ValueIdSet AppliedStatMan::getPotentialOutputs(
const CANodeIdSet & jbbcsNodeSet)
{
ValueIdSet potentialOutputs;
for (CANodeId jbbc = jbbcsNodeSet.init();
jbbcsNodeSet.next(jbbc);
jbbcsNodeSet.advance(jbbc))
{
if (NodeAnalysis * jbbcNodeAnalysis = jbbc.getNodeAnalysis())
{
ValueIdSet outputs;
const Join * jbbcParentJoin = jbbcNodeAnalysis->getJBBC()->
getOriginalParentJoin();
if((!jbbcParentJoin) ||
(jbbcParentJoin && jbbcParentJoin->isInnerNonSemiJoin()))
outputs = jbbcNodeAnalysis->getOriginalExpr()->\
getGroupAttr()->getCharacteristicOutputs();
else if (jbbcParentJoin->isLeftJoin())
outputs = jbbcParentJoin->nullInstantiatedOutput();
potentialOutputs.insert(outputs);
}
}
return potentialOutputs;
} // AppliedStatMan::getPotentialOutputs
void
IndexDesc::getNonKeyColumnSet(ValueIdSet& nonKeyColumnSet) const
{
const ValueIdList
&indexColumns = getIndexColumns(),
&keyColumns = getIndexKey();
// clean up input:
nonKeyColumnSet.clear();
// Add all index columns
CollIndex i = 0;
for (i=0;
i < indexColumns.entries();
i++)
{
nonKeyColumnSet.insert(indexColumns[i]);
}
// And remove all key columns:
for (i=0;
i < keyColumns.entries();
i++)
{
nonKeyColumnSet.remove(keyColumns[i]);
// if this is a secondary index, the base column
// which is part of the index,
// may also be present, remove it:
const ItemExpr *colPtr = keyColumns[i].getItemExpr();
if (colPtr->getOperatorType()
==
ITM_INDEXCOLUMN)
{
const ValueId & colDef = ((IndexColumn *)(colPtr))->getDefinition();
nonKeyColumnSet.remove(colDef);
}
}
} // IndexDesc::getNonKeyColumnSet(ValueIdSet& nonKeyColumnSet) const
// this method sets the primary key columns. It goes through all the columns
// of the table, and collects the columns which are marked as primary keys
void TableDesc::setPrimaryKeyColumns()
{
ValueIdSet primaryColumns;
for ( CollIndex j = 0 ; j < colList_.entries() ; j++ )
{
ValueId valId = colList_[j];
NAColumn *column = valId.getNAColumn();
if ( column->isPrimaryKey() )
{
primaryColumns.insert(valId) ;
// mark column as referenced for histogram, as we may need its histogram
// during plan generation
if ((column->isUserColumn() || column->isSaltColumn() ) &&
(column->getNATable()->getSpecialType() == ExtendedQualName::NORMAL_TABLE) )
column->setReferencedForMultiIntHist();
}
}
primaryKeyColumns_ = primaryColumns;
}
/********************************************************************
* Input: Selection predicates for the scan node, boolean indicating if
* it is a indexOnlyIndex, reference parameter that will indicate if
* IndexJoin is viable or not, GroupAttributes for the group and characteristic
* inputs
* Output: MdamFlag indicating if the index key access is good enough for
* MDAM access (if a index does not have good MDAM access we have to
* scan the whole index because single subset also will not have any
* keys to apply)
* IndexJoin flag indicating if index join cost would exceed base table
* access or not.
********************************************************************/
MdamFlags IndexDesc::pruneMdam(const ValueIdSet& preds,
NABoolean indexOnlyIndex,
IndexJoinSelectivityEnum&
selectivityEnum /* out*/ ,
const GroupAttributes * groupAttr,
const ValueIdSet * inputValues) const
{
CollIndex numEmptyColumns=0;
CostScalar numSkips = csOne;
ValueIdSet emptyColumns;
ValueId vid;
if(indexOnlyIndex)
selectivityEnum = INDEX_ONLY_INDEX;
else
selectivityEnum = INDEX_JOIN_VIABLE;
if(preds.isEmpty()) return MDAM_OFF;
//calculate how many key columns don't have any predicates
for(CollIndex i=0;i<indexKey_.entries();i++)
{
if(preds.referencesTheGivenValue(indexKey_[i],vid))
break;
else
numEmptyColumns++;
}
//if we don't have any empty columns or we don't have to evaluate if index
//join is promising or not then just return
if(numEmptyColumns>=1 OR NOT indexOnlyIndex)
{
IndexDescHistograms ixHistogram(*this,
(indexOnlyIndex?numEmptyColumns:indexKey_.entries()));
NABoolean multiColUecAvail = ixHistogram.isMultiColUecInfoAvail();
ColumnOrderList keyPredsByCol(indexKey_);
for(CollIndex j=0;j<numEmptyColumns;j++)
{
emptyColumns.insert(indexKey_[j]);
if(j==0 OR multiColUecAvail == FALSE)
{
//no MCUec so just multiply the empty columns UEC count to
//calculate MDAM skips
numSkips *=(ixHistogram.getColStatsForColumn(indexKey_[j])).
