/** Extract a minimal set of WindowGroups from analyticExprs. */
private List<WindowGroup> collectWindowGroups() {
List<Expr> analyticExprs = analyticInfo_.getAnalyticExprs();
List<WindowGroup> groups = Lists.newArrayList();
for (int i = 0; i < analyticExprs.size(); ++i) {
AnalyticExpr analyticExpr = (AnalyticExpr) analyticExprs.get(i);
// Do not generate the plan for non-materialized analytic exprs.
if (!analyticInfo_.getOutputTupleDesc().getSlots().get(i).isMaterialized()) {
continue;
}
boolean match = false;
for (WindowGroup group : groups) {
if (group.isCompatible(analyticExpr)) {
group.add(
(AnalyticExpr) analyticInfo_.getAnalyticExprs().get(i),
analyticInfo_.getOutputTupleDesc().getSlots().get(i),
analyticInfo_.getIntermediateTupleDesc().getSlots().get(i));
match = true;
break;
}
}
if (!match) {
groups.add(
new WindowGroup(
(AnalyticExpr) analyticInfo_.getAnalyticExprs().get(i),
analyticInfo_.getOutputTupleDesc().getSlots().get(i),
analyticInfo_.getIntermediateTupleDesc().getSlots().get(i)));
}
}
return groups;
}
public AnalyticPlanner(
List<TupleId> stmtTupleIds,
AnalyticInfo analyticInfo,
Analyzer analyzer,
IdGenerator<PlanNodeId> idGenerator) {
Preconditions.checkState(!stmtTupleIds.isEmpty());
TupleId lastStmtTupleId = stmtTupleIds.get(stmtTupleIds.size() - 1);
Preconditions.checkState(
stmtTupleIds.size() == 1 || lastStmtTupleId.equals(analyticInfo.getOutputTupleId()));
stmtTupleIds_ = stmtTupleIds;
analyticInfo_ = analyticInfo;
analyzer_ = analyzer;
idGenerator_ = idGenerator;
}
/**
* Create plan tree for the entire sort group, including all contained window groups. Marks the
* SortNode as requiring its input to be partitioned if partitionExprs is not null (partitionExprs
* represent the data partition of the entire partition group of which this sort group is a part).
*/
private PlanNode createSortGroupPlan(
PlanNode root, SortGroup sortGroup, List<Expr> partitionExprs) throws ImpalaException {
List<Expr> partitionByExprs = sortGroup.partitionByExprs;
List<OrderByElement> orderByElements = sortGroup.orderByElements;
ExprSubstitutionMap sortSmap = null;
TupleId sortTupleId = null;
TupleDescriptor bufferedTupleDesc = null;
// map from input to buffered tuple
ExprSubstitutionMap bufferedSmap = new ExprSubstitutionMap();
// sort on partition by (pb) + order by (ob) exprs and create pb/ob predicates
if (!partitionByExprs.isEmpty() || !orderByElements.isEmpty()) {
// first sort on partitionExprs (direction doesn't matter)
List<Expr> sortExprs = Lists.newArrayList(partitionByExprs);
List<Boolean> isAsc =
Lists.newArrayList(Collections.nCopies(sortExprs.size(), new Boolean(true)));
// TODO: utilize a direction and nulls/first last that has benefit
// for subsequent sort groups
List<Boolean> nullsFirst =
Lists.newArrayList(Collections.nCopies(sortExprs.size(), new Boolean(true)));
// then sort on orderByExprs
for (OrderByElement orderByElement : sortGroup.orderByElements) {
sortExprs.add(orderByElement.getExpr());
isAsc.add(orderByElement.isAsc());
nullsFirst.add(orderByElement.getNullsFirstParam());
}
SortInfo sortInfo = createSortInfo(root, sortExprs, isAsc, nullsFirst);
SortNode sortNode = new SortNode(idGenerator_.getNextId(), root, sortInfo, false, 0);
// if this sort group does not have partitioning exprs, we want the sort
// to be executed like a regular distributed sort
if (!partitionByExprs.isEmpty()) sortNode.setIsAnalyticSort(true);
if (partitionExprs != null) {
// create required input partition
DataPartition inputPartition = DataPartition.UNPARTITIONED;
if (!partitionExprs.isEmpty()) {
inputPartition = new DataPartition(TPartitionType.HASH_PARTITIONED, partitionExprs);
}
sortNode.setInputPartition(inputPartition);
}
root = sortNode;
root.init(analyzer_);
sortSmap = sortNode.getOutputSmap();
// create bufferedTupleDesc and bufferedSmap
sortTupleId = sortNode.tupleIds_.get(0);
bufferedTupleDesc = analyzer_.getDescTbl().copyTupleDescriptor(sortTupleId, "buffered-tuple");
LOG.trace("desctbl: " + analyzer_.getDescTbl().debugString());
List<SlotDescriptor> inputSlots = analyzer_.getTupleDesc(sortTupleId).getSlots();
List<SlotDescriptor> bufferedSlots = bufferedTupleDesc.getSlots();
for (int i = 0; i < inputSlots.size(); ++i) {
bufferedSmap.put(new SlotRef(inputSlots.get(i)), new SlotRef(bufferedSlots.get(i)));
}
}
// create one AnalyticEvalNode per window group
for (WindowGroup windowGroup : sortGroup.windowGroups) {
// Create partition-by (pb) and order-by (ob) less-than predicates between the
// input tuple (the output of the preceding sort) and a buffered tuple that is
// identical to the input tuple. We need a different tuple descriptor for the
// buffered tuple because the generated predicates should compare two different
// tuple instances from the same input stream (i.e., the predicates should be
// evaluated over a row that is composed of the input and the buffered tuple).
// we need to remap the pb/ob exprs to a) the sort output, b) our buffer of the
// sort input
Expr partitionByEq = null;
if (!windowGroup.partitionByExprs.isEmpty()) {
partitionByEq =
createNullMatchingEquals(
Expr.substituteList(windowGroup.partitionByExprs, sortSmap, analyzer_, false),
sortTupleId,
bufferedSmap);
LOG.trace("partitionByEq: " + partitionByEq.debugString());
}
Expr orderByEq = null;
if (!windowGroup.orderByElements.isEmpty()) {
orderByEq =
createNullMatchingEquals(
OrderByElement.getOrderByExprs(
OrderByElement.substitute(windowGroup.orderByElements, sortSmap, analyzer_)),
sortTupleId,
bufferedSmap);
LOG.trace("orderByEq: " + orderByEq.debugString());
}
root =
new AnalyticEvalNode(
idGenerator_.getNextId(),
root,
stmtTupleIds_,
windowGroup.analyticFnCalls,
windowGroup.partitionByExprs,
windowGroup.orderByElements,
windowGroup.window,
analyticInfo_.getOutputTupleDesc(),
windowGroup.physicalIntermediateTuple,
windowGroup.physicalOutputTuple,
windowGroup.logicalToPhysicalSmap,
partitionByEq,
orderByEq,
bufferedTupleDesc);
root.init(analyzer_);
}
return root;
}