github.com/whtcorpsinc/milevadb-prod@v0.0.0-20211104133533-f57f4be3b597/interlock/index_lookup_hash_join.go

// Copyright 2020 WHTCORPS INC, Inc.
//
// Licensed 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,
// See the License for the specific language governing permissions and
// limitations under the License.

package interlock

import (
	"context"
	"fmt"
	"hash"
	"hash/fnv"
	"runtime/trace"
	"sync"
	"sync/atomic"
	"time"

	"github.com/whtcorpsinc/errors"
	"github.com/whtcorpsinc/failpoint"
	causetembedded "github.com/whtcorpsinc/milevadb/causet/embedded"
	"github.com/whtcorpsinc/milevadb/memex"
	"github.com/whtcorpsinc/milevadb/soliton"
	"github.com/whtcorpsinc/milevadb/soliton/chunk"
	"github.com/whtcorpsinc/milevadb/soliton/codec"
	"github.com/whtcorpsinc/milevadb/soliton/memory"
	"github.com/whtcorpsinc/milevadb/soliton/ranger"
)

// numResChkHold indicates the number of resource chunks that an inner worker
// holds at the same time.
// It's used in 2 cases individually:
// 1. IndexMergeJoin
// 2. IndexNestedLoopHashJoin:
//    It's used when IndexNestedLoopHashJoin.keepOuterOrder is true.
//    Otherwise, there will be at most `concurrency` resource chunks throughout
//    the execution of IndexNestedLoopHashJoin.
const numResChkHold = 4

// IndexNestedLoopHashJoin employs one outer worker and N inner workers to
// execute concurrently. The output order is not promised.
//
// The execution flow is very similar to IndexLookUpReader:
// 1. The outer worker reads N outer rows, builds a task and sends it to the
// inner worker channel.
// 2. The inner worker receives the tasks and does 3 things for every task:
//    1. builds hash causet from the outer rows
//    2. builds key ranges from outer rows and fetches inner rows
//    3. probes the hash causet and sends the join result to the main thread channel.
//    Note: step 1 and step 2 runs concurrently.
// 3. The main thread receives the join results.
type IndexNestedLoopHashJoin struct {
	IndexLookUpJoin
	resultCh          chan *indexHashJoinResult
	joinChkResourceCh []chan *chunk.Chunk
	// We build individual joiner for each inner worker when using chunk-based
	// execution, to avoid the concurrency of joiner.chk and joiner.selected.
	joiners        []joiner
	keepOuterOrder bool
	curTask        *indexHashJoinTask
	// taskCh is only used when `keepOuterOrder` is true.
	taskCh chan *indexHashJoinTask

	stats *indexLookUpJoinRuntimeStats
}

type indexHashJoinOuterWorker struct {
	outerWorker
	innerCh        chan *indexHashJoinTask
	keepOuterOrder bool
	// taskCh is only used when the outer order needs to be promised.
	taskCh chan *indexHashJoinTask
}

type indexHashJoinInnerWorker struct {
	innerWorker
	matchedOuterPtrs  []chunk.EventPtr
	joiner            joiner
	joinChkResourceCh chan *chunk.Chunk
	// resultCh is valid only when indexNestedLoopHashJoin do not need to keep
	// order. Otherwise, it will be nil.
	resultCh         chan *indexHashJoinResult
	taskCh           <-chan *indexHashJoinTask
	wg               *sync.WaitGroup
	joinKeyBuf       []byte
	outerEventStatus []outerEventStatusFlag
}

type indexHashJoinResult struct {
	chk *chunk.Chunk
	err error
	src chan<- *chunk.Chunk
}

type indexHashJoinTask struct {
	*lookUpJoinTask
	outerEventStatus [][]outerEventStatusFlag
	lookupMap        baseHashBlock
	err              error
	keepOuterOrder   bool
	// resultCh is only used when the outer order needs to be promised.
	resultCh chan *indexHashJoinResult
	// matchedInnerEventPtrs is only valid when the outer order needs to be
	// promised. Otherwise, it will be nil.
	// len(matchedInnerEventPtrs) equals to
	// lookUpJoinTask.outerResult.NumChunks(), and the elements of every
	// matchedInnerEventPtrs[chkIdx][rowIdx] indicates the matched inner event ptrs
	// of the corresponding outer event.
	matchedInnerEventPtrs [][][]chunk.EventPtr
}

