Improve INTERSECT/EXCEPT hashing by realizing that we don't need to make any

hashtable entries for tuples that are found only in the second input: they
can never contribute to the output.  Furthermore, this implies that the
planner should endeavor to put first the smaller (in number of groups) input
relation for an INTERSECT.  Implement that, and upgrade prepunion's estimation
of the number of rows returned by setops so that there's some amount of sanity
in the estimate of which one is smaller.

3 files changed, 181 insertions, 99 deletions

diff --git a/src/backend/optimizer/plan/createplan.c b/src/backend/optimizer/plan/createplan.c
index 4cf2a5208b1..0c200e05795 100644
--- a/src/backend/optimizer/plan/createplan.c
+++ b/src/backend/optimizer/plan/createplan.c
@@ -10,7 +10,7 @@
  *
  *
  * IDENTIFICATION
- *	  $PostgreSQL: pgsql/src/backend/optimizer/plan/createplan.c,v 1.243 2008/08/07 03:04:03 tgl Exp $
+ *	  $PostgreSQL: pgsql/src/backend/optimizer/plan/createplan.c,v 1.244 2008/08/07 19:35:02 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -3109,8 +3109,8 @@ make_unique(Plan *lefttree, List *distinctList)
  */
 SetOp *
 make_setop(SetOpCmd cmd, SetOpStrategy strategy, Plan *lefttree,
-		   List *distinctList, AttrNumber flagColIdx, long numGroups,
-		   double outputRows)
+		   List *distinctList, AttrNumber flagColIdx, int firstFlag,
+		   long numGroups, double outputRows)
 {
 	SetOp	   *node = makeNode(SetOp);
 	Plan	   *plan = &node->plan;
@@ -3159,6 +3159,7 @@ make_setop(SetOpCmd cmd, SetOpStrategy strategy, Plan *lefttree,
 	node->dupColIdx = dupColIdx;
 	node->dupOperators = dupOperators;
 	node->flagColIdx = flagColIdx;
+	node->firstFlag = firstFlag;
 	node->numGroups = numGroups;
 
 	return node;
diff --git a/src/backend/optimizer/prep/prepunion.c b/src/backend/optimizer/prep/prepunion.c
index 3a155c0d0ad..dda41770c8d 100644
--- a/src/backend/optimizer/prep/prepunion.c
+++ b/src/backend/optimizer/prep/prepunion.c
@@ -22,7 +22,7 @@
  *
  *
  * IDENTIFICATION
- *	  $PostgreSQL: pgsql/src/backend/optimizer/prep/prepunion.c,v 1.151 2008/08/07 03:04:03 tgl Exp $
+ *	  $PostgreSQL: pgsql/src/backend/optimizer/prep/prepunion.c,v 1.152 2008/08/07 19:35:02 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -49,19 +49,22 @@
 #include "parser/parsetree.h"
 #include "utils/lsyscache.h"
 #include "utils/rel.h"
+#include "utils/selfuncs.h"
 
 
 static Plan *recurse_set_operations(Node *setOp, PlannerInfo *root,
 					   double tuple_fraction,
 					   List *colTypes, bool junkOK,
 					   int flag, List *refnames_tlist,
-					   List **sortClauses);
+					   List **sortClauses, double *pNumGroups);
 static Plan *generate_union_plan(SetOperationStmt *op, PlannerInfo *root,
 					double tuple_fraction,
-					List *refnames_tlist, List **sortClauses);
+					List *refnames_tlist,
+					List **sortClauses, double *pNumGroups);
 static Plan *generate_nonunion_plan(SetOperationStmt *op, PlannerInfo *root,
 					   double tuple_fraction,
-					   List *refnames_tlist, List **sortClauses);
+					   List *refnames_tlist,
+					   List **sortClauses, double *pNumGroups);
 static List *recurse_union_children(Node *setOp, PlannerInfo *root,
 					   double tuple_fraction,
 					   SetOperationStmt *top_union,
@@ -71,7 +74,8 @@ static Plan *make_union_unique(SetOperationStmt *op, Plan *plan,
 				  List **sortClauses);
 static bool choose_hashed_setop(PlannerInfo *root, List *groupClauses,
 					Plan *input_plan,
-					double tuple_fraction, double dNumDistinctRows,
+					double dNumGroups, double dNumOutputRows,
+					double tuple_fraction,
 					const char *construct);
 static List *generate_setop_tlist(List *colTypes, int flag,
 					 Index varno,
@@ -153,7 +157,7 @@ plan_set_operations(PlannerInfo *root, double tuple_fraction,
 	return recurse_set_operations((Node *) topop, root, tuple_fraction,
 								  topop->colTypes, true, -1,
 								  leftmostQuery->targetList,
-								  sortClauses);
+								  sortClauses, NULL);
 }
 
