Search in sources :

Example 41 with BufferAllocator

use of org.apache.drill.exec.memory.BufferAllocator in project drill by apache.

the class TestLoad method testLoadValueVectorEmptyVarCharArray.

@Test
public void testLoadValueVectorEmptyVarCharArray() throws Exception {
    try (BufferAllocator allocator = RootAllocatorFactory.newRoot(drillConfig)) {
        TupleMetadata schema = new SchemaBuilder().addArray("chars", MinorType.VARCHAR).build();
        ResultSetLoaderImpl.ResultSetOptions options = new ResultSetOptionBuilder().readerSchema(schema).build();
        ResultSetLoader resultSetLoader = new ResultSetLoaderImpl(allocator, options);
        resultSetLoader.startBatch();
        RowSetLoader rowWriter = resultSetLoader.writer();
        rowWriter.addRow(new Object[] { null });
        VectorContainer harvest = resultSetLoader.harvest();
        // Create vectors
        List<ValueVector> vectors = StreamSupport.stream(harvest.spliterator(), false).map(VectorWrapper::getValueVector).collect(Collectors.toList());
        // Writeable batch now owns vector buffers
        WritableBatch writableBatch = WritableBatch.getBatchNoHV(1, vectors, false);
        // Serialize the vectors
        DrillBuf byteBuf = serializeBatch(allocator, writableBatch);
        // Batch loader does NOT take ownership of the serialized buffer
        RecordBatchLoader batchLoader = new RecordBatchLoader(allocator);
        batchLoader.load(writableBatch.getDef(), byteBuf);
        // Release the serialized buffer.
        byteBuf.release();
        assertEquals(1, batchLoader.getRecordCount());
        // Free the original vectors
        writableBatch.clear();
        // Free the deserialized vectors
        batchLoader.clear();
    }
}
Also used : ResultSetLoaderImpl(org.apache.drill.exec.physical.resultSet.impl.ResultSetLoaderImpl) RecordBatchLoader(org.apache.drill.exec.record.RecordBatchLoader) BufferAllocator(org.apache.drill.exec.memory.BufferAllocator) VectorContainer(org.apache.drill.exec.record.VectorContainer) ResultSetOptionBuilder(org.apache.drill.exec.physical.resultSet.impl.ResultSetOptionBuilder) ValueVector(org.apache.drill.exec.vector.ValueVector) ResultSetLoader(org.apache.drill.exec.physical.resultSet.ResultSetLoader) TupleMetadata(org.apache.drill.exec.record.metadata.TupleMetadata) SchemaBuilder(org.apache.drill.exec.record.metadata.SchemaBuilder) BatchSchemaBuilder(org.apache.drill.exec.record.BatchSchemaBuilder) WritableBatch(org.apache.drill.exec.record.WritableBatch) RowSetLoader(org.apache.drill.exec.physical.resultSet.RowSetLoader) DrillBuf(io.netty.buffer.DrillBuf) ExecTest(org.apache.drill.exec.ExecTest) Test(org.junit.Test) VectorTest(org.apache.drill.categories.VectorTest)

Example 42 with BufferAllocator

use of org.apache.drill.exec.memory.BufferAllocator in project drill by apache.

the class BufferedBatches method bufferBatch.

private void bufferBatch(VectorContainer convertedBatch, SelectionVector2 sv2, long netSize) {
    BufferAllocator allocator = context.getAllocator();
    RecordBatchData rbd = new RecordBatchData(convertedBatch, allocator);
    try {
        rbd.setSv2(sv2);
        bufferedBatches.add(new InputBatch(rbd.getContainer(), rbd.getSv2(), allocator, netSize));
    } catch (Throwable t) {
        rbd.clear();
        throw t;
    }
}
Also used : RecordBatchData(org.apache.drill.exec.physical.impl.sort.RecordBatchData) BufferAllocator(org.apache.drill.exec.memory.BufferAllocator)