// Open implements the IndexNestedLoopHashJoin InterlockingDirectorate interface.
func (e *IndexNestedLoopHashJoin) Open(ctx context.Context) error {
	// Be careful, very dirty replog in this line!!!
	// IndexLookUpJoin need to rebuild interlock (the dataReaderBuilder) during
	// executing. However `interlock.Next()` is lazy evaluation when the RecordSet
	// result is drained.
	// Lazy evaluation means the saved stochastik context may change during interlock's
	// building and its running.
	// A specific sequence for example:
	//
	// e := buildInterlockingDirectorate()   // txn at build time
	// recordSet := runStmt(e)
	// stochastik.CommitTxn()    // txn closed
	// recordSet.Next()
	// e.dataReaderBuilder.Build() // txn is used again, which is already closed
	//
	// The trick here is `getSnapshotTS` will cache snapshot ts in the dataReaderBuilder,
	// so even txn is destroyed later, the dataReaderBuilder could still use the
	// cached snapshot ts to construct PosetDag.
	_, err := e.innerCtx.readerBuilder.getSnapshotTS()
	if err != nil {
		return err
	}

	err = e.children[0].Open(ctx)
	if err != nil {
		return err
	}
	e.memTracker = memory.NewTracker(e.id, -1)
	e.memTracker.AttachTo(e.ctx.GetStochastikVars().StmtCtx.MemTracker)
	e.innerPtrBytes = make([][]byte, 0, 8)
	if e.runtimeStats != nil {
		e.stats = &indexLookUpJoinRuntimeStats{}
		e.ctx.GetStochastikVars().StmtCtx.RuntimeStatsDefCausl.RegisterStats(e.id, e.stats)
	}
	e.startWorkers(ctx)
	return nil
}

func (e *IndexNestedLoopHashJoin) startWorkers(ctx context.Context) {
	concurrency := e.ctx.GetStochastikVars().IndexLookupJoinConcurrency()
	if e.stats != nil {
		e.stats.concurrency = concurrency
	}
	workerCtx, cancelFunc := context.WithCancel(ctx)
	e.cancelFunc = cancelFunc
	innerCh := make(chan *indexHashJoinTask, concurrency)
	if e.keepOuterOrder {
		e.taskCh = make(chan *indexHashJoinTask, concurrency)
	}
	e.workerWg.Add(1)
	ow := e.newOuterWorker(innerCh)
	go soliton.WithRecovery(func() { ow.run(workerCtx) }, e.finishJoinWorkers)

	if !e.keepOuterOrder {
		e.resultCh = make(chan *indexHashJoinResult, concurrency)
	} else {
		// When `keepOuterOrder` is true, each task holds their own `resultCh`
		// individually, thus we do not need a global resultCh.
		e.resultCh = nil
	}
	e.joinChkResourceCh = make([]chan *chunk.Chunk, concurrency)
	for i := 0; i < concurrency; i++ {
		if !e.keepOuterOrder {
			e.joinChkResourceCh[i] = make(chan *chunk.Chunk, 1)
			e.joinChkResourceCh[i] <- newFirstChunk(e)
		} else {
			e.joinChkResourceCh[i] = make(chan *chunk.Chunk, numResChkHold)
			for j := 0; j < numResChkHold; j++ {
				e.joinChkResourceCh[i] <- newFirstChunk(e)
			}
		}
	}

