*: fix bug when unsigned histogram meets signed ranges in feedback

distsql/request_builder.go

@@ -269,7 +269,10 @@ func IndexRangesToKVRanges(sc *stmtctx.StatementContext, tid, idxID int64, range
 		feedbackRanges = append(feedbackRanges, &ranger.Range{LowVal: []types.Datum{types.NewBytesDatum(low)},
 			HighVal: []types.Datum{types.NewBytesDatum(high)}, LowExclude: false, HighExclude: true})
 	}
-	feedbackRanges = fb.Hist.SplitRange(sc, feedbackRanges, true)
+	feedbackRanges, ok := fb.Hist.SplitRange(sc, feedbackRanges, true)
+	if !ok {
+		fb.Invalidate()
+	}
 	krs := make([]kv.KeyRange, 0, len(feedbackRanges))
 	for _, ran := range feedbackRanges {
 		low, high := ran.LowVal[0].GetBytes(), ran.HighVal[0].GetBytes()

executor/executor_test.go

@@ -49,6 +49,7 @@ import (
 	"github.com/pingcap/tidb/sessionctx"
 	"github.com/pingcap/tidb/sessionctx/stmtctx"
 	"github.com/pingcap/tidb/sessionctx/variable"
+	"github.com/pingcap/tidb/statistics"
 	"github.com/pingcap/tidb/store/mockstore"
 	"github.com/pingcap/tidb/store/mockstore/mocktikv"
 	"github.com/pingcap/tidb/store/tikv"
@@ -3757,6 +3758,21 @@ func (s *testSuite) TestSplitIndexRegion(c *C) {
 	c.Assert(terr.Code(), Equals, terror.ErrCode(mysql.WarnDataTruncated))
 }
 
+func (s *testSuite) TestUnsignedFeedback(c *C) {
+	tk := testkit.NewTestKit(c, s.store)
+	oriProbability := statistics.FeedbackProbability.Load()
+	statistics.FeedbackProbability.Store(1.0)
+	defer func() { statistics.FeedbackProbability.Store(oriProbability) }()
+	tk.MustExec("use test")
+	tk.MustExec("drop table if exists t")
+	tk.MustExec("create table t(a bigint unsigned, b int, primary key(a))")
+	tk.MustExec("insert into t values (1,1),(2,2)")
+	tk.MustExec("analyze table t")
+	tk.MustQuery("select count(distinct b) from t").Check(testkit.Rows("2"))
+	result := tk.MustQuery("explain analyze select count(distinct b) from t")
+	c.Assert(result.Rows()[2][3], Equals, "table:t, range:[0,+inf], keep order:false")
+}
+
 type testOOMSuite struct {
 	store kv.Storage
 	do    *domain.Domain

executor/table_reader.go

@@ -103,7 +103,11 @@ func (e *TableReaderExecutor) Open(ctx context.Context) error {
 	e.resultHandler = &tableResultHandler{}
 	if e.feedback != nil && e.feedback.Hist != nil {
 		// EncodeInt don't need *statement.Context.
-		e.ranges = e.feedback.Hist.SplitRange(nil, e.ranges, false)
+		var ok bool
+		e.ranges, ok = e.feedback.Hist.SplitRange(nil, e.ranges, false)
+		if !ok {
+			e.feedback.Invalidate()
+		}
 	}
 	firstPartRanges, secondPartRanges := splitRanges(e.ranges, e.keepOrder, e.desc)
 	firstResult, err := e.buildResp(ctx, firstPartRanges)

planner/core/logical_plans.go

@@ -376,19 +376,21 @@ func (ds *DataSource) deriveTablePathStats(path *accessPath) (bool, error) {
 	path.tableFilters = ds.pushedDownConds
 	var pkCol *expression.Column
 	columnLen := len(ds.schema.Columns)
-	if columnLen > 0 && ds.schema.Columns[columnLen-1].ID == model.ExtraHandleID {
-		pkCol = ds.schema.Columns[columnLen-1]
-	} else if ds.tableInfo.PKIsHandle {
+	isUnsigned := false
+	if ds.tableInfo.PKIsHandle {
 		if pkColInfo := ds.tableInfo.GetPkColInfo(); pkColInfo != nil {
+			isUnsigned = mysql.HasUnsignedFlag(pkColInfo.Flag)
 			pkCol = expression.ColInfo2Col(ds.schema.Columns, pkColInfo)
 		}
+	} else if columnLen > 0 && ds.schema.Columns[columnLen-1].ID == model.ExtraHandleID {
+		pkCol = ds.schema.Columns[columnLen-1]
 	}
 	if pkCol == nil {
-		path.ranges = ranger.FullIntRange(false)
+		path.ranges = ranger.FullIntRange(isUnsigned)
 		return false, nil
 	}
 
