public static RedistFeatureStep buildLookupJoinRedist(
PlannerContext pc,
RedistFeatureStep lookupTable,
List<ExpressionNode> lookupJoinColumns,
PEStorageGroup targetGroup,
JoinEntry origEntry,
List<ExpressionNode> origJoinColumns,
PartitionEntry nonLookupSide)
throws PEException {
// we will modify the existing non lookup side projecting feature step, and add the lookup table
// as a requirement to it.
ProjectingFeatureStep nonLookupStep = (ProjectingFeatureStep) nonLookupSide.getStep(null);
nonLookupSide.maybeForceDoublePrecision(nonLookupStep);
SelectStatement lookupSelect = lookupTable.getTargetTempTable().buildSelect(pc.getContext());
SelectStatement nonLookupSelect = (SelectStatement) nonLookupStep.getPlannedStatement();
SelectStatement actualJoinStatement =
DMLStatementUtils.compose(pc.getContext(), nonLookupSelect, lookupSelect);
List<ExpressionNode> ands =
ExpressionUtils.decomposeAndClause(actualJoinStatement.getWhereClause());
// instead of building the join spec directly, map forward the original join condition from the
// joined table if it's available
IndexCollector ic = new IndexCollector();
if (origEntry.getJoin().getJoin() != null) {
FunctionCall mapped =
(FunctionCall)
actualJoinStatement
.getMapper()
.copyForward(origEntry.getJoin().getJoin().getJoinOn());
ands.add(mapped);
ListSet<ColumnInstance> cols = ColumnInstanceCollector.getColumnInstances(mapped);
for (ColumnInstance ci : cols) ic.addColumnInstance(ci);
} else {
Map<RewriteKey, ExpressionNode> projEntries = null;
for (int i = 0; i < origJoinColumns.size(); i++) {
ColumnKey mck =
actualJoinStatement.getMapper().mapExpressionToColumn(origJoinColumns.get(i));
ColumnKey muck =
actualJoinStatement.getMapper().mapExpressionToColumn(lookupJoinColumns.get(i));
ExpressionNode mc = null;
ExpressionNode muc = null;
if (mck == null || muck == null) {
if (projEntries == null) {
projEntries = new HashMap<RewriteKey, ExpressionNode>();
for (ExpressionNode en : actualJoinStatement.getProjectionEdge()) {
ExpressionNode actual = ExpressionUtils.getTarget(en);
if (actual instanceof ColumnInstance) {
projEntries.put(((ColumnInstance) actual).getColumnKey(), actual);
} else {
projEntries.put(new ExpressionKey(actual), actual);
}
}
}
if (mck == null)
mc =
(ExpressionNode)
projEntries.get(new ExpressionKey(origJoinColumns.get(i))).copy(null);
if (muck == null)
mc =
(ExpressionNode)
projEntries.get(new ExpressionKey(lookupJoinColumns.get(i))).copy(null);
}
if (mc == null) mc = mck.toInstance();
if (muc == null) muc = muck.toInstance();
if (mc instanceof ColumnInstance) ic.addColumnInstance((ColumnInstance) mc);
if (muc instanceof ColumnInstance) ic.addColumnInstance((ColumnInstance) muc);
FunctionCall eq = new FunctionCall(FunctionName.makeEquals(), mc, muc);
ands.add(eq);
}
}
ic.setIndexes(origEntry.getSchemaContext());
TempTable lookupEntryTarget = lookupTable.getTargetTempTable();
// everything from the lhs that's in the projection should be cleared - it's invisible
// note that we do it after building the where clause so that the original join condition can be
// mapped
for (Iterator<ExpressionNode> iter = actualJoinStatement.getProjectionEdge().iterator();
iter.hasNext(); ) {
ExpressionNode en = iter.next();
ExpressionNode targ = ExpressionUtils.getTarget(en);
if (targ instanceof ColumnInstance) {
ColumnInstance ci = (ColumnInstance) targ;
if (ci.getTableInstance().getAbstractTable() == lookupEntryTarget) iter.remove();
}
}
actualJoinStatement.setWhereClause(ExpressionUtils.safeBuildAnd(ands));