getTotalUec().getCeiling();
}
else // otherwise try to use MCUec
{
NABoolean uecFound = FALSE;
CostScalar correctUec = csOne;
CostScalar combinedUECCount = csOne;
// first let's see if there is multiColUec count for the skipped columns
// so far. If there is that will be number of skips. If there isn't then
// get the best estimate of UEC count for the current column using MCUec
// if possible otherwise just using single column histograms.
combinedUECCount = ixHistogram.getUecCountForColumns(emptyColumns);
if(combinedUECCount >0)
{
numSkips = combinedUECCount;
}
else
{
uecFound = ixHistogram.estimateUecUsingMultiColUec(keyPredsByCol,j,correctUec);
if(uecFound==TRUE)
{
numSkips *= correctUec;
}
else
{
numSkips *=(ixHistogram.getColStatsForColumn(indexKey_[j])).
getTotalUec().getCeiling();
}
}
}
}
CostScalar rowCount = ixHistogram.getRowCount();
CostScalar numIndexBlocks = rowCount /getEstimatedRecordsPerBlock();
CostScalar numProbes = csOne;
CostScalar numBaseTableBlocks = csOne;
CostScalar inputProbes = csOne;
// Pass any selectivity hint provided by the user
const SelectivityHint * selHint = tableDesc_->getSelectivityHint();
const CardinalityHint * cardHint = tableDesc_->getCardinalityHint();
// If it is an index join then compute the number probes into the base
// table. If the alternate index is not selective enough, we will have
// lots of them making the index quite expensive.
if(NOT indexOnlyIndex)
{
if((groupAttr->getInputLogPropList()).entries() >0)
{
//if there are incoming probes to the index. i.e. if the index join
//is under another nested join or TSJ then compute result for all
//probes. We are using the initial inputEstLogProp to compute the
//resulting cardinality. It is possible that for the same group and
//different inputEstLogProp would provide less row count per probe.
//So in FileScanRule::nextSubstitute() we make sure that the context
//inputEstLogProp is in the error range of this inputEstLogProp.
// Ex. select * from lineitem, customer, nation
// where l_custkey < c_custkey and c_custkey = n_nationkey;
//Now if we were evaluating lineitem indexes where the outer was customer
//we would want to exclude alternate index on custkey whereas if nation got
//pushed below customer then range of values would be fewer and max value
//being less would make alternate index on custkey quite attractive.
ixHistogram.
applyPredicatesWhenMultipleProbes(preds,
*((groupAttr->getInputLogPropList())[0]),
*inputValues,
TRUE,
selHint,
cardHint,
NULL,
REL_SCAN);
inputProbes = MIN_ONE((groupAttr->getInputLogPropList())[0]->getResultCardinality());
}
else
{
RelExpr * dummyExpr = new (STMTHEAP) RelExpr(ITM_FIRST_ITEM_OP,
NULL,
NULL,
STMTHEAP);
ixHistogram.applyPredicates(preds, *dummyExpr, selHint, cardHint, REL_SCAN);
}
numProbes = ixHistogram.getRowCount();
numBaseTableBlocks = rowCount / tableDesc_->getClusteringIndex()->
getEstimatedRecordsPerBlock();
double readAhead = CURRSTMT_OPTDEFAULTS->readAheadMaxBlocks();
// although we compute cardinality from the index for all probes we
// do the comparison for per probe. The assumption is that per probe
// the upper bound of cost is scanning the whole base table.
