planner: improve skyline pruning

planner/core/exhaust_physical_plans.go

@@ -1251,7 +1251,7 @@ func (ijHelper *indexJoinBuildHelper) findUsefulEqAndInFilters(innerPlan *DataSo
 	var remainedEqOrIn []expression.Expression
 	// Extract the eq/in functions of possible join key.
 	// you can see the comment of ExtractEqAndInCondition to get the meaning of the second return value.
-	usefulEqOrInFilters, remainedEqOrIn, remainingRangeCandidates, _ = ranger.ExtractEqAndInCondition(
+	usefulEqOrInFilters, remainedEqOrIn, remainingRangeCandidates, _, _ = ranger.ExtractEqAndInCondition(
 		innerPlan.ctx, innerPlan.pushedDownConds,
 		ijHelper.curNotUsedIndexCols,
 		ijHelper.curNotUsedColLens,

planner/core/find_best_task.go

@@ -414,10 +414,11 @@ func (ds *DataSource) tryToGetDualTask() (task, error) {
 
 // candidatePath is used to maintain required info for skyline pruning.
 type candidatePath struct {
-	path         *util.AccessPath
-	columnSet    *intsets.Sparse // columnSet is the set of columns that occurred in the access conditions.
-	isSingleScan bool
-	isMatchProp  bool
+	path               *util.AccessPath
+	accessCondsColSet  *intsets.Sparse // accessCondsColSet is the set of columns that occurred in the access conditions.
+	indexFiltersColSet *intsets.Sparse // indexFiltersColSet is the set of columns that occurred in the index filters.
+	isSingleScan       bool
+	isMatchProp        bool
 }
 
 // compareColumnSet will compares the two set. The last return value is used to indicate
@@ -450,18 +451,31 @@ func compareBool(l, r bool) int {
 	return 1
 }
 
+func compareIndexBack(lhs, rhs *candidatePath) (int, bool) {
+	result := compareBool(lhs.isSingleScan, rhs.isSingleScan)
+	if result == 0 && !lhs.isSingleScan {
+		// if both lhs and rhs need to access table after IndexScan, we use the set of columns that occurred in IndexFilters
+		// to compare how many table rows will be accessed.
+		return compareColumnSet(lhs.indexFiltersColSet, rhs.indexFiltersColSet)
+	}
+	return result, true
+}
+
 // compareCandidates is the core of skyline pruning. It compares the two candidate paths on three dimensions:
 // (1): the set of columns that occurred in the access condition,
 // (2): whether or not it matches the physical property
 // (3): does it require a double scan.
 // If `x` is not worse than `y` at all factors,
 // and there exists one factor that `x` is better than `y`, then `x` is better than `y`.
 func compareCandidates(lhs, rhs *candidatePath) int {
-	setsResult, comparable := compareColumnSet(lhs.columnSet, rhs.columnSet)
+	setsResult, comparable := compareColumnSet(lhs.accessCondsColSet, rhs.accessCondsColSet)
+	if !comparable {
+		return 0
+	}
+	scanResult, comparable := compareIndexBack(lhs, rhs)
 	if !comparable {
 		return 0
 	}
-	scanResult := compareBool(lhs.isSingleScan, rhs.isSingleScan)
 	matchResult := compareBool(lhs.isMatchProp, rhs.isMatchProp)
 	sum := setsResult + scanResult + matchResult
 	if setsResult >= 0 && scanResult >= 0 && matchResult >= 0 && sum > 0 {
@@ -473,52 +487,70 @@ func compareCandidates(lhs, rhs *candidatePath) int {
 	return 0
 }
 