	e.workerWg.Add(concurrency)
	for i := 0; i < concurrency; i++ {
		workerID := i
		go soliton.WithRecovery(func() { e.newInnerWorker(innerCh, workerID).run(workerCtx, cancelFunc) }, e.finishJoinWorkers)
	}
	go e.wait4JoinWorkers()
}

func (e *IndexNestedLoopHashJoin) finishJoinWorkers(r interface{}) {
	if r != nil {
		e.resultCh <- &indexHashJoinResult{
			err: errors.New(fmt.Sprintf("%v", r)),
		}
		if e.cancelFunc != nil {
			e.cancelFunc()
		}
	}
	e.workerWg.Done()
}

func (e *IndexNestedLoopHashJoin) wait4JoinWorkers() {
	e.workerWg.Wait()
	if e.resultCh != nil {
		close(e.resultCh)
	}
	if e.taskCh != nil {
		close(e.taskCh)
	}
}

// Next implements the IndexNestedLoopHashJoin InterlockingDirectorate interface.
func (e *IndexNestedLoopHashJoin) Next(ctx context.Context, req *chunk.Chunk) error {
	req.Reset()
	if e.keepOuterOrder {
		return e.runInOrder(ctx, req)
	}
	// unordered run
	var (
		result *indexHashJoinResult
		ok     bool
	)
	select {
	case result, ok = <-e.resultCh:
		if !ok {
			return nil
		}
		if result.err != nil {
			return result.err
		}
	case <-ctx.Done():
		return ctx.Err()
	}
	req.SwapDeferredCausets(result.chk)
	result.src <- result.chk
	return nil
}

func (e *IndexNestedLoopHashJoin) runInOrder(ctx context.Context, req *chunk.Chunk) error {
	var (
		result *indexHashJoinResult
		ok     bool
	)
	for {
		if e.isDryUpTasks(ctx) {
			return nil
		}
		select {
		case result, ok = <-e.curTask.resultCh:
			if !ok {
				e.curTask = nil
				continue
			}
			if result.err != nil {
				return result.err
			}
		case <-ctx.Done():
			return ctx.Err()
		}
		req.SwapDeferredCausets(result.chk)
		result.src <- result.chk
		return nil
	}
}

// isDryUpTasks indicates whether all the tasks have been processed.
func (e *IndexNestedLoopHashJoin) isDryUpTasks(ctx context.Context) bool {
	if e.curTask != nil {
		return false
	}
	var ok bool
	select {
	case e.curTask, ok = <-e.taskCh:
		if !ok {
			return true
		}
	case <-ctx.Done():
		return true
	}
	return false
}

// Close implements the IndexNestedLoopHashJoin InterlockingDirectorate interface.
func (e *IndexNestedLoopHashJoin) Close() error {
	if e.cancelFunc != nil {
		e.cancelFunc()
		e.cancelFunc = nil
	}
	if e.resultCh != nil {
		for range e.resultCh {
		}
		e.resultCh = nil
	}
	if e.taskCh != nil {
		for range e.taskCh {
		}
		e.taskCh = nil
	}
	for i := range e.joinChkResourceCh {
		close(e.joinChkResourceCh[i])
	}
	e.joinChkResourceCh = nil
	return e.baseInterlockingDirectorate.Close()
}

func (ow *indexHashJoinOuterWorker) run(ctx context.Context) {
	defer trace.StartRegion(ctx, "IndexHashJoinOuterWorker").End()
	defer close(ow.innerCh)
	for {
		task, err := ow.buildTask(ctx)
		failpoint.Inject("testIndexHashJoinOuterWorkerErr", func() {
			err = errors.New("mocHoTTexHashJoinOuterWorkerErr")
		})
		if err != nil {
			task = &indexHashJoinTask{err: err}
			if ow.keepOuterOrder {
				task.keepOuterOrder, task.resultCh = true, make(chan *indexHashJoinResult, 1)
				ow.pushToChan(ctx, task, ow.taskCh)
			}
			ow.pushToChan(ctx, task, ow.innerCh)
			return
		}
		if task == nil {
			return
		}
		if finished := ow.pushToChan(ctx, task, ow.innerCh); finished {
			return
		}
		if ow.keepOuterOrder {
			if finished := ow.pushToChan(ctx, task, ow.taskCh); finished {
				return
			}
		}
	}
}

