Examples with VectorContainer org.apache.drill.exec.record.VectorContainer used on opensource projects

Example 11 with VectorContainer

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

the class ExternalSortBatch method mergeAndSpill.

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;
        }
        @SuppressWarnings("resource") 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()).build(logger);
    } finally {
        hyperBatch.clear();
    }
    logger.debug("mergeAndSpill: final total size in memory = {}", oAllocator.getAllocatedMemory());
    logger.info("Completed spilling to {}", outputFile);
    return newGroup;
}

Example 12 with VectorContainer

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

the class ExternalSortBatch method constructHyperBatch.

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;
}

Example 13 with VectorContainer

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

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;
}

Example 14 with VectorContainer

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

the class TestPartitionSender method testPartitionSenderCostToThreads.

@Test
public /**
   * Main test to go over different scenarios
   * @throws Exception
   */
void testPartitionSenderCostToThreads() throws Exception {
    final VectorContainer container = new VectorContainer();
    container.buildSchema(SelectionVectorMode.FOUR_BYTE);
    final SelectionVector4 sv = Mockito.mock(SelectionVector4.class, "SelectionVector4");
    Mockito.when(sv.getCount()).thenReturn(100);
    Mockito.when(sv.getTotalCount()).thenReturn(100);
    for (int i = 0; i < 100; i++) {
        Mockito.when(sv.get(i)).thenReturn(i);
    }
    final TopNBatch.SimpleRecordBatch incoming = new TopNBatch.SimpleRecordBatch(container, sv, null);
    updateTestCluster(DRILLBITS_COUNT, null);
    test("ALTER SESSION SET `planner.slice_target`=1");
    String plan = getPlanInString("EXPLAIN PLAN FOR " + groupByQuery, JSON_FORMAT);
    System.out.println("Plan: " + plan);
    final DrillbitContext drillbitContext = getDrillbitContext();
    final PhysicalPlanReader planReader = drillbitContext.getPlanReader();
    final PhysicalPlan physicalPlan = planReader.readPhysicalPlan(plan);
    final Fragment rootFragment = PopUnitTestBase.getRootFragmentFromPlanString(planReader, plan);
    final PlanningSet planningSet = new PlanningSet();
    final FunctionImplementationRegistry registry = new FunctionImplementationRegistry(config);
    // Create a planningSet to get the assignment of major fragment ids to fragments.
    PARALLELIZER.initFragmentWrappers(rootFragment, planningSet);
    final List<PhysicalOperator> operators = physicalPlan.getSortedOperators(false);
    // get HashToRandomExchange physical operator
    HashToRandomExchange hashToRandomExchange = null;
    for (PhysicalOperator operator : operators) {
        if (operator instanceof HashToRandomExchange) {
            hashToRandomExchange = (HashToRandomExchange) operator;
            break;
        }
    }
    final OptionList options = new OptionList();
    // try multiple scenarios with different set of options
    options.add(OptionValue.createLong(OptionType.SESSION, "planner.slice_target", 1));
    testThreadsHelper(hashToRandomExchange, drillbitContext, options, incoming, registry, planReader, planningSet, rootFragment, 1);
    options.clear();
    options.add(OptionValue.createLong(OptionType.SESSION, "planner.slice_target", 1));
    options.add(OptionValue.createLong(OptionType.SESSION, "planner.partitioner_sender_max_threads", 10));
    hashToRandomExchange.setCost(1000);
    testThreadsHelper(hashToRandomExchange, drillbitContext, options, incoming, registry, planReader, planningSet, rootFragment, 10);
    options.clear();
    options.add(OptionValue.createLong(OptionType.SESSION, "planner.slice_target", 1000));
    options.add(OptionValue.createLong(OptionType.SESSION, "planner.partitioner_sender_threads_factor", 2));
    hashToRandomExchange.setCost(14000);
    testThreadsHelper(hashToRandomExchange, drillbitContext, options, incoming, registry, planReader, planningSet, rootFragment, 2);
}

Example 15 with VectorContainer

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

the class PruneScanRule method doOnMatch.

