Examples with GenericUDAFEvaluator org.apache.hadoop.hive.ql.udf.generic.GenericUDAFEvaluator used on opensource projects

Example 6 with GenericUDAFEvaluator

use of org.apache.hadoop.hive.ql.udf.generic.GenericUDAFEvaluator in project hive by apache.

the class SemanticAnalyzer method genGroupByPlanGroupByOperator2MR.

/**
   * Generate the second GroupByOperator for the Group By Plan
   * (parseInfo.getXXX(dest)). The new GroupByOperator will do the second
   * aggregation based on the partial aggregation results.
   *
   * @param mode
   *          the mode of aggregation (FINAL)
   * @param genericUDAFEvaluators
   *          The mapping from Aggregation StringTree to the
   *          genericUDAFEvaluator.
   * @return the new GroupByOperator
   * @throws SemanticException
   */
@SuppressWarnings("nls")
private Operator genGroupByPlanGroupByOperator2MR(QBParseInfo parseInfo, String dest, Operator reduceSinkOperatorInfo2, GroupByDesc.Mode mode, Map<String, GenericUDAFEvaluator> genericUDAFEvaluators, boolean groupingSetsPresent) throws SemanticException {
    RowResolver groupByInputRowResolver2 = opParseCtx.get(reduceSinkOperatorInfo2).getRowResolver();
    RowResolver groupByOutputRowResolver2 = new RowResolver();
    groupByOutputRowResolver2.setIsExprResolver(true);
    ArrayList<ExprNodeDesc> groupByKeys = new ArrayList<ExprNodeDesc>();
    ArrayList<AggregationDesc> aggregations = new ArrayList<AggregationDesc>();
    Map<String, ExprNodeDesc> colExprMap = new HashMap<String, ExprNodeDesc>();
    List<ASTNode> grpByExprs = getGroupByForClause(parseInfo, dest);
    ArrayList<String> outputColumnNames = new ArrayList<String>();
    for (int i = 0; i < grpByExprs.size(); ++i) {
        ASTNode grpbyExpr = grpByExprs.get(i);
        ColumnInfo exprInfo = groupByInputRowResolver2.getExpression(grpbyExpr);
        if (exprInfo == null) {
            throw new SemanticException(ErrorMsg.INVALID_COLUMN.getMsg(grpbyExpr));
        }
        String expression = exprInfo.getInternalName();
        groupByKeys.add(new ExprNodeColumnDesc(exprInfo.getType(), expression, exprInfo.getTabAlias(), exprInfo.getIsVirtualCol()));
        String field = getColumnInternalName(i);
        outputColumnNames.add(field);
        ColumnInfo oColInfo = new ColumnInfo(field, exprInfo.getType(), "", false);
        groupByOutputRowResolver2.putExpression(grpbyExpr, oColInfo);
        addAlternateGByKeyMappings(grpbyExpr, oColInfo, reduceSinkOperatorInfo2, groupByOutputRowResolver2);
        colExprMap.put(field, groupByKeys.get(groupByKeys.size() - 1));
    }
    int groupingSetsPosition = -1;
    // For grouping sets, add a dummy grouping key
    if (groupingSetsPresent) {
        groupingSetsPosition = groupByKeys.size();
        addGroupingSetKey(groupByKeys, groupByInputRowResolver2, groupByOutputRowResolver2, outputColumnNames, colExprMap);
    }
    HashMap<String, ASTNode> aggregationTrees = parseInfo.getAggregationExprsForClause(dest);
    boolean containsDistinctAggr = false;
    for (Map.Entry<String, ASTNode> entry : aggregationTrees.entrySet()) {
        ArrayList<ExprNodeDesc> aggParameters = new ArrayList<ExprNodeDesc>();
        ASTNode value = entry.getValue();
        ColumnInfo paraExprInfo = groupByInputRowResolver2.getExpression(value);
        if (paraExprInfo == null) {
            throw new SemanticException(ErrorMsg.INVALID_COLUMN.getMsg(value));
        }
        String paraExpression = paraExprInfo.getInternalName();
        assert (paraExpression != null);
        aggParameters.add(new ExprNodeColumnDesc(paraExprInfo.getType(), paraExpression, paraExprInfo.getTabAlias(), paraExprInfo.getIsVirtualCol()));
        String aggName = unescapeIdentifier(value.getChild(0).getText());
        boolean isDistinct = value.getType() == HiveParser.TOK_FUNCTIONDI;
        containsDistinctAggr = containsDistinctAggr || isDistinct;
        boolean isStar = value.getType() == HiveParser.TOK_FUNCTIONSTAR;
        Mode amode = groupByDescModeToUDAFMode(mode, isDistinct);
        GenericUDAFEvaluator genericUDAFEvaluator = genericUDAFEvaluators.get(entry.getKey());
        assert (genericUDAFEvaluator != null);
        GenericUDAFInfo udaf = getGenericUDAFInfo(genericUDAFEvaluator, amode, aggParameters);
        aggregations.add(new AggregationDesc(aggName.toLowerCase(), udaf.genericUDAFEvaluator, udaf.convertedParameters, (mode != GroupByDesc.Mode.FINAL && value.getToken().getType() == HiveParser.TOK_FUNCTIONDI), amode));
        String field = getColumnInternalName(groupByKeys.size() + aggregations.size() - 1);
        outputColumnNames.add(field);
        groupByOutputRowResolver2.putExpression(value, new ColumnInfo(field, udaf.returnType, "", false));
    }
    float groupByMemoryUsage = HiveConf.getFloatVar(conf, HiveConf.ConfVars.HIVEMAPAGGRHASHMEMORY);
    float memoryThreshold = HiveConf.getFloatVar(conf, HiveConf.ConfVars.HIVEMAPAGGRMEMORYTHRESHOLD);
    Operator op = putOpInsertMap(OperatorFactory.getAndMakeChild(new GroupByDesc(mode, outputColumnNames, groupByKeys, aggregations, false, groupByMemoryUsage, memoryThreshold, null, false, groupingSetsPosition, containsDistinctAggr), new RowSchema(groupByOutputRowResolver2.getColumnInfos()), reduceSinkOperatorInfo2), groupByOutputRowResolver2);
    op.setColumnExprMap(colExprMap);
    return op;
}

