ddl: enhance row size estimation before adding index (pingcap#52086)

close pingcap#52085

pkg/ddl/BUILD.bazel

@@ -117,7 +117,6 @@ go_library(
         "//pkg/sessiontxn",
         "//pkg/statistics",
         "//pkg/statistics/handle",
-        "//pkg/statistics/handle/cache",
         "//pkg/statistics/handle/util",
         "//pkg/store/copr",
         "//pkg/store/driver/backoff",

pkg/ddl/backfilling_dist_executor.go

@@ -44,6 +44,7 @@ type BackfillTaskMeta struct {
 	EleTypeKey []byte `json:"ele_type_key"`
 
 	CloudStorageURI string `json:"cloud_storage_uri"`
+	EstimateRowSize int    `json:"estimate_row_size"`
 }
 
 // BackfillSubTaskMeta is the sub-task meta for backfilling index.
@@ -107,15 +108,16 @@ func (s *backfillDistExecutor) newBackfillSubtaskExecutor(
 		indexInfos = append(indexInfos, indexInfo)
 	}
 	cloudStorageURI := s.taskMeta.CloudStorageURI
+	estRowSize := s.taskMeta.EstimateRowSize
 
 	switch stage {
 	case proto.BackfillStepReadIndex:
 		jc := ddlObj.jobContext(jobMeta.ID, jobMeta.ReorgMeta)
 		ddlObj.setDDLLabelForTopSQL(jobMeta.ID, jobMeta.Query)
 		ddlObj.setDDLSourceForDiagnosis(jobMeta.ID, jobMeta.Type)
-		return newReadIndexExecutor(ddlObj, jobMeta, indexInfos, tbl, jc, s.getBackendCtx, cloudStorageURI)
+		return newReadIndexExecutor(ddlObj, jobMeta, indexInfos, tbl, jc, s.getBackendCtx, cloudStorageURI, estRowSize)
 	case proto.BackfillStepMergeSort:
-		return newMergeSortExecutor(jobMeta.ID, len(indexInfos), tbl, cloudStorageURI)
+		return newMergeSortExecutor(jobMeta.ID, len(indexInfos), tbl, cloudStorageURI, estRowSize)
 	case proto.BackfillStepWriteAndIngest:
 		if len(cloudStorageURI) == 0 {
 			return nil, errors.Errorf("local import does not have write & ingest step")

pkg/ddl/backfilling_merge_sort.go

@@ -33,10 +33,12 @@ import (
 
 type mergeSortExecutor struct {
 	taskexecutor.EmptyStepExecutor
-	jobID               int64
-	idxNum              int
-	ptbl                table.PhysicalTable
-	cloudStoreURI       string
+	jobID         int64
+	idxNum        int
+	ptbl          table.PhysicalTable
+	avgRowSize    int
+	cloudStoreURI string
+
 	mu                  sync.Mutex
 	subtaskSortedKVMeta *external.SortedKVMeta
 }
@@ -46,12 +48,14 @@ func newMergeSortExecutor(
 	idxNum int,
 	ptbl table.PhysicalTable,
 	cloudStoreURI string,
+	avgRowSize int,
 ) (*mergeSortExecutor, error) {
 	return &mergeSortExecutor{
 		jobID:         jobID,
 		idxNum:        idxNum,
 		ptbl:          ptbl,
 		cloudStoreURI: cloudStoreURI,
+		avgRowSize:    avgRowSize,
 	}, nil
 }
 
@@ -85,7 +89,7 @@ func (m *mergeSortExecutor) RunSubtask(ctx context.Context, subtask *proto.Subta
 	}
 
 	prefix := path.Join(strconv.Itoa(int(m.jobID)), strconv.Itoa(int(subtask.ID)))
-	partSize, err := getMergeSortPartSize(m.ptbl.Meta(), int(variable.GetDDLReorgWorkerCounter()), m.idxNum)
+	partSize, err := getMergeSortPartSize(m.avgRowSize, int(variable.GetDDLReorgWorkerCounter()), m.idxNum)
 	if err != nil {
 		return err
 	}

pkg/ddl/backfilling_operators.go

@@ -112,11 +112,11 @@ func (ctx *OperatorCtx) OperatorErr() error {
 	return *err
 }
 
-func getWriterMemSize(tblInfo *model.TableInfo, idxNum int) (uint64, error) {
+func getWriterMemSize(avgRowSize int, idxNum int) (uint64, error) {
 	failpoint.Inject("mockWriterMemSize", func() {
 		failpoint.Return(1*size.GB, nil)
 	})
-	_, writerCnt := expectedIngestWorkerCnt(tblInfo)
+	_, writerCnt := expectedIngestWorkerCnt(avgRowSize)
 	memTotal, err := memory.MemTotal()
 	if err != nil {
 		return 0, err
@@ -131,8 +131,8 @@ func getWriterMemSize(tblInfo *model.TableInfo, idxNum int) (uint64, error) {
 	return memSize, nil
 }
 
