Examples with GroupByOperator org.apache.hadoop.hive.ql.exec.GroupByOperator used on opensource projects

Example 11 with GroupByOperator

use of org.apache.hadoop.hive.ql.exec.GroupByOperator in project hive by apache.

the class DynamicPartitionPruningOptimization method generateSemiJoinOperatorPlan.

// Generates plan for min/max when dynamic partition pruning is ruled out.
private boolean generateSemiJoinOperatorPlan(DynamicListContext ctx, ParseContext parseContext, TableScanOperator ts, String keyBaseAlias, String internalColName, String colName, SemiJoinHint sjHint) throws SemanticException {
    // we will put a fork in the plan at the source of the reduce sink
    Operator<? extends OperatorDesc> parentOfRS = ctx.generator.getParentOperators().get(0);
    // we need the expr that generated the key of the reduce sink
    ExprNodeDesc key = ctx.generator.getConf().getKeyCols().get(ctx.desc.getKeyIndex());
    assert colName != null;
    // Fetch the TableScan Operator.
    Operator<?> op = parentOfRS;
    while (!(op == null || op instanceof TableScanOperator || op instanceof ReduceSinkOperator)) {
        op = op.getParentOperators().get(0);
    }
    Preconditions.checkNotNull(op);
    if (op instanceof TableScanOperator) {
        Table table = ((TableScanOperator) op).getConf().getTableMetadata();
        if (table.isPartitionKey(colName)) {
            // The column is partition column, skip the optimization.
            return false;
        }
    }
    // Check if there already exists a semijoin branch
    GroupByOperator gb = parseContext.getColExprToGBMap().get(key);
    if (gb != null) {
        // Already an existing semijoin branch, reuse it
        createFinalRsForSemiJoinOp(parseContext, ts, gb, key, keyBaseAlias, ctx.parent.getChildren().get(0), sjHint != null);
        // done!
        return true;
    }
    List<ExprNodeDesc> keyExprs = new ArrayList<ExprNodeDesc>();
    keyExprs.add(key);
    // group by requires "ArrayList", don't ask.
    ArrayList<String> outputNames = new ArrayList<String>();
    outputNames.add(HiveConf.getColumnInternalName(0));
    // project the relevant key column
    SelectDesc select = new SelectDesc(keyExprs, outputNames);
    // Create the new RowSchema for the projected column
    ColumnInfo columnInfo = parentOfRS.getSchema().getColumnInfo(internalColName);
    ArrayList<ColumnInfo> signature = new ArrayList<ColumnInfo>();
    signature.add(columnInfo);
    RowSchema rowSchema = new RowSchema(signature);
    // Create the column expr map
    Map<String, ExprNodeDesc> colExprMap = new HashMap<String, ExprNodeDesc>();
    ExprNodeDesc exprNode = null;
    if (parentOfRS.getColumnExprMap() != null) {
        exprNode = parentOfRS.getColumnExprMap().get(internalColName).clone();
    } else {
        exprNode = new ExprNodeColumnDesc(columnInfo);
    }
    if (exprNode instanceof ExprNodeColumnDesc) {
        ExprNodeColumnDesc encd = (ExprNodeColumnDesc) exprNode;
        encd.setColumn(internalColName);
    }
    colExprMap.put(internalColName, exprNode);
    // Create the Select Operator
    SelectOperator selectOp = (SelectOperator) OperatorFactory.getAndMakeChild(select, rowSchema, colExprMap, parentOfRS);
    // do a group by to aggregate min,max and bloom filter.
    float groupByMemoryUsage = HiveConf.getFloatVar(parseContext.getConf(), HiveConf.ConfVars.HIVEMAPAGGRHASHMEMORY);
    float memoryThreshold = HiveConf.getFloatVar(parseContext.getConf(), HiveConf.ConfVars.HIVEMAPAGGRMEMORYTHRESHOLD);
    // Add min/max and bloom filter aggregations
    List<ObjectInspector> aggFnOIs = new ArrayList<ObjectInspector>();
    aggFnOIs.add(key.getWritableObjectInspector());
    ArrayList<ExprNodeDesc> params = new ArrayList<ExprNodeDesc>();
