From bae195cf176f98cff94a8139e433f89403485f23 Mon Sep 17 00:00:00 2001 From: bigsheeper Date: Wed, 4 Dec 2024 21:05:14 +0800 Subject: [PATCH] Reduce mutex contention in datacoord meta Signed-off-by: bigsheeper --- .../datacoord/compaction_policy_clustering.go | 7 +- .../datacoord/compaction_policy_single.go | 7 +- internal/datacoord/compaction_trigger.go | 13 +- internal/datacoord/compaction_trigger_test.go | 190 +++++++++++------- internal/datacoord/meta.go | 21 +- internal/datacoord/meta_test.go | 4 +- 6 files changed, 145 insertions(+), 97 deletions(-) diff --git a/internal/datacoord/compaction_policy_clustering.go b/internal/datacoord/compaction_policy_clustering.go index 1e7d652c9b791..ff33b327c5880 100644 --- a/internal/datacoord/compaction_policy_clustering.go +++ b/internal/datacoord/compaction_policy_clustering.go @@ -120,15 +120,14 @@ func (policy *clusteringCompactionPolicy) triggerOneCollection(ctx context.Conte return nil, 0, err } - partSegments := policy.meta.GetSegmentsChanPart(func(segment *SegmentInfo) bool { - return segment.CollectionID == collectionID && - isSegmentHealthy(segment) && + partSegments := GetSegmentsChanPart(policy.meta, WithCollection(collectionID), SegmentFilterFunc(func(segment *SegmentInfo) bool { + return isSegmentHealthy(segment) && isFlush(segment) && !segment.isCompacting && // not compacting now !segment.GetIsImporting() && // not importing now segment.GetLevel() != datapb.SegmentLevel_L0 && // ignore level zero segments !segment.GetIsInvisible() - }) + })) views := make([]CompactionView, 0) // partSegments is list of chanPartSegments, which is channel-partition organized segments diff --git a/internal/datacoord/compaction_policy_single.go b/internal/datacoord/compaction_policy_single.go index 6fabd2f610520..a262229e71484 100644 --- a/internal/datacoord/compaction_policy_single.go +++ b/internal/datacoord/compaction_policy_single.go @@ -87,15 +87,14 @@ func (policy *singleCompactionPolicy) triggerOneCollection(ctx context.Context, return nil, 0, err } - partSegments := policy.meta.GetSegmentsChanPart(func(segment *SegmentInfo) bool { - return segment.CollectionID == collectionID && - isSegmentHealthy(segment) && + partSegments := GetSegmentsChanPart(policy.meta, WithCollection(collectionID), SegmentFilterFunc(func(segment *SegmentInfo) bool { + return isSegmentHealthy(segment) && isFlush(segment) && !segment.isCompacting && // not compacting now !segment.GetIsImporting() && // not importing now segment.GetLevel() == datapb.SegmentLevel_L2 && // only support L2 for now !segment.GetIsInvisible() - }) + })) views := make([]CompactionView, 0) for _, group := range partSegments { diff --git a/internal/datacoord/compaction_trigger.go b/internal/datacoord/compaction_trigger.go index 103afc9570de0..84472b4814743 100644 --- a/internal/datacoord/compaction_trigger.go +++ b/internal/datacoord/compaction_trigger.go @@ -291,16 +291,21 @@ func (t *compactionTrigger) handleGlobalSignal(signal *compactionSignal) error { zap.Int64("signal.collectionID", signal.collectionID), zap.Int64("signal.partitionID", signal.partitionID), zap.Int64("signal.segmentID", signal.segmentID)) - partSegments := t.meta.GetSegmentsChanPart(func(segment *SegmentInfo) bool { - return (signal.collectionID == 0 || segment.CollectionID == signal.collectionID) && - isSegmentHealthy(segment) && + filters := []SegmentFilter{SegmentFilterFunc(func(segment *SegmentInfo) bool { + return isSegmentHealthy(segment) && isFlush(segment) && !segment.isCompacting && // not compacting now !segment.GetIsImporting() && // not importing now segment.GetLevel() != datapb.SegmentLevel_L0 && // ignore level zero segments segment.GetLevel() != datapb.SegmentLevel_L2 && // ignore l2 segment !segment.GetIsInvisible() - }) // partSegments is list of chanPartSegments, which is channel-partition organized segments + })} // partSegments is list of chanPartSegments, which is channel-partition organized segments + + // get all segments if signal.collection == 0, otherwise get collection segments + if signal.collectionID != 0 { + filters = append(filters, WithCollection(signal.collectionID)) + } + partSegments := GetSegmentsChanPart(t.meta, filters...) if len(partSegments) == 0 { log.Info("the length of SegmentsChanPart is 0, skip to handle compaction") diff --git a/internal/datacoord/compaction_trigger_test.go b/internal/datacoord/compaction_trigger_test.go index a5d1df1e318a2..0f03a74fa95bc 100644 --- a/internal/datacoord/compaction_trigger_test.go +++ b/internal/datacoord/compaction_trigger_test.go @@ -122,24 +122,34 @@ func Test_compactionTrigger_force_without_index(t *testing.T) { }, } + segInfo := &datapb.SegmentInfo{ + ID: 1, + CollectionID: collectionID, + PartitionID: 1, + LastExpireTime: 100, + NumOfRows: 100, + MaxRowNum: 300, + InsertChannel: "ch1", + State: commonpb.SegmentState_Flushed, + Binlogs: binlogs, + Deltalogs: deltaLogs, + IsSorted: true, + } m := &meta{ catalog: catalog, channelCPs: newChannelCps(), segments: &SegmentsInfo{ segments: map[int64]*SegmentInfo{ 1: { - SegmentInfo: &datapb.SegmentInfo{ - ID: 1, - CollectionID: collectionID, - PartitionID: 1, - LastExpireTime: 100, - NumOfRows: 100, - MaxRowNum: 300, - InsertChannel: "ch1", - State: commonpb.SegmentState_Flushed, - Binlogs: binlogs, - Deltalogs: deltaLogs, - IsSorted: true, + SegmentInfo: segInfo, + }, + }, + secondaryIndexes: segmentInfoIndexes{ + coll2Segments: map[UniqueID]map[UniqueID]*SegmentInfo{ + collectionID: { + 1: { + SegmentInfo: segInfo, + }, }, }, }, @@ -214,6 +224,76 @@ func Test_compactionTrigger_force(t *testing.T) { mock0Allocator := newMock0Allocator(t) + seg1 := &SegmentInfo{ + SegmentInfo: &datapb.SegmentInfo{ + ID: 1, + CollectionID: 2, + PartitionID: 1, + LastExpireTime: 100, + NumOfRows: 100, + MaxRowNum: 300, + InsertChannel: "ch1", + State: commonpb.SegmentState_Flushed, + Binlogs: []*datapb.FieldBinlog{ + { + Binlogs: []*datapb.Binlog{ + {EntriesNum: 5, LogID: 1}, + }, + }, + }, + Deltalogs: []*datapb.FieldBinlog{ + { + Binlogs: []*datapb.Binlog{ + {EntriesNum: 5, LogID: 1}, + }, + }, + }, + IsSorted: true, + }, + } + + seg2 := &SegmentInfo{ + SegmentInfo: &datapb.SegmentInfo{ + ID: 2, + CollectionID: 2, + PartitionID: 1, + LastExpireTime: 100, + NumOfRows: 100, + MaxRowNum: 300, + InsertChannel: "ch1", + State: commonpb.SegmentState_Flushed, + Binlogs: []*datapb.FieldBinlog{ + { + Binlogs: []*datapb.Binlog{ + {EntriesNum: 5, LogID: 2}, + }, + }, + }, + Deltalogs: []*datapb.FieldBinlog{ + { + Binlogs: []*datapb.Binlog{ + {EntriesNum: 5, LogID: 2}, + }, + }, + }, + IsSorted: true, + }, + } + + seg3 := &SegmentInfo{ + SegmentInfo: &datapb.SegmentInfo{ + ID: 3, + CollectionID: 1111, + PartitionID: 1, + LastExpireTime: 100, + NumOfRows: 100, + MaxRowNum: 300, + InsertChannel: "ch1", + State: commonpb.SegmentState_Flushed, + IsSorted: true, + }, + } + tests := []struct { name string fields fields @@ -230,71 +310,18 @@ func Test_compactionTrigger_force(t *testing.T) { channelCPs: newChannelCps(), segments: &SegmentsInfo{ segments: map[int64]*SegmentInfo{ - 1: { - SegmentInfo: &datapb.SegmentInfo{ - ID: 1, - CollectionID: 2, - PartitionID: 1, - LastExpireTime: 100, - NumOfRows: 100, - MaxRowNum: 300, - InsertChannel: "ch1", - State: commonpb.SegmentState_Flushed, - Binlogs: []*datapb.FieldBinlog{ - { - Binlogs: []*datapb.Binlog{ - {EntriesNum: 5, LogID: 1}, - }, - }, - }, - Deltalogs: []*datapb.FieldBinlog{ - { - Binlogs: []*datapb.Binlog{ - {EntriesNum: 5, LogID: 1}, - }, - }, - }, - IsSorted: true, - }, - }, - 2: { - SegmentInfo: &datapb.SegmentInfo{ - ID: 2, - CollectionID: 2, - PartitionID: 1, - LastExpireTime: 100, - NumOfRows: 100, - MaxRowNum: 300, - InsertChannel: "ch1", - State: commonpb.SegmentState_Flushed, - Binlogs: []*datapb.FieldBinlog{ - { - Binlogs: []*datapb.Binlog{ - {EntriesNum: 5, LogID: 2}, - }, - }, - }, - Deltalogs: []*datapb.FieldBinlog{ - { - Binlogs: []*datapb.Binlog{ - {EntriesNum: 5, LogID: 2}, - }, - }, - }, - IsSorted: true, + 1: seg1, + 2: seg2, + 3: seg3, + }, + secondaryIndexes: segmentInfoIndexes{ + coll2Segments: map[UniqueID]map[UniqueID]*SegmentInfo{ + 2: { + seg1.GetID(): seg1, + seg2.GetID(): seg2, }, - }, - 3: { - SegmentInfo: &datapb.SegmentInfo{ - ID: 3, - CollectionID: 1111, - PartitionID: 1, - LastExpireTime: 100, - NumOfRows: 100, - MaxRowNum: 300, - InsertChannel: "ch1", - State: commonpb.SegmentState_Flushed, - IsSorted: true, + 1111: { + seg3.GetID(): seg3, }, }, }, @@ -617,7 +644,13 @@ func Test_compactionTrigger_force(t *testing.T) { t.Run(tt.name+" with DiskANN index", func(t *testing.T) { for _, segment := range tt.fields.meta.segments.GetSegments() { // Collection 1000 means it has DiskANN index + delete(tt.fields.meta.segments.secondaryIndexes.coll2Segments[segment.GetCollectionID()], segment.GetID()) segment.CollectionID = 1000 + _, ok := tt.fields.meta.segments.secondaryIndexes.coll2Segments[segment.GetCollectionID()] + if !ok { + tt.fields.meta.segments.secondaryIndexes.coll2Segments[segment.GetCollectionID()] = make(map[UniqueID]*SegmentInfo) + } + tt.fields.meta.segments.secondaryIndexes.coll2Segments[segment.GetCollectionID()][segment.GetID()] = segment } tr := &compactionTrigger{ meta: tt.fields.meta, @@ -725,6 +758,9 @@ func Test_compactionTrigger_force_maxSegmentLimit(t *testing.T) { vecFieldID := int64(201) segmentInfos := &SegmentsInfo{ segments: make(map[UniqueID]*SegmentInfo), + secondaryIndexes: segmentInfoIndexes{ + coll2Segments: make(map[UniqueID]map[UniqueID]*SegmentInfo), + }, } indexMeta := newSegmentIndexMeta(nil) @@ -751,6 +787,7 @@ func Test_compactionTrigger_force_maxSegmentLimit(t *testing.T) { }, } + segmentInfos.secondaryIndexes.coll2Segments[2] = make(map[UniqueID]*SegmentInfo) for i := UniqueID(0); i < 50; i++ { info := &SegmentInfo{ SegmentInfo: &datapb.SegmentInfo{ @@ -793,6 +830,7 @@ func Test_compactionTrigger_force_maxSegmentLimit(t *testing.T) { }) segmentInfos.segments[i] = info + segmentInfos.secondaryIndexes.coll2Segments[2][i] = info } mock0Allocator := newMockAllocator(t) diff --git a/internal/datacoord/meta.go b/internal/datacoord/meta.go index 20f72ec454562..7980c93f79539 100644 --- a/internal/datacoord/meta.go +++ b/internal/datacoord/meta.go @@ -344,9 +344,8 @@ func (m *meta) GetClonedCollectionInfo(collectionID UniqueID) *collectionInfo { } // GetSegmentsChanPart returns segments organized in Channel-Partition dimension with selector applied -func (m *meta) GetSegmentsChanPart(selector SegmentInfoSelector) []*chanPartSegments { - m.RLock() - defer m.RUnlock() +// TODO: Move this function to the compaction module after reorganizing the DataCoord modules. +func GetSegmentsChanPart(m *meta, filters ...SegmentFilter) []*chanPartSegments { type dim struct { partitionID int64 channelName string @@ -354,10 +353,8 @@ func (m *meta) GetSegmentsChanPart(selector SegmentInfoSelector) []*chanPartSegm mDimEntry := make(map[dim]*chanPartSegments) - for _, si := range m.segments.segments { - if !selector(si) { - continue - } + candidates := m.SelectSegments(context.Background(), filters...) + for _, si := range candidates { d := dim{si.PartitionID, si.InsertChannel} entry, ok := mDimEntry[d] if !ok { @@ -1359,6 +1356,16 @@ func (m *meta) SetAllocations(segmentID UniqueID, allocations []*Allocation) { m.segments.SetAllocations(segmentID, allocations) } +// SetSegmentsAllocations set Segments allocations, will overwrite ALL original allocations +// Note that allocations is not persisted in KV store +func (m *meta) SetSegmentsAllocations(segmentsAllocations map[int64][]*Allocation) { + m.Lock() + defer m.Unlock() + for segmentID, allocations := range segmentsAllocations { + m.segments.SetAllocations(segmentID, allocations) + } +} + // SetCurrentRows set current row count for segment with provided `segmentID` // Note that currRows is not persisted in KV store func (m *meta) SetCurrentRows(segmentID UniqueID, rows int64) { diff --git a/internal/datacoord/meta_test.go b/internal/datacoord/meta_test.go index 2144060ddc166..c5d1f7ce709eb 100644 --- a/internal/datacoord/meta_test.go +++ b/internal/datacoord/meta_test.go @@ -609,7 +609,7 @@ func TestMeta_Basic(t *testing.T) { }) t.Run("Test GetSegmentsChanPart", func(t *testing.T) { - result := meta.GetSegmentsChanPart(func(*SegmentInfo) bool { return true }) + result := GetSegmentsChanPart(meta, SegmentFilterFunc(func(segment *SegmentInfo) bool { return true })) assert.Equal(t, 2, len(result)) for _, entry := range result { assert.Equal(t, "c1", entry.channelName) @@ -620,7 +620,7 @@ func TestMeta_Basic(t *testing.T) { assert.Equal(t, 1, len(entry.segments)) } } - result = meta.GetSegmentsChanPart(func(seg *SegmentInfo) bool { return seg.GetCollectionID() == 10 }) + result = GetSegmentsChanPart(meta, WithCollection(10)) assert.Equal(t, 0, len(result)) })