-func getMergeSortPartSize(tblInfo *model.TableInfo, concurrency int, idxNum int) (uint64, error) {
-	writerMemSize, err := getWriterMemSize(tblInfo, idxNum)
+func getMergeSortPartSize(avgRowSize int, concurrency int, idxNum int) (uint64, error) {
+	writerMemSize, err := getWriterMemSize(avgRowSize, idxNum)
 	if err != nil {
 		return 0, nil
 	}
@@ -155,6 +155,7 @@ func NewAddIndexIngestPipeline(
 	totalRowCount *atomic.Int64,
 	metricCounter prometheus.Counter,
 	reorgMeta *model.DDLReorgMeta,
+	avgRowSize int,
 ) (*operator.AsyncPipeline, error) {
 	indexes := make([]table.Index, 0, len(idxInfos))
 	for _, idxInfo := range idxInfos {
@@ -171,7 +172,7 @@ func NewAddIndexIngestPipeline(
 	for i := 0; i < poolSize; i++ {
 		srcChkPool <- chunk.NewChunkWithCapacity(copCtx.GetBase().FieldTypes, copReadBatchSize())
 	}
-	readerCnt, writerCnt := expectedIngestWorkerCnt(tbl.Meta())
+	readerCnt, writerCnt := expectedIngestWorkerCnt(avgRowSize)
 
 	srcOp := NewTableScanTaskSource(ctx, store, tbl, startKey, endKey)
 	scanOp := NewTableScanOperator(ctx, sessPool, copCtx, srcChkPool, readerCnt)
@@ -207,6 +208,7 @@ func NewWriteIndexToExternalStoragePipeline(
 	metricCounter prometheus.Counter,
 	onClose external.OnCloseFunc,
 	reorgMeta *model.DDLReorgMeta,
+	avgRowSize int,
 ) (*operator.AsyncPipeline, error) {
 	indexes := make([]table.Index, 0, len(idxInfos))
 	for _, idxInfo := range idxInfos {
@@ -223,7 +225,7 @@ func NewWriteIndexToExternalStoragePipeline(
 	for i := 0; i < poolSize; i++ {
 		srcChkPool <- chunk.NewChunkWithCapacity(copCtx.GetBase().FieldTypes, copReadBatchSize())
 	}
-	readerCnt, writerCnt := expectedIngestWorkerCnt(tbl.Meta())
+	readerCnt, writerCnt := expectedIngestWorkerCnt(avgRowSize)
 
 	backend, err := storage.ParseBackend(extStoreURI, nil)
 	if err != nil {
@@ -234,7 +236,7 @@ func NewWriteIndexToExternalStoragePipeline(
 		return nil, err
 	}
 
-	memSize, err := getWriterMemSize(tbl.Meta(), len(indexes))
+	memSize, err := getWriterMemSize(avgRowSize, len(indexes))
 	if err != nil {
 		return nil, err
 	}

pkg/ddl/backfilling_read_index.go

@@ -44,6 +44,7 @@ type readIndexExecutor struct {
 	ptbl    table.PhysicalTable
 	jc      *JobContext
 
+	avgRowSize      int
 	cloudStorageURI string
 
 	bc          ingest.BackendCtx
@@ -65,6 +66,7 @@ func newReadIndexExecutor(
 	jc *JobContext,
 	bcGetter func() (ingest.BackendCtx, error),
 	cloudStorageURI string,
+	avgRowSize int,
 ) (*readIndexExecutor, error) {
 	bc, err := bcGetter()
 	if err != nil {
@@ -78,6 +80,7 @@ func newReadIndexExecutor(
 		jc:              jc,
 		bc:              bc,
 		cloudStorageURI: cloudStorageURI,
+		avgRowSize:      avgRowSize,
 		curRowCount:     &atomic.Int64{},
 	}, nil
 }
@@ -241,7 +244,7 @@ func (r *readIndexExecutor) buildLocalStorePipeline(
 		metrics.GenerateReorgLabel("add_idx_rate", r.job.SchemaName, tbl.Meta().Name.O))
 	return NewAddIndexIngestPipeline(
 		opCtx, d.store, d.sessPool, r.bc, engines, sessCtx, r.job.ID,
-		tbl, r.indexes, start, end, totalRowCount, counter, r.job.ReorgMeta)
+		tbl, r.indexes, start, end, totalRowCount, counter, r.job.ReorgMeta, r.avgRowSize)
 }
 
 func (r *readIndexExecutor) buildExternalStorePipeline(
@@ -282,5 +285,7 @@ func (r *readIndexExecutor) buildExternalStorePipeline(
 		totalRowCount,
 		counter,
 		onClose,
-		r.job.ReorgMeta)
+		r.job.ReorgMeta,
+		r.avgRowSize,
+	)
 }

pkg/ddl/backfilling_scheduler.go

@@ -34,7 +34,6 @@ import (
 	poolutil "github.com/pingcap/tidb/pkg/resourcemanager/util"
 	"github.com/pingcap/tidb/pkg/sessionctx"
 	"github.com/pingcap/tidb/pkg/sessionctx/variable"
-	"github.com/pingcap/tidb/pkg/statistics/handle/cache"
 	"github.com/pingcap/tidb/pkg/table"
 	"github.com/pingcap/tidb/pkg/types"
 	"github.com/pingcap/tidb/pkg/util"
@@ -88,7 +87,7 @@ func newBackfillScheduler(ctx context.Context, info *reorgInfo, sessPool *sess.P
 	jobCtx *JobContext) (backfillScheduler, error) {
 	if tp == typeAddIndexWorker && info.ReorgMeta.ReorgTp == model.ReorgTypeLitMerge {
 		ctx = logutil.WithCategory(ctx, "ddl-ingest")
-		return newIngestBackfillScheduler(ctx, info, sessPool, tbl, false), nil
+		return newIngestBackfillScheduler(ctx, info, sessPool, tbl, false)
 	}
 	return newTxnBackfillScheduler(ctx, info, sessPool, tp, tbl, sessCtx, jobCtx)
 }
@@ -314,6 +313,7 @@ type ingestBackfillScheduler struct {
 	sessPool   *sess.Pool
 	tbl        table.PhysicalTable
 	distribute bool
+	avgRowSize int
 
 	closed bool
 
@@ -329,18 +329,30 @@ type ingestBackfillScheduler struct {
 	checkpointMgr *ingest.CheckpointManager
 }
 
-func newIngestBackfillScheduler(ctx context.Context, info *reorgInfo,
-	sessPool *sess.Pool, tbl table.PhysicalTable, distribute bool) *ingestBackfillScheduler {
+func newIngestBackfillScheduler(
+	ctx context.Context,
+	info *reorgInfo,
+	sessPool *sess.Pool,
+	tbl table.PhysicalTable,
+	distribute bool,
+) (*ingestBackfillScheduler, error) {
+	sctx, err := sessPool.Get()
+	if err != nil {
+		return nil, err
+	}
+	defer sessPool.Put(sctx)
+	avgRowSize := estimateTableRowSize(ctx, info.d.store, sctx.GetRestrictedSQLExecutor(), tbl)
 	return &ingestBackfillScheduler{
 		ctx:        ctx,
 		reorgInfo:  info,
 		sessPool:   sessPool,
 		tbl:        tbl,
+		avgRowSize: avgRowSize,
 		taskCh:     make(chan *reorgBackfillTask, backfillTaskChanSize),
 		resultCh:   make(chan *backfillResult, backfillTaskChanSize),
 		poolErr:    make(chan error),
 		distribute: distribute,
-	}
+	}, nil
 }
 
 func (b *ingestBackfillScheduler) setupWorkers() error {
@@ -506,13 +518,12 @@ func (b *ingestBackfillScheduler) createCopReqSenderPool() (*copReqSenderPool, e
 }
 
 func (b *ingestBackfillScheduler) expectedWorkerSize() (readerSize int, writerSize int) {
-	return expectedIngestWorkerCnt(b.tbl.Meta())
+	return expectedIngestWorkerCnt(b.avgRowSize)
 }
 
-func expectedIngestWorkerCnt(tblInfo *model.TableInfo) (readerCnt, writerCnt int) {
+func expectedIngestWorkerCnt(avgRowSize int) (readerCnt, writerCnt int) {
 	workerCnt := int(variable.GetDDLReorgWorkerCounter())
-	rowCount, avgRowLen, _, _ := cache.TableRowStatsCache.EstimateDataLength(tblInfo)
-	if rowCount == 0 {
+	if avgRowSize == 0 {
 		// Statistic data not exist, use default concurrency.
 		readerCnt = min(workerCnt/2, maxBackfillWorkerSize)
 		readerCnt = max(readerCnt, 1)
@@ -523,7 +534,7 @@ func expectedIngestWorkerCnt(tblInfo *model.TableInfo) (readerCnt, writerCnt int
 	readerRatio := []float64{0.5, 1, 2, 4, 8}
 	rowSize := []uint64{200, 500, 1000, 3000, math.MaxUint64}
 	for i, s := range rowSize {
-		if avgRowLen <= s {
+		if uint64(avgRowSize) <= s {
 			readerCnt = max(int(float64(workerCnt)*readerRatio[i]), 1)
 			writerCnt = max(workerCnt, 1)
 			break