actualJoinStatement.normalize(origEntry.getSchemaContext());
// build a new projecting feature step
ProjectingFeatureStep lookupJoinStep =
DefaultFeatureStepBuilder.INSTANCE.buildProjectingStep(
origEntry.getPlannerContext(),
origEntry.getFeaturePlanner(),
actualJoinStatement,
new ExecutionCost(false, false, null, -1),
nonLookupStep.getSourceGroup(),
actualJoinStatement.getDatabase(origEntry.getSchemaContext()),
nonLookupStep.getDistributionVector(),
null,
DMLExplainReason.LOOKUP_JOIN.makeRecord());
// arrange for the children of the nonlookup side to become my children
// and add the lookup table step as well
lookupJoinStep.getSelfChildren().addAll(nonLookupStep.getAllChildren());
lookupJoinStep.getSelfChildren().add(lookupTable);
// children must be sequential - no need to modify here
List<Integer> mappedRedistOn =
nonLookupSide.mapDistributedOn(origJoinColumns, actualJoinStatement);
// now remove the lookup table from the mapper - have to do this pretty late - at this point
// the query won't be manipulated any more
actualJoinStatement.getMapper().remove(lookupEntryTarget);
RedistFeatureStep out =
lookupJoinStep.redist(
origEntry.getPlannerContext(),
origEntry.getFeaturePlanner(),
new TempTableCreateOptions(Model.STATIC, targetGroup)
.distributeOn(mappedRedistOn)
.withRowCount(origEntry.getScore().getRowCount()),
null,
DMLExplainReason.LOOKUP_JOIN.makeRecord());
return out;
}
@Override
public JoinedPartitionEntry build() throws PEException {
ExecutionCost combinedCost = buildCombinedCost();
PEStorageGroup tempGroup =
getPlannerContext().getTempGroupManager().getGroup(combinedCost.getGroupScore());
StrategyTable constrained = null;
StrategyTable unconstrained = null;
if (constrainedSide == left) {
constrained = left;
unconstrained = right;
} else {
constrained = right;
unconstrained = left;
}
// we've found the constrained and unconstrained sides - now find the columns in the join
List<ExpressionNode> constrainedJC =
constrained
.getEntry()
.mapDistributedOn(rje.getJoin().getRedistJoinExpressions(constrained.getSingleTable()));
List<ExpressionNode> unconstrainedJC =
unconstrained
.getEntry()
.mapDistributedOn(
rje.getJoin().getRedistJoinExpressions(unconstrained.getSingleTable()));
// redistribute the constrained side onto the temp group.
RedistFeatureStep redistConstrained =
colocateViaRedist(
getPlannerContext(),
getJoin(),
constrained.getSingleTable(),
constrained.getEntry(),
Model.STATIC,
tempGroup,
true,
explain,
rje.getFeaturePlanner());
// the constrained side is represented by this temp table
constrained.getEntry().setStep(redistConstrained);
// next, we have to build the select distinct(id) redist'd to the pg
RedistFeatureStep lookupTableStep =
buildLookupTableRedist(
getPlannerContext(),
constrained.getEntry(),
redistConstrained,
rje,
unconstrained.getGroup(),
true);
// now we have the unconstrained side still on the pers group
// and we have a new temp table (lookupTableStep) also on the pers group.
// this temp table is the result of the join between the lookup table
// and the unconstrained side. the unconstrained entry will be represented by
// this temp table.
RedistFeatureStep redistUnconstrained =
buildLookupJoinRedist(
getPlannerContext(),
lookupTableStep,
constrainedJC,
tempGroup,
rje,
unconstrainedJC,
unconstrained.getEntry());
unconstrained.getEntry().setStep(redistUnconstrained);
// so now we have redistConstrained and redistUnconstrained, we can build the ipe.
// the two RedistFeatureSteps end up being the children for the new ipe, but we need to build
// out the join kernel.