-func (ds *DataSource) getTableCandidate(path *util.AccessPath, prop *property.PhysicalProperty) *candidatePath {
-	candidate := &candidatePath{path: path}
+func (ds *DataSource) isMatchProp(path *util.AccessPath, prop *property.PhysicalProperty) bool {
+	var isMatchProp bool
 	if path.IsIntHandlePath {
 		pkCol := ds.getPKIsHandleCol()
 		if len(prop.SortItems) == 1 && pkCol != nil {
-			candidate.isMatchProp = prop.SortItems[0].Col.Equal(nil, pkCol)
+			isMatchProp = prop.SortItems[0].Col.Equal(nil, pkCol)
 			if path.StoreType == kv.TiFlash {
-				candidate.isMatchProp = candidate.isMatchProp && !prop.SortItems[0].Desc
+				isMatchProp = isMatchProp && !prop.SortItems[0].Desc
 			}
 		}
-	} else {
-		all, _ := prop.AllSameOrder()
-		// When the prop is empty or `all` is false, `isMatchProp` is better to be `false` because
-		// it needs not to keep order for index scan.
-		if !prop.IsEmpty() && all {
-			for i, col := range path.IdxCols {
-				if col.Equal(nil, prop.SortItems[0].Col) {
-					candidate.isMatchProp = matchIndicesProp(path.IdxCols[i:], path.IdxColLens[i:], prop.SortItems)
+		return isMatchProp
+	}
+	all, _ := prop.AllSameOrder()
+	// When the prop is empty or `all` is false, `isMatchProp` is better to be `false` because
+	// it needs not to keep order for index scan.
+
+	// Basically, if `prop.SortItems` is the prefix of `path.IdxCols`, then `isMatchProp` is true. However, we need to consider
+	// the situations when some columns of `path.IdxCols` are evaluated as constant. For example:
+	// ```
+	// create table t(a int, b int, c int, d int, index idx_a_b_c(a, b, c), index idx_d_c_b_a(d, c, b, a));
+	// select * from t where a = 1 order by b, c;
+	// select * from t where b = 1 order by a, c;
+	// select * from t where d = 1 and b = 2 order by c, a;
+	// select * from t where d = 1 and b = 2 order by c, b, a;
+	// ```
+	// In the first two `SELECT` statements, `idx_a_b_c` matches the sort order. In the last two `SELECT` statements, `idx_d_c_b_a`
+	// matches the sort order. Hence, we use `path.ConstCols` to deal with the above situations.
+	if !prop.IsEmpty() && all && len(path.IdxCols) >= len(prop.SortItems) {
+		isMatchProp = true
+		i := 0
+		for _, sortItem := range prop.SortItems {
+			found := false
+			for ; i < len(path.IdxCols); i++ {
+				if path.IdxColLens[i] == types.UnspecifiedLength && sortItem.Col.Equal(nil, path.IdxCols[i]) {
+					found = true
+					i++
 					break
-				} else if i >= path.EqCondCount {
+				}
+				if path.ConstCols == nil || i >= len(path.ConstCols) || !path.ConstCols[i] {
 					break
 				}
 			}
+			if !found {
+				isMatchProp = false
+				break
+			}
 		}
 	}
-	candidate.columnSet = expression.ExtractColumnSet(path.AccessConds)
+	return isMatchProp
+}
+
+func (ds *DataSource) getTableCandidate(path *util.AccessPath, prop *property.PhysicalProperty) *candidatePath {
+	candidate := &candidatePath{path: path}
+	candidate.isMatchProp = ds.isMatchProp(path, prop)
+	candidate.accessCondsColSet = expression.ExtractColumnSet(path.AccessConds)
 	candidate.isSingleScan = true
 	return candidate
 }
 
 func (ds *DataSource) getIndexCandidate(path *util.AccessPath, prop *property.PhysicalProperty, isSingleScan bool) *candidatePath {
 	candidate := &candidatePath{path: path}
-	all, _ := prop.AllSameOrder()
-	// When the prop is empty or `all` is false, `isMatchProp` is better to be `false` because
-	// it needs not to keep order for index scan.
-	if !prop.IsEmpty() && all {
-		for i, col := range path.IdxCols {
-			if col.Equal(nil, prop.SortItems[0].Col) {
-				candidate.isMatchProp = matchIndicesProp(path.IdxCols[i:], path.IdxColLens[i:], prop.SortItems)
-				break
-			} else if i >= path.EqCondCount {
-				break
-			}
-		}
-	}
-	candidate.columnSet = expression.ExtractColumnSet(path.AccessConds)
+	candidate.isMatchProp = ds.isMatchProp(path, prop)
+	candidate.accessCondsColSet = expression.ExtractColumnSet(path.AccessConds)
+	candidate.indexFiltersColSet = expression.ExtractColumnSet(path.IndexFilters)
 	candidate.isSingleScan = isSingleScan
 	return candidate
 }

planner/core/integration_test.go

@@ -3971,6 +3971,30 @@ func (s *testIntegrationSerialSuite) TestSelectIgnoreTemporaryTableInView(c *C)
 