func (ow *indexHashJoinOuterWorker) buildTask(ctx context.Context) (*indexHashJoinTask, error) {
	task, err := ow.outerWorker.buildTask(ctx)
	if task == nil || err != nil {
		return nil, err
	}
	var (
		resultCh              chan *indexHashJoinResult
		matchedInnerEventPtrs [][][]chunk.EventPtr
	)
	if ow.keepOuterOrder {
		resultCh = make(chan *indexHashJoinResult, numResChkHold)
		matchedInnerEventPtrs = make([][][]chunk.EventPtr, task.outerResult.NumChunks())
		for i := range matchedInnerEventPtrs {
			matchedInnerEventPtrs[i] = make([][]chunk.EventPtr, task.outerResult.GetChunk(i).NumEvents())
		}
	}
	numChks := task.outerResult.NumChunks()
	outerEventStatus := make([][]outerEventStatusFlag, numChks)
	for i := 0; i < numChks; i++ {
		outerEventStatus[i] = make([]outerEventStatusFlag, task.outerResult.GetChunk(i).NumEvents())
	}
	return &indexHashJoinTask{
		lookUpJoinTask:        task,
		outerEventStatus:      outerEventStatus,
		keepOuterOrder:        ow.keepOuterOrder,
		resultCh:              resultCh,
		matchedInnerEventPtrs: matchedInnerEventPtrs,
	}, nil
}

func (ow *indexHashJoinOuterWorker) pushToChan(ctx context.Context, task *indexHashJoinTask, dst chan<- *indexHashJoinTask) bool {
	select {
	case <-ctx.Done():
		return true
	case dst <- task:
	}
	return false
}

func (e *IndexNestedLoopHashJoin) newOuterWorker(innerCh chan *indexHashJoinTask) *indexHashJoinOuterWorker {
	ow := &indexHashJoinOuterWorker{
		outerWorker: outerWorker{
			outerCtx:         e.outerCtx,
			ctx:              e.ctx,
			interlock:        e.children[0],
			batchSize:        32,
			maxBatchSize:     e.ctx.GetStochastikVars().IndexJoinBatchSize,
			parentMemTracker: e.memTracker,
			lookup:           &e.IndexLookUpJoin,
		},
		innerCh:        innerCh,
		keepOuterOrder: e.keepOuterOrder,
		taskCh:         e.taskCh,
	}
	return ow
}

func (e *IndexNestedLoopHashJoin) newInnerWorker(taskCh chan *indexHashJoinTask, workerID int) *indexHashJoinInnerWorker {
	// Since multiple inner workers run concurrently, we should copy join's indexRanges for every worker to avoid data race.
	copiedRanges := make([]*ranger.Range, 0, len(e.indexRanges))
	for _, ran := range e.indexRanges {
		copiedRanges = append(copiedRanges, ran.Clone())
	}
	var innerStats *innerWorkerRuntimeStats
	if e.stats != nil {
		innerStats = &e.stats.innerWorker
	}
	iw := &indexHashJoinInnerWorker{
		innerWorker: innerWorker{
			innerCtx:      e.innerCtx,
			outerCtx:      e.outerCtx,
			ctx:           e.ctx,
			interlockChk:  chunk.NewChunkWithCapacity(e.innerCtx.rowTypes, e.maxChunkSize),
			indexRanges:   copiedRanges,
			keyOff2IdxOff: e.keyOff2IdxOff,
			stats:         innerStats,
		},
		taskCh:            taskCh,
		joiner:            e.joiners[workerID],
		joinChkResourceCh: e.joinChkResourceCh[workerID],
		resultCh:          e.resultCh,
		matchedOuterPtrs:  make([]chunk.EventPtr, 0, e.maxChunkSize),
		joinKeyBuf:        make([]byte, 1),
		outerEventStatus:  make([]outerEventStatusFlag, 0, e.maxChunkSize),
	}
	if e.lastDefCausHelper != nil {
		// nextCwf.TmpConstant needs to be reset for every individual
		// inner worker to avoid data race when the inner workers is running
		// concurrently.
		nextCwf := *e.lastDefCausHelper
		nextCwf.TmpConstant = make([]*memex.Constant, len(e.lastDefCausHelper.TmpConstant))
		for i := range e.lastDefCausHelper.TmpConstant {
			nextCwf.TmpConstant[i] = &memex.Constant{RetType: nextCwf.TargetDefCaus.RetType}
		}
		iw.nextDefCausCompareFilters = &nextCwf
	}
	return iw
}

