Beispiel #1
0
 private VectorContainer getBatch() throws IOException {
   assert fs != null;
   assert path != null;
   if (inputStream == null) {
     inputStream = fs.open(path);
   }
   VectorAccessibleSerializable vas = new VectorAccessibleSerializable(allocator);
   Stopwatch watch = Stopwatch.createStarted();
   vas.readFromStream(inputStream);
   VectorContainer c = vas.get();
   if (schema != null) {
     c = SchemaUtil.coerceContainer(c, schema, context);
   }
   //    logger.debug("Took {} us to read {} records", watch.elapsed(TimeUnit.MICROSECONDS),
   // c.getRecordCount());
   spilledBatches--;
   currentContainer.zeroVectors();
   Iterator<VectorWrapper<?>> wrapperIterator = c.iterator();
   for (VectorWrapper w : currentContainer) {
     TransferPair pair =
         wrapperIterator.next().getValueVector().makeTransferPair(w.getValueVector());
     pair.transfer();
   }
   currentContainer.setRecordCount(c.getRecordCount());
   c.zeroVectors();
   return c;
 }
Beispiel #2
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;
  }
  private void buildTable()
      throws SchemaChangeException, ClassTransformationException, IOException {

    // Get all samples from distributed map

    SortRecordBatchBuilder containerBuilder =
        new SortRecordBatchBuilder(context.getAllocator(), MAX_SORT_BYTES);
    for (CachedVectorContainer w : mmap.get(mapKey)) {
      containerBuilder.add(w.get());
    }
    VectorContainer allSamplesContainer = new VectorContainer();
    containerBuilder.build(context, allSamplesContainer);

    List<Ordering> orderDefs = Lists.newArrayList();
    int i = 0;
    for (Ordering od : popConfig.getOrderings()) {
      SchemaPath sp = SchemaPath.getSimplePath("f" + i++);
      orderDefs.add(new Ordering(od.getDirection(), new FieldReference(sp)));
    }

    // sort the data incoming samples.
    SelectionVector4 newSv4 = containerBuilder.getSv4();
    Sorter sorter = SortBatch.createNewSorter(context, orderDefs, allSamplesContainer);
    sorter.setup(context, newSv4, allSamplesContainer);
    sorter.sort(newSv4, allSamplesContainer);

    // Copy every Nth record from the samples into a candidate partition table, where N =
    // totalSampledRecords/partitions
    // Attempt to push this to the distributed map. Only the first candidate to get pushed will be
    // used.
    VectorContainer candidatePartitionTable = new VectorContainer();
    SampleCopier copier = null;
    List<ValueVector> localAllocationVectors = Lists.newArrayList();
    copier =
        getCopier(
            newSv4,
            allSamplesContainer,
            candidatePartitionTable,
            orderDefs,
            localAllocationVectors);
    int allocationSize = 50;
    while (true) {
      for (ValueVector vv : localAllocationVectors) {
        AllocationHelper.allocate(vv, samplingFactor * partitions, allocationSize);
      }
      int skipRecords = containerBuilder.getSv4().getTotalCount() / partitions;
      if (copier.copyRecords(skipRecords, skipRecords, partitions - 1)) {
        assert copier.getOutputRecords() == partitions - 1
            : String.format(
                "output records: %d partitions: %d", copier.getOutputRecords(), partitions);
        for (VectorWrapper<?> vw : candidatePartitionTable) {
          vw.getValueVector().getMutator().setValueCount(copier.getOutputRecords());
        }
        break;
      } else {
        candidatePartitionTable.zeroVectors();
        allocationSize *= 2;
      }
    }
    candidatePartitionTable.setRecordCount(copier.getOutputRecords());
    WritableBatch batch =
        WritableBatch.getBatchNoHVWrap(
            candidatePartitionTable.getRecordCount(), candidatePartitionTable, false);
    CachedVectorContainer wrap =
        new CachedVectorContainer(batch, context.getDrillbitContext().getAllocator());
    tableMap.putIfAbsent(mapKey + "final", wrap, 1, TimeUnit.MINUTES);

    candidatePartitionTable.clear();
    allSamplesContainer.clear();
    containerBuilder.clear();
    wrap.clear();
  }
  private boolean saveSamples()
      throws SchemaChangeException, ClassTransformationException, IOException {
    recordsSampled = 0;
    IterOutcome upstream;

    // Start collecting batches until recordsToSample records have been collected

    SortRecordBatchBuilder builder =
        new SortRecordBatchBuilder(oContext.getAllocator(), MAX_SORT_BYTES);
    builder.add(incoming);

    recordsSampled += incoming.getRecordCount();

    outer:
    while (recordsSampled < recordsToSample) {
      upstream = next(incoming);
      switch (upstream) {
        case NONE:
        case NOT_YET:
        case STOP:
          upstreamNone = true;
          break outer;
        default:
          // fall through
      }
      builder.add(incoming);
      recordsSampled += incoming.getRecordCount();
      if (upstream == IterOutcome.NONE) break;
    }
    VectorContainer sortedSamples = new VectorContainer();
    builder.build(context, sortedSamples);

    // Sort the records according the orderings given in the configuration

    Sorter sorter = SortBatch.createNewSorter(context, popConfig.getOrderings(), sortedSamples);
    SelectionVector4 sv4 = builder.getSv4();
    sorter.setup(context, sv4, sortedSamples);
    sorter.sort(sv4, sortedSamples);

    // Project every Nth record to a new vector container, where N = recordsSampled/(samplingFactor
    // * partitions).
    // Uses the
    // the expressions from the Orderings to populate each column. There is one column for each
    // Ordering in
    // popConfig.orderings.

    VectorContainer containerToCache = new VectorContainer();
    List<ValueVector> localAllocationVectors = Lists.newArrayList();
    SampleCopier copier =
        getCopier(
            sv4, sortedSamples, containerToCache, popConfig.getOrderings(), localAllocationVectors);
    int allocationSize = 50;
    while (true) {
      for (ValueVector vv : localAllocationVectors) {
        AllocationHelper.allocate(vv, samplingFactor * partitions, allocationSize);
      }
      if (copier.copyRecords(
          recordsSampled / (samplingFactor * partitions), 0, samplingFactor * partitions)) {
        break;
      } else {
        containerToCache.zeroVectors();
        allocationSize *= 2;
      }
    }
    for (VectorWrapper<?> vw : containerToCache) {
      vw.getValueVector().getMutator().setValueCount(copier.getOutputRecords());
    }
    containerToCache.setRecordCount(copier.getOutputRecords());

    // Get a distributed multimap handle from the distributed cache, and put the vectors from the
    // new vector container
    // into a serializable wrapper object, and then add to distributed map

    WritableBatch batch =
        WritableBatch.getBatchNoHVWrap(containerToCache.getRecordCount(), containerToCache, false);
    CachedVectorContainer sampleToSave = new CachedVectorContainer(batch, context.getAllocator());

    mmap.put(mapKey, sampleToSave);
    this.sampledIncomingBatches = builder.getHeldRecordBatches();
    builder.clear();
    batch.clear();
    containerToCache.clear();
    sampleToSave.clear();
    return true;
  }