opt: add session setting for join elimination optimization

We recently added support for column remapping in the join elimination
rules that allows columns from the eliminated input of the join to be
mapped to equivalent columns from the preserved input. This allows
joins to be eliminated in more cases - in particular, the self-join
patterns that can arise from an `UPDATE ... FROM` statement where the
table in the `FROM` clause is the same as the table being updated.

This patch adds a setting, `optimizer_use_improved_join_elimination`,
which gates the column-remapping logic for the join-elimination rules.
The plan is to backport the column-remapping changes to 23.1 behind
this setting turned off by default.

Informs cockroachdb#102614

Release note: None

pkg/sql/exec_util.go

@@ -3608,6 +3608,10 @@ func (m *sessionDataMutator) SetReplicationMode(val sessiondatapb.ReplicationMod
 	m.data.ReplicationMode = val
 }
 
+func (m *sessionDataMutator) SetOptimizerUseImprovedJoinElimination(val bool) {
+	m.data.OptimizerUseImprovedJoinElimination = val
+}
+
 // Utility functions related to scrubbing sensitive information on SQL Stats.
 
 // quantizeCounts ensures that the Count field in the

pkg/sql/logictest/testdata/logic_test/information_schema

@@ -5350,6 +5350,7 @@ optimizer_use_forecasts                                    on
 optimizer_use_histograms                                   on
 optimizer_use_improved_computed_column_filters_derivation  on
 optimizer_use_improved_disjunction_stats                   on
+optimizer_use_improved_join_elimination                    on
 optimizer_use_improved_split_disjunction_for_joins         on
 optimizer_use_limit_ordering_for_streaming_group_by        on
 optimizer_use_multicol_stats                               on

pkg/sql/logictest/testdata/logic_test/pg_catalog

@@ -2785,6 +2785,7 @@ optimizer_use_forecasts                                    on                  N
 optimizer_use_histograms                                   on                  NULL      NULL        NULL        string
 optimizer_use_improved_computed_column_filters_derivation  on                  NULL      NULL        NULL        string
 optimizer_use_improved_disjunction_stats                   on                  NULL      NULL        NULL        string
+optimizer_use_improved_join_elimination                    on                  NULL      NULL        NULL        string
 optimizer_use_improved_split_disjunction_for_joins         on                  NULL      NULL        NULL        string
 optimizer_use_limit_ordering_for_streaming_group_by        on                  NULL      NULL        NULL        string
 optimizer_use_multicol_stats                               on                  NULL      NULL        NULL        string
@@ -2944,6 +2945,7 @@ optimizer_use_forecasts                                    on                  N
 optimizer_use_histograms                                   on                  NULL  user     NULL      on                  on
 optimizer_use_improved_computed_column_filters_derivation  on                  NULL  user     NULL      on                  on
 optimizer_use_improved_disjunction_stats                   on                  NULL  user     NULL      on                  on
+optimizer_use_improved_join_elimination                    on                  NULL  user     NULL      on                  on
 optimizer_use_improved_split_disjunction_for_joins         on                  NULL  user     NULL      on                  on
 optimizer_use_limit_ordering_for_streaming_group_by        on                  NULL  user     NULL      on                  on
 optimizer_use_multicol_stats                               on                  NULL  user     NULL      on                  on
@@ -3102,6 +3104,7 @@ optimizer_use_forecasts                                    NULL    NULL     NULL
 optimizer_use_histograms                                   NULL    NULL     NULL     NULL        NULL
 optimizer_use_improved_computed_column_filters_derivation  NULL    NULL     NULL     NULL        NULL
 optimizer_use_improved_disjunction_stats                   NULL    NULL     NULL     NULL        NULL
+optimizer_use_improved_join_elimination                    NULL    NULL     NULL     NULL        NULL
 optimizer_use_improved_split_disjunction_for_joins         NULL    NULL     NULL     NULL        NULL
 optimizer_use_limit_ordering_for_streaming_group_by        NULL    NULL     NULL     NULL        NULL
 optimizer_use_multicol_stats                               NULL    NULL     NULL     NULL        NULL

pkg/sql/logictest/testdata/logic_test/show_source

@@ -120,6 +120,7 @@ optimizer_use_forecasts                                    on
 optimizer_use_histograms                                   on
 optimizer_use_improved_computed_column_filters_derivation  on
 optimizer_use_improved_disjunction_stats                   on
+optimizer_use_improved_join_elimination                    on
 optimizer_use_improved_split_disjunction_for_joins         on
 optimizer_use_limit_ordering_for_streaming_group_by        on
 optimizer_use_multicol_stats                               on

pkg/sql/opt/memo/logical_props_builder.go

@@ -2258,6 +2258,7 @@ func addOuterColsToFuncDep(outerCols opt.ColSet, fdset *props.FuncDepSet) {
 // joins that are used internally when deriving logical properties and
 // statistics.
 type joinPropsHelper struct {
+	evalCtx  *eval.Context
 	join     RelExpr
 	joinType opt.Operator
 
