From fbe0b94cc535bcdc85eabd916f211d44aa411fca Mon Sep 17 00:00:00 2001
From: Brian Cloutier <github@briancloutier.com>
Date: Wed, 27 Apr 2016 05:05:02 -0700
Subject: [PATCH] Query Planning Performance Improvments

- Only look at pruned shards when determining AnchorTable
- Use cached shardIntervalCompareFunction during copartition check
---
 .../planner/multi_logical_optimizer.c          | 18 ++++++++++--------
 .../planner/multi_physical_planner.c           | 15 ++++++++++++++-
 2 files changed, 24 insertions(+), 9 deletions(-)

diff --git a/src/backend/distributed/planner/multi_logical_optimizer.c b/src/backend/distributed/planner/multi_logical_optimizer.c
index e68804d8c..784982894 100644
--- a/src/backend/distributed/planner/multi_logical_optimizer.c
+++ b/src/backend/distributed/planner/multi_logical_optimizer.c
@@ -139,7 +139,8 @@ static bool JoinOnPartitionColumn(Query *query);
 static void ErrorIfUnsupportedShardDistribution(Query *query);
 static List * RelationIdList(Query *query);
 static bool CoPartitionedTables(Oid firstRelationId, Oid secondRelationId);
-static bool ShardIntervalsEqual(ShardInterval *firstInterval,
+static bool ShardIntervalsEqual(FmgrInfo *comparisonFunction,
+								ShardInterval *firstInterval,
 								ShardInterval *secondInterval);
 static void ErrorIfUnsupportedFilters(Query *subquery);
 static bool EqualOpExpressionLists(List *firstOpExpressionList,
@@ -3496,6 +3497,7 @@ CoPartitionedTables(Oid firstRelationId, Oid secondRelationId)
 		secondTableCache->sortedShardIntervalArray;
 	uint32 firstListShardCount = firstTableCache->shardIntervalArrayLength;
 	uint32 secondListShardCount = secondTableCache->shardIntervalArrayLength;
+	FmgrInfo *comparisonFunction = firstTableCache->shardIntervalCompareFunction;
 
 	if (firstListShardCount != secondListShardCount)
 	{
@@ -3508,12 +3510,16 @@ CoPartitionedTables(Oid firstRelationId, Oid secondRelationId)
 		return true;
 	}
 
+	Assert(comparisonFunction != NULL);
+
 	for (intervalIndex = 0; intervalIndex < firstListShardCount; intervalIndex++)
 	{
 		ShardInterval *firstInterval = sortedFirstIntervalArray[intervalIndex];
 		ShardInterval *secondInterval = sortedSecondIntervalArray[intervalIndex];
 
-		bool shardIntervalsEqual = ShardIntervalsEqual(firstInterval, secondInterval);
+		bool shardIntervalsEqual = ShardIntervalsEqual(comparisonFunction,
+													   firstInterval,
+													   secondInterval);
 		if (!shardIntervalsEqual)
 		{
 			coPartitionedTables = false;
@@ -3529,19 +3535,15 @@ CoPartitionedTables(Oid firstRelationId, Oid secondRelationId)
  * ShardIntervalsEqual checks if given shard intervals have equal min/max values.
  */
 static bool
-ShardIntervalsEqual(ShardInterval *firstInterval, ShardInterval *secondInterval)
+ShardIntervalsEqual(FmgrInfo *comparisonFunction, ShardInterval *firstInterval,
+					ShardInterval *secondInterval)
 {
-	Oid typeId = InvalidOid;
-	FmgrInfo *comparisonFunction = NULL;
 	bool shardIntervalsEqual = false;
 	Datum firstMin = 0;
 	Datum firstMax = 0;
 	Datum secondMin = 0;
 	Datum secondMax = 0;
 
-	typeId = firstInterval->valueTypeId;
-	comparisonFunction = GetFunctionInfo(typeId, BTREE_AM_OID, BTORDER_PROC);
-
 	firstMin = firstInterval->minValue;
 	firstMax = firstInterval->maxValue;
 	secondMin = secondInterval->minValue;
diff --git a/src/backend/distributed/planner/multi_physical_planner.c b/src/backend/distributed/planner/multi_physical_planner.c
index 45f364c3d..88675c1b6 100644
--- a/src/backend/distributed/planner/multi_physical_planner.c
+++ b/src/backend/distributed/planner/multi_physical_planner.c
@@ -1938,6 +1938,7 @@ SubquerySqlTaskList(Job *job)
 	uint32 anchorRangeTableId = 0;
 	uint32 rangeTableIndex = 0;
 	const uint32 fragmentSize = sizeof(RangeTableFragment);
+	uint64 largestTableSize = 0;
 
 	/* find filters on partition columns */
 	ExtractQueryWalker((Node *) subquery, &queryList);
@@ -1961,7 +1962,6 @@ SubquerySqlTaskList(Job *job)
 
 	/* get list of all range tables in subquery tree */
 	ExtractRangeTableRelationWalker((Node *) subquery, &rangeTableList);
-	anchorRangeTableId = AnchorRangeTableId(rangeTableList);
 
 	/*
 	 * For each range table entry, first we prune shards for the relation
@@ -1979,6 +1979,7 @@ SubquerySqlTaskList(Job *job)
 		ListCell *shardIntervalCell = NULL;
 		uint32 tableId = rangeTableIndex + 1; /* tableId starts from 1 */
 		uint32 finalShardCount = 0;
+		uint64 tableSize = 0;
 
 		if (opExpressionList != NIL)
 		{
@@ -2006,6 +2007,8 @@ SubquerySqlTaskList(Job *job)
 		{
 			ShardInterval *shardInterval = (ShardInterval *) lfirst(shardIntervalCell);
 
+			tableSize += ShardLength(shardInterval->shardId);
+
 			RangeTableFragment *shardFragment = palloc0(fragmentSize);
 			shardFragment->fragmentReference = &(shardInterval->shardId);
 			shardFragment->fragmentType = CITUS_RTE_RELATION;
@@ -2027,6 +2030,16 @@ SubquerySqlTaskList(Job *job)
 			}
 		}
 
+		/*
+		 * Determine anchor table using shards which survive pruning instead of calling
+		 * AnchorRangeTableId
+		 */
+		if (anchorRangeTableId == 0 || tableSize > largestTableSize)
+		{
+			largestTableSize = tableSize;
+			anchorRangeTableId = tableId;
+		}
+
 		rangeTableIndex++;
 	}