Example 7 with GenericUDAFEvaluator

use of org.apache.hadoop.hive.ql.udf.generic.GenericUDAFEvaluator in project hive by apache.

the class SemanticAnalyzer method genGroupByPlanGroupByOperator1.

/**
   * Generate the GroupByOperator for the Query Block (parseInfo.getXXX(dest)).
   * The new GroupByOperator will be a child of the reduceSinkOperatorInfo.
   *
   * @param parseInfo
   * @param dest
   * @param reduceSinkOperatorInfo
   * @param mode
   *          The mode of the aggregation (MERGEPARTIAL, PARTIAL2)
   * @param genericUDAFEvaluators
   *          The mapping from Aggregation StringTree to the
   *          genericUDAFEvaluator.
   * @param groupingSets
   *          list of grouping sets
   * @param groupingSetsPresent
   *          whether grouping sets are present in this query
   * @param groupingSetsNeedAdditionalMRJob
   *          whether grouping sets are consumed by this group by
   * @return the new GroupByOperator
   */
@SuppressWarnings("nls")
private Operator genGroupByPlanGroupByOperator1(QBParseInfo parseInfo, String dest, Operator reduceSinkOperatorInfo, GroupByDesc.Mode mode, Map<String, GenericUDAFEvaluator> genericUDAFEvaluators, List<Integer> groupingSets, boolean groupingSetsPresent, boolean groupingSetsNeedAdditionalMRJob) throws SemanticException {
    ArrayList<String> outputColumnNames = new ArrayList<String>();
    RowResolver groupByInputRowResolver = opParseCtx.get(reduceSinkOperatorInfo).getRowResolver();
    RowResolver groupByOutputRowResolver = new RowResolver();
    groupByOutputRowResolver.setIsExprResolver(true);
    ArrayList<ExprNodeDesc> groupByKeys = new ArrayList<ExprNodeDesc>();
    ArrayList<AggregationDesc> aggregations = new ArrayList<AggregationDesc>();
    List<ASTNode> grpByExprs = getGroupByForClause(parseInfo, dest);
    Map<String, ExprNodeDesc> colExprMap = new HashMap<String, ExprNodeDesc>();
    for (int i = 0; i < grpByExprs.size(); ++i) {
        ASTNode grpbyExpr = grpByExprs.get(i);
        ColumnInfo exprInfo = groupByInputRowResolver.getExpression(grpbyExpr);
        if (exprInfo == null) {
            throw new SemanticException(ErrorMsg.INVALID_COLUMN.getMsg(grpbyExpr));
        }
        groupByKeys.add(new ExprNodeColumnDesc(exprInfo));
        String field = getColumnInternalName(i);
        outputColumnNames.add(field);
        ColumnInfo oColInfo = new ColumnInfo(field, exprInfo.getType(), "", false);
        groupByOutputRowResolver.putExpression(grpbyExpr, oColInfo);
        addAlternateGByKeyMappings(grpbyExpr, oColInfo, reduceSinkOperatorInfo, groupByOutputRowResolver);
        colExprMap.put(field, groupByKeys.get(groupByKeys.size() - 1));
    }
    // This is only needed if a new grouping set key is being created
    int groupingSetsPosition = -1;
    // For grouping sets, add a dummy grouping key
    if (groupingSetsPresent) {
        groupingSetsPosition = groupByKeys.size();
        // This function is called for GroupBy2 to add grouping id as part of the groupby keys
        if (!groupingSetsNeedAdditionalMRJob) {
            addGroupingSetKey(groupByKeys, groupByInputRowResolver, groupByOutputRowResolver, outputColumnNames, colExprMap);
        } else {
            // The grouping set has not yet been processed. Create a new grouping key
            // Consider the query: select a,b, count(1) from T group by a,b with cube;
            // where it is being executed in 2 map-reduce jobs
            // The plan for 1st MR is TableScan -> GroupBy1 -> ReduceSink -> GroupBy2 -> FileSink
            // GroupBy1/ReduceSink worked as if grouping sets were not present
            // This function is called for GroupBy2 to create new rows for grouping sets
            // For each input row (a,b), 4 rows are created for the example above:
            // (a,b), (a,null), (null, b), (null, null)
            createNewGroupingKey(groupByKeys, outputColumnNames, groupByOutputRowResolver, colExprMap);
        }
    }
    HashMap<String, ASTNode> aggregationTrees = parseInfo.getAggregationExprsForClause(dest);
    // get the last colName for the reduce KEY
    // it represents the column name corresponding to distinct aggr, if any
    String lastKeyColName = null;
    List<ExprNodeDesc> reduceValues = null;
    if (reduceSinkOperatorInfo.getConf() instanceof ReduceSinkDesc) {
        List<String> inputKeyCols = ((ReduceSinkDesc) reduceSinkOperatorInfo.getConf()).getOutputKeyColumnNames();
        if (inputKeyCols.size() > 0) {