func (iw *indexHashJoinInnerWorker) run(ctx context.Context, cancelFunc context.CancelFunc) {
	defer trace.StartRegion(ctx, "IndexHashJoinInnerWorker").End()
	var task *indexHashJoinTask
	joinResult, ok := iw.getNewJoinResult(ctx)
	if !ok {
		cancelFunc()
		return
	}
	h, resultCh := fnv.New64(), iw.resultCh
	for {
		select {
		case <-ctx.Done():
			return
		case task, ok = <-iw.taskCh:
		}
		if !ok {
			break
		}
		// We need to init resultCh before the err is returned.
		if task.keepOuterOrder {
			resultCh = task.resultCh
		}
		if task.err != nil {
			joinResult.err = task.err
			break
		}
		err := iw.handleTask(ctx, task, joinResult, h, resultCh)
		if err != nil {
			joinResult.err = err
			break
		}
		if task.keepOuterOrder {
			// We need to get a new result holder here because the old
			// `joinResult` hash been sent to the `resultCh` or to the
			// `joinChkResourceCh`.
			joinResult, ok = iw.getNewJoinResult(ctx)
			if !ok {
				cancelFunc()
				return
			}
		}
	}
	failpoint.Inject("testIndexHashJoinInnerWorkerErr", func() {
		joinResult.err = errors.New("mocHoTTexHashJoinInnerWorkerErr")
	})
	if joinResult.err != nil {
		resultCh <- joinResult
		return
	}
	// When task.keepOuterOrder is TRUE(resultCh != iw.resultCh), the last
	// joinResult will be checked when the a task has been processed, thus we do
	// not need to check it here again.
	if resultCh == iw.resultCh && joinResult.chk != nil && joinResult.chk.NumEvents() > 0 {
		select {
		case resultCh <- joinResult:
		case <-ctx.Done():
			return
		}
	}
}

func (iw *indexHashJoinInnerWorker) getNewJoinResult(ctx context.Context) (*indexHashJoinResult, bool) {
	joinResult := &indexHashJoinResult{
		src: iw.joinChkResourceCh,
	}
	ok := true
	select {
	case joinResult.chk, ok = <-iw.joinChkResourceCh:
	case <-ctx.Done():
		return nil, false
	}
	return joinResult, ok
}

