnowhereBuildIndexLatency.WithLabelValues(strconv.FormatInt(paramtable.GetNodeID(), 10)).Observe(float64(buildIndexLatency.Milliseconds()))
fileInfos, err := it.index.GetIndexFileInfo()
if err != nil {
log.Ctx(ctx).Error("IndexNode index Serialize failed", zap.Error(err))
return err
}
log.Ctx(ctx).Info("index serialize done", zap.Int64("buildID", it.BuildID),
zap.Duration("duration", it.tr.RecordSpan()))
// use serialized size before encoding
it.serializedSize = 0
for _, info := range fileInfos {
it.serializedSize += uint64(info.FileSize)
it.indexBlobs = append(it.indexBlobs, &storage.Blob{
Key: info.FileName,
Size: info.FileSize,
})
}
// early release index for gc, and we can ensure that Delete is idempotent.
if err := it.index.Delete(); err != nil {
log.Ctx(it.ctx).Error("IndexNode indexBuildTask Execute CIndexDelete failed", zap.Error(err))
}
encodeIndexFileDur := it.tr.RecordSpan()
metrics.IndexNodeEncodeIndexFileLatency.WithLabelValues(strconv.FormatInt(paramtable.GetNodeID(), 10)).Observe(float64(encodeIndexFileDur.Milliseconds()))
return nil
}
func (it *indexBuildTask) SaveIndexFiles(ctx context.Context) error {
// support build diskann index
indexType := it.newIndexParams["index_type"]
if indexType == indexparamcheck.IndexDISKANN {
return it.SaveDiskAnnIndexFiles(ctx)
}
blobCnt := len(it.indexBlobs)
savePaths := make([]string, blobCnt)
saveFileKeys := make([]string, blobCnt)
saveIndexFile := func(idx int) error {
blob := it.indexBlobs[idx]
savePath := metautil.BuildSegmentIndexFilePath(it.cm.RootPath(), it.req.BuildID,
it.req.IndexVersion, it.partitionID, it.segmentID, blob.Key)
saveFn := func() error {
return it.cm.Write(ctx, savePath, blob.Value)
}
if err := retry.Do(ctx, saveFn, retry.Attempts(5)); err != nil {
log.Ctx(ctx).Warn("index node save index file failed", zap.Error(err), zap.String("savePath", savePath))
return err
}
savePaths[idx] = savePath
saveFileKeys[idx] = blob.Key
return nil
}
// If an error occurs, return the error that the task state will be set to retry.
if err := funcutil.ProcessFuncParallel(blobCnt, runtime.NumCPU(), saveIndexFile, "saveIndexFile"); err != nil {
log.Ctx(ctx).Error("saveIndexFile fail")
return err
}
it.savePaths = savePaths
it.statistic.EndTime = time.Now().UnixMicro()
it.node.storeIndexFilesAndStatistic(it.ClusterID, it.BuildID, saveFileKeys, it.serializedSize, &it.statistic)
log.Ctx(ctx).Info("save index files done", zap.Strings("IndexFiles", savePaths))
saveIndexFileDur := it.tr.RecordSpan()
metrics.IndexNodeSaveIndexFileLatency.WithLabelValues(strconv.FormatInt(paramtable.GetNodeID(), 10)).Observe(float64(saveIndexFileDur.Milliseconds()))
it.tr.Elapse("index building all done")
log.Ctx(ctx).Info("Successfully save index files", zap.Int64("buildID", it.BuildID), zap.Int64("Collection", it.collectionID),
zap.Int64("partition", it.partitionID), zap.Int64("SegmentId", it.segmentID))
return nil
}
func (it *indexBuildTask) SaveDiskAnnIndexFiles(ctx context.Context) error {
savePaths := make([]string, len(it.indexBlobs))
saveFileKeys := make([]string, len(it.indexBlobs))
for i, blob := range it.indexBlobs {
savePath := blob.Key
savePaths[i] = savePath
// TODO: unify blob key to file key instead of full path
parts := strings.Split(blob.Key, "/")
if len(parts) == 0 {
return fmt.Errorf("invaild blob key: %s", blob.Key)
}
fileKey := parts[len(parts)-1]
saveFileKeys[i] = fileKey
}
// add indexparams file
codec := storage.NewIndexFileBinlogCodec()
indexParamBlob, err := codec.SerializeIndexParams(
it.req.GetBuildID(),
it.req.GetIndexVersion(),
it.collectionID,
it.partitionID,
it.segmentID,
it.fieldID,
it.newIndexParams,
it.req.IndexName,
it.req.IndexID,
)
if err != nil {
return err
}
indexParamPath := metautil.BuildSegmentIndexFilePath(it.cm.RootPath(), it.req.BuildID, it.req.IndexVersion,
it.partitionID, it.segmentID, indexParamBlob.Key)
saveFn := func() error {
return it.cm.Write(ctx, indexParamPath, indexParamBlob.Value)
}
if err := retry.Do(ctx, saveFn, retry.Attempts(5)); err != nil {
log.Ctx(ctx).Warn("index node save index param file failed", zap.Error(err), zap.String("savePath", indexParamPath))
return err
}
saveFileKeys = append(saveFileKeys, indexParamBlob.Key)
savePaths = append(savePaths, indexParamPath)
it.savePaths = savePaths
it.statistic.EndTime = time.Now().UnixMicro()
it.node.storeIndexFilesAndStatistic(it.ClusterID, it.BuildID, saveFileKeys, it.serializedSize, &it.statistic)
log.Ctx(ctx).Info("save index files done", zap.Strings("IndexFiles", savePaths))
saveIndexFileDur := it.tr.RecordSpan()
metrics.IndexNodeSaveIndexFileLatency.WithLabelValues(strconv.FormatInt(paramtable.GetNodeID(), 10)).Observe(float64(saveIndexFileDur.Milliseconds()))
it.tr.Elapse("index building all done")
log.Ctx(ctx).Info("IndexNode CreateIndex successfully ", zap.Int64("collect", it.collectionID),
zap.Int64("partition", it.partitionID), zap.Int64("segment", it.segmentID))
return nil
}
func (it *indexBuildTask) decodeBlobs(ctx context.Context, blobs []*storage.Blob) error {
var insertCodec storage.InsertCodec
collectionID, partitionID, segmentID, insertData, err2 := insertCodec.DeserializeAll(blobs)
if err2 != nil {
return err2
}
decodeDuration := it.tr.RecordSpan().Milliseconds()
metrics.IndexNodeDecodeFieldLatency.WithLabelValues(strconv.FormatInt(paramtable.GetNodeID(), 10)).Observe(float64(decodeDuration))
if len(insertData.Data) != 1 {
return errors.New("we expect only one field in deserialized insert data")
}
it.collectionID = collectionID
it.partitionID = partitionID
it.segmentID = segmentID
deserializeDur := it.tr.RecordSpan()
log.Ctx(ctx).Info("IndexNode deserialize data success",
zap.Int64("index id", it.req.IndexID),
zap.String("index name", it.req.IndexName),
zap.Int64("collectionID", it.collectionID),
zap.Int64("partitionID", it.partitionID),
zap.Int64("segmentID", it.segmentID),
zap.Duration("deserialize duration", deserializeDur))
// we can ensure that there blobs are in one Field
var data storage.FieldData
var fieldID storage.FieldID
for fID, value := range insertData.Data {
data = value
fieldID = fID
break
}
it.statistic.NumRows = int64(data.RowNum())
it.fieldID = fieldID
it.fieldData = data
return nil
}