    params.add(new ExprNodeColumnDesc(key.getTypeInfo(), outputNames.get(0), "", false));
    ArrayList<AggregationDesc> aggs = new ArrayList<AggregationDesc>();
    try {
        AggregationDesc min = new AggregationDesc("min", FunctionRegistry.getGenericUDAFEvaluator("min", aggFnOIs, false, false), params, false, Mode.PARTIAL1);
        AggregationDesc max = new AggregationDesc("max", FunctionRegistry.getGenericUDAFEvaluator("max", aggFnOIs, false, false), params, false, Mode.PARTIAL1);
        AggregationDesc bloomFilter = new AggregationDesc("bloom_filter", FunctionRegistry.getGenericUDAFEvaluator("bloom_filter", aggFnOIs, false, false), params, false, Mode.PARTIAL1);
        GenericUDAFBloomFilterEvaluator bloomFilterEval = (GenericUDAFBloomFilterEvaluator) bloomFilter.getGenericUDAFEvaluator();
        bloomFilterEval.setSourceOperator(selectOp);
        if (sjHint != null && sjHint.getNumEntries() > 0) {
            LOG.debug("Setting size for " + keyBaseAlias + " to " + sjHint.getNumEntries() + " based on the hint");
            bloomFilterEval.setHintEntries(sjHint.getNumEntries());
        }
        bloomFilterEval.setMaxEntries(parseContext.getConf().getLongVar(ConfVars.TEZ_MAX_BLOOM_FILTER_ENTRIES));
        bloomFilterEval.setMinEntries(parseContext.getConf().getLongVar(ConfVars.TEZ_MIN_BLOOM_FILTER_ENTRIES));
        bloomFilterEval.setFactor(parseContext.getConf().getFloatVar(ConfVars.TEZ_BLOOM_FILTER_FACTOR));
        bloomFilter.setGenericUDAFWritableEvaluator(bloomFilterEval);
        aggs.add(min);
        aggs.add(max);
        aggs.add(bloomFilter);
    } catch (SemanticException e) {
        LOG.error("Error creating min/max aggregations on key", e);
        throw new IllegalStateException("Error creating min/max aggregations on key", e);
    }
    // Create the Group by Operator
    ArrayList<String> gbOutputNames = new ArrayList<String>();
    gbOutputNames.add(SemanticAnalyzer.getColumnInternalName(0));
    gbOutputNames.add(SemanticAnalyzer.getColumnInternalName(1));
    gbOutputNames.add(SemanticAnalyzer.getColumnInternalName(2));
    GroupByDesc groupBy = new GroupByDesc(GroupByDesc.Mode.HASH, gbOutputNames, new ArrayList<ExprNodeDesc>(), aggs, false, groupByMemoryUsage, memoryThreshold, null, false, -1, false);
    ArrayList<ColumnInfo> groupbyColInfos = new ArrayList<ColumnInfo>();
    groupbyColInfos.add(new ColumnInfo(gbOutputNames.get(0), key.getTypeInfo(), "", false));
    groupbyColInfos.add(new ColumnInfo(gbOutputNames.get(1), key.getTypeInfo(), "", false));
    groupbyColInfos.add(new ColumnInfo(gbOutputNames.get(2), key.getTypeInfo(), "", false));
    GroupByOperator groupByOp = (GroupByOperator) OperatorFactory.getAndMakeChild(groupBy, new RowSchema(groupbyColInfos), selectOp);
    groupByOp.setColumnExprMap(new HashMap<String, ExprNodeDesc>());
    // Get the column names of the aggregations for reduce sink
    int colPos = 0;
    ArrayList<ExprNodeDesc> rsValueCols = new ArrayList<ExprNodeDesc>();
    for (int i = 0; i < aggs.size() - 1; i++) {
        ExprNodeColumnDesc colExpr = new ExprNodeColumnDesc(key.getTypeInfo(), gbOutputNames.get(colPos++), "", false);
        rsValueCols.add(colExpr);
    }
    // Bloom Filter uses binary
    ExprNodeColumnDesc colExpr = new ExprNodeColumnDesc(TypeInfoFactory.binaryTypeInfo, gbOutputNames.get(colPos++), "", false);
    rsValueCols.add(colExpr);
    // Create the reduce sink operator
    ReduceSinkDesc rsDesc = PlanUtils.getReduceSinkDesc(new ArrayList<ExprNodeDesc>(), rsValueCols, gbOutputNames, false, -1, 0, 1, Operation.NOT_ACID);
    ReduceSinkOperator rsOp = (ReduceSinkOperator) OperatorFactory.getAndMakeChild(rsDesc, new RowSchema(groupByOp.getSchema()), groupByOp);
    Map<String, ExprNodeDesc> columnExprMap = new HashMap<String, ExprNodeDesc>();