func (iw *indexHashJoinInnerWorker) buildHashBlockForOuterResult(ctx context.Context, task *indexHashJoinTask, h hash.Hash64) {
	if iw.stats != nil {
		start := time.Now()
		defer func() {
			atomic.AddInt64(&iw.stats.build, int64(time.Since(start)))
		}()
	}
	buf, numChks := make([]byte, 1), task.outerResult.NumChunks()
	task.lookupMap = newUnsafeHashBlock(task.outerResult.Len())
	for chkIdx := 0; chkIdx < numChks; chkIdx++ {
		chk := task.outerResult.GetChunk(chkIdx)
		numEvents := chk.NumEvents()
	OUTER:
		for rowIdx := 0; rowIdx < numEvents; rowIdx++ {
			if task.outerMatch != nil && !task.outerMatch[chkIdx][rowIdx] {
				continue
			}
			event := chk.GetEvent(rowIdx)
			keyDefCausIdx := iw.outerCtx.keyDefCauss
			for _, i := range keyDefCausIdx {
				if event.IsNull(i) {
					continue OUTER
				}
			}
			h.Reset()
			err := codec.HashChunkEvent(iw.ctx.GetStochastikVars().StmtCtx, h, event, iw.outerCtx.rowTypes, keyDefCausIdx, buf)
			failpoint.Inject("testIndexHashJoinBuildErr", func() {
				err = errors.New("mocHoTTexHashJoinBuildErr")
			})
			if err != nil {
				// This panic will be recovered by the invoker.
				panic(err.Error())
			}
			rowPtr := chunk.EventPtr{ChkIdx: uint32(chkIdx), EventIdx: uint32(rowIdx)}
			task.lookupMap.Put(h.Sum64(), rowPtr)
		}
	}
}

func (iw *indexHashJoinInnerWorker) fetchInnerResults(ctx context.Context, task *lookUpJoinTask) error {
	lookUpContents, err := iw.constructLookupContent(task)
	if err != nil {
		return err
	}
	return iw.innerWorker.fetchInnerResults(ctx, task, lookUpContents)
}

func (iw *indexHashJoinInnerWorker) handleHashJoinInnerWorkerPanic(r interface{}) {
	if r != nil {
		iw.resultCh <- &indexHashJoinResult{err: errors.Errorf("%v", r)}
	}
	iw.wg.Done()
}

func (iw *indexHashJoinInnerWorker) handleTask(ctx context.Context, task *indexHashJoinTask, joinResult *indexHashJoinResult, h hash.Hash64, resultCh chan *indexHashJoinResult) error {
	var joinStartTime time.Time
	if iw.stats != nil {
		start := time.Now()
		defer func() {
			endTime := time.Now()
			atomic.AddInt64(&iw.stats.totalTime, int64(endTime.Sub(start)))
			atomic.AddInt64(&iw.stats.join, int64(endTime.Sub(joinStartTime)))
		}()
	}

	iw.wg = &sync.WaitGroup{}
	iw.wg.Add(1)
	// TODO(XuHuaiyu): we may always use the smaller side to build the hashblock.
	go soliton.WithRecovery(func() { iw.buildHashBlockForOuterResult(ctx, task, h) }, iw.handleHashJoinInnerWorkerPanic)
	err := iw.fetchInnerResults(ctx, task.lookUpJoinTask)
	if err != nil {
		return err
	}
	iw.wg.Wait()

	joinStartTime = time.Now()
	if !task.keepOuterOrder {
		return iw.doJoinUnordered(ctx, task, joinResult, h, resultCh)
	}
	return iw.doJoinInOrder(ctx, task, joinResult, h, resultCh)
}

func (iw *indexHashJoinInnerWorker) doJoinUnordered(ctx context.Context, task *indexHashJoinTask, joinResult *indexHashJoinResult, h hash.Hash64, resultCh chan *indexHashJoinResult) error {
	var ok bool
	iter := chunk.NewIterator4List(task.innerResult)
	for event := iter.Begin(); event != iter.End(); event = iter.Next() {
		ok, joinResult = iw.joinMatchedInnerEvent2Chunk(ctx, event, task, joinResult, h, iw.joinKeyBuf)
		if !ok {
			return errors.New("indexHashJoinInnerWorker.doJoinUnordered failed")
		}
	}
	for chkIdx, outerEventStatus := range task.outerEventStatus {
		chk := task.outerResult.GetChunk(chkIdx)
		for rowIdx, val := range outerEventStatus {
			if val == outerEventMatched {
				continue
			}
			iw.joiner.onMissMatch(val == outerEventHasNull, chk.GetEvent(rowIdx), joinResult.chk)
			if joinResult.chk.IsFull() {
				select {
				case resultCh <- joinResult:
				case <-ctx.Done():
					return ctx.Err()
				}
				joinResult, ok = iw.getNewJoinResult(ctx)
				if !ok {
					return errors.New("indexHashJoinInnerWorker.doJoinUnordered failed")
				}
			}
		}
	}
	return nil
}