if(numProbes/inputProbes + MINOF((numIndexBlocks / readAhead),numSkips)
> (numBaseTableBlocks/readAhead))
{
selectivityEnum = EXCEEDS_BT_SCAN;
}
}
//Does the number of skips exceed the cost of scanning the index.
if((indexOnlyIndex AND numSkips <=
(numIndexBlocks * CURRSTMT_OPTDEFAULTS->mdamSelectionDefault())) OR
(NOT indexOnlyIndex AND numSkips + numProbes/inputProbes <=
(numBaseTableBlocks * CURRSTMT_OPTDEFAULTS->mdamSelectionDefault())))
return MDAM_ON;
}
else
return MDAM_ON;
return MDAM_OFF;
}
// compress the histograms based on query predicates on this table
void TableDesc::compressHistogramsForCurrentQuery()
{
// if there are some column statistics
if ((colStats_.entries() != 0) &&
(table_) &&
(table_->getExtendedQualName().getSpecialType() == ExtendedQualName::NORMAL_TABLE))
{ // if 1
// check if query analysis info is available
if(QueryAnalysis::Instance()->isAnalysisON())
{ // if 2
// get a handle to the query analysis
QueryAnalysis* queryAnalysis = QueryAnalysis::Instance();
// get a handle to the table analysis
const TableAnalysis * tableAnalysis = getTableAnalysis();
if(!tableAnalysis)
return;
// iterate over statistics for each column
for(CollIndex i = 0; i < colStats_.entries(); i++)
{ // for 1
// Get a handle to the column's statistics descriptor
ColStatDescSharedPtr columnStatDesc = colStats_[i];
// get a handle to the ColStats
ColStatsSharedPtr colStats = columnStatDesc->getColStats();
// if this is a single column, as opposed to a multicolumn
if(colStats->getStatColumns().entries() == 1)
{ // if 3
// get column's value id
const ValueId columnId = columnStatDesc->getColumn();
// get column analysis
ColAnalysis* colAnalysis = queryAnalysis->getColAnalysis(columnId);
if(!colAnalysis) continue;
ValueIdSet predicatesOnColumn =
colAnalysis->getReferencingPreds();
// we can compress this column's histogram if there
// is a equality predicate against a constant
ItemExpr *constant = NULL;
NABoolean colHasEqualityAgainstConst =
colAnalysis->getConstValue(constant);
// if a equality predicate with a constant was found
// i.e. predicate of the form col = 5
if (colHasEqualityAgainstConst)
{ // if 4
if (constant)
// compress the histogram
columnStatDesc->compressColStatsForQueryPreds(constant,constant);
} // if 4
else { // else 4
// since there is no equality predicates we might still
// be able to compress the column's histogram based on
// range predicates against a constant. Following are
// examples of such predicates
// * col > 1 <-- predicate defines a lower bound
// * col < 3 <-- predicate defines a upper bound
// * col >1 and col < 30 <-- window predicate, define both bounds
ItemExpr * lowerBound = NULL;
ItemExpr * upperBound = NULL;
// Extract predicates from range spec and add it to the
// original predicate set otherwise isARangePredicate() will
// return FALSE, so histgram compression won't happen.
ValueIdSet rangeSpecPred(predicatesOnColumn);
for (ValueId predId= rangeSpecPred.init();
rangeSpecPred.next(predId);
rangeSpecPred.advance(predId))
{
ItemExpr * pred = predId.getItemExpr();
if ( pred->getOperatorType() == ITM_RANGE_SPEC_FUNC )
{
ValueIdSet vs;
((RangeSpecRef *)pred)->getValueIdSetForReconsItemExpr(vs);
// remove rangespec vid from the original set
predicatesOnColumn.remove(predId);
// add preds extracted from rangespec to the original set
predicatesOnColumn.insert(vs);
}
}
// in the following loop we iterate over all the predicates
// on this column. If there is a range predicate e.g. a > 2
// or a < 3, then we use that to define upper and lower bounds.
// Given predicate a > 2, we get a lower bound of 2.
// Given predicate a < 3, we get a upper bound of 3.
// The bound are then passed down to the histogram
// compression methods.
