Esempio n. 1
0
  private void getIndex(ClassGenerator<StreamingAggregator> g) {
    switch (incoming.getSchema().getSelectionVectorMode()) {
      case FOUR_BYTE:
        {
          JVar var = g.declareClassField("sv4_", g.getModel()._ref(SelectionVector4.class));
          g.getBlock("setupInterior")
              .assign(var, JExpr.direct("incoming").invoke("getSelectionVector4"));
          g.getBlock("getVectorIndex")._return(var.invoke("get").arg(JExpr.direct("recordIndex")));
          ;
          return;
        }
      case NONE:
        {
          g.getBlock("getVectorIndex")._return(JExpr.direct("recordIndex"));
          ;
          return;
        }
      case TWO_BYTE:
        {
          JVar var = g.declareClassField("sv2_", g.getModel()._ref(SelectionVector2.class));
          g.getBlock("setupInterior")
              .assign(var, JExpr.direct("incoming").invoke("getSelectionVector2"));
          g.getBlock("getVectorIndex")
              ._return(var.invoke("getIndex").arg(JExpr.direct("recordIndex")));
          ;
          return;
        }

      default:
        throw new IllegalStateException();
    }
  }
  @Override
  public void setup(
      FragmentContext context, RecordBatch incoming, RecordBatch outgoing, TransferPair[] transfers)
      throws SchemaChangeException {
    this.transfers = transfers;
    this.outgoingSelectionVector = outgoing.getSelectionVector2();
    this.svMode = incoming.getSchema().getSelectionVectorMode();

    switch (svMode) {
      case NONE:
        break;
      case TWO_BYTE:
        this.incomingSelectionVector = incoming.getSelectionVector2();
        break;
      default:
        throw new UnsupportedOperationException();
    }
    doSetup(context, incoming, outgoing);
  }
Esempio n. 3
0
  @Override
  public IterOutcome innerNext() {
    if (schema != null) {
      if (spillCount == 0) {
        return (getSelectionVector4().next()) ? IterOutcome.OK : IterOutcome.NONE;
      } else {
        Stopwatch w = Stopwatch.createStarted();
        int count = copier.next(targetRecordCount);
        if (count > 0) {
          long t = w.elapsed(TimeUnit.MICROSECONDS);
          logger.debug("Took {} us to merge {} records", t, count);
          container.setRecordCount(count);
          return IterOutcome.OK;
        } else {
          logger.debug("copier returned 0 records");
          return IterOutcome.NONE;
        }
      }
    }

    int totalCount = 0;
    int totalBatches = 0; // total number of batches received so far

    try {
      container.clear();
      outer:
      while (true) {
        IterOutcome upstream;
        if (first) {
          upstream = IterOutcome.OK_NEW_SCHEMA;
        } else {
          upstream = next(incoming);
        }
        if (upstream == IterOutcome.OK && sorter == null) {
          upstream = IterOutcome.OK_NEW_SCHEMA;
        }
        switch (upstream) {
          case NONE:
            if (first) {
              return upstream;
            }
            break outer;
          case NOT_YET:
            throw new UnsupportedOperationException();
          case STOP:
            return upstream;
          case OK_NEW_SCHEMA:
          case OK:
            VectorContainer convertedBatch;
            // only change in the case that the schema truly changes.  Artificial schema changes are
            // ignored.
            if (upstream == IterOutcome.OK_NEW_SCHEMA && !incoming.getSchema().equals(schema)) {
              if (schema != null) {
                if (unionTypeEnabled) {
                  this.schema = SchemaUtil.mergeSchemas(schema, incoming.getSchema());
                } else {
                  throw new SchemaChangeException(
                      "Schema changes not supported in External Sort. Please enable Union type");
                }
              } else {
                schema = incoming.getSchema();
              }
              convertedBatch = SchemaUtil.coerceContainer(incoming, schema, oContext);
              for (BatchGroup b : batchGroups) {
                b.setSchema(schema);
              }
              for (BatchGroup b : spilledBatchGroups) {
                b.setSchema(schema);
              }
              this.sorter = createNewSorter(context, convertedBatch);
            } else {
              convertedBatch = SchemaUtil.coerceContainer(incoming, schema, oContext);
            }
            if (first) {
              first = false;
            }
            if (convertedBatch.getRecordCount() == 0) {
              for (VectorWrapper<?> w : convertedBatch) {
                w.clear();
              }
              break;
            }
            SelectionVector2 sv2;
            if (incoming.getSchema().getSelectionVectorMode()
                == BatchSchema.SelectionVectorMode.TWO_BYTE) {
              sv2 = incoming.getSelectionVector2().clone();
            } else {
              try {
                sv2 = newSV2();
              } catch (InterruptedException e) {
                return IterOutcome.STOP;
              } catch (OutOfMemoryException e) {
                throw new OutOfMemoryException(e);
              }
            }

            int count = sv2.getCount();
            totalCount += count;
            totalBatches++;
            sorter.setup(context, sv2, convertedBatch);
            sorter.sort(sv2);
            RecordBatchData rbd = new RecordBatchData(convertedBatch, oAllocator);
            boolean success = false;
            try {
              rbd.setSv2(sv2);
              batchGroups.add(new BatchGroup(rbd.getContainer(), rbd.getSv2(), oContext));
              if (peakNumBatches < batchGroups.size()) {
                peakNumBatches = batchGroups.size();
                stats.setLongStat(Metric.PEAK_BATCHES_IN_MEMORY, peakNumBatches);
              }

              batchesSinceLastSpill++;
              if ( // If we haven't spilled so far, do we have enough memory for MSorter if this
              // turns out to be the last incoming batch?
              (spillCount == 0 && !hasMemoryForInMemorySort(totalCount))
                  ||
                  // If we haven't spilled so far, make sure we don't exceed the maximum number of
                  // batches SV4 can address
                  (spillCount == 0 && totalBatches > Character.MAX_VALUE)
                  ||
                  // TODO(DRILL-4438) - consider setting this threshold more intelligently,
                  // lowering caused a failing low memory condition (test in
                  // BasicPhysicalOpUnitTest)
                  // to complete successfully (although it caused perf decrease as there was more
                  // spilling)