 /*
@@ -165,7 +169,11 @@ plan_set_operations(PlannerInfo *root, double tuple_fraction,
  * junkOK: if true, child resjunk columns may be left in the result
  * flag: if >= 0, add a resjunk output column indicating value of flag
  * refnames_tlist: targetlist to take column names from
+ *
+ * Returns a plan for the subtree, as well as these output parameters:
  * *sortClauses: receives list of SortGroupClauses for result plan, if any
+ * *pNumGroups: if not NULL, we estimate the number of distinct groups
+ *		in the result, and store it there
  *
  * We don't have to care about typmods here: the only allowed difference
  * between set-op input and output typmods is input is a specific typmod
@@ -176,7 +184,7 @@ recurse_set_operations(Node *setOp, PlannerInfo *root,
 					   double tuple_fraction,
 					   List *colTypes, bool junkOK,
 					   int flag, List *refnames_tlist,
-					   List **sortClauses)
+					   List **sortClauses, double *pNumGroups)
 {
 	if (IsA(setOp, RangeTblRef))
 	{
@@ -198,6 +206,22 @@ recurse_set_operations(Node *setOp, PlannerInfo *root,
 								   &subroot);
 
 		/*
+		 * Estimate number of groups if caller wants it.  If the subquery
+		 * used grouping or aggregation, its output is probably mostly
+		 * unique anyway; otherwise do statistical estimation.
+		 */
+		if (pNumGroups)
+		{
+			if (subquery->groupClause || subquery->distinctClause ||
+				subroot->hasHavingQual || subquery->hasAggs)
+				*pNumGroups = subplan->plan_rows;
+			else
+				*pNumGroups = estimate_num_groups(subroot,
+												  get_tlist_exprs(subquery->targetList, false),
+												  subplan->plan_rows);
+		}
+
+		/*
 		 * Add a SubqueryScan with the caller-requested targetlist
 		 */
 		plan = (Plan *)
@@ -228,11 +252,11 @@ recurse_set_operations(Node *setOp, PlannerInfo *root,
 		if (op->op == SETOP_UNION)
 			plan = generate_union_plan(op, root, tuple_fraction,
 									   refnames_tlist,
-									   sortClauses);
+									   sortClauses, pNumGroups);
 		else
 			plan = generate_nonunion_plan(op, root, tuple_fraction,
 										  refnames_tlist,
-										  sortClauses);
+										  sortClauses, pNumGroups);
 
 		/*
 		 * If necessary, add a Result node to project the caller-requested
@@ -277,7 +301,7 @@ static Plan *
 generate_union_plan(SetOperationStmt *op, PlannerInfo *root,
 					double tuple_fraction,
 					List *refnames_tlist,
-					List **sortClauses)
+					List **sortClauses, double *pNumGroups)
 {
 	List	   *planlist;
 	List	   *tlist;
@@ -334,6 +358,14 @@ generate_union_plan(SetOperationStmt *op, PlannerInfo *root,
 	else
 		plan = make_union_unique(op, plan, root, tuple_fraction, sortClauses);
 
+	/*
+	 * Estimate number of groups if caller wants it.  For now we just
+	 * assume the output is unique --- this is certainly true for the
+	 * UNION case, and we want worst-case estimates anyway.
+	 */
+	if (pNumGroups)
+		*pNumGroups = plan->plan_rows;
+
 	return plan;
 }
 
@@ -344,7 +376,7 @@ static Plan *
 generate_nonunion_plan(SetOperationStmt *op, PlannerInfo *root,
 					   double tuple_fraction,
 					   List *refnames_tlist,
-					   List **sortClauses)
+					   List **sortClauses, double *pNumGroups)
 {
 	Plan	   *lplan,
 			   *rplan,
@@ -353,24 +385,43 @@ generate_nonunion_plan(SetOperationStmt *op, PlannerInfo *root,
 			   *groupList,
 			   *planlist,
 			   *child_sortclauses;
-	double		dNumDistinctRows;
-	double		dNumOutputRows;
-	long		numDistinctRows;
+	double		dLeftGroups,
+				dRightGroups,
+				dNumGroups,
+				dNumOutputRows;
+	long		numGroups;
 	bool		use_hash;
 	SetOpCmd	cmd;
+	int			firstFlag;
 
