Search in sources :

Example 71 with VectorContainer

use of org.apache.drill.exec.record.VectorContainer in project drill by axbaretto.

the class ExternalSortBatch method innerNext.

@SuppressWarnings("resource")
@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;
    // total number of batches received so far
    int totalBatches = 0;
    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) || // current memory used is more than 95% of memory usage limit of this operator
                        (oAllocator.getAllocatedMemory() > .95 * oAllocator.getLimit()) || // 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(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.getExecutorState().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);
            // no need to cleanup spilledBatchGroups, all it's batches are in batchGroups now
            spilledBatchGroups = null;
            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.getExecutorState().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.getExecutorState().fail(ex);
        return IterOutcome.STOP;
    } catch (UnsupportedOperationException e) {
        throw new RuntimeException(e);
    }
}
Also used : ClassTransformationException(org.apache.drill.exec.exception.ClassTransformationException) RecordBatchData(org.apache.drill.exec.physical.impl.sort.RecordBatchData) VectorWrapper(org.apache.drill.exec.record.VectorWrapper) Stopwatch(com.google.common.base.Stopwatch) SortRecordBatchBuilder(org.apache.drill.exec.physical.impl.sort.SortRecordBatchBuilder) IOException(java.io.IOException) VectorContainer(org.apache.drill.exec.record.VectorContainer) SchemaChangeException(org.apache.drill.exec.exception.SchemaChangeException) SelectionVector2(org.apache.drill.exec.record.selection.SelectionVector2) OutOfMemoryException(org.apache.drill.exec.exception.OutOfMemoryException)

Example 72 with VectorContainer

use of org.apache.drill.exec.record.VectorContainer in project drill by axbaretto.

the class BufferedBatches method add.

@SuppressWarnings("resource")
public void add(VectorAccessible incoming, long batchSize) {
    // Convert the incoming batch to the agreed-upon schema.
    // No converted batch means we got an empty input batch.
    // Converting the batch transfers memory ownership to our
    // allocator. This gives a round-about way to learn the batch
    // size: check the before and after memory levels, then use
    // the difference as the batch size, in bytes.
    VectorContainer convertedBatch = convertBatch(incoming);
    if (convertedBatch == null) {
        return;
    }
    SelectionVector2 sv2;
    try {
        sv2 = makeSelectionVector(incoming);
    } catch (Exception e) {
        convertedBatch.clear();
        throw e;
    }
    // Sort the incoming batch using either the original selection vector,
    // or a new one created here.
    sorterWrapper.sortBatch(convertedBatch, sv2);
    bufferBatch(convertedBatch, sv2, batchSize);
}
Also used : SelectionVector2(org.apache.drill.exec.record.selection.SelectionVector2) UserException(org.apache.drill.common.exceptions.UserException) OutOfMemoryException(org.apache.drill.exec.exception.OutOfMemoryException) VectorContainer(org.apache.drill.exec.record.VectorContainer)

Example 73 with VectorContainer

use of org.apache.drill.exec.record.VectorContainer in project drill by axbaretto.

the class FrameSupportTemplate method setup.

@Override
public void setup(final List<WindowDataBatch> batches, final VectorContainer container, final OperatorContext oContext, final boolean requireFullPartition, final WindowPOP popConfig) throws SchemaChangeException {
    this.container = container;
    this.batches = batches;
    internal = new VectorContainer(oContext);
    allocateInternal();
    outputCount = 0;
    this.requireFullPartition = requireFullPartition;
    this.popConfig = popConfig;
}
Also used : VectorContainer(org.apache.drill.exec.record.VectorContainer)

Example 74 with VectorContainer

use of org.apache.drill.exec.record.VectorContainer in project drill by axbaretto.

the class NoFrameSupportTemplate method setup.

@Override
public void setup(final List<WindowDataBatch> batches, final VectorContainer container, final OperatorContext oContext, final boolean requireFullPartition, final WindowPOP popConfig) throws SchemaChangeException {
    this.container = container;
    this.batches = batches;
    internal = new VectorContainer(oContext);
    allocateInternal();
    lagCopiedToInternal = false;
    outputCount = 0;
    partition = null;
    this.requireFullPartition = requireFullPartition;
}
Also used : VectorContainer(org.apache.drill.exec.record.VectorContainer)

Example 75 with VectorContainer

use of org.apache.drill.exec.record.VectorContainer in project drill by axbaretto.

the class BatchGroup method getBatch.

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.trace("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;
}
Also used : TransferPair(org.apache.drill.exec.record.TransferPair) VectorAccessibleSerializable(org.apache.drill.exec.cache.VectorAccessibleSerializable) VectorWrapper(org.apache.drill.exec.record.VectorWrapper) Stopwatch(com.google.common.base.Stopwatch) VectorContainer(org.apache.drill.exec.record.VectorContainer)

Aggregations

VectorContainer (org.apache.drill.exec.record.VectorContainer)178 Test (org.junit.Test)75 TupleMetadata (org.apache.drill.exec.record.metadata.TupleMetadata)63 SchemaBuilder (org.apache.drill.exec.record.metadata.SchemaBuilder)62 SubOperatorTest (org.apache.drill.test.SubOperatorTest)60 ValueVector (org.apache.drill.exec.vector.ValueVector)44 RowSetTest (org.apache.drill.categories.RowSetTest)41 RowSet (org.apache.drill.exec.physical.rowSet.RowSet)39 SingleRowSet (org.apache.drill.exec.physical.rowSet.RowSet.SingleRowSet)31 BatchSchema (org.apache.drill.exec.record.BatchSchema)27 ArrayList (java.util.ArrayList)23 MaterializedField (org.apache.drill.exec.record.MaterializedField)23 ResultSetLoader (org.apache.drill.exec.physical.resultSet.ResultSetLoader)18 SchemaChangeException (org.apache.drill.exec.exception.SchemaChangeException)17 RowSetComparison (org.apache.drill.test.rowSet.RowSetComparison)16 UserException (org.apache.drill.common.exceptions.UserException)15 RowSetLoader (org.apache.drill.exec.physical.resultSet.RowSetLoader)15 SelectionVector4 (org.apache.drill.exec.record.selection.SelectionVector4)15 OperatorTest (org.apache.drill.categories.OperatorTest)14 MockRecordBatch (org.apache.drill.exec.physical.impl.MockRecordBatch)14