// figure out which is left and right now
PartitionEntry leftEntry = null;
PartitionEntry rightEntry = null;
if (constrainedSide == left) {
leftEntry = constrained.getEntry();
rightEntry = unconstrained.getEntry();
} else {
leftEntry = unconstrained.getEntry();
rightEntry = constrained.getEntry();
}
RedistFeatureStep leftStep = (RedistFeatureStep) leftEntry.getStep(null);
RedistFeatureStep rightStep = (RedistFeatureStep) rightEntry.getStep(null);
return buildResultEntry(
leftStep.buildNewSelect(getPlannerContext()),
leftStep.getTargetTempTable().getDistributionVector(getSchemaContext()),
rightStep.buildNewSelect(getPlannerContext()),
rightStep.getTargetTempTable().getDistributionVector(getSchemaContext()),
combinedCost,
tempGroup,
false // never parallel because we have an ordering issue
);
}
// we have the temp table on the temp group for the constrained side
// we need to build the bcast temp table on the pers group
public static RedistFeatureStep buildLookupTableRedist(
PlannerContext pc,
PartitionEntry srcEntry,
RedistFeatureStep constrainedOnTempGroup,
JoinEntry rje,
PEStorageGroup targetGroup,
boolean indexJoinColumns)
throws PEException {
ProjectingFeatureStep selectDistinct =
constrainedOnTempGroup.buildNewProjectingStep(
pc,
rje.getFeaturePlanner(),
null,
DMLExplainReason.LOOKUP_JOIN_LOOKUP_TABLE.makeRecord());
SelectStatement ss = (SelectStatement) selectDistinct.getPlannedStatement();
DistributionVector distVect =
constrainedOnTempGroup.getTargetTempTable().getDistributionVector(pc.getContext());
ListSet<ColumnKey> mappedColumnsInJoin = null;
if (rje.getJoin().getJoin() != null) {
ListSet<ColumnKey> columnsInJoin =
ColumnInstanceCollector.getColumnKeys(
ColumnInstanceCollector.getColumnInstances(rje.getJoin().getJoin().getJoinOn()));
// build the set of columns in the src entry projection
ListSet<ColumnKey> srcColumns = new ListSet<ColumnKey>();
for (BufferEntry be : srcEntry.getBufferEntries()) {
ExpressionNode targ = ExpressionUtils.getTarget(be.getTarget());
if (targ instanceof ColumnInstance) {
ColumnInstance ci = (ColumnInstance) targ;
srcColumns.add(ci.getColumnKey());
}
}
columnsInJoin.retainAll(srcColumns);
mappedColumnsInJoin = new ListSet<ColumnKey>();
for (ColumnKey ck : columnsInJoin) {
mappedColumnsInJoin.add(ss.getMapper().copyColumnKeyForward(ck));
}
}
for (Iterator<ExpressionNode> iter = ss.getProjectionEdge().iterator(); iter.hasNext(); ) {
ExpressionNode en = iter.next();
if (en instanceof ColumnInstance) {
ColumnInstance ci = (ColumnInstance) en;
PEColumn pec = ci.getPEColumn();
if (!distVect.contains(srcEntry.getSchemaContext(), pec)
&& (mappedColumnsInJoin == null || !mappedColumnsInJoin.contains(ci.getColumnKey())))
iter.remove();
}
}
ss.setSetQuantifier(SetQuantifier.DISTINCT);
ss.normalize(srcEntry.getSchemaContext());
// now that we have the select distinct set up
// redistribute it bcast back onto the pers group
RedistFeatureStep out =
selectDistinct.redist(
pc,
rje.getFeaturePlanner(),
new TempTableCreateOptions(Model.BROADCAST, targetGroup),
null,
DMLExplainReason.LOOKUP_JOIN_LOOKUP_TABLE.makeRecord());
if (indexJoinColumns)
out.getTargetTempTable()
.noteJoinedColumns(pc.getContext(), out.getTargetTempTable().getColumns(pc.getContext()));
return out;
}