@@ -2276,7 +2277,7 @@ type joinPropsHelper struct {
 func (h *joinPropsHelper) init(b *logicalPropsBuilder, joinExpr RelExpr) {
 	// This initialization pattern ensures that fields are not unwittingly
 	// reused. Field reuse must be explicit.
-	*h = joinPropsHelper{join: joinExpr}
+	*h = joinPropsHelper{evalCtx: b.evalCtx, join: joinExpr}
 
 	switch join := joinExpr.(type) {
 	case *LookupJoinExpr:
@@ -2514,7 +2515,9 @@ func (h *joinPropsHelper) setFuncDeps(rel *props.Relational) {
 		// created new possibilities for simplifying removed columns.
 		rel.FuncDeps.ProjectCols(rel.OutputCols)
 	}
-	h.addSelfJoinImpliedFDs(rel)
+	if h.evalCtx.SessionData().OptimizerUseImprovedJoinElimination {
+		h.addSelfJoinImpliedFDs(rel)
+	}
 }
 
 // addSelfJoinImpliedFDs adds any extra equality FDs that are implied by a self

pkg/sql/opt/memo/memo.go

@@ -167,6 +167,7 @@ type Memo struct {
 	alwaysUseHistograms                        bool
 	hoistUncorrelatedEqualitySubqueries        bool
 	useImprovedComputedColumnFiltersDerivation bool
+	useImprovedJoinElimination                 bool
 
 	// curRank is the highest currently in-use scalar expression rank.
 	curRank opt.ScalarRank
@@ -228,6 +229,7 @@ func (m *Memo) Init(ctx context.Context, evalCtx *eval.Context) {
 		alwaysUseHistograms:                        evalCtx.SessionData().OptimizerAlwaysUseHistograms,
 		hoistUncorrelatedEqualitySubqueries:        evalCtx.SessionData().OptimizerHoistUncorrelatedEqualitySubqueries,
 		useImprovedComputedColumnFiltersDerivation: evalCtx.SessionData().OptimizerUseImprovedComputedColumnFiltersDerivation,
+		useImprovedJoinElimination:                 evalCtx.SessionData().OptimizerUseImprovedJoinElimination,
 	}
 	m.metadata.Init()
 	m.logPropsBuilder.init(ctx, evalCtx, m)
@@ -372,7 +374,8 @@ func (m *Memo) IsStale(
 		m.useImprovedSplitDisjunctionForJoins != evalCtx.SessionData().OptimizerUseImprovedSplitDisjunctionForJoins ||
 		m.alwaysUseHistograms != evalCtx.SessionData().OptimizerAlwaysUseHistograms ||
 		m.hoistUncorrelatedEqualitySubqueries != evalCtx.SessionData().OptimizerHoistUncorrelatedEqualitySubqueries ||
-		m.useImprovedComputedColumnFiltersDerivation != evalCtx.SessionData().OptimizerUseImprovedComputedColumnFiltersDerivation {
+		m.useImprovedComputedColumnFiltersDerivation != evalCtx.SessionData().OptimizerUseImprovedComputedColumnFiltersDerivation ||
+		m.useImprovedJoinElimination != evalCtx.SessionData().OptimizerUseImprovedJoinElimination {
 		return true, nil
 	}
 

pkg/sql/opt/memo/memo_test.go

@@ -373,6 +373,12 @@ func TestMemoIsStale(t *testing.T) {
 	evalCtx.SessionData().OptimizerUseImprovedComputedColumnFiltersDerivation = false
 	notStale()
 
+	// Stale optimizer_use_improved_join_elimination.
+	evalCtx.SessionData().OptimizerUseImprovedJoinElimination = true
+	stale()
+	evalCtx.SessionData().OptimizerUseImprovedJoinElimination = 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)

pkg/sql/opt/norm/project_funcs.go

@@ -908,3 +908,10 @@ func (c *CustomFuncs) RemapProjectionCols(
 	}
 	return *(replace(&projections).(*memo.ProjectionsExpr))
 }
+
+// CanUseImprovedJoinElimination returns true if either no column remapping is
+// required in order to eliminate the join, or column remapping is enabled by
+// OptimizerUseImprovedJoinElimination.
+func (c *CustomFuncs) CanUseImprovedJoinElimination(from, to opt.ColSet) bool {
+	return c.f.evalCtx.SessionData().OptimizerUseImprovedJoinElimination || from.SubsetOf(to)
+}

pkg/sql/opt/norm/rules/groupby.opt

@@ -79,6 +79,7 @@
             $leftCols
             $fds:(FuncDeps $input)
         ) &
+        (CanUseImprovedJoinElimination $toRemap $leftCols) &
         (CanEliminateJoinUnderGroupByLeft $input $aggs)
 )
 =>
@@ -114,6 +115,7 @@
             $rightCols
             $fds:(FuncDeps $input)
         ) &
+        (CanUseImprovedJoinElimination $toRemap $rightCols) &
         (CanEliminateJoinUnderGroupByRight $input $aggs)
 )
 =>