func (iw *indexHashJoinInnerWorker) getMatchedOuterEvents(innerEvent chunk.Event, task *indexHashJoinTask, h hash.Hash64, buf []byte) (matchedEvents []chunk.Event, matchedEventPtr []chunk.EventPtr, err error) {
	h.Reset()
	err = codec.HashChunkEvent(iw.ctx.GetStochastikVars().StmtCtx, h, innerEvent, iw.rowTypes, iw.keyDefCauss, buf)
	if err != nil {
		return nil, nil, err
	}
	iw.matchedOuterPtrs = task.lookupMap.Get(h.Sum64())
	if len(iw.matchedOuterPtrs) == 0 {
		return nil, nil, nil
	}
	joinType := JoinerType(iw.joiner)
	isSemiJoin := joinType == causetembedded.SemiJoin || joinType == causetembedded.LeftOuterSemiJoin
	matchedEvents = make([]chunk.Event, 0, len(iw.matchedOuterPtrs))
	matchedEventPtr = make([]chunk.EventPtr, 0, len(iw.matchedOuterPtrs))
	for _, ptr := range iw.matchedOuterPtrs {
		outerEvent := task.outerResult.GetEvent(ptr)
		ok, err := codec.EqualChunkEvent(iw.ctx.GetStochastikVars().StmtCtx, innerEvent, iw.rowTypes, iw.keyDefCauss, outerEvent, iw.outerCtx.rowTypes, iw.outerCtx.keyDefCauss)
		if err != nil {
			return nil, nil, err
		}
		if !ok || (task.outerEventStatus[ptr.ChkIdx][ptr.EventIdx] == outerEventMatched && isSemiJoin) {
			continue
		}
		matchedEvents = append(matchedEvents, outerEvent)
		matchedEventPtr = append(matchedEventPtr, chunk.EventPtr{ChkIdx: ptr.ChkIdx, EventIdx: ptr.EventIdx})
	}
	return matchedEvents, matchedEventPtr, nil
}

func (iw *indexHashJoinInnerWorker) joinMatchedInnerEvent2Chunk(ctx context.Context, innerEvent chunk.Event, task *indexHashJoinTask,
	joinResult *indexHashJoinResult, h hash.Hash64, buf []byte) (bool, *indexHashJoinResult) {
	matchedOuterEvents, matchedOuterEventPtr, err := iw.getMatchedOuterEvents(innerEvent, task, h, buf)
	if err != nil {
		joinResult.err = err
		return false, joinResult
	}
	if len(matchedOuterEvents) == 0 {
		return true, joinResult
	}
	var (
		ok     bool
		iter   = chunk.NewIterator4Slice(matchedOuterEvents)
		cursor = 0
	)
	for iter.Begin(); iter.Current() != iter.End(); {
		iw.outerEventStatus, err = iw.joiner.tryToMatchOuters(iter, innerEvent, joinResult.chk, iw.outerEventStatus)
		if err != nil {
			joinResult.err = err
			return false, joinResult
		}
		for _, status := range iw.outerEventStatus {
			chkIdx, rowIdx := matchedOuterEventPtr[cursor].ChkIdx, matchedOuterEventPtr[cursor].EventIdx
			if status == outerEventMatched || task.outerEventStatus[chkIdx][rowIdx] == outerEventUnmatched {
				task.outerEventStatus[chkIdx][rowIdx] = status
			}
			cursor++
		}
		if joinResult.chk.IsFull() {
			select {
			case iw.resultCh <- joinResult:
			case <-ctx.Done():
			}
			joinResult, ok = iw.getNewJoinResult(ctx)
			if !ok {
				return false, joinResult
			}
		}
	}
	return true, joinResult
}

func (iw *indexHashJoinInnerWorker) defCauslectMatchedInnerPtrs4OuterEvents(ctx context.Context, innerEvent chunk.Event, innerEventPtr chunk.EventPtr,
	task *indexHashJoinTask, h hash.Hash64, buf []byte) error {
	_, matchedOuterEventIdx, err := iw.getMatchedOuterEvents(innerEvent, task, h, buf)
	if err != nil {
		return err
	}
	for _, outerEventPtr := range matchedOuterEventIdx {
		chkIdx, rowIdx := outerEventPtr.ChkIdx, outerEventPtr.EventIdx
		task.matchedInnerEventPtrs[chkIdx][rowIdx] = append(task.matchedInnerEventPtrs[chkIdx][rowIdx], innerEventPtr)
	}
	return nil
}

// doJoinInOrder follows the following steps:
// 1. defCauslect all the matched inner event ptrs for every outer event
// 2. do the join work
//   2.1 defCauslect all the matched inner rows using the defCauslected ptrs for every outer event
//   2.2 call tryToMatchInners for every outer event
//   2.3 call onMissMatch when no inner rows are matched
func (iw *indexHashJoinInnerWorker) doJoinInOrder(ctx context.Context, task *indexHashJoinTask, joinResult *indexHashJoinResult, h hash.Hash64, resultCh chan *indexHashJoinResult) (err error) {
	defer func() {
		if err == nil && joinResult.chk != nil {
			if joinResult.chk.NumEvents() > 0 {
				select {
				case resultCh <- joinResult:
				case <-ctx.Done():
					return
				}
			} else {
				joinResult.src <- joinResult.chk
			}
		}
		close(resultCh)
	}()
	for i, numChunks := 0, task.innerResult.NumChunks(); i < numChunks; i++ {
		for j, chk := 0, task.innerResult.GetChunk(i); j < chk.NumEvents(); j++ {
			event := chk.GetEvent(j)
			ptr := chunk.EventPtr{ChkIdx: uint32(i), EventIdx: uint32(j)}
			err = iw.defCauslectMatchedInnerPtrs4OuterEvents(ctx, event, ptr, task, h, iw.joinKeyBuf)
			if err != nil {
				return err
			}
		}
	}
	// TODO: matchedInnerEventPtrs and matchedInnerEvents can be moved to inner worker.
	matchedInnerEvents := make([]chunk.Event, len(task.matchedInnerEventPtrs))
	var hasMatched, hasNull, ok bool
	for chkIdx, innerEventPtrs4Chk := range task.matchedInnerEventPtrs {
		for outerEventIdx, innerEventPtrs := range innerEventPtrs4Chk {
			matchedInnerEvents, hasMatched, hasNull = matchedInnerEvents[:0], false, false
			outerEvent := task.outerResult.GetChunk(chkIdx).GetEvent(outerEventIdx)
			for _, ptr := range innerEventPtrs {
				matchedInnerEvents = append(matchedInnerEvents, task.innerResult.GetEvent(ptr))
			}
			iter := chunk.NewIterator4Slice(matchedInnerEvents)
			for iter.Begin(); iter.Current() != iter.End(); {
				matched, isNull, err := iw.joiner.tryToMatchInners(outerEvent, iter, joinResult.chk)
				if err != nil {
					return err
				}
				hasMatched, hasNull = matched || hasMatched, isNull || hasNull
				if joinResult.chk.IsFull() {
					select {
					case resultCh <- joinResult:
					case <-ctx.Done():
						return ctx.Err()
					}
					joinResult, ok = iw.getNewJoinResult(ctx)
					if !ok {
						return errors.New("indexHashJoinInnerWorker.doJoinInOrder failed")
					}
				}
			}
			if !hasMatched {
				iw.joiner.onMissMatch(hasNull, outerEvent, joinResult.chk)
			}
		}
	}
	return nil
}