Cost model favors parallel union all [#130121327]

We enable both `serial` and `parallel union all` in the search space
and we favor `parallel union all` if the `optimizer_parallel_union`
GUC is enabled. `parallel union all` is costed as the same as maximum
costed child instead of adding the children's cost all together.

This makes the framework to choose the parallel union all instead of
serial union all when both operators are part of valid plan
alternatives.

libgpdbcost/include/gpdbcost/CCostModelGPDB.h

@@ -95,6 +95,10 @@ namespace gpdbcost
 			static
 			CCost CostChildren(IMemoryPool *pmp, CExpressionHandle &exprhdl, const SCostingInfo *pci, ICostModelParams *pcp);
 
+			// returns cost of highest costed child
+			static
+			CCost CostMaxChild(IMemoryPool *pmp, CExpressionHandle &exprhdl, const SCostingInfo *pci, ICostModelParams *pcp);
+
 			// check if given operator is unary
 			static
 			BOOL FUnary(COperator::EOperatorId eopid);

libgpdbcost/src/CCostModelGPDB.cpp

@@ -342,6 +342,40 @@ CCostModelGPDB::CostChildren
 	return CCost(dRes);
 }
 
+//---------------------------------------------------------------------------
+//	@function:
+//		CCostModelGPDB::CostMaxChild
+//
+//	@doc:
+//		Returns cost of highest costed child
+//
+//---------------------------------------------------------------------------
+
+CCost
+CCostModelGPDB::CostMaxChild
+(
+	IMemoryPool *,
+	CExpressionHandle &,
+	const SCostingInfo *pci,
+	ICostModelParams *
+	)
+{
+	GPOS_ASSERT(NULL != pci);
+
+	DOUBLE *pdCost = pci->PdCost();
+	const ULONG ulSize = pci->UlChildren();
+
+	DOUBLE dRes = 0.0;
+	for (ULONG ul = 0; ul < ulSize; ul++)
+	{
+		if (pdCost[ul] > dRes)
+		{
+			dRes = pdCost[ul];
+		}
+	}
+
+	return CCost(dRes);
+}
 
 //---------------------------------------------------------------------------
 //	@function:
@@ -695,6 +729,11 @@ CCostModelGPDB::CostUnionAll
 	GPOS_ASSERT(NULL != pci);
 	GPOS_ASSERT(NULL != CPhysicalUnionAll::PopConvert(exprhdl.Pop()));
 
+	if(COperator::EopPhysicalParallelUnionAll == exprhdl.Pop()->Eopid())
+	{
+		return CostMaxChild(pmp, exprhdl, pci, pcmgpdb->Pcp());
+	}
+
 	CCost costLocal = CCost(pci->DRebinds() * CostTupleProcessing(pci->DRows(), pci->DWidth(), pcmgpdb->Pcp()).DVal());
 	CCost costChild = CostChildren(pmp, exprhdl, pci, pcmgpdb->Pcp());