    rsOp.setColumnExprMap(columnExprMap);
    rsOp.getConf().setReducerTraits(EnumSet.of(ReduceSinkDesc.ReducerTraits.QUICKSTART));
    // Create the final Group By Operator
    ArrayList<AggregationDesc> aggsFinal = new ArrayList<AggregationDesc>();
    try {
        List<ObjectInspector> minFinalFnOIs = new ArrayList<ObjectInspector>();
        List<ObjectInspector> maxFinalFnOIs = new ArrayList<ObjectInspector>();
        List<ObjectInspector> bloomFilterFinalFnOIs = new ArrayList<ObjectInspector>();
        ArrayList<ExprNodeDesc> minFinalParams = new ArrayList<ExprNodeDesc>();
        ArrayList<ExprNodeDesc> maxFinalParams = new ArrayList<ExprNodeDesc>();
        ArrayList<ExprNodeDesc> bloomFilterFinalParams = new ArrayList<ExprNodeDesc>();
        // Use the expressions from Reduce Sink.
        minFinalFnOIs.add(rsValueCols.get(0).getWritableObjectInspector());
        maxFinalFnOIs.add(rsValueCols.get(1).getWritableObjectInspector());
        bloomFilterFinalFnOIs.add(rsValueCols.get(2).getWritableObjectInspector());
        // Coming from a ReduceSink the aggregations would be in the form VALUE._col0, VALUE._col1
        minFinalParams.add(new ExprNodeColumnDesc(rsValueCols.get(0).getTypeInfo(), Utilities.ReduceField.VALUE + "." + gbOutputNames.get(0), "", false));
        maxFinalParams.add(new ExprNodeColumnDesc(rsValueCols.get(1).getTypeInfo(), Utilities.ReduceField.VALUE + "." + gbOutputNames.get(1), "", false));
        bloomFilterFinalParams.add(new ExprNodeColumnDesc(rsValueCols.get(2).getTypeInfo(), Utilities.ReduceField.VALUE + "." + gbOutputNames.get(2), "", false));
        AggregationDesc min = new AggregationDesc("min", FunctionRegistry.getGenericUDAFEvaluator("min", minFinalFnOIs, false, false), minFinalParams, false, Mode.FINAL);
        AggregationDesc max = new AggregationDesc("max", FunctionRegistry.getGenericUDAFEvaluator("max", maxFinalFnOIs, false, false), maxFinalParams, false, Mode.FINAL);
        AggregationDesc bloomFilter = new AggregationDesc("bloom_filter", FunctionRegistry.getGenericUDAFEvaluator("bloom_filter", bloomFilterFinalFnOIs, false, false), bloomFilterFinalParams, false, Mode.FINAL);
        GenericUDAFBloomFilterEvaluator bloomFilterEval = (GenericUDAFBloomFilterEvaluator) bloomFilter.getGenericUDAFEvaluator();
        bloomFilterEval.setSourceOperator(selectOp);
        if (sjHint != null && sjHint.getNumEntries() > 0) {
            bloomFilterEval.setHintEntries(sjHint.getNumEntries());
        }
        bloomFilterEval.setMaxEntries(parseContext.getConf().getLongVar(ConfVars.TEZ_MAX_BLOOM_FILTER_ENTRIES));
        bloomFilterEval.setMinEntries(parseContext.getConf().getLongVar(ConfVars.TEZ_MIN_BLOOM_FILTER_ENTRIES));
        bloomFilterEval.setFactor(parseContext.getConf().getFloatVar(ConfVars.TEZ_BLOOM_FILTER_FACTOR));
        bloomFilter.setGenericUDAFWritableEvaluator(bloomFilterEval);
        aggsFinal.add(min);
        aggsFinal.add(max);
        aggsFinal.add(bloomFilter);
    } catch (SemanticException e) {
        LOG.error("Error creating min/max aggregations on key", e);
        throw new IllegalStateException("Error creating min/max aggregations on key", e);
    }
    GroupByDesc groupByDescFinal = new GroupByDesc(GroupByDesc.Mode.FINAL, gbOutputNames, new ArrayList<ExprNodeDesc>(), aggsFinal, false, groupByMemoryUsage, memoryThreshold, null, false, 0, false);
    GroupByOperator groupByOpFinal = (GroupByOperator) OperatorFactory.getAndMakeChild(groupByDescFinal, new RowSchema(rsOp.getSchema()), rsOp);
    groupByOpFinal.setColumnExprMap(new HashMap<String, ExprNodeDesc>());
    createFinalRsForSemiJoinOp(parseContext, ts, groupByOpFinal, key, keyBaseAlias, ctx.parent.getChildren().get(0), sjHint != null);
    return true;
}