protected void doOnMatch(RelOptRuleCall call, Filter filterRel, Project projectRel, TableScan scanRel) {
    final String pruningClassName = getClass().getName();
    logger.info("Beginning partition pruning, pruning class: {}", pruningClassName);
    Stopwatch totalPruningTime = Stopwatch.createStarted();
    final PlannerSettings settings = PrelUtil.getPlannerSettings(call.getPlanner());
    PartitionDescriptor descriptor = getPartitionDescriptor(settings, scanRel);
    final BufferAllocator allocator = optimizerContext.getAllocator();
    final Object selection = getDrillTable(scanRel).getSelection();
    MetadataContext metaContext = null;
    if (selection instanceof FormatSelection) {
        metaContext = ((FormatSelection) selection).getSelection().getMetaContext();
    }
    RexNode condition = null;
    if (projectRel == null) {
        condition = filterRel.getCondition();
    } else {
        // get the filter as if it were below the projection.
        condition = RelOptUtil.pushFilterPastProject(filterRel.getCondition(), projectRel);
    }
    RewriteAsBinaryOperators visitor = new RewriteAsBinaryOperators(true, filterRel.getCluster().getRexBuilder());
    condition = condition.accept(visitor);
    Map<Integer, String> fieldNameMap = Maps.newHashMap();
    List<String> fieldNames = scanRel.getRowType().getFieldNames();
    BitSet columnBitset = new BitSet();
    BitSet partitionColumnBitSet = new BitSet();
    Map<Integer, Integer> partitionMap = Maps.newHashMap();
    int relColIndex = 0;
    for (String field : fieldNames) {
        final Integer partitionIndex = descriptor.getIdIfValid(field);
        if (partitionIndex != null) {
            fieldNameMap.put(partitionIndex, field);
            partitionColumnBitSet.set(partitionIndex);
            columnBitset.set(relColIndex);
            // mapping between the relColIndex and partitionIndex
            partitionMap.put(relColIndex, partitionIndex);
        }
        relColIndex++;
    }
    if (partitionColumnBitSet.isEmpty()) {
        logger.info("No partition columns are projected from the scan..continue. " + "Total pruning elapsed time: {} ms", totalPruningTime.elapsed(TimeUnit.MILLISECONDS));
        setPruneStatus(metaContext, PruneStatus.NOT_PRUNED);
        return;
    }
    // stop watch to track how long we spend in different phases of pruning
    Stopwatch miscTimer = Stopwatch.createUnstarted();
    // track how long we spend building the filter tree
    miscTimer.start();
    FindPartitionConditions c = new FindPartitionConditions(columnBitset, filterRel.getCluster().getRexBuilder());
    c.analyze(condition);
    RexNode pruneCondition = c.getFinalCondition();
    BitSet referencedDirsBitSet = c.getReferencedDirs();
    logger.info("Total elapsed time to build and analyze filter tree: {} ms", miscTimer.elapsed(TimeUnit.MILLISECONDS));
    miscTimer.reset();
    if (pruneCondition == null) {
        logger.info("No conditions were found eligible for partition pruning." + "Total pruning elapsed time: {} ms", totalPruningTime.elapsed(TimeUnit.MILLISECONDS));
        setPruneStatus(metaContext, PruneStatus.NOT_PRUNED);
        return;
    }
    // set up the partitions
    List<PartitionLocation> newPartitions = Lists.newArrayList();
    // total number of partitions
    long numTotal = 0;
    int batchIndex = 0;
    PartitionLocation firstLocation = null;
    LogicalExpression materializedExpr = null;
    String[] spInfo = null;
    int maxIndex = -1;
    BitSet matchBitSet = new BitSet();
    // Outer loop: iterate over a list of batches of PartitionLocations
    for (List<PartitionLocation> partitions : descriptor) {
        numTotal += partitions.size();
        logger.debug("Evaluating partition pruning for batch {}", batchIndex);
        if (batchIndex == 0) {
            // save the first location in case everything is pruned
            firstLocation = partitions.get(0);
        }
        final NullableBitVector output = new NullableBitVector(MaterializedField.create("", Types.optional(MinorType.BIT)), allocator);
        final VectorContainer container = new VectorContainer();
        try {
            final ValueVector[] vectors = new ValueVector[descriptor.getMaxHierarchyLevel()];
            for (int partitionColumnIndex : BitSets.toIter(partitionColumnBitSet)) {
                SchemaPath column = SchemaPath.getSimplePath(fieldNameMap.get(partitionColumnIndex));
                MajorType type = descriptor.getVectorType(column, settings);
                MaterializedField field = MaterializedField.create(column.getAsUnescapedPath(), type);
                ValueVector v = TypeHelper.getNewVector(field, allocator);