            lastKeyColName = inputKeyCols.get(inputKeyCols.size() - 1);
        }
        reduceValues = ((ReduceSinkDesc) reduceSinkOperatorInfo.getConf()).getValueCols();
    }
    int numDistinctUDFs = 0;
    boolean containsDistinctAggr = false;
    for (Map.Entry<String, ASTNode> entry : aggregationTrees.entrySet()) {
        ASTNode value = entry.getValue();
        String aggName = unescapeIdentifier(value.getChild(0).getText());
        ArrayList<ExprNodeDesc> aggParameters = new ArrayList<ExprNodeDesc>();
        boolean isDistinct = (value.getType() == HiveParser.TOK_FUNCTIONDI);
        containsDistinctAggr = containsDistinctAggr || isDistinct;
        // side, so always look for the parameters: d+e
        if (isDistinct) {
            // 0 is the function name
            for (int i = 1; i < value.getChildCount(); i++) {
                ASTNode paraExpr = (ASTNode) value.getChild(i);
                ColumnInfo paraExprInfo = groupByInputRowResolver.getExpression(paraExpr);
                if (paraExprInfo == null) {
                    throw new SemanticException(ErrorMsg.INVALID_COLUMN.getMsg(paraExpr));
                }
                String paraExpression = paraExprInfo.getInternalName();
                assert (paraExpression != null);
                if (isDistinct && lastKeyColName != null) {
                    // if aggr is distinct, the parameter is name is constructed as
                    // KEY.lastKeyColName:<tag>._colx
                    paraExpression = Utilities.ReduceField.KEY.name() + "." + lastKeyColName + ":" + numDistinctUDFs + "." + getColumnInternalName(i - 1);
                }
                ExprNodeDesc expr = new ExprNodeColumnDesc(paraExprInfo.getType(), paraExpression, paraExprInfo.getTabAlias(), paraExprInfo.getIsVirtualCol());
                ExprNodeDesc reduceValue = isConstantParameterInAggregationParameters(paraExprInfo.getInternalName(), reduceValues);
                if (reduceValue != null) {
                    // this parameter is a constant
                    expr = reduceValue;
                }
                aggParameters.add(expr);
            }
        } else {
            ColumnInfo paraExprInfo = groupByInputRowResolver.getExpression(value);
            if (paraExprInfo == null) {
                throw new SemanticException(ErrorMsg.INVALID_COLUMN.getMsg(value));
            }
            String paraExpression = paraExprInfo.getInternalName();
            assert (paraExpression != null);
            aggParameters.add(new ExprNodeColumnDesc(paraExprInfo.getType(), paraExpression, paraExprInfo.getTabAlias(), paraExprInfo.getIsVirtualCol()));
        }
        if (isDistinct) {
            numDistinctUDFs++;
        }
        Mode amode = groupByDescModeToUDAFMode(mode, isDistinct);
        GenericUDAFEvaluator genericUDAFEvaluator = null;
        genericUDAFEvaluator = genericUDAFEvaluators.get(entry.getKey());
        assert (genericUDAFEvaluator != null);
        GenericUDAFInfo udaf = getGenericUDAFInfo(genericUDAFEvaluator, amode, aggParameters);
        aggregations.add(new AggregationDesc(aggName.toLowerCase(), udaf.genericUDAFEvaluator, udaf.convertedParameters, (mode != GroupByDesc.Mode.FINAL && isDistinct), amode));
        String field = getColumnInternalName(groupByKeys.size() + aggregations.size() - 1);
        outputColumnNames.add(field);
        groupByOutputRowResolver.putExpression(value, new ColumnInfo(field, udaf.returnType, "", false));
    }
    float groupByMemoryUsage = HiveConf.getFloatVar(conf, HiveConf.ConfVars.HIVEMAPAGGRHASHMEMORY);
    float memoryThreshold = HiveConf.getFloatVar(conf, HiveConf.ConfVars.HIVEMAPAGGRMEMORYTHRESHOLD);
    // Nothing special needs to be done for grouping sets if
    // this is the final group by operator, and multiple rows corresponding to the
    // grouping sets have been generated upstream.
    // However, if an addition MR job has been created to handle grouping sets,
    // additional rows corresponding to grouping sets need to be created here.
    Operator op = putOpInsertMap(OperatorFactory.getAndMakeChild(new GroupByDesc(mode, outputColumnNames, groupByKeys, aggregations, groupByMemoryUsage, memoryThreshold, groupingSets, groupingSetsPresent && groupingSetsNeedAdditionalMRJob, groupingSetsPosition, containsDistinctAggr), new RowSchema(groupByOutputRowResolver.getColumnInfos()), reduceSinkOperatorInfo), groupByOutputRowResolver);
    op.setColumnExprMap(colExprMap);
    return op;
}

