/**
   * Creates a copier that does a project for every Nth record from a VectorContainer incoming into
   * VectorContainer outgoing. Each Ordering in orderings generates a column, and evaluation of the
   * expression associated with each Ordering determines the value of each column. These records
   * will later be sorted based on the values in each column, in the same order as the orderings.
   *
   * @param sv4
   * @param incoming
   * @param outgoing
   * @param orderings
   * @return
   * @throws SchemaChangeException
   */
  private SampleCopier getCopier(
      SelectionVector4 sv4,
      VectorContainer incoming,
      VectorContainer outgoing,
      List<Ordering> orderings,
      List<ValueVector> localAllocationVectors)
      throws SchemaChangeException {
    final ErrorCollector collector = new ErrorCollectorImpl();
    final ClassGenerator<SampleCopier> cg =
        CodeGenerator.getRoot(SampleCopier.TEMPLATE_DEFINITION, context.getFunctionRegistry());

    int i = 0;
    for (Ordering od : orderings) {
      final LogicalExpression expr =
          ExpressionTreeMaterializer.materialize(
              od.getExpr(), incoming, collector, context.getFunctionRegistry());
      SchemaPath schemaPath = SchemaPath.getSimplePath("f" + i++);
      TypeProtos.MajorType.Builder builder =
          TypeProtos.MajorType.newBuilder()
              .mergeFrom(expr.getMajorType())
              .clearMode()
              .setMode(TypeProtos.DataMode.REQUIRED);
      TypeProtos.MajorType newType = builder.build();
      MaterializedField outputField = MaterializedField.create(schemaPath, newType);
      if (collector.hasErrors()) {
        throw new SchemaChangeException(
            String.format(
                "Failure while trying to materialize incoming schema.  Errors:\n %s.",
                collector.toErrorString()));
      }

      ValueVector vector = TypeHelper.getNewVector(outputField, oContext.getAllocator());
      localAllocationVectors.add(vector);
      TypedFieldId fid = outgoing.add(vector);
      ValueVectorWriteExpression write = new ValueVectorWriteExpression(fid, expr, true);
      HoldingContainer hc = cg.addExpr(write);
      cg.getEvalBlock()._if(hc.getValue().eq(JExpr.lit(0)))._then()._return(JExpr.FALSE);
    }
    cg.rotateBlock();
    cg.getEvalBlock()._return(JExpr.TRUE);
    outgoing.buildSchema(BatchSchema.SelectionVectorMode.NONE);
    try {
      SampleCopier sampleCopier = context.getImplementationClass(cg);
      sampleCopier.setupCopier(context, sv4, incoming, outgoing);
      return sampleCopier;
    } catch (ClassTransformationException | IOException e) {
      throw new SchemaChangeException(e);
    }
  }
Ejemplo n.º 2
0
 private VectorContainer constructHyperBatch(List<BatchGroup> batchGroupList) {
   VectorContainer cont = new VectorContainer();
   for (MaterializedField field : schema) {
     ValueVector[] vectors = new ValueVector[batchGroupList.size()];
     int i = 0;
     for (BatchGroup group : batchGroupList) {
       vectors[i++] =
           group
               .getValueAccessorById(
                   field.getValueClass(),
                   group.getValueVectorId(SchemaPath.getSimplePath(field.getPath())).getFieldIds())
               .getValueVector();
     }
     cont.add(vectors);
   }
   cont.buildSchema(BatchSchema.SelectionVectorMode.FOUR_BYTE);
   return cont;
 }
 @Override
 public void setNewSchema() throws SchemaChangeException {
   container.buildSchema(SelectionVectorMode.NONE);
 }
Ejemplo n.º 4
0
  public BatchGroup mergeAndSpill(LinkedList<BatchGroup> batchGroups) throws SchemaChangeException {
    logger.debug("Copier allocator current allocation {}", copierAllocator.getAllocatedMemory());
    logger.debug(
        "mergeAndSpill: starting total size in memory = {}", oAllocator.getAllocatedMemory());
    VectorContainer outputContainer = new VectorContainer();
    List<BatchGroup> batchGroupList = Lists.newArrayList();
    int batchCount = batchGroups.size();
    for (int i = 0; i < batchCount / 2; i++) {
      if (batchGroups.size() == 0) {
        break;
      }
      BatchGroup batch = batchGroups.pollLast();
      assert batch != null : "Encountered a null batch during merge and spill operation";
      batchGroupList.add(batch);
    }

    if (batchGroupList.size() == 0) {
      return null;
    }
    int estimatedRecordSize = 0;
    for (VectorWrapper<?> w : batchGroupList.get(0)) {
      try {
        estimatedRecordSize += TypeHelper.getSize(w.getField().getType());
      } catch (UnsupportedOperationException e) {
        estimatedRecordSize += 50;
      }
    }
    int targetRecordCount = Math.max(1, COPIER_BATCH_MEM_LIMIT / estimatedRecordSize);
    VectorContainer hyperBatch = constructHyperBatch(batchGroupList);
    createCopier(hyperBatch, batchGroupList, outputContainer, true);

    int count = copier.next(targetRecordCount);
    assert count > 0;

    logger.debug(
        "mergeAndSpill: estimated record size = {}, target record count = {}",
        estimatedRecordSize,
        targetRecordCount);

    // 1 output container is kept in memory, so we want to hold on to it and transferClone
    // allows keeping ownership
    VectorContainer c1 = VectorContainer.getTransferClone(outputContainer, oContext);
    c1.buildSchema(BatchSchema.SelectionVectorMode.NONE);
    c1.setRecordCount(count);

    String spillDir = dirs.next();
    Path currSpillPath = new Path(Joiner.on("/").join(spillDir, fileName));
    currSpillDirs.add(currSpillPath);
    String outputFile = Joiner.on("/").join(currSpillPath, spillCount++);
    try {
      fs.deleteOnExit(currSpillPath);
    } catch (IOException e) {
      // since this is meant to be used in a batches's spilling, we don't propagate the exception
      logger.warn("Unable to mark spill directory " + currSpillPath + " for deleting on exit", e);
    }
    stats.setLongStat(Metric.SPILL_COUNT, spillCount);
    BatchGroup newGroup = new BatchGroup(c1, fs, outputFile, oContext);
    try (AutoCloseable a = AutoCloseables.all(batchGroupList)) {
      logger.info("Merging and spilling to {}", outputFile);
      while ((count = copier.next(targetRecordCount)) > 0) {
        outputContainer.buildSchema(BatchSchema.SelectionVectorMode.NONE);
        outputContainer.setRecordCount(count);
        // note that addBatch also clears the outputContainer
        newGroup.addBatch(outputContainer);
      }
      injector.injectChecked(
          context.getExecutionControls(), INTERRUPTION_WHILE_SPILLING, IOException.class);
      newGroup.closeOutputStream();
    } catch (Throwable e) {
      // we only need to cleanup newGroup if spill failed
      try {
        AutoCloseables.close(e, newGroup);
      } catch (Throwable t) {
        /* close() may hit the same IO issue; just ignore */
      }
      throw UserException.resourceError(e)
          .message("External Sort encountered an error while spilling to disk")
          .addContext(e.getMessage() /* more detail */)
          .build(logger);
    } finally {
      hyperBatch.clear();
    }
    logger.debug("mergeAndSpill: final total size in memory = {}", oAllocator.getAllocatedMemory());
    logger.info("Completed spilling to {}", outputFile);
    return newGroup;
  }