//---------------------------------------------------------------------------
// @function:
// CExpressionHandle::CopyCostCtxtProps
//
// @doc:
// Cache plan properties of cost context and its children on the handle
//
//---------------------------------------------------------------------------
void
CExpressionHandle::CopyCostCtxtProps()
{
GPOS_ASSERT(NULL != m_pcc);
GPOS_ASSERT(NULL == m_pdrgpdp);
GPOS_ASSERT(NULL == m_pdp);
// add-ref context properties
CDrvdProp *pdp = m_pcc->Pdpplan();
pdp->AddRef();
m_pdp = pdp;
// add-ref child group expressions' properties
const ULONG ulArity = UlArity();
m_pdrgpdp = GPOS_NEW(m_pmp) DrgPdp(m_pmp, ulArity);
for (ULONG ul = 0; ul < ulArity; ul++)
{
CGroup *pgroupChild = (*m_pgexpr)[ul];
if (!pgroupChild->FScalar())
{
COptimizationContext *pocChild = (*m_pcc->Pdrgpoc())[ul];
GPOS_ASSERT(NULL != pocChild);
CCostContext *pccChild = pocChild->PccBest();
GPOS_ASSERT(NULL != pccChild);
pdp = pccChild->Pdpplan();
pdp->AddRef();
m_pdrgpdp->Append(pdp);
}
}
}
//---------------------------------------------------------------------------
// @function:
// CPhysicalMotion::FValidContext
//
// @doc:
// Check if optimization context is valid
//
//---------------------------------------------------------------------------
BOOL
CPhysicalMotion::FValidContext
(
IMemoryPool *pmp,
COptimizationContext *poc,
DrgPoc *pdrgpocChild
)
const
{
GPOS_ASSERT(NULL != pdrgpocChild);
GPOS_ASSERT(1 == pdrgpocChild->UlLength());
COptimizationContext *pocChild = (*pdrgpocChild)[0];
CCostContext *pccBest = pocChild->PccBest();
GPOS_ASSERT(NULL != pccBest);
CDrvdPropPlan *pdpplanChild = pccBest->Pdpplan();
if (pdpplanChild->Ppim()->FContainsUnresolved())
{
return false;
}
CExpressionHandle exprhdl(pmp);
exprhdl.Attach(pccBest);
exprhdl.DeriveProps(NULL /*CDrvdPropCtxt*/);
if (exprhdl.FHasOuterRefs())
{
// disallow plans with outer references below motion operator
return false;
}
CEnfdDistribution *ped = poc->Prpp()->Ped();
if (ped->FCompatible(this->Pds()) && ped->FCompatible(pdpplanChild->Pds()))
{
// required distribution is compatible with the distribution delivered by Motion and its child plan,
// in this case, Motion is redundant since child plan delivers the required distribution
return false;
}
return true;
}
//---------------------------------------------------------------------------
// @function:
// CCostContext::ComputeCost
//
// @doc:
// Compute cost of current context,
//
// the function extracts cardinality and row width of owner operator
// and child operators, and then adjusts row estimate obtained from
// statistics based on data distribution obtained from plan properties,
//
// statistics row estimate is computed on logical expressions by
// estimating the size of the whole relation regardless data
// distribution, on the other hand, optimizer's cost model computes
// the cost of a plan instance on some segment,
//
// when a plan produces tuples distributed to multiple segments, we
// need to divide statistics row estimate by the number segments to
// provide a per-segment row estimate for cost computation,
//
// Note that this scaling of row estimate cannot happen during
// statistics derivation since plans are not created yet at this point
//
// this function also extracts number of rebinds of owner operator child
// operators, if statistics are computed using predicates with external
// parameters (outer references), number of rebinds is the total number
// of external parameters' values
//
//---------------------------------------------------------------------------
CCost
CCostContext::CostCompute
(
IMemoryPool *pmp,
DrgPcost *pdrgpcostChildren
)
{
// derive context stats
DeriveStats();
ULONG ulArity = 0;
if (NULL != m_pdrgpoc)
{
ulArity = Pdrgpoc()->UlLength();
}
m_pstats->AddRef();
ICostModel::SCostingInfo ci(pmp, ulArity, GPOS_NEW(pmp) ICostModel::CCostingStats(m_pstats));
ICostModel *pcm = COptCtxt::PoctxtFromTLS()->Pcm();
CExpressionHandle exprhdl(pmp);
exprhdl.Attach(this);
// extract local costing info
DOUBLE dRows = m_pstats->DRows().DVal();
if (CDistributionSpec::EdptPartitioned == Pdpplan()->Pds()->Edpt())
{
// scale statistics row estimate by number of segments
dRows = DRowsPerHost().DVal();
}
ci.SetRows(dRows);
DOUBLE dWidth = m_pstats->DWidth(pmp, m_poc->Prpp()->PcrsRequired()).DVal();
ci.SetWidth(dWidth);
DOUBLE dRebinds = m_pstats->DRebinds().DVal();
ci.SetRebinds(dRebinds);
GPOS_ASSERT_IMP(!exprhdl.FHasOuterRefs(), GPOPT_DEFAULT_REBINDS == (ULONG) (dRebinds) && "invalid number of rebinds when there are no outer references");
// extract children costing info
for (ULONG ul = 0; ul < ulArity; ul++)
{
COptimizationContext *pocChild = (*m_pdrgpoc)[ul];
CCostContext *pccChild = pocChild->PccBest();
GPOS_ASSERT(NULL != pccChild);