Example 43 with BufferAllocator

use of org.apache.drill.exec.memory.BufferAllocator 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.debug("Beginning partition pruning, pruning class: {}", pruningClassName);
    Stopwatch totalPruningTime = logger.isDebugEnabled() ? Stopwatch.createStarted() : null;
    final PlannerSettings settings = PrelUtil.getPlannerSettings(call.getPlanner());
    PartitionDescriptor descriptor = getPartitionDescriptor(settings, scanRel);
    final BufferAllocator allocator = optimizerContext.getAllocator();
    final Object selection = DrillRelOptUtil.getDrillTable(scanRel).getSelection();
    MetadataContext metaContext = null;
    if (selection instanceof FormatSelection) {
        metaContext = ((FormatSelection) selection).getSelection().getMetaContext();
    }
    RexNode condition;
    if (projectRel == null) {
        condition = filterRel.getCondition();
    } else {
        // get the filter as if it were below the projection.
        condition = RelOptUtil.pushPastProject(filterRel.getCondition(), projectRel);
    }
    RewriteAsBinaryOperators visitor = new RewriteAsBinaryOperators(true, filterRel.getCluster().getRexBuilder());
    condition = condition.accept(visitor);
    Map<Integer, String> fieldNameMap = new HashMap<>();
    List<String> fieldNames = scanRel.getRowType().getFieldNames();
    BitSet columnBitset = new BitSet();
    BitSet partitionColumnBitSet = new BitSet();
    Map<Integer, Integer> partitionMap = new HashMap<>();
    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()) {
        if (totalPruningTime != null) {
            logger.debug("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
    // first track how long we spend building the filter tree
    Stopwatch miscTimer = logger.isDebugEnabled() ? Stopwatch.createStarted() : null;
    FindPartitionConditions c = new FindPartitionConditions(columnBitset, filterRel.getCluster().getRexBuilder());
    c.analyze(condition);
    RexNode pruneCondition = c.getFinalCondition();
    BitSet referencedDirsBitSet = c.getReferencedDirs();
    if (miscTimer != null) {
        logger.debug("Total elapsed time to build and analyze filter tree: {} ms", miscTimer.elapsed(TimeUnit.MILLISECONDS));
        miscTimer.reset();
    }
    if (pruneCondition == null) {
        if (totalPruningTime != null) {
            logger.debug("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 = new ArrayList<>();
    // 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));
                // ParquetPartitionDescriptor.populatePruningVector() expects nullable value vectors,
                // so force nullability here to avoid class cast exceptions
                MajorType type = descriptor.getVectorType(column, settings).toBuilder().setMode(TypeProtos.DataMode.OPTIONAL).build();
                MaterializedField field = MaterializedField.create(column.getLastSegment().getNameSegment().getPath(), type);
                ValueVector v = TypeHelper.getNewVector(field, allocator);
                v.allocateNew();
                vectors[partitionColumnIndex] = v;
                container.add(v);
            }
            if (miscTimer != null) {
                // track how long we spend populating partition column vectors
                miscTimer.start();
            }
            // populate partition vectors.
            descriptor.populatePartitionVectors(vectors, partitions, partitionColumnBitSet, fieldNameMap);
            if (miscTimer != null) {
                logger.debug("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) {
                    // materializePruneExpr logs it already
                    if (totalPruningTime != null) {
                        logger.debug("Total pruning elapsed time: {} ms", totalPruningTime.elapsed(TimeUnit.MILLISECONDS));
                    }
                    setPruneStatus(metaContext, PruneStatus.NOT_PRUNED);
                    return;
                }
            }
            output.allocateNew(partitions.size());
            if (miscTimer != null) {
                // start the timer to evaluate how long we spend in the interpreter evaluation
                miscTimer.start();
            }
            InterpreterEvaluator.evaluate(partitions.size(), optimizerContext, container, output, materializedExpr);
            if (miscTimer != null) {
                logger.debug("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);
            if (totalPruningTime != null) {
                logger.debug("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.debug("No partitions were eligible for pruning");
            return;
        }
        // handle the case all partitions are filtered out.
        boolean canDropFilter = true;
        boolean wasAllPartitionsPruned = false;
        Path 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.debug("All {} partitions were pruned; added back a single partition to allow creating a schema", numTotal);
            // set the cacheFileRoot appropriately
            if (firstLocation.isCompositePartition()) {
                cacheFileRoot = Path.mergePaths(descriptor.getBaseTableLocation(), firstLocation.getCompositePartitionPath());
            }
        }
        logger.debug("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()) {
                StringBuilder path = new StringBuilder();
                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.append("/").append(spInfo[j]);
                }
                cacheFileRoot = Path.mergePaths(descriptor.getBaseTableLocation(), DrillFileSystemUtil.createPathSafe(path.toString()));
            }
            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 {
        if (totalPruningTime != null) {
            logger.debug("Total pruning elapsed time: {} ms", totalPruningTime.elapsed(TimeUnit.MILLISECONDS));
        }
    }
}
Also used : PlannerSettings(org.apache.drill.exec.planner.physical.PlannerSettings) HashMap(java.util.HashMap) Stopwatch(org.apache.drill.shaded.guava.com.google.common.base.Stopwatch) ArrayList(java.util.ArrayList) FormatSelection(org.apache.drill.exec.store.dfs.FormatSelection) LogicalExpression(org.apache.drill.common.expression.LogicalExpression) NullableBitVector(org.apache.drill.exec.vector.NullableBitVector) SchemaPath(org.apache.drill.common.expression.SchemaPath) PartitionDescriptor(org.apache.drill.exec.planner.PartitionDescriptor) FileSystemPartitionDescriptor(org.apache.drill.exec.planner.FileSystemPartitionDescriptor) PartitionLocation(org.apache.drill.exec.planner.PartitionLocation) Path(org.apache.hadoop.fs.Path) SchemaPath(org.apache.drill.common.expression.SchemaPath) MajorType(org.apache.drill.common.types.TypeProtos.MajorType) BitSet(java.util.BitSet) MaterializedField(org.apache.drill.exec.record.MaterializedField) BufferAllocator(org.apache.drill.exec.memory.BufferAllocator) VectorContainer(org.apache.drill.exec.record.VectorContainer) ValueVector(org.apache.drill.exec.vector.ValueVector) RelNode(org.apache.calcite.rel.RelNode) MetadataContext(org.apache.drill.exec.store.dfs.MetadataContext) RexNode(org.apache.calcite.rex.RexNode)