-	path.ranges = ranger.FullIntRange(mysql.HasUnsignedFlag(pkCol.RetType.Flag))
+	path.ranges = ranger.FullIntRange(isUnsigned)
 	if len(ds.pushedDownConds) == 0 {
 		return false, nil
 	}

statistics/handle/update.go

@@ -972,7 +972,11 @@ func (h *Handle) dumpRangeFeedback(sc *stmtctx.StatementContext, ran *ranger.Ran
 			ran.HighVal[0] = statistics.GetMaxValue(q.Hist.Tp)
 		}
 	}
-	ranges := q.Hist.SplitRange(sc, []*ranger.Range{ran}, q.Tp == statistics.IndexType)
+	ranges, ok := q.Hist.SplitRange(sc, []*ranger.Range{ran}, q.Tp == statistics.IndexType)
+	if !ok {
+		logutil.Logger(context.Background()).Debug("type of histogram and ranges mismatch")
+		return nil
+	}
 	counts := make([]float64, 0, len(ranges))
 	sum := 0.0
 	for i, r := range ranges {

statistics/handle/update_test.go

@@ -708,7 +708,7 @@ func (s *testStatsSuite) TestSplitRange(c *C) {
 				HighExclude: t.exclude[i+1],
 			})
 		}
-		ranges = h.SplitRange(nil, ranges, false)
+		ranges, _ = h.SplitRange(nil, ranges, false)
 		var ranStrs []string
 		for _, ran := range ranges {
 			ranStrs = append(ranStrs, ran.String())

statistics/histogram.go

@@ -395,10 +395,48 @@ func validRange(sc *stmtctx.StatementContext, ran *ranger.Range, encoded bool) b
 	return bytes.Compare(low, high) < 0
 }
 
+func checkKind(vals []types.Datum, kind byte) bool {
+	if kind == types.KindString {
+		kind = types.KindBytes
+	}
+	for _, val := range vals {
+		valKind := val.Kind()
+		if valKind == types.KindNull || valKind == types.KindMinNotNull || valKind == types.KindMaxValue {
+			continue
+		}
+		if valKind == types.KindString {
+			valKind = types.KindBytes
+		}
+		if valKind != kind {
+			return false
+		}
+		// Only check the first non-null value.
+		break
+	}
+	return true
+}
+
+func (hg *Histogram) typeMatch(ranges []*ranger.Range) bool {
+	kind := hg.GetLower(0).Kind()
+	for _, ran := range ranges {
+		if !checkKind(ran.LowVal, kind) || !checkKind(ran.HighVal, kind) {
+			return false
+		}
+	}
+	return true
+}
+
 // SplitRange splits the range according to the histogram upper bound. Note that we treat last bucket's upper bound
 // as inf, so all the split Ranges will totally fall in one of the (-inf, u(0)], (u(0), u(1)],...(u(n-3), u(n-2)],
 // (u(n-2), +inf), where n is the number of buckets, u(i) is the i-th bucket's upper bound.
-func (hg *Histogram) SplitRange(sc *stmtctx.StatementContext, ranges []*ranger.Range, encoded bool) []*ranger.Range {
+func (hg *Histogram) SplitRange(sc *stmtctx.StatementContext, oldRanges []*ranger.Range, encoded bool) ([]*ranger.Range, bool) {
+	if !hg.typeMatch(oldRanges) {
+		return oldRanges, false
+	}
+	ranges := make([]*ranger.Range, 0, len(oldRanges))
+	for _, ran := range oldRanges {
+		ranges = append(ranges, ran.Clone())
+	}
 	split := make([]*ranger.Range, 0, len(ranges))
 	for len(ranges) > 0 {
 		// Find the last bound that greater or equal to the LowVal.
@@ -447,7 +485,7 @@ func (hg *Histogram) SplitRange(sc *stmtctx.StatementContext, ranges []*ranger.R
 			}
 		}
 	}
-	return split
+	return split, true
 }
 
 func (hg *Histogram) bucketCount(idx int) int64 {
@@ -954,7 +992,11 @@ func (coll *HistColl) NewHistCollBySelectivity(sc *stmtctx.StatementContext, sta
 		}
 		newCol.Histogram = *NewHistogram(oldCol.ID, int64(float64(oldCol.NDV)*node.Selectivity), 0, 0, oldCol.Tp, chunk.InitialCapacity, 0)
 		var err error
-		splitRanges := oldCol.Histogram.SplitRange(sc, node.Ranges, false)
+		splitRanges, ok := oldCol.Histogram.SplitRange(sc, node.Ranges, false)
+		if !ok {
+			logutil.Logger(context.Background()).Warn("[Histogram-in-plan]: the type of histogram and ranges mismatch")
+			continue
+		}
 		// Deal with some corner case.
 		if len(splitRanges) > 0 {
 			// Deal with NULL values.