Example 8 with GenericUDAFEvaluator

use of org.apache.hadoop.hive.ql.udf.generic.GenericUDAFEvaluator in project hive by apache.

the class SemanticAnalyzer method genGroupByPlan2MR.

/**
   * Generate a Group-By plan using a 2 map-reduce jobs (5 operators will be
   * inserted):
   *
   * ReduceSink ( keys = (K1_EXP, K2_EXP, DISTINCT_EXP), values = (A1_EXP,
   * A2_EXP) ) NOTE: If DISTINCT_EXP is null, partition by rand() SortGroupBy
   * (keys = (KEY.0,KEY.1), aggregations = (count_distinct(KEY.2), sum(VALUE.0),
   * count(VALUE.1))) ReduceSink ( keys = (0,1), values=(2,3,4)) SortGroupBy
   * (keys = (KEY.0,KEY.1), aggregations = (sum(VALUE.0), sum(VALUE.1),
   * sum(VALUE.2))) Select (final selects).
   *
   * @param dest
   * @param qb
   * @param input
   * @return
   * @throws SemanticException
   *
   *           Generate a Group-By plan using a 2 map-reduce jobs. Spray by the
   *           grouping key and distinct key (or a random number, if no distinct
   *           is present) in hope of getting a uniform distribution, and
   *           compute partial aggregates grouped by the reduction key (grouping
   *           key + distinct key). Evaluate partial aggregates first, and spray
   *           by the grouping key to compute actual aggregates in the second
   *           phase. The aggregation evaluation functions are as follows:
   *           Partitioning Key: random() if no DISTINCT grouping + distinct key
   *           if DISTINCