opt: add optimizer_prefer_bounded_cardinality session setting

Release note (sql change): The `optimizer_prefer_bounded_cardinality` session setting has been added which instructs the optimizer to prefer query plans where every expression has a guaranteed upper-bound on the number of rows it will process. This may help the optimizer produce better query plans in some cases. This setting is disabled by default.
cockroachdb · Jan 31, 2025 · 5918578 · 5918578
1 parent f28b341
commit 5918578
Show file tree

Hide file tree

Showing 14 changed files with 554 additions and 3 deletions.
diff --git a/pkg/sql/exec_util.go b/pkg/sql/exec_util.go
@@ -3851,6 +3851,10 @@ func (m *sessionDataMutator) SetLegacyVarcharTyping(val bool) {
 	m.data.LegacyVarcharTyping = val
 }
 
+func (m *sessionDataMutator) SetOptimizerPreferBoundedCardinality(b bool) {
+	m.data.OptimizerPreferBoundedCardinality = b
+}
+
 // Utility functions related to scrubbing sensitive information on SQL Stats.
 
 // quantizeCounts ensures that the Count field in the

diff --git a/pkg/sql/logictest/testdata/logic_test/information_schema b/pkg/sql/logictest/testdata/logic_test/information_schema
@@ -6172,6 +6172,7 @@ optimizer                                                  on
 optimizer_always_use_histograms                            on
 optimizer_hoist_uncorrelated_equality_subqueries           on
 optimizer_merge_joins_enabled                              on
+optimizer_prefer_bounded_cardinality                       off
 optimizer_prove_implication_with_virtual_computed_columns  on
 optimizer_push_limit_into_project_filtered_scan            off
 optimizer_push_offset_into_index_join                      on

diff --git a/pkg/sql/logictest/testdata/logic_test/pg_catalog b/pkg/sql/logictest/testdata/logic_test/pg_catalog
@@ -2947,6 +2947,7 @@ opt_split_scan_limit                                       2048                N
 optimizer_always_use_histograms                            on                  NULL      NULL        NULL        string
 optimizer_hoist_uncorrelated_equality_subqueries           on                  NULL      NULL        NULL        string
 optimizer_merge_joins_enabled                              on                  NULL      NULL        NULL        string
+optimizer_prefer_bounded_cardinality                       off                 NULL      NULL        NULL        string
 optimizer_prove_implication_with_virtual_computed_columns  on                  NULL      NULL        NULL        string
 optimizer_push_limit_into_project_filtered_scan            off                 NULL      NULL        NULL        string
 optimizer_push_offset_into_index_join                      on                  NULL      NULL        NULL        string
@@ -3138,6 +3139,7 @@ opt_split_scan_limit                                       2048                N
 optimizer_always_use_histograms                            on                  NULL  user     NULL      on                  on
 optimizer_hoist_uncorrelated_equality_subqueries           on                  NULL  user     NULL      on                  on
 optimizer_merge_joins_enabled                              on                  NULL  user     NULL      on                  on
+optimizer_prefer_bounded_cardinality                       off                 NULL  user     NULL      off                 off
 optimizer_prove_implication_with_virtual_computed_columns  on                  NULL  user     NULL      on                  on
 optimizer_push_limit_into_project_filtered_scan            off                 NULL  user     NULL      off                 off
 optimizer_push_offset_into_index_join                      on                  NULL  user     NULL      on                  on
@@ -3328,6 +3330,7 @@ optimizer                                                  NULL    NULL     NULL
 optimizer_always_use_histograms                            NULL    NULL     NULL     NULL        NULL
 optimizer_hoist_uncorrelated_equality_subqueries           NULL    NULL     NULL     NULL        NULL
 optimizer_merge_joins_enabled                              NULL    NULL     NULL     NULL        NULL
+optimizer_prefer_bounded_cardinality                       NULL    NULL     NULL     NULL        NULL
 optimizer_prove_implication_with_virtual_computed_columns  NULL    NULL     NULL     NULL        NULL
 optimizer_push_limit_into_project_filtered_scan            NULL    NULL     NULL     NULL        NULL
 optimizer_push_offset_into_index_join                      NULL    NULL     NULL     NULL        NULL