Example 12 with GroupByOperator

use of org.apache.hadoop.hive.ql.exec.GroupByOperator in project hive by apache.

the class HiveGBOpConvUtil method genMapSideGBRS.

private static OpAttr genMapSideGBRS(OpAttr inputOpAf, GBInfo gbInfo) throws SemanticException {
    Map<String, ExprNodeDesc> colExprMap = new HashMap<String, ExprNodeDesc>();
    List<String> outputKeyColumnNames = new ArrayList<String>();
    List<String> outputValueColumnNames = new ArrayList<String>();
    ArrayList<ColumnInfo> colInfoLst = new ArrayList<ColumnInfo>();
    GroupByOperator mapGB = (GroupByOperator) inputOpAf.inputs.get(0);
    ArrayList<ExprNodeDesc> reduceKeys = getReduceKeysForRS(mapGB, 0, gbInfo.gbKeys.size() - 1, outputKeyColumnNames, false, colInfoLst, colExprMap, false, false);
    int keyLength = reduceKeys.size();
    if (inclGrpSetInMapSide(gbInfo)) {
        addGrpSetCol(false, SemanticAnalyzer.getColumnInternalName(reduceKeys.size()), true, reduceKeys, outputKeyColumnNames, colInfoLst, colExprMap);
        keyLength++;
    }
    if (mapGB.getConf().getKeys().size() > reduceKeys.size()) {
        // NOTE: All dist cols have single output col name;
        reduceKeys.addAll(getReduceKeysForRS(mapGB, reduceKeys.size(), mapGB.getConf().getKeys().size() - 1, outputKeyColumnNames, true, colInfoLst, colExprMap, false, false));
    } else if (!gbInfo.distColIndices.isEmpty()) {
        // This is the case where distinct cols are part of GB Keys in which case
        // we still need to add it to out put col names
        outputKeyColumnNames.add(SemanticAnalyzer.getColumnInternalName(reduceKeys.size()));
    }
    ArrayList<ExprNodeDesc> reduceValues = getValueKeysForRS(mapGB, mapGB.getConf().getKeys().size(), outputValueColumnNames, colInfoLst, colExprMap, false, false);
    ReduceSinkOperator rsOp = (ReduceSinkOperator) OperatorFactory.getAndMakeChild(PlanUtils.getReduceSinkDesc(reduceKeys, keyLength, reduceValues, gbInfo.distColIndices, outputKeyColumnNames, outputValueColumnNames, true, -1, getNumPartFieldsForMapSideRS(gbInfo), getParallelismForMapSideRS(gbInfo), AcidUtils.Operation.NOT_ACID), new RowSchema(colInfoLst), mapGB);
    rsOp.setColumnExprMap(colExprMap);
    return new OpAttr("", new HashSet<Integer>(), rsOp);
}

Example 13 with GroupByOperator

use of org.apache.hadoop.hive.ql.exec.GroupByOperator in project hive by apache.

the class DynamicPartitionPruningOptimization method generateEventOperatorPlan.