                v.allocateNew();
                vectors[partitionColumnIndex] = v;
                container.add(v);
            }
            // track how long we spend populating partition column vectors
            miscTimer.start();
            // populate partition vectors.
            descriptor.populatePartitionVectors(vectors, partitions, partitionColumnBitSet, fieldNameMap);
            logger.info("Elapsed time to populate partitioning column vectors: {} ms within batchIndex: {}", miscTimer.elapsed(TimeUnit.MILLISECONDS), batchIndex);
            miscTimer.reset();
            // materialize the expression; only need to do this once
            if (batchIndex == 0) {
                materializedExpr = materializePruneExpr(pruneCondition, settings, scanRel, container);
                if (materializedExpr == null) {
                    // continue without partition pruning; no need to log anything here since
                    // materializePruneExpr logs it already
                    logger.info("Total pruning elapsed time: {} ms", totalPruningTime.elapsed(TimeUnit.MILLISECONDS));
                    setPruneStatus(metaContext, PruneStatus.NOT_PRUNED);
                    return;
                }
            }
            output.allocateNew(partitions.size());
            // start the timer to evaluate how long we spend in the interpreter evaluation
            miscTimer.start();
            InterpreterEvaluator.evaluate(partitions.size(), optimizerContext, container, output, materializedExpr);
            logger.info("Elapsed time in interpreter evaluation: {} ms within batchIndex: {} with # of partitions : {}", miscTimer.elapsed(TimeUnit.MILLISECONDS), batchIndex, partitions.size());
            miscTimer.reset();
            int recordCount = 0;
            int qualifiedCount = 0;
            if (descriptor.supportsMetadataCachePruning() && partitions.get(0).isCompositePartition()) /* apply single partition check only for composite partitions */
            {
                // Inner loop: within each batch iterate over the PartitionLocations
                for (PartitionLocation part : partitions) {
                    assert part.isCompositePartition();
                    if (!output.getAccessor().isNull(recordCount) && output.getAccessor().get(recordCount) == 1) {
                        newPartitions.add(part);
                        // Rather than using the PartitionLocation, get the array of partition values for the directories that are
                        // referenced by the filter since we are not interested in directory references in other parts of the query.
                        Pair<String[], Integer> p = composePartition(referencedDirsBitSet, partitionMap, vectors, recordCount);
                        String[] parts = p.getLeft();
                        int tmpIndex = p.getRight();
                        maxIndex = Math.max(maxIndex, tmpIndex);
                        if (spInfo == null) {
                            // initialization
                            spInfo = parts;
                            for (int j = 0; j <= tmpIndex; j++) {
                                if (parts[j] != null) {
                                    matchBitSet.set(j);
                                }
                            }
                        } else {
                            // compare the new partition with existing partition
                            for (int j = 0; j <= tmpIndex; j++) {
                                if (parts[j] == null || spInfo[j] == null) {
                                    // nulls don't match
                                    matchBitSet.clear(j);
                                } else {
                                    if (!parts[j].equals(spInfo[j])) {
                                        matchBitSet.clear(j);
                                    }
                                }
                            }
                        }
                        qualifiedCount++;
                    }
                    recordCount++;
                }
            } else {
                // Inner loop: within each batch iterate over the PartitionLocations
                for (PartitionLocation part : partitions) {
                    if (!output.getAccessor().isNull(recordCount) && output.getAccessor().get(recordCount) == 1) {
                        newPartitions.add(part);
                        qualifiedCount++;
                    }
                    recordCount++;
                }
            }
            logger.debug("Within batch {}: total records: {}, qualified records: {}", batchIndex, recordCount, qualifiedCount);