diff --git a/pkg/sql/logictest/testdata/logic_test/show_source b/pkg/sql/logictest/testdata/logic_test/show_source
@@ -128,6 +128,7 @@ opt_split_scan_limit                                       2048
 optimizer_always_use_histograms                            on
 optimizer_hoist_uncorrelated_equality_subqueries           on
 optimizer_merge_joins_enabled                              on
+optimizer_prefer_bounded_cardinality                       off
 optimizer_prove_implication_with_virtual_computed_columns  on
 optimizer_push_limit_into_project_filtered_scan            off
 optimizer_push_offset_into_index_join                      on

diff --git a/pkg/sql/opt/memo/cost.go b/pkg/sql/opt/memo/cost.go
@@ -19,8 +19,12 @@ type Cost struct {
 // group members during testing, by setting their cost so high that any other
 // member will have a lower cost.
 var MaxCost = Cost{
-	C:     math.Inf(+1),
-	Flags: CostFlags{FullScanPenalty: true, HugeCostPenalty: true},
+	C: math.Inf(+1),
+	Flags: CostFlags{
+		FullScanPenalty:      true,
+		HugeCostPenalty:      true,
+		UnboundedCardinality: true,
+	},
 }
 
 // Less returns true if this cost is lower than the given cost.
@@ -57,6 +61,10 @@ type CostFlags struct {
 	// used when the optimizer is forced to use a particular plan, and will error
 	// if it cannot be used.
 	HugeCostPenalty bool
+	// UnboundedCardinality is true if the operator or any of its descendants
+	// have no guaranteed upperbound on the number of rows that they can
+	// produce. See props.AnyCardinality.
+	UnboundedCardinality bool
 }
 
 // Less returns true if these flags indicate a lower penalty than the other
@@ -71,16 +79,20 @@ func (c CostFlags) Less(other CostFlags) bool {
 	if c.FullScanPenalty != other.FullScanPenalty {
 		return !c.FullScanPenalty
 	}
+	if c.UnboundedCardinality != other.UnboundedCardinality {
+		return !c.UnboundedCardinality
+	}
 	return false
 }
 
 // Add adds the other flags to these flags.
 func (c *CostFlags) Add(other CostFlags) {
 	c.FullScanPenalty = c.FullScanPenalty || other.FullScanPenalty
 	c.HugeCostPenalty = c.HugeCostPenalty || other.HugeCostPenalty
+	c.UnboundedCardinality = c.UnboundedCardinality || other.UnboundedCardinality
 }
 
 // Empty returns true if these flags are empty.
 func (c CostFlags) Empty() bool {
-	return !c.FullScanPenalty && !c.HugeCostPenalty
+	return !c.FullScanPenalty && !c.HugeCostPenalty && !c.UnboundedCardinality
 }
diff --git a/pkg/sql/opt/memo/cost_test.go b/pkg/sql/opt/memo/cost_test.go
@@ -35,6 +35,8 @@ func TestCostLess(t *testing.T) {
 		{memo.MaxCost, memo.MaxCost, false},
 		{memo.MaxCost, memo.Cost{C: 1.0, Flags: memo.CostFlags{FullScanPenalty: true}}, false},
 		{memo.Cost{C: 1.0, Flags: memo.CostFlags{HugeCostPenalty: true}}, memo.MaxCost, true},
+		{memo.Cost{C: 2.0, Flags: memo.CostFlags{}}, memo.Cost{C: 1.0, Flags: memo.CostFlags{UnboundedCardinality: true}}, true},
+		{memo.Cost{C: 1.0, Flags: memo.CostFlags{UnboundedCardinality: true}}, memo.Cost{C: 2.0, Flags: memo.CostFlags{}}, false},
 	}
 	for _, tc := range testCases {
 		if tc.left.Less(tc.right) != tc.expected {
@@ -72,6 +74,8 @@ func TestCostFlagsLess(t *testing.T) {
 		{memo.CostFlags{FullScanPenalty: true, HugeCostPenalty: true}, memo.CostFlags{FullScanPenalty: true, HugeCostPenalty: true}, false},
 		{memo.CostFlags{FullScanPenalty: false}, memo.CostFlags{FullScanPenalty: true}, true},
 		{memo.CostFlags{HugeCostPenalty: false}, memo.CostFlags{HugeCostPenalty: true}, true},
+		{memo.CostFlags{UnboundedCardinality: false}, memo.CostFlags{UnboundedCardinality: true}, true},
+		{memo.CostFlags{UnboundedCardinality: true}, memo.CostFlags{UnboundedCardinality: false}, false},
 	}
 	for _, tc := range testCases {
 		if tc.left.Less(tc.right) != tc.expected {