 }
 
+// TestIsMatchProp is used to test https://github.com/pingcap/tidb/issues/26017.
+func (s *testIntegrationSuite) TestIsMatchProp(c *C) {
+	tk := testkit.NewTestKit(c, s.store)
+
+	tk.MustExec("use test")
+	tk.MustExec("drop table if exists t1, t2")
+	tk.MustExec("create table t1(a int, b int, c int, d int, index idx_a_b_c(a, b, c))")
+	tk.MustExec("create table t2(a int, b int, c int, d int, index idx_a_b_c_d(a, b, c, d))")
+
+	var input []string
+	var output []struct {
+		SQL  string
+		Plan []string
+	}
+	s.testData.GetTestCases(c, &input, &output)
+	for i, tt := range input {
+		s.testData.OnRecord(func() {
+			output[i].SQL = tt
+			output[i].Plan = s.testData.ConvertRowsToStrings(tk.MustQuery("explain format = 'brief' " + tt).Rows())
+		})
+		tk.MustQuery("explain format = 'brief' " + tt).Check(testkit.Rows(output[i].Plan...))
+	}
+}
+
 func (s *testIntegrationSerialSuite) TestIssue26250(c *C) {
 	tk := testkit.NewTestKit(c, s.store)
 	tk.MustExec("use test")

planner/core/logical_plan_test.go

@@ -1694,12 +1694,28 @@ func (s *testPlanSuite) TestSkylinePruning(c *C) {
 		},
 		{
 			sql:    "select * from t where f > 1 and g > 1",
-			result: "PRIMARY_KEY,f,g,f_g",
+			result: "PRIMARY_KEY,g,f_g",
 		},
 		{
 			sql:    "select count(1) from t",
 			result: "PRIMARY_KEY,c_d_e,f,g,f_g,c_d_e_str,e_d_c_str_prefix",
 		},
+		{
+			sql:    "select * from t where f > 3 and g = 5",
+			result: "PRIMARY_KEY,g,f_g",
+		},
+		{
+			sql:    "select * from t where g = 5 order by f",
+			result: "PRIMARY_KEY,g,f_g",
+		},
+		{
+			sql:    "select * from t where d = 3 order by c, e",
+			result: "PRIMARY_KEY,c_d_e",
+		},
+		{
+			sql:    "select * from t where d = 1 and f > 1 and g > 1 order by c, e",
+			result: "PRIMARY_KEY,c_d_e,g,f_g",
+		},
 	}
 	ctx := context.TODO()
 	for i, tt := range tests {

planner/core/logical_plans.go

@@ -675,6 +675,12 @@ func (ds *DataSource) deriveCommonHandleTablePathStats(path *util.AccessPath, co
 		path.EqCondCount = res.EqCondCount
 		path.EqOrInCondCount = res.EqOrInCount
 		path.IsDNFCond = res.IsDNFCond
+		path.ConstCols = make([]bool, len(path.IdxCols))
+		if res.ColumnValues != nil {
+			for i := range path.ConstCols {
+				path.ConstCols[i] = res.ColumnValues[i] != nil
+			}
+		}
 		path.CountAfterAccess, err = ds.tableStats.HistColl.GetRowCountByIndexRanges(sc, path.Index.ID, path.Ranges)
 		if err != nil {
 			return false, err
@@ -854,6 +860,12 @@ func (ds *DataSource) fillIndexPath(path *util.AccessPath, conds []expression.Ex
 		path.EqCondCount = res.EqCondCount
 		path.EqOrInCondCount = res.EqOrInCount
 		path.IsDNFCond = res.IsDNFCond
+		path.ConstCols = make([]bool, len(path.IdxCols))
+		if res.ColumnValues != nil {
+			for i := range path.ConstCols {
+				path.ConstCols[i] = res.ColumnValues[i] != nil
+			}
+		}
 		path.CountAfterAccess, err = ds.tableStats.HistColl.GetRowCountByIndexRanges(sc, path.Index.ID, path.Ranges)
 		if err != nil {
 			return err

planner/core/mock.go

@@ -43,9 +43,9 @@ func newDateType() types.FieldType {
 
 // MockSignedTable is only used for plan related tests.
 func MockSignedTable() *model.TableInfo {
-	// column: a, b, c, d, e, c_str, d_str, e_str, f, g
+	// column: a, b, c, d, e, c_str, d_str, e_str, f, g, h, i_date
 	// PK: a
-	// indices: c_d_e, e, f, g, f_g, c_d_e_str, c_d_e_str_prefix
+	// indices: c_d_e, e, f, g, f_g, c_d_e_str, e_d_c_str_prefix
 	indices := []*model.IndexInfo{
 		{
 			Name: model.NewCIStr("c_d_e"),