            batchIndex++;
        } catch (Exception e) {
            logger.warn("Exception while trying to prune partition.", e);
            logger.info("Total pruning elapsed time: {} ms", totalPruningTime.elapsed(TimeUnit.MILLISECONDS));
            setPruneStatus(metaContext, PruneStatus.NOT_PRUNED);
            // continue without partition pruning
            return;
        } finally {
            container.clear();
            if (output != null) {
                output.clear();
            }
        }
    }
    try {
        if (newPartitions.size() == numTotal) {
            logger.info("No partitions were eligible for pruning");
            return;
        }
        // handle the case all partitions are filtered out.
        boolean canDropFilter = true;
        boolean wasAllPartitionsPruned = false;
        String cacheFileRoot = null;
        if (newPartitions.isEmpty()) {
            assert firstLocation != null;
            // Add the first non-composite partition location, since execution requires schema.
            // In such case, we should not drop filter.
            newPartitions.add(firstLocation.getPartitionLocationRecursive().get(0));
            canDropFilter = false;
            // NOTE: with DRILL-4530, the PruneScanRule may be called with only a list of
            // directories first and the non-composite partition location will still return
            // directories, not files.  So, additional processing is done depending on this flag
            wasAllPartitionsPruned = true;
            logger.info("All {} partitions were pruned; added back a single partition to allow creating a schema", numTotal);
            // set the cacheFileRoot appropriately
            if (firstLocation.isCompositePartition()) {
                cacheFileRoot = descriptor.getBaseTableLocation() + firstLocation.getCompositePartitionPath();
            }
        }
        logger.info("Pruned {} partitions down to {}", numTotal, newPartitions.size());
        List<RexNode> conjuncts = RelOptUtil.conjunctions(condition);
        List<RexNode> pruneConjuncts = RelOptUtil.conjunctions(pruneCondition);
        conjuncts.removeAll(pruneConjuncts);
        RexNode newCondition = RexUtil.composeConjunction(filterRel.getCluster().getRexBuilder(), conjuncts, false);
        RewriteCombineBinaryOperators reverseVisitor = new RewriteCombineBinaryOperators(true, filterRel.getCluster().getRexBuilder());
        condition = condition.accept(reverseVisitor);
        pruneCondition = pruneCondition.accept(reverseVisitor);
        if (descriptor.supportsMetadataCachePruning() && !wasAllPartitionsPruned) {
            // if metadata cache file could potentially be used, then assign a proper cacheFileRoot
            int index = -1;
            if (!matchBitSet.isEmpty()) {
                String path = "";
                index = matchBitSet.length() - 1;
                for (int j = 0; j < matchBitSet.length(); j++) {
                    if (!matchBitSet.get(j)) {
                        // stop at the first index with no match and use the immediate
                        // previous index
                        index = j - 1;
                        break;
                    }
                }
                for (int j = 0; j <= index; j++) {
                    path += "/" + spInfo[j];
                }
                cacheFileRoot = descriptor.getBaseTableLocation() + path;
            }
            if (index != maxIndex) {
                // if multiple partitions are being selected, we should not drop the filter
                // since we are reading the cache file at a parent/ancestor level
                canDropFilter = false;
            }
        }
        RelNode inputRel = descriptor.supportsMetadataCachePruning() ? descriptor.createTableScan(newPartitions, cacheFileRoot, wasAllPartitionsPruned, metaContext) : descriptor.createTableScan(newPartitions, wasAllPartitionsPruned);
        if (projectRel != null) {
            inputRel = projectRel.copy(projectRel.getTraitSet(), Collections.singletonList(inputRel));
        }
        if (newCondition.isAlwaysTrue() && canDropFilter) {
            call.transformTo(inputRel);
        } else {
            final RelNode newFilter = filterRel.copy(filterRel.getTraitSet(), Collections.singletonList(inputRel));
            call.transformTo(newFilter);
        }
        setPruneStatus(metaContext, PruneStatus.PRUNED);
    } catch (Exception e) {
        logger.warn("Exception while using the pruned partitions.", e);
    } finally {
        logger.info("Total pruning elapsed time: {} ms", totalPruningTime.elapsed(TimeUnit.MILLISECONDS));
    }
}