                  // current memory used is more than 95% of memory usage limit of this operator
                  (oAllocator.getAllocatedMemory() > .95 * oAllocator.getLimit())
                  ||
                  // Number of incoming batches (BatchGroups) exceed the limit and number of
                  // incoming batches accumulated
                  // since the last spill exceed the defined limit
                  (batchGroups.size() > SPILL_THRESHOLD
                      && batchesSinceLastSpill >= SPILL_BATCH_GROUP_SIZE)) {

                if (firstSpillBatchCount == 0) {
                  firstSpillBatchCount = batchGroups.size();
                }

                if (spilledBatchGroups.size() > firstSpillBatchCount / 2) {
                  logger.info("Merging spills");
                  final BatchGroup merged = mergeAndSpill(spilledBatchGroups);
                  if (merged != null) {
                    spilledBatchGroups.addFirst(merged);
                  }
                }
                final BatchGroup merged = mergeAndSpill(batchGroups);
                if (merged != null) { // make sure we don't add null to spilledBatchGroups
                  spilledBatchGroups.add(merged);
                  batchesSinceLastSpill = 0;
                }
              }
              success = true;
            } finally {
              if (!success) {
                rbd.clear();
              }
            }
            break;
          case OUT_OF_MEMORY:
            logger.debug("received OUT_OF_MEMORY, trying to spill");
            if (batchesSinceLastSpill > 2) {
              final BatchGroup merged = mergeAndSpill(batchGroups);
              if (merged != null) {
                spilledBatchGroups.add(merged);
                batchesSinceLastSpill = 0;
              }
            } else {
              logger.debug("not enough batches to spill, sending OUT_OF_MEMORY downstream");
              return IterOutcome.OUT_OF_MEMORY;
            }
            break;
          default:
            throw new UnsupportedOperationException();
        }
      }

      if (totalCount == 0) {
        return IterOutcome.NONE;
      }
      if (spillCount == 0) {

        if (builder != null) {
          builder.clear();
          builder.close();
        }
        builder = new SortRecordBatchBuilder(oAllocator);

        for (BatchGroup group : batchGroups) {
          RecordBatchData rbd = new RecordBatchData(group.getContainer(), oAllocator);
          rbd.setSv2(group.getSv2());
          builder.add(rbd);
        }

        builder.build(context, container);
        sv4 = builder.getSv4();
        mSorter = createNewMSorter();
        mSorter.setup(context, oAllocator, getSelectionVector4(), this.container);

        // For testing memory-leak purpose, inject exception after mSorter finishes setup
        injector.injectUnchecked(context.getExecutionControls(), INTERRUPTION_AFTER_SETUP);
        mSorter.sort(this.container);

        // sort may have prematurely exited due to should continue returning false.
        if (!context.shouldContinue()) {
          return IterOutcome.STOP;
        }

        // For testing memory-leak purpose, inject exception after mSorter finishes sorting
        injector.injectUnchecked(context.getExecutionControls(), INTERRUPTION_AFTER_SORT);
        sv4 = mSorter.getSV4();

        container.buildSchema(SelectionVectorMode.FOUR_BYTE);
      } else { // some batches were spilled
        final BatchGroup merged = mergeAndSpill(batchGroups);
        if (merged != null) {
          spilledBatchGroups.add(merged);
        }
        batchGroups.addAll(spilledBatchGroups);
        spilledBatchGroups =
            null; // no need to cleanup spilledBatchGroups, all it's batches are in batchGroups now

        logger.warn(
            "Starting to merge. {} batch groups. Current allocated memory: {}",
            batchGroups.size(),
            oAllocator.getAllocatedMemory());
        VectorContainer hyperBatch = constructHyperBatch(batchGroups);
        createCopier(hyperBatch, batchGroups, container, false);

        int estimatedRecordSize = 0;
        for (VectorWrapper<?> w : batchGroups.get(0)) {
          try {
            estimatedRecordSize += TypeHelper.getSize(w.getField().getType());
          } catch (UnsupportedOperationException e) {
            estimatedRecordSize += 50;
          }
        }
        targetRecordCount =
            Math.min(MAX_BATCH_SIZE, Math.max(1, COPIER_BATCH_MEM_LIMIT / estimatedRecordSize));
        int count = copier.next(targetRecordCount);
        container.buildSchema(SelectionVectorMode.NONE);
        container.setRecordCount(count);
      }

      return IterOutcome.OK_NEW_SCHEMA;

    } catch (SchemaChangeException ex) {
      kill(false);
      context.fail(
          UserException.unsupportedError(ex)
              .message("Sort doesn't currently support sorts with changing schemas")
              .build(logger));
      return IterOutcome.STOP;
    } catch (ClassTransformationException | IOException ex) {
      kill(false);
      context.fail(ex);
      return IterOutcome.STOP;
    } catch (UnsupportedOperationException e) {
      throw new RuntimeException(e);
    }
  }