Aggregations

BufferAllocator (org.apache.drill.exec.memory.BufferAllocator)43 Test (org.junit.Test)24 DrillBuf (io.netty.buffer.DrillBuf)11 RecordBatchLoader (org.apache.drill.exec.record.RecordBatchLoader)11 BatchSchema (org.apache.drill.exec.record.BatchSchema)10 DrillConfig (org.apache.drill.common.config.DrillConfig)9 ExecTest (org.apache.drill.exec.ExecTest)9 MaterializedField (org.apache.drill.exec.record.MaterializedField)8 ValueVector (org.apache.drill.exec.vector.ValueVector)8 IOException (java.io.IOException)7 VectorTest (org.apache.drill.categories.VectorTest)7 OutOfMemoryException (org.apache.drill.exec.exception.OutOfMemoryException)7 SchemaPath (org.apache.drill.common.expression.SchemaPath)6 VectorContainer (org.apache.drill.exec.record.VectorContainer)6 TestOutputMutator (org.apache.drill.exec.store.TestOutputMutator)6 LogFixture (org.apache.drill.test.LogFixture)6 LogFixtureBuilder (org.apache.drill.test.LogFixture.LogFixtureBuilder)6 SubOperatorTest (org.apache.drill.test.SubOperatorTest)6 HashMap (java.util.HashMap)4 UserException (org.apache.drill.common.exceptions.UserException)4