diff --git a/pkg/sql/opt/memo/expr_format.go b/pkg/sql/opt/memo/expr_format.go
@@ -861,6 +861,9 @@ func (f *ExprFmtCtx) formatRelational(e RelExpr, tp treeprinter.Node) {
 			if cost.Flags.HugeCostPenalty {
 				b.WriteString(" huge-cost-penalty")
 			}
+			if cost.Flags.UnboundedCardinality {
+				b.WriteString(" unbounded-cardinality")
+			}
 			tp.Child(b.String())
 		}
 	}

diff --git a/pkg/sql/opt/memo/memo.go b/pkg/sql/opt/memo/memo.go
@@ -196,6 +196,7 @@ type Memo struct {
 	usePolymorphicParameterFix                 bool
 	pushLimitIntoProjectFilteredScan           bool
 	legacyVarcharTyping                        bool
+	preferBoundedCardinality                   bool
 
 	// txnIsoLevel is the isolation level under which the plan was created. This
 	// affects the planning of some locking operations, so it must be included in
@@ -283,6 +284,7 @@ func (m *Memo) Init(ctx context.Context, evalCtx *eval.Context) {
 		usePolymorphicParameterFix:                 evalCtx.SessionData().OptimizerUsePolymorphicParameterFix,
 		pushLimitIntoProjectFilteredScan:           evalCtx.SessionData().OptimizerPushLimitIntoProjectFilteredScan,
 		legacyVarcharTyping:                        evalCtx.SessionData().LegacyVarcharTyping,
+		preferBoundedCardinality:                   evalCtx.SessionData().OptimizerPreferBoundedCardinality,
 		txnIsoLevel:                                evalCtx.TxnIsoLevel,
 	}
 	m.metadata.Init()
@@ -448,6 +450,7 @@ func (m *Memo) IsStale(
 		m.usePolymorphicParameterFix != evalCtx.SessionData().OptimizerUsePolymorphicParameterFix ||
 		m.pushLimitIntoProjectFilteredScan != evalCtx.SessionData().OptimizerPushLimitIntoProjectFilteredScan ||
 		m.legacyVarcharTyping != evalCtx.SessionData().LegacyVarcharTyping ||
+		m.preferBoundedCardinality != evalCtx.SessionData().OptimizerPreferBoundedCardinality ||
 		m.txnIsoLevel != evalCtx.TxnIsoLevel {
 		return true, nil
 	}

diff --git a/pkg/sql/opt/memo/memo_test.go b/pkg/sql/opt/memo/memo_test.go
@@ -507,6 +507,11 @@ func TestMemoIsStale(t *testing.T) {
 	evalCtx.SessionData().LegacyVarcharTyping = false
 	notStale()
 
+	evalCtx.SessionData().OptimizerPreferBoundedCardinality = true
+	stale()
+	evalCtx.SessionData().OptimizerPreferBoundedCardinality = false
+	notStale()
+
 	// User no longer has access to view.
 	catalog.View(tree.NewTableNameWithSchema("t", catconstants.PublicSchemaName, "abcview")).Revoked = true
 	_, err = o.Memo().IsStale(ctx, &evalCtx, catalog)

diff --git a/pkg/sql/opt/xform/coster.go b/pkg/sql/opt/xform/coster.go
@@ -619,8 +619,13 @@ func (c *coster) ComputeCost(candidate memo.RelExpr, required *physical.Required
 	// Add a one-time cost for any operator with unbounded cardinality. This
 	// ensures we prefer plans that push limits as far down the tree as possible,
 	// all else being equal.
+	//
+	// Also add a cost flag for unbounded cardinality.
 	if candidate.Relational().Cardinality.IsUnbounded() {
 		cost.C += cpuCostFactor
+		if c.evalCtx.SessionData().OptimizerPreferBoundedCardinality {
+			cost.Flags.UnboundedCardinality = true
+		}
 	}
 
 	if !cost.Less(memo.MaxCost) {