Exemplo n.º 1
0
  /** Create SortInfo, including sort tuple, to sort entire input row on sortExprs. */
  private SortInfo createSortInfo(
      PlanNode input, List<Expr> sortExprs, List<Boolean> isAsc, List<Boolean> nullsFirst) {
    // create tuple for sort output = the entire materialized input in a single tuple
    TupleDescriptor sortTupleDesc = analyzer_.getDescTbl().createTupleDescriptor("sort-tuple");
    ExprSubstitutionMap sortSmap = new ExprSubstitutionMap();
    List<Expr> sortSlotExprs = Lists.newArrayList();
    sortTupleDesc.setIsMaterialized(true);
    for (TupleId tid : input.getTupleIds()) {
      TupleDescriptor tupleDesc = analyzer_.getTupleDesc(tid);
      for (SlotDescriptor inputSlotDesc : tupleDesc.getSlots()) {
        if (!inputSlotDesc.isMaterialized()) continue;
        SlotDescriptor sortSlotDesc = analyzer_.copySlotDescriptor(inputSlotDesc, sortTupleDesc);
        // all output slots need to be materialized
        sortSlotDesc.setIsMaterialized(true);
        sortSmap.put(new SlotRef(inputSlotDesc), new SlotRef(sortSlotDesc));
        sortSlotExprs.add(new SlotRef(inputSlotDesc));
      }
    }

    SortInfo sortInfo =
        new SortInfo(Expr.substituteList(sortExprs, sortSmap, analyzer_, false), isAsc, nullsFirst);
    LOG.trace("sortinfo exprs: " + Expr.debugString(sortInfo.getOrderingExprs()));
    sortInfo.setMaterializedTupleInfo(sortTupleDesc, sortSlotExprs);
    return sortInfo;
  }
Exemplo n.º 2
0
 /** Returns an smap that combines the childrens' smaps. */
 protected ExprSubstitutionMap getCombinedChildSmap() {
   if (getChildren().size() == 0) return new ExprSubstitutionMap();
   if (getChildren().size() == 1) return getChild(0).getOutputSmap();
   ExprSubstitutionMap result =
       ExprSubstitutionMap.combine(getChild(0).getOutputSmap(), getChild(1).getOutputSmap());
   for (int i = 2; i < getChildren().size(); ++i) {
     result = ExprSubstitutionMap.combine(result, getChild(i).getOutputSmap());
   }
   return result;
 }
Exemplo n.º 3
0
    /**
     * Creates the physical output and intermediate tuples as well as the logical to physical smap
     * for this window group. Computes the mem layout for the tuple descriptors.
     */
    public void init(Analyzer analyzer, String tupleName) {
      Preconditions.checkState(physicalOutputTuple == null);
      Preconditions.checkState(physicalIntermediateTuple == null);
      Preconditions.checkState(analyticFnCalls.size() == analyticExprs.size());

      // If needed, create the intermediate tuple first to maintain
      // intermediateTupleId < outputTupleId for debugging purposes and consistency with
      // tuple creation for aggregations.
      boolean requiresIntermediateTuple =
          AggregateInfoBase.requiresIntermediateTuple(analyticFnCalls);
      if (requiresIntermediateTuple) {
        physicalIntermediateTuple =
            analyzer.getDescTbl().createTupleDescriptor(tupleName + "intermed");
        physicalOutputTuple = analyzer.getDescTbl().createTupleDescriptor(tupleName + "out");
      } else {
        physicalOutputTuple = analyzer.getDescTbl().createTupleDescriptor(tupleName + "out");
        physicalIntermediateTuple = physicalOutputTuple;
      }

      Preconditions.checkState(analyticExprs.size() == logicalIntermediateSlots.size());
      Preconditions.checkState(analyticExprs.size() == logicalOutputSlots.size());
      for (int i = 0; i < analyticExprs.size(); ++i) {
        SlotDescriptor logicalOutputSlot = logicalOutputSlots.get(i);
        SlotDescriptor physicalOutputSlot =
            analyzer.copySlotDescriptor(logicalOutputSlot, physicalOutputTuple);
        physicalOutputSlot.setIsMaterialized(true);
        if (requiresIntermediateTuple) {
          SlotDescriptor logicalIntermediateSlot = logicalIntermediateSlots.get(i);
          SlotDescriptor physicalIntermediateSlot =
              analyzer.copySlotDescriptor(logicalIntermediateSlot, physicalIntermediateTuple);
          physicalIntermediateSlot.setIsMaterialized(true);
        }
        logicalToPhysicalSmap.put(new SlotRef(logicalOutputSlot), new SlotRef(physicalOutputSlot));
      }
      physicalOutputTuple.computeMemLayout();
      if (requiresIntermediateTuple) physicalIntermediateTuple.computeMemLayout();
    }
Exemplo n.º 4
0
 /**
  * Sets outputSmap_ to compose(existing smap, combined child smap). Also substitutes conjuncts_
  * using the combined child smap.
  */
 protected void createDefaultSmap(Analyzer analyzer) {
   ExprSubstitutionMap combinedChildSmap = getCombinedChildSmap();
   outputSmap_ = ExprSubstitutionMap.compose(outputSmap_, combinedChildSmap, analyzer);
   conjuncts_ = Expr.substituteList(conjuncts_, outputSmap_, analyzer, false);
 }
Exemplo n.º 5
0
  /**
   * 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;
  }