IStatistics *pstatsChild = pccChild->Pstats();
DOUBLE dRowsChild = pstatsChild->DRows().DVal();
if (CDistributionSpec::EdptPartitioned == pccChild->Pdpplan()->Pds()->Edpt())
{
// scale statistics row estimate by number of segments
dRowsChild = pccChild->DRowsPerHost().DVal();
}
ci.SetChildRows(ul, dRowsChild);
DOUBLE dWidthChild = pstatsChild->DWidth(pmp, pocChild->Prpp()->PcrsRequired()).DVal();
ci.SetChildWidth(ul, dWidthChild);
DOUBLE dRebindsChild = pstatsChild->DRebinds().DVal();
ci.SetChildRebinds(ul, dRebindsChild);
GPOS_ASSERT_IMP(!exprhdl.FHasOuterRefs(ul), GPOPT_DEFAULT_REBINDS == (ULONG) (dRebindsChild) && "invalid number of rebinds when there are no outer references");
DOUBLE dCostChild = (*pdrgpcostChildren)[ul]->DVal();
ci.SetChildCost(ul, dCostChild);
}
// compute cost using the underlying cost model
return pcm->Cost(exprhdl, &ci);
}
//---------------------------------------------------------------------------
// @function:
// COptimizationContext::PrppCTEProducer
//
// @doc:
// Compute required properties to CTE producer based on plan properties
// of CTE consumer
//
//---------------------------------------------------------------------------
CReqdPropPlan *
COptimizationContext::PrppCTEProducer
(
IMemoryPool *mp,
COptimizationContext *poc,
ULONG ulSearchStages
)
{
GPOS_ASSERT(NULL != poc);
GPOS_ASSERT(NULL != poc->PccBest());
CCostContext *pccBest = poc->PccBest();
CGroupExpression *pgexpr = pccBest->Pgexpr();
BOOL fOptimizeCTESequence =
(
COperator::EopPhysicalSequence == pgexpr->Pop()->Eopid() &&
(*pgexpr)[0]->FHasCTEProducer()
);
if (!fOptimizeCTESequence)
{
// best group expression is not a CTE sequence
return NULL;
}
COptimizationContext *pocProducer = (*pgexpr)[0]->PocLookupBest(mp, ulSearchStages, (*pccBest->Pdrgpoc())[0]->Prpp());
if (NULL == pocProducer)
{
return NULL;
}
CCostContext *pccProducer = pocProducer->PccBest();
if (NULL == pccProducer)
{
return NULL;
}
COptimizationContext *pocConsumer = (*pgexpr)[1]->PocLookupBest(mp, ulSearchStages, (*pccBest->Pdrgpoc())[1]->Prpp());
if (NULL == pocConsumer)
{
return NULL;
}
CCostContext *pccConsumer = pocConsumer->PccBest();
if (NULL == pccConsumer)
{
return NULL;
}
CColRefSet *pcrsInnerOutput = CDrvdPropRelational::GetRelationalProperties((*pgexpr)[1]->Pdp())->PcrsOutput();
CPhysicalCTEProducer *popProducer = CPhysicalCTEProducer::PopConvert(pccProducer->Pgexpr()->Pop());
UlongToColRefMap *colref_mapping = COptCtxt::PoctxtFromTLS()->Pcteinfo()->PhmulcrConsumerToProducer(mp, popProducer->UlCTEId(), pcrsInnerOutput, popProducer->Pdrgpcr());
CReqdPropPlan *prppProducer = CReqdPropPlan::PrppRemap(mp, pocProducer->Prpp(), pccConsumer->Pdpplan(), colref_mapping);
colref_mapping->Release();
if (prppProducer->Equals(pocProducer->Prpp()))
{
prppProducer->Release();
return NULL;
}
return prppProducer;
}
BOOL
CPhysicalSpool::FValidContext
(
IMemoryPool *,
COptimizationContext *poc,
COptimizationContextArray *pdrgpocChild
)
const
{
GPOS_ASSERT(NULL != pdrgpocChild);
GPOS_ASSERT(1 == pdrgpocChild->Size());
COptimizationContext *pocChild = (*pdrgpocChild)[0];
CCostContext *pccBest = pocChild->PccBest();
GPOS_ASSERT(NULL != pccBest);
// partition selections that happen outside of a physical spool does not do
// any good on rescan: a physical spool blocks the rescan from the entire
// subtree (in particular, any dynamic scan) underneath it. That means when
// we have a dynamic scan under a spool, and a corresponding partition
// selector outside the spool, we run the risk of materializing the wrong
// results.
// For example, the following plan is invalid because the partition selector
// won't be able to influence inner side of the nested loop join as intended
// ("blocked" by the spool):
// +--CPhysicalMotionGather(master)
// +--CPhysicalInnerNLJoin
// |--CPhysicalPartitionSelector
// | +--CPhysicalMotionBroadcast
// | +--CPhysicalTableScan "foo" ("foo")
// |--CPhysicalSpool
// | +--CPhysicalLeftOuterHashJoin
// | |--CPhysicalDynamicTableScan "pt" ("pt")
// | |--CPhysicalMotionHashDistribute
// | | +--CPhysicalTableScan "bar" ("bar")
// | +--CScalarCmp (=)
// | |--CScalarIdent "d" (19)
// | +--CScalarIdent "dk" (9)
// +--CScalarCmp (<)
// |--CScalarIdent "a" (0)
// +--CScalarIdent "partkey" (10)
CDrvdPropPlan *pdpplanChild = pccBest->Pdpplan();
if (pdpplanChild->Ppim()->FContainsUnresolved())
{
return false;
}
// Discard any context that is requesting for rewindability with motion hazard handling and
// the physical spool is streaming with a motion underneath it.
// We do not want to add a blocking spool over a spool as spooling twice will be expensive,
// hence invalidate this context.
CEnfdRewindability *per = poc->Prpp()->Per();
if(per->PrsRequired()->HasMotionHazard() &&
pdpplanChild->Prs()->HasMotionHazard())
{
return FEager();
}
return true;
}