private void generateEventOperatorPlan(DynamicListContext ctx, ParseContext parseContext, TableScanOperator ts, String column, String columnType) {
    // we will put a fork in the plan at the source of the reduce sink
    Operator<? extends OperatorDesc> parentOfRS = ctx.generator.getParentOperators().get(0);
    // we need the expr that generated the key of the reduce sink
    ExprNodeDesc key = ctx.generator.getConf().getKeyCols().get(ctx.desc.getKeyIndex());
    // we also need the expr for the partitioned table
    ExprNodeDesc partKey = ctx.parent.getChildren().get(0);
    if (LOG.isDebugEnabled()) {
        LOG.debug("key expr: " + key);
        LOG.debug("partition key expr: " + partKey);
    }
    List<ExprNodeDesc> keyExprs = new ArrayList<ExprNodeDesc>();
    keyExprs.add(key);
    // group by requires "ArrayList", don't ask.
    ArrayList<String> outputNames = new ArrayList<String>();
    outputNames.add(HiveConf.getColumnInternalName(0));
    // project the relevant key column
    SelectDesc select = new SelectDesc(keyExprs, outputNames);
    SelectOperator selectOp = (SelectOperator) OperatorFactory.getAndMakeChild(select, parentOfRS);
    // do a group by on the list to dedup
    float groupByMemoryUsage = HiveConf.getFloatVar(parseContext.getConf(), HiveConf.ConfVars.HIVEMAPAGGRHASHMEMORY);
    float memoryThreshold = HiveConf.getFloatVar(parseContext.getConf(), HiveConf.ConfVars.HIVEMAPAGGRMEMORYTHRESHOLD);
    ArrayList<ExprNodeDesc> groupByExprs = new ArrayList<ExprNodeDesc>();
    ExprNodeDesc groupByExpr = new ExprNodeColumnDesc(key.getTypeInfo(), outputNames.get(0), null, false);
    groupByExprs.add(groupByExpr);
    GroupByDesc groupBy = new GroupByDesc(GroupByDesc.Mode.HASH, outputNames, groupByExprs, new ArrayList<AggregationDesc>(), false, groupByMemoryUsage, memoryThreshold, null, false, -1, true);
    GroupByOperator groupByOp = (GroupByOperator) OperatorFactory.getAndMakeChild(groupBy, selectOp);
    Map<String, ExprNodeDesc> colMap = new HashMap<String, ExprNodeDesc>();
    colMap.put(outputNames.get(0), groupByExpr);
    groupByOp.setColumnExprMap(colMap);
    // finally add the event broadcast operator
    if (HiveConf.getVar(parseContext.getConf(), ConfVars.HIVE_EXECUTION_ENGINE).equals("tez")) {
        DynamicPruningEventDesc eventDesc = new DynamicPruningEventDesc();
        eventDesc.setTableScan(ts);
        eventDesc.setGenerator(ctx.generator);
        eventDesc.setTable(PlanUtils.getReduceValueTableDesc(PlanUtils.getFieldSchemasFromColumnList(keyExprs, "key")));
        eventDesc.setTargetColumnName(column);
        eventDesc.setTargetColumnType(columnType);
        eventDesc.setPartKey(partKey);
        OperatorFactory.getAndMakeChild(eventDesc, groupByOp);
    } else {
        // Must be spark branch
        SparkPartitionPruningSinkDesc desc = new SparkPartitionPruningSinkDesc();
        desc.setTable(PlanUtils.getReduceValueTableDesc(PlanUtils.getFieldSchemasFromColumnList(keyExprs, "key")));
        desc.addTarget(column, columnType, partKey, null, ts);
        SparkPartitionPruningSinkOperator dppSink = (SparkPartitionPruningSinkOperator) OperatorFactory.getAndMakeChild(desc, groupByOp);
        if (HiveConf.getBoolVar(parseContext.getConf(), ConfVars.HIVE_COMBINE_EQUIVALENT_WORK_OPTIMIZATION)) {
            mayReuseExistingDPPSink(parentOfRS, Arrays.asList(selectOp, groupByOp, dppSink));
        }
    }
}

Example 14 with GroupByOperator

use of org.apache.hadoop.hive.ql.exec.GroupByOperator in project hive by apache.

the class GlobalLimitOptimizer method checkQbpForGlobalLimit.