 	/* Recurse on children, ensuring their outputs are marked */
 	lplan = recurse_set_operations(op->larg, root,
 								   0.0 /* all tuples needed */ ,
 								   op->colTypes, false, 0,
 								   refnames_tlist,
-								   &child_sortclauses);
+								   &child_sortclauses, &dLeftGroups);
 	rplan = recurse_set_operations(op->rarg, root,
 								   0.0 /* all tuples needed */ ,
 								   op->colTypes, false, 1,
 								   refnames_tlist,
-								   &child_sortclauses);
-	planlist = list_make2(lplan, rplan);
+								   &child_sortclauses, &dRightGroups);
+
+	/*
+	 * For EXCEPT, we must put the left input first.  For INTERSECT, either
+	 * order should give the same results, and we prefer to put the smaller
+	 * input first in order to minimize the size of the hash table in the
+	 * hashing case.  "Smaller" means the one with the fewer groups.
+	 */
+	if (op->op == SETOP_EXCEPT || dLeftGroups <= dRightGroups)
+	{
+		planlist = list_make2(lplan, rplan);
+		firstFlag = 0;
+	}
+	else
+	{
+		planlist = list_make2(rplan, lplan);
+		firstFlag = 1;
+	}
 
 	/*
 	 * Generate tlist for Append plan node.
@@ -400,27 +451,32 @@ generate_nonunion_plan(SetOperationStmt *op, PlannerInfo *root,
 	}
 
 	/*
-	 * XXX for the moment, take the number of distinct groups as being the
-	 * total input size, ie, the worst case.  This is too conservative, but
-	 * we don't want to risk having the hashtable overrun memory; also,
-	 * it's not clear how to get a decent estimate of the true size.
+	 * Estimate number of distinct groups that we'll need hashtable entries
+	 * for; this is the size of the left-hand input for EXCEPT, or the smaller
+	 * input for INTERSECT.  Also estimate the number of eventual output rows.
+	 * In non-ALL cases, we estimate each group produces one output row;
+	 * in ALL cases use the relevant relation size.  These are worst-case
+	 * estimates, of course, but we need to be conservative.
 	 */
-	dNumDistinctRows = plan->plan_rows;
+	if (op->op == SETOP_EXCEPT)
+	{
+		dNumGroups = dLeftGroups;
+		dNumOutputRows = op->all ? lplan->plan_rows : dNumGroups;
+	}
+	else
+	{
+		dNumGroups = Min(dLeftGroups, dRightGroups);
+		dNumOutputRows = op->all ? Min(lplan->plan_rows, rplan->plan_rows) : dNumGroups;
+	}
 
 	/* Also convert to long int --- but 'ware overflow! */
-	numDistinctRows = (long) Min(dNumDistinctRows, (double) LONG_MAX);
-
-	/*
-	 * The output size is taken as 10% of that, which is a completely bogus
-	 * guess, but it's what we've used historically.
-	 */
-	dNumOutputRows = ceil(dNumDistinctRows * 0.1);
+	numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 
 	/*
 	 * Decide whether to hash or sort, and add a sort node if needed.
 	 */
 	use_hash = choose_hashed_setop(root, groupList, plan,
-								   tuple_fraction, dNumDistinctRows,
+								   dNumGroups, dNumOutputRows, tuple_fraction,
 								   (op->op == SETOP_INTERSECT) ? "INTERSECT" : "EXCEPT");
 
 	if (!use_hash)
@@ -443,12 +499,17 @@ generate_nonunion_plan(SetOperationStmt *op, PlannerInfo *root,
 			break;
 	}
 	plan = (Plan *) make_setop(cmd, use_hash ? SETOP_HASHED : SETOP_SORTED,
-							   plan, groupList, list_length(op->colTypes) + 1,
-							   numDistinctRows, dNumOutputRows);
+							   plan, groupList,
+							   list_length(op->colTypes) + 1,
+							   use_hash ? firstFlag : -1,
+							   numGroups, dNumOutputRows);
 
 	/* Result is sorted only if we're not hashing */
 	*sortClauses = use_hash ? NIL : groupList;
 
+	if (pNumGroups)
+		*pNumGroups = dNumGroups;
+
 	return plan;
 }
 
@@ -499,7 +560,7 @@ recurse_union_children(Node *setOp, PlannerInfo *root,
 											 tuple_fraction,
 											 top_union->colTypes, false,
 											 -1, refnames_tlist,
-											 &child_sortclauses));
+											 &child_sortclauses, NULL));
 }
 
 /*
@@ -511,8 +572,8 @@ make_union_unique(SetOperationStmt *op, Plan *plan,
 				  List **sortClauses)
 {
 	List	   *groupList;
-	double		dNumDistinctRows;
-	long		numDistinctRows;
+	double		dNumGroups;
+	long		numGroups;
 