pkg/sql/opt/norm/rules/project.opt

@@ -27,13 +27,14 @@
     $projections:*
     $passthrough:* &
         (CanRemapCols
-            (UnionCols
+            $fromCols:(UnionCols
                 $passthrough
                 (ProjectionOuterCols $projections)
             )
             $leftCols:(OutputCols $left)
             $fds:(FuncDeps $join)
-        )
+        ) &
+        (CanUseImprovedJoinElimination $fromCols $leftCols)
 )
 =>
 (Project
@@ -56,13 +57,14 @@
     $projections:*
     $passthrough:* &
         (CanRemapCols
-            (UnionCols
+            $fromCols:(UnionCols
                 $passthrough
                 (ProjectionOuterCols $projections)
             )
             $rightCols:(OutputCols $right)
             $fds:(FuncDeps $join)
-        )
+        ) &
+        (CanUseImprovedJoinElimination $fromCols $rightCols)
 )
 =>
 (Project

pkg/sql/opt/norm/testdata/rules/project

@@ -266,6 +266,40 @@ project
  └── projections
       └── child.parent_id:2 [as=parent_id:9, outer=(2)]
 
+# The join can be eliminated as long as it doesn't require remapping with
+# OptimizerUseImprovedJoinElimination disabled.
+norm set=optimizer_use_improved_join_elimination=false expect=EliminateJoinUnderProjectLeft
+SELECT b.x, b.z FROM b LEFT JOIN a ON b.x = a.x
+----
+scan b
+ ├── columns: x:1!null z:2
+ ├── key: (1)
+ └── fd: (1)-->(2)
+
+# The join cannot be eliminated with OptimizerUseImprovedJoinElimination
+# disabled.
+norm set=optimizer_use_improved_join_elimination=false expect-not=EliminateJoinUnderProjectLeft
+SELECT b.j, b1.x FROM b INNER JOIN b AS b1 ON b.x = b1.x
+----
+project
+ ├── columns: j:3 x:6!null
+ ├── key: (6)
+ ├── fd: (6)-->(3)
+ └── inner-join (hash)
+      ├── columns: b.x:1!null b.j:3 b1.x:6!null
+      ├── multiplicity: left-rows(exactly-one), right-rows(exactly-one)
+      ├── key: (6)
+      ├── fd: (1)-->(3), (1)==(6), (6)==(1)
+      ├── scan b
+      │    ├── columns: b.x:1!null b.j:3
+      │    ├── key: (1)
+      │    └── fd: (1)-->(3)
+      ├── scan b [as=b1]
+      │    ├── columns: b1.x:6!null
+      │    └── key: (6)
+      └── filters
+           └── b.x:1 = b1.x:6 [outer=(1,6), constraints=(/1: (/NULL - ]; /6: (/NULL - ]), fd=(1)==(6), (6)==(1)]
+
 # --------------------------------------------------
 # EliminateJoinUnderProjectRight
 # --------------------------------------------------

pkg/sql/opt/testutils/opttester/opt_tester.go

@@ -301,6 +301,7 @@ func New(catalog cat.Catalog, sql string) *OptTester {
 	ot.evalCtx.SessionData().OptimizerAlwaysUseHistograms = true
 	ot.evalCtx.SessionData().OptimizerHoistUncorrelatedEqualitySubqueries = true
 	ot.evalCtx.SessionData().OptimizerUseImprovedComputedColumnFiltersDerivation = true
+	ot.evalCtx.SessionData().OptimizerUseImprovedJoinElimination = true
 
 	return ot
 }

pkg/sql/sessiondatapb/local_only_session_data.proto

@@ -405,6 +405,10 @@ message LocalOnlySessionData {
   // ReplicationMode represents the replication parameter passed in during
   // connection time.
   ReplicationMode replication_mode = 106;
+  // OptimizerUseImprovedJoinElimination, when true, allows the optimizer to
+  // eliminate joins in more cases by remapping columns from the eliminated
+  // input of the join to equivalent columns from the preserved input.
+  bool optimizer_use_improved_join_elimination = 107;
 
   ///////////////////////////////////////////////////////////////////////////
   // WARNING: consider whether a session parameter you're adding needs to  //

pkg/sql/vars.go

@@ -2800,6 +2800,23 @@ var varGen = map[string]sessionVar{
 			return strconv.FormatInt(maxConn, 10), nil
 		},
 	},
+
+	// CockroachDB extension.
+	`optimizer_use_improved_join_elimination`: {
+		GetStringVal: makePostgresBoolGetStringValFn(`optimizer_use_improved_join_elimination`),
+		Set: func(_ context.Context, m sessionDataMutator, s string) error {
+			b, err := paramparse.ParseBoolVar("optimizer_use_improved_join_elimination", s)
+			if err != nil {
+				return err
+			}
+			m.SetOptimizerUseImprovedJoinElimination(b)
+			return nil
+		},
+		Get: func(evalCtx *extendedEvalContext, _ *kv.Txn) (string, error) {
+			return formatBoolAsPostgresSetting(evalCtx.SessionData().OptimizerUseImprovedJoinElimination), nil
+		},
+		GlobalDefault: globalTrue,
+	},
 }
 
 func ReplicationModeFromString(s string) (sessiondatapb.ReplicationMode, error) {