/**
 * Check the limit number in all sub queries
 *
 * @return if there is one and only one limit for all subqueries, return the limit
 *         if there is no limit, return 0
 *         otherwise, return null
 */
private static LimitOperator checkQbpForGlobalLimit(TableScanOperator ts) {
    Set<Class<? extends Operator<?>>> searchedClasses = new ImmutableSet.Builder<Class<? extends Operator<?>>>().add(ReduceSinkOperator.class).add(GroupByOperator.class).add(FilterOperator.class).add(LimitOperator.class).build();
    Multimap<Class<? extends Operator<?>>, Operator<?>> ops = OperatorUtils.classifyOperators(ts, searchedClasses);
    // existsOrdering AND existsPartitioning should be false.
    for (Operator<?> op : ops.get(ReduceSinkOperator.class)) {
        ReduceSinkDesc reduceSinkConf = ((ReduceSinkOperator) op).getConf();
        if (reduceSinkConf.isOrdering() || reduceSinkConf.isPartitioning()) {
            return null;
        }
    }
    // - There cannot exist any (distinct) aggregate.
    for (Operator<?> op : ops.get(GroupByOperator.class)) {
        GroupByDesc groupByConf = ((GroupByOperator) op).getConf();
        if (groupByConf.isAggregate() || groupByConf.isDistinct()) {
            return null;
        }
    }
    // - There cannot exist any sampling predicate.
    for (Operator<?> op : ops.get(FilterOperator.class)) {
        FilterDesc filterConf = ((FilterOperator) op).getConf();
        if (filterConf.getIsSamplingPred()) {
            return null;
        }
    }
    // If there is one and only one limit starting at op, return the limit
    // If there is no limit, return 0
    // Otherwise, return null
    Collection<Operator<?>> limitOps = ops.get(LimitOperator.class);
    if (limitOps.size() == 1) {
        return (LimitOperator) limitOps.iterator().next();
    } else if (limitOps.size() == 0) {
        return null;
    }
    return null;
}

Example 15 with GroupByOperator

use of org.apache.hadoop.hive.ql.exec.GroupByOperator in project hive by apache.

the class SemanticAnalyzer method genGroupByPlan1MR.

/**
 * Generate a Group-By plan using a single map-reduce job (3 operators will be
 * inserted):
 *
 * ReduceSink ( keys = (K1_EXP, K2_EXP, DISTINCT_EXP), values = (A1_EXP,
 * A2_EXP) ) SortGroupBy (keys = (KEY.0,KEY.1), aggregations =
 * (count_distinct(KEY.2), sum(VALUE.0), count(VALUE.1))) Select (final
 * selects).
 *
 * @param dest
 * @param qb
 * @param input
 * @return
 * @throws SemanticException
 *
 *           Generate a Group-By plan using 1 map-reduce job. Spray by the
 *           group by key, and sort by the distinct key (if any), and compute
 *           aggregates * The aggregation evaluation functions are as
 *           follows: Partitioning Key: grouping key
 *
 *           Sorting Key: grouping key if no DISTINCT grouping + distinct key
 *           if DISTINCT
 *
 *           Reducer: iterate/merge (mode = COMPLETE)
 */
@SuppressWarnings({ "nls" })
private Operator genGroupByPlan1MR(String dest, QB qb, Operator input) throws SemanticException {
    QBParseInfo parseInfo = qb.getParseInfo();
    int numReducers = -1;
    ObjectPair<List<ASTNode>, List<Long>> grpByExprsGroupingSets = getGroupByGroupingSetsForClause(parseInfo, dest);
    List<ASTNode> grpByExprs = grpByExprsGroupingSets.getFirst();
    List<Long> groupingSets = grpByExprsGroupingSets.getSecond();
    if (grpByExprs.isEmpty()) {
        numReducers = 1;
    }
    // Grouping sets are not allowed
    if (!groupingSets.isEmpty()) {
        throw new SemanticException(ErrorMsg.HIVE_GROUPING_SETS_AGGR_NOMAPAGGR.getMsg());
    }
    // ////// 1. Generate ReduceSinkOperator
    ReduceSinkOperator reduceSinkOperatorInfo = genGroupByPlanReduceSinkOperator(qb, dest, input, grpByExprs, grpByExprs.size(), false, numReducers, false, false);
    // ////// 2. Generate GroupbyOperator
    Operator groupByOperatorInfo = genGroupByPlanGroupByOperator(parseInfo, dest, reduceSinkOperatorInfo, reduceSinkOperatorInfo, GroupByDesc.Mode.COMPLETE, null);
    return groupByOperatorInfo;
}