 	/* Identify the grouping semantics */
 	groupList = generate_setop_grouplist(op, plan->targetlist);
@@ -525,21 +586,21 @@ make_union_unique(SetOperationStmt *op, Plan *plan,
 	}
 
 	/*
-	 * XXX for the moment, take the number of distinct groups as being the
-	 * total input size, ie, the worst case.  This is too conservative, but
-	 * we don't want to risk having the hashtable overrun memory; also,
+	 * XXX for the moment, take the number of distinct groups as equal to
+	 * the total input size, ie, the worst case.  This is too conservative,
+	 * but we don't want to risk having the hashtable overrun memory; also,
 	 * it's not clear how to get a decent estimate of the true size.  One
 	 * should note as well the propensity of novices to write UNION rather
 	 * than UNION ALL even when they don't expect any duplicates...
 	 */
-	dNumDistinctRows = plan->plan_rows;
+	dNumGroups = plan->plan_rows;
 
 	/* Also convert to long int --- but 'ware overflow! */
-	numDistinctRows = (long) Min(dNumDistinctRows, (double) LONG_MAX);
+	numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 
 	/* Decide whether to hash or sort */
 	if (choose_hashed_setop(root, groupList, plan,
-							tuple_fraction, dNumDistinctRows,
+							dNumGroups, dNumGroups, tuple_fraction,
 							"UNION"))
 	{
 		/* Hashed aggregate plan --- no sort needed */
@@ -551,7 +612,7 @@ make_union_unique(SetOperationStmt *op, Plan *plan,
 								 extract_grouping_cols(groupList,
 													   plan->targetlist),
 								 extract_grouping_ops(groupList),
-								 numDistinctRows,
+								 numGroups,
 								 0,
 								 plan);
 		/* Hashed aggregation produces randomly-ordered results */
@@ -562,7 +623,7 @@ make_union_unique(SetOperationStmt *op, Plan *plan,
 		/* Sort and Unique */
 		plan = (Plan *) make_sort_from_sortclauses(root, groupList, plan);
 		plan = (Plan *) make_unique(plan, groupList);
-		plan->plan_rows = dNumDistinctRows;
+		plan->plan_rows = dNumGroups;
 		/* We know the sort order of the result */
 		*sortClauses = groupList;
 	}
@@ -576,14 +637,14 @@ make_union_unique(SetOperationStmt *op, Plan *plan,
 static bool
 choose_hashed_setop(PlannerInfo *root, List *groupClauses,
 					Plan *input_plan,
-					double tuple_fraction, double dNumDistinctRows,
+					double dNumGroups, double dNumOutputRows,
+					double tuple_fraction,
 					const char *construct)
 {
-	int			numDistinctCols = list_length(groupClauses);
+	int			numGroupCols = list_length(groupClauses);
 	bool		can_sort;
 	bool		can_hash;
 	Size		hashentrysize;
-	List	   *needed_pathkeys;
 	Path		hashed_p;
 	Path		sorted_p;
 
@@ -615,7 +676,7 @@ choose_hashed_setop(PlannerInfo *root, List *groupClauses,
 	 */
 	hashentrysize = MAXALIGN(input_plan->plan_width) + MAXALIGN(sizeof(MinimalTupleData));
 
-	if (hashentrysize * dNumDistinctRows > work_mem * 1024L)
+	if (hashentrysize * dNumGroups > work_mem * 1024L)
 		return false;
 
 	/*
@@ -630,7 +691,7 @@ choose_hashed_setop(PlannerInfo *root, List *groupClauses,
 	 * make actual Paths for these steps.
 	 */
 	cost_agg(&hashed_p, root, AGG_HASHED, 0,
-			 numDistinctCols, dNumDistinctRows,
+			 numGroupCols, dNumGroups,
 			 input_plan->startup_cost, input_plan->total_cost,
 			 input_plan->plan_rows);
 
@@ -640,14 +701,10 @@ choose_hashed_setop(PlannerInfo *root, List *groupClauses,
 	 */
 	sorted_p.startup_cost = input_plan->startup_cost;
 	sorted_p.total_cost = input_plan->total_cost;
-	/* XXX this is more expensive than cost_sort really needs: */
-	needed_pathkeys = make_pathkeys_for_sortclauses(root,
-													groupClauses,
-													input_plan->targetlist,
-													true);
-	cost_sort(&sorted_p, root, needed_pathkeys, sorted_p.total_cost,
+	/* XXX cost_sort doesn't actually look at pathkeys, so just pass NIL */
+	cost_sort(&sorted_p, root, NIL, sorted_p.total_cost,
 			  input_plan->plan_rows, input_plan->plan_width, -1.0);
-	cost_group(&sorted_p, root, numDistinctCols, dNumDistinctRows,
+	cost_group(&sorted_p, root, numGroupCols, dNumGroups,
 			   sorted_p.startup_cost, sorted_p.total_cost,
 			   input_plan->plan_rows);
 