planner/core/testdata/integration_suite_in.json

@@ -310,5 +310,15 @@
       "select sum(1) from s1",
       "select count(1) as cnt from s1 union select count(1) as cnt from s2"
     ]
+  },
+  {
+    "name": "TestIsMatchProp",
+    "cases": [
+      "select a, b, c from t1 where a > 3 and b = 4 order by a, c",
+      "select * from t2 where a = 1 and c = 2 order by b, d",
+      "select a, b, c from t1 where (a = 1 and b = 1 and c = 1) or (a = 1 and b = 1 and c = 2) order by c",
+      "select a, b, c from t1 where (a = 1 and b = 1 and c < 3) or (a = 1 and b = 1 and c > 6) order by c",
+      "select * from t2 where ((a = 1 and b = 1 and d < 3) or (a = 1 and b = 1 and d > 6)) and c = 3 order by d"
+    ]
   }
 ]

planner/core/testdata/integration_suite_out.json

@@ -1636,5 +1636,48 @@
         ]
       }
     ]
+  },
+  {
+    "Name": "TestIsMatchProp",
+    "Cases": [
+      {
+        "SQL": "select a, b, c from t1 where a > 3 and b = 4 order by a, c",
+        "Plan": [
+          "IndexReader 3.33 root  index:Selection",
+          "└─Selection 3.33 cop[tikv]  eq(test.t1.b, 4)",
+          "  └─IndexRangeScan 3333.33 cop[tikv] table:t1, index:idx_a_b_c(a, b, c) range:(3,+inf], keep order:true, stats:pseudo"
+        ]
+      },
+      {
+        "SQL": "select * from t2 where a = 1 and c = 2 order by b, d",
+        "Plan": [
+          "IndexReader 0.01 root  index:Selection",
+          "└─Selection 0.01 cop[tikv]  eq(test.t2.c, 2)",
+          "  └─IndexRangeScan 10.00 cop[tikv] table:t2, index:idx_a_b_c_d(a, b, c, d) range:[1,1], keep order:true, stats:pseudo"
+        ]
+      },
+      {
+        "SQL": "select a, b, c from t1 where (a = 1 and b = 1 and c = 1) or (a = 1 and b = 1 and c = 2) order by c",
+        "Plan": [
+          "IndexReader 0.03 root  index:IndexRangeScan",
+          "└─IndexRangeScan 0.03 cop[tikv] table:t1, index:idx_a_b_c(a, b, c) range:[1 1 1,1 1 2], keep order:true, stats:pseudo"
+        ]
+      },
+      {
+        "SQL": "select a, b, c from t1 where (a = 1 and b = 1 and c < 3) or (a = 1 and b = 1 and c > 6) order by c",
+        "Plan": [
+          "IndexReader 0.67 root  index:IndexRangeScan",
+          "└─IndexRangeScan 0.67 cop[tikv] table:t1, index:idx_a_b_c(a, b, c) range:[1 1 -inf,1 1 3), (1 1 6,1 1 +inf], keep order:true, stats:pseudo"
+        ]
+      },
+      {
+        "SQL": "select * from t2 where ((a = 1 and b = 1 and d < 3) or (a = 1 and b = 1 and d > 6)) and c = 3 order by d",
+        "Plan": [
+          "IndexReader 0.00 root  index:Selection",
+          "└─Selection 0.00 cop[tikv]  eq(test.t2.c, 3), or(and(eq(test.t2.a, 1), and(eq(test.t2.b, 1), lt(test.t2.d, 3))), and(eq(test.t2.a, 1), and(eq(test.t2.b, 1), gt(test.t2.d, 6))))",
+          "  └─IndexRangeScan 10.00 cop[tikv] table:t2, index:idx_a_b_c_d(a, b, c, d) range:[1,1], keep order:true, stats:pseudo"
+        ]
+      }
+    ]
   }
 ]

planner/util/path.go

@@ -32,7 +32,9 @@ type AccessPath struct {
 	FullIdxColLens []int
 	IdxCols        []*expression.Column
 	IdxColLens     []int
-	Ranges         []*ranger.Range
+	// ConstCols indicates whether the column is constant under the given conditions for all index columns.
+	ConstCols []bool
+	Ranges    []*ranger.Range
 	// CountAfterAccess is the row count after we apply range seek and before we use other filter to filter data.
 	// For index merge path, CountAfterAccess is the row count after partial paths and before we apply table filters.
 	CountAfterAccess float64