// iterate over predicates to see if any of them is a range
// predicate e.g. a > 2
for (ValueId predId= predicatesOnColumn.init();
predicatesOnColumn.next(predId);
predicatesOnColumn.advance(predId))
{ // for 2
// check if this predicate is a range predicate
ItemExpr * predicateOnColumn = predId.getItemExpr();
if (predicateOnColumn->isARangePredicate())
{ // if 5
// if a predicate is a range predicate we need to find out more
// information regarding the predicate to see if it can be used
// to compress the columns histogram. We look for the following:
// * The predicate is against a constant e.g. a > 3 and not against
// another column e.g. a > b
// Also give a predicate we need to find out what side is the column
// and what side is the constant. Normally people write a range predicate
// as a > 3, but the same could be written as 3 < a.
// Also either on of the operands of the range predicate might be
// a VEG, if so then we need to dig into the VEG to see where is
// the constant and where is the column.
// check the right and left children of this predicate to
// see if one of them is a constant
ItemExpr * leftChildItemExpr = (ItemExpr *) predicateOnColumn->getChild(0);
ItemExpr * rightChildItemExpr = (ItemExpr *) predicateOnColumn->getChild(1);
// by default assume the literal is at right i.e. predicate of
// the form a > 2
NABoolean columnAtRight = FALSE;
// check if right child of predicate is a VEG
if ( rightChildItemExpr->getOperatorType() == ITM_VEG_REFERENCE)
{ // if 6
// if child is a VEG
VEGReference * rightChildVEG = (VEGReference *) rightChildItemExpr;
// check if the VEG contains the current column
// if it does contain the current column then
// the predicate has the column on right and potentially
// a constant on the left.
if(rightChildVEG->getVEG()->getAllValues().contains(columnId))
{ // if 7
// column is at right i.e. predicate is of the form
// 2 < a
columnAtRight = TRUE;
} // if 7
} // if 6
else { // else 6
// child is not a VEG
if ( columnId == rightChildItemExpr->getValueId() )
{ // if 8
// literals are at left i.e. predicate is of the form
// (1,2) < (a, b)
columnAtRight = TRUE;
} // if 8
} // else 6
ItemExpr * potentialConstantExpr = NULL;
// check if the range predicate is against a constant
if (columnAtRight)
{ // if 9
// the left child is potentially a constant
potentialConstantExpr = leftChildItemExpr;
} // if 9
else { // else 9
// the right child is potentially a constant
potentialConstantExpr = rightChildItemExpr;
} // else 9
// initialize constant to NULL before
// looking for next constant
constant = NULL;
// check if potentialConstantExpr contains a constant.
// we need to see if this range predicate is a predicate
// against a constant e.g col > 1 and not a predicate
// against another column e.g. col > anothercol
// if the expression is a VEG
if ( potentialConstantExpr->getOperatorType() == ITM_VEG_REFERENCE)
{ // if 10
// expression is a VEG, dig into the VEG to
// get see if it contains a constant
VEGReference * potentialConstantExprVEG =
(VEGReference *) potentialConstantExpr;
potentialConstantExprVEG->getVEG()->\
getAllValues().referencesAConstValue(&constant);
} // if 10
else { // else 10
// express is not a VEG, it is a constant
if ( potentialConstantExpr->getOperatorType() == ITM_CONSTANT )
constant = potentialConstantExpr;
} // else 10
// if predicate involves a constant, does the constant imply
// a upper bound or lower bound
if (constant)
{ // if 11
// if range predicate has column at right e.g. 3 > a
if (columnAtRight)
{ // if 12
if ( predicateOnColumn->getOperatorType() == ITM_GREATER ||
predicateOnColumn->getOperatorType() == ITM_GREATER_EQ)
{ // if 13
if (!upperBound)
upperBound = constant;
} // if 13
else
{ // else 13
if (!lowerBound)
lowerBound = constant;
} // else 13
} // if 12
else { // else 12
// range predicate has column at left e.g. a < 3
if ( predicateOnColumn->getOperatorType() == ITM_LESS ||
predicateOnColumn->getOperatorType() == ITM_LESS_EQ)
{ // if 14
if (!upperBound)
upperBound = constant;
} // if 14
else
{ // else 14
if (!lowerBound)
lowerBound = constant;
} // else 14
} // else 12
} // if 11
} // if 5
} // for 2
// if we found a upper bound or a lower bound
if (lowerBound || upperBound)
{
// compress the histogram based on range predicates
columnStatDesc->compressColStatsForQueryPreds(lowerBound, upperBound);
}
} // else 4
} // if 3
} // for 1
} // if 2
} // if 1
// All histograms compressed. Set the histCompressed flag to TRUE
histsCompressed(TRUE);
}