planner, sessionctx: reintroduce #41996 through optimizer fix control (#44865)

planner/core/exhaust_physical_plans.go

@@ -32,11 +32,13 @@ import (
 	"github.com/pingcap/tidb/planner/property"
 	"github.com/pingcap/tidb/planner/util"
 	"github.com/pingcap/tidb/sessionctx"
+	"github.com/pingcap/tidb/sessionctx/variable"
 	"github.com/pingcap/tidb/statistics"
 	"github.com/pingcap/tidb/types"
 	"github.com/pingcap/tidb/util/chunk"
 	"github.com/pingcap/tidb/util/collate"
 	"github.com/pingcap/tidb/util/logutil"
+	"github.com/pingcap/tidb/util/mathutil"
 	"github.com/pingcap/tidb/util/plancodec"
 	"github.com/pingcap/tidb/util/ranger"
 	"github.com/pingcap/tidb/util/set"
@@ -965,7 +967,7 @@ func (p *LogicalJoin) buildIndexJoinInner2IndexScan(
 			maxOneRow = ok && (sf.FuncName.L == ast.EQ)
 		}
 	}
-	innerTask := p.constructInnerIndexScanTask(ds, helper.chosenPath, helper.chosenRanges.Range(), helper.chosenRemained, outerJoinKeys, us, rangeInfo, false, false, avgInnerRowCnt, maxOneRow)
+	innerTask := p.constructInnerIndexScanTask(ds, helper.chosenPath, helper.chosenRanges.Range(), helper.chosenRemained, outerJoinKeys, us, helper.idxOff2KeyOff, rangeInfo, false, false, avgInnerRowCnt, maxOneRow)
 	failpoint.Inject("MockOnlyEnableIndexHashJoin", func(val failpoint.Value) {
 		if val.(bool) && !p.ctx.GetSessionVars().InRestrictedSQL {
 			failpoint.Return(p.constructIndexHashJoin(prop, outerIdx, innerTask, helper.chosenRanges, keyOff2IdxOff, helper.chosenPath, helper.lastColManager))
@@ -980,7 +982,7 @@ func (p *LogicalJoin) buildIndexJoinInner2IndexScan(
 	// Because we can't keep order for union scan, if there is a union scan in inner task,
 	// we can't construct index merge join.
 	if us == nil {
-		innerTask2 := p.constructInnerIndexScanTask(ds, helper.chosenPath, helper.chosenRanges.Range(), helper.chosenRemained, outerJoinKeys, us, rangeInfo, true, !prop.IsSortItemEmpty() && prop.SortItems[0].Desc, avgInnerRowCnt, maxOneRow)
+		innerTask2 := p.constructInnerIndexScanTask(ds, helper.chosenPath, helper.chosenRanges.Range(), helper.chosenRemained, outerJoinKeys, us, helper.idxOff2KeyOff, rangeInfo, true, !prop.IsSortItemEmpty() && prop.SortItems[0].Desc, avgInnerRowCnt, maxOneRow)
 		if innerTask2 != nil {
 			joins = append(joins, p.constructIndexMergeJoin(prop, outerIdx, innerTask2, helper.chosenRanges, keyOff2IdxOff, helper.chosenPath, helper.lastColManager)...)
 		}
@@ -1164,6 +1166,55 @@ func (p *LogicalJoin) constructInnerUnionScan(us *LogicalUnionScan, reader Physi
 	return physicalUnionScan
 }
 