@@ -656,7 +713,7 @@ choose_hashed_setop(PlannerInfo *root, List *groupClauses,
 	 * have to convert an absolute count (LIMIT) into fractional form.
 	 */
 	if (tuple_fraction >= 1.0)
-		tuple_fraction /= dNumDistinctRows;
+		tuple_fraction /= dNumOutputRows;
 
 	if (compare_fractional_path_costs(&hashed_p, &sorted_p,
 									  tuple_fraction) < 0)
diff --git a/src/backend/optimizer/util/tlist.c b/src/backend/optimizer/util/tlist.c
index a2c627fd4d5..b2fb112ebef 100644
--- a/src/backend/optimizer/util/tlist.c
+++ b/src/backend/optimizer/util/tlist.c
@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *	  $PostgreSQL: pgsql/src/backend/optimizer/util/tlist.c,v 1.80 2008/08/07 01:11:50 tgl Exp $
+ *	  $PostgreSQL: pgsql/src/backend/optimizer/util/tlist.c,v 1.81 2008/08/07 19:35:02 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -134,6 +134,69 @@ add_to_flat_tlist(List *tlist, List *vars)
 
 
 /*
+ * get_tlist_exprs
+ *		Get just the expression subtrees of a tlist
+ *
+ * Resjunk columns are ignored unless includeJunk is true
+ */
+List *
+get_tlist_exprs(List *tlist, bool includeJunk)
+{
+	List	   *result = NIL;
+	ListCell   *l;
+
+	foreach(l, tlist)
+	{
+		TargetEntry *tle = (TargetEntry *) lfirst(l);
+
+		if (tle->resjunk && !includeJunk)
+			continue;
+
+		result = lappend(result, tle->expr);
+	}
+	return result;
+}
+
+
+/*
+ * Does tlist have same output datatypes as listed in colTypes?
+ *
+ * Resjunk columns are ignored if junkOK is true; otherwise presence of
+ * a resjunk column will always cause a 'false' result.
+ *
+ * Note: currently no callers care about comparing typmods.
+ */
+bool
+tlist_same_datatypes(List *tlist, List *colTypes, bool junkOK)
+{
+	ListCell   *l;
+	ListCell   *curColType = list_head(colTypes);
+
+	foreach(l, tlist)
+	{
+		TargetEntry *tle = (TargetEntry *) lfirst(l);
+
+		if (tle->resjunk)
+		{
+			if (!junkOK)
+				return false;
+		}
+		else
+		{
+			if (curColType == NULL)
+				return false;	/* tlist longer than colTypes */
+			if (exprType((Node *) tle->expr) != lfirst_oid(curColType))
+				return false;
+			curColType = lnext(curColType);
+		}
+	}
+	if (curColType != NULL)
+		return false;			/* tlist shorter than colTypes */
+	return true;
+}
+
+
+/*
  * get_sortgroupref_tle
  *		Find the targetlist entry matching the given SortGroupRef index,
  *		and return it.
@@ -303,42 +366,3 @@ grouping_is_hashable(List *groupClause)
 	}
 	return true;
 }
-
-
-
-/*
- * Does tlist have same output datatypes as listed in colTypes?
- *
- * Resjunk columns are ignored if junkOK is true; otherwise presence of
- * a resjunk column will always cause a 'false' result.
- *
- * Note: currently no callers care about comparing typmods.
- */
-bool
-tlist_same_datatypes(List *tlist, List *colTypes, bool junkOK)
-{
-	ListCell   *l;
-	ListCell   *curColType = list_head(colTypes);
-
-	foreach(l, tlist)
-	{
-		TargetEntry *tle = (TargetEntry *) lfirst(l);
-
-		if (tle->resjunk)
-		{
-			if (!junkOK)
-				return false;
-		}
-		else
-		{
-			if (curColType == NULL)
-				return false;	/* tlist longer than colTypes */
-			if (exprType((Node *) tle->expr) != lfirst_oid(curColType))
-				return false;
-			curColType = lnext(curColType);
-		}
-	}
-	if (curColType != NULL)
-		return false;			/* tlist shorter than colTypes */
-	return true;
-}