+// getColsNDVLowerBoundFromHistColl tries to get a lower bound of the NDV of columns (whose uniqueIDs are colUIDs).
+func getColsNDVLowerBoundFromHistColl(colUIDs []int64, histColl *statistics.HistColl) int64 {
+	if len(colUIDs) == 0 || histColl == nil {
+		return -1
+	}
+
+	// 1. Try to get NDV from column stats if it's a single column.
+	if len(colUIDs) == 1 && histColl.Columns != nil {
+		uid := colUIDs[0]
+		if colStats, ok := histColl.Columns[uid]; ok && colStats != nil && colStats.IsStatsInitialized() {
+			return colStats.NDV
+		}
+	}
+
+	slices.Sort(colUIDs)
+
+	// 2. Try to get NDV from index stats.
+	// Note that we don't need to specially handle prefix index here, because the NDV of a prefix index is
+	// equal or less than the corresponding normal index, and that's safe here since we want a lower bound.
+	for idxID, idxCols := range histColl.Idx2ColumnIDs {
+		if len(idxCols) != len(colUIDs) {
+			continue
+		}
+		orderedIdxCols := make([]int64, len(idxCols))
+		copy(orderedIdxCols, idxCols)
+		slices.Sort(orderedIdxCols)
+		if !slices.Equal(orderedIdxCols, colUIDs) {
+			continue
+		}
+		if idxStats, ok := histColl.Indices[idxID]; ok && idxStats != nil && idxStats.IsStatsInitialized() {
+			return idxStats.NDV
+		}
+	}
+
+	// TODO: if there's an index that contains the expected columns, we can also make use of its NDV.
+	// For example, NDV(a,b,c) / NDV(c) is a safe lower bound of NDV(a,b).
+
+	// 3. If we still haven't got an NDV, we use the maximum NDV in the column stats as a lower bound.
+	maxNDV := int64(-1)
+	for _, uid := range colUIDs {
+		colStats := histColl.Columns[uid]
+		if colStats == nil || !colStats.IsStatsInitialized() {
+			continue
+		}
+		maxNDV = mathutil.Max(maxNDV, colStats.NDV)
+	}
+	return maxNDV
+}
+
 // constructInnerIndexScanTask is specially used to construct the inner plan for PhysicalIndexJoin.
 func (p *LogicalJoin) constructInnerIndexScanTask(
 	ds *DataSource,
@@ -1172,6 +1223,7 @@ func (p *LogicalJoin) constructInnerIndexScanTask(
 	filterConds []expression.Expression,
 	_ []*expression.Column,
 	us *LogicalUnionScan,
+	idxOffset2joinKeyOffset []int,
 	rangeInfo string,
 	keepOrder bool,
 	desc bool,
@@ -1257,6 +1309,35 @@ func (p *LogicalJoin) constructInnerIndexScanTask(
 	}
 	is.initSchema(append(path.FullIdxCols, ds.commonHandleCols...), cop.tablePlan != nil)
 	indexConds, tblConds := ds.splitIndexFilterConditions(filterConds, path.FullIdxCols, path.FullIdxColLens)
+
+	// Note: due to a regression in JOB workload, we use the optimizer fix control to enable this for now.
+	//
+	// Because we are estimating an average row count of the inner side corresponding to each row from the outer side,
+	// the estimated row count of the IndexScan should be no larger than (total row count / NDV of join key columns).
+	// We can calculate the lower bound of the NDV therefore we can get an upper bound of the row count here.
+	rowCountUpperBound := -1.0
+	fixValue, ok := ds.ctx.GetSessionVars().GetOptimizerFixControlValue(variable.TiDBOptFixControl44855)
+	if ok && variable.TiDBOptOn(fixValue) && ds.tableStats != nil {
+		usedColIDs := make([]int64, 0)
+		// We only consider columns in this index that (1) are used to probe as join key,
+		// and (2) are not prefix column in the index (for which we can't easily get a lower bound)
+		for idxOffset, joinKeyOffset := range idxOffset2joinKeyOffset {
+			if joinKeyOffset < 0 ||
+				path.FullIdxColLens[idxOffset] != types.UnspecifiedLength ||
+				path.FullIdxCols[idxOffset] == nil {
+				continue
+			}
+			usedColIDs = append(usedColIDs, path.FullIdxCols[idxOffset].UniqueID)
+		}
+		joinKeyNDV := getColsNDVLowerBoundFromHistColl(usedColIDs, ds.tableStats.HistColl)
+		if joinKeyNDV > 0 {
+			rowCountUpperBound = ds.tableStats.RowCount / float64(joinKeyNDV)
+		}
+	}
+
+	if rowCountUpperBound > 0 {
+		rowCount = math.Min(rowCount, rowCountUpperBound)
+	}
 	if maxOneRow {
 		// Theoretically, this line is unnecessary because row count estimation of join should guarantee rowCount is not larger
 		// than 1.0; however, there may be rowCount larger than 1.0 in reality, e.g, pseudo statistics cases, which does not reflect
@@ -1279,6 +1360,9 @@ func (p *LogicalJoin) constructInnerIndexScanTask(
 		// rowCount is computed from result row count of join, which has already accounted the filters on DataSource,
 		// i.e, rowCount equals to `countAfterIndex * selectivity`.
 		cnt := rowCount / selectivity
+		if rowCountUpperBound > 0 {
+			cnt = math.Min(cnt, rowCountUpperBound)
+		}
 		if maxOneRow {
 			cnt = math.Min(cnt, 1.0)
 		}
@@ -1292,6 +1376,9 @@ func (p *LogicalJoin) constructInnerIndexScanTask(
 			selectivity = SelectionFactor
 		}
 		cnt := tmpPath.CountAfterIndex / selectivity
+		if rowCountUpperBound > 0 {
+			cnt = math.Min(cnt, rowCountUpperBound)
+		}
 		if maxOneRow {
 			cnt = math.Min(cnt, 1.0)
 		}

sessionctx/variable/session.go

@@ -1340,6 +1340,9 @@ var (
 	TiDBOptFixControl44262 uint64 = 44262
 	// TiDBOptFixControl44389 controls whether to consider non-point ranges of some CNF item when building ranges.
 	TiDBOptFixControl44389 uint64 = 44389
+	// TiDBOptFixControl44855 controls whether to use a more accurate upper bound when estimating row count of index
+	// range scan under inner side of index join.
+	TiDBOptFixControl44855 uint64 = 44855
 )
 
 // GetOptimizerFixControlValue returns the specified value of the optimizer fix control.

statistics/integration_test.go

@@ -24,10 +24,12 @@ import (
 
 	"github.com/pingcap/failpoint"
 	"github.com/pingcap/tidb/parser/model"
+	"github.com/pingcap/tidb/parser/mysql"
 	"github.com/pingcap/tidb/statistics"
 	"github.com/pingcap/tidb/statistics/handle"
 	"github.com/pingcap/tidb/testkit"
 	"github.com/pingcap/tidb/testkit/testdata"
+	"github.com/pingcap/tidb/types"
 	"github.com/stretchr/testify/require"
 )
 
@@ -708,3 +710,64 @@ func TestUpdateNotLoadIndexFMSketch(t *testing.T) {
 	require.Nil(t, h.GetPartitionStats(tblInfo, p0.ID).Indices[idxInfo.ID].FMSketch)
 	require.Nil(t, h.GetPartitionStats(tblInfo, p1.ID).Indices[idxInfo.ID].FMSketch)
 }
+
+func TestIndexJoinInnerRowCountUpperBound(t *testing.T) {
+	store, dom := testkit.CreateMockStoreAndDomain(t)
+	testKit := testkit.NewTestKit(t, store)
+	h := dom.StatsHandle()
+
+	testKit.MustExec("use test")
+	testKit.MustExec("drop table if exists t")
+	testKit.MustExec("create table t(a int, b int, index idx(b))")
+	require.NoError(t, h.HandleDDLEvent(<-h.DDLEventCh()))
+	is := dom.InfoSchema()
+	tb, err := is.TableByName(model.NewCIStr("test"), model.NewCIStr("t"))
+	require.NoError(t, err)
+	tblInfo := tb.Meta()
+
+	// Mock the stats:
+	// The two columns are the same.
+	// From 0 to 499, each value has 1000 rows. Therefore, NDV is 500 and total row count is 500000.
+	mockStatsTbl := mockStatsTable(tblInfo, 500000)
+	colValues, err := generateIntDatum(1, 500)
+	require.NoError(t, err)
+	for i := 1; i <= 2; i++ {
+		mockStatsTbl.Columns[int64(i)] = &statistics.Column{
+			Histogram:         *mockStatsHistogram(int64(i), colValues, 1000, types.NewFieldType(mysql.TypeLonglong)),
+			Info:              tblInfo.Columns[i-1],
+			StatsLoadedStatus: statistics.NewStatsFullLoadStatus(),
+			StatsVer:          2,
+		}
+	}
+	generateMapsForMockStatsTbl(mockStatsTbl)
+	stat := h.GetTableStats(tblInfo)
+	stat.HistColl = mockStatsTbl.HistColl
+
+	query := "explain format = 'brief' " +
+		"select /*+ inl_join(t2) */ * from (select * from t where t.a < 1) as t1 join t t2 where t2.a = 0 and t1.a = t2.b"
+
+	testKit.MustQuery(query).Check(testkit.Rows(
+		"IndexJoin 1000000.00 root  inner join, inner:IndexLookUp, outer key:test.t.a, inner key:test.t.b, equal cond:eq(test.t.a, test.t.b)",
+		"├─TableReader(Build) 1000.00 root  data:Selection",
+		"│ └─Selection 1000.00 cop[tikv]  lt(test.t.a, 1), not(isnull(test.t.a))",
+		"│   └─TableFullScan 500000.00 cop[tikv] table:t keep order:false, stats:pseudo",
+		"└─IndexLookUp(Probe) 1000000.00 root  ",
+		"  ├─Selection(Build) 500000000.00 cop[tikv]  not(isnull(test.t.b))",
+		"  │ └─IndexRangeScan 500000000.00 cop[tikv] table:t2, index:idx(b) range: decided by [eq(test.t.b, test.t.a)], keep order:false, stats:pseudo",
+		"  └─Selection(Probe) 1000000.00 cop[tikv]  eq(test.t.a, 0)",
+		"    └─TableRowIDScan 500000000.00 cop[tikv] table:t2 keep order:false, stats:pseudo",
+	))
+
+	testKit.MustExec("set @@tidb_opt_fix_control = '44855:ON'")
+	testKit.MustQuery(query).Check(testkit.Rows(
+		"IndexJoin 1000000.00 root  inner join, inner:IndexLookUp, outer key:test.t.a, inner key:test.t.b, equal cond:eq(test.t.a, test.t.b)",
+		"├─TableReader(Build) 1000.00 root  data:Selection",
+		"│ └─Selection 1000.00 cop[tikv]  lt(test.t.a, 1), not(isnull(test.t.a))",
+		"│   └─TableFullScan 500000.00 cop[tikv] table:t keep order:false, stats:pseudo",
+		"└─IndexLookUp(Probe) 1000000.00 root  ",
+		"  ├─Selection(Build) 1000000.00 cop[tikv]  not(isnull(test.t.b))",
+		"  │ └─IndexRangeScan 1000000.00 cop[tikv] table:t2, index:idx(b) range: decided by [eq(test.t.b, test.t.a)], keep order:false, stats:pseudo",
+		"  └─Selection(Probe) 1000000.00 cop[tikv]  eq(test.t.a, 0)",
+		"    └─TableRowIDScan 1000000.00 cop[tikv] table:t2 keep order:false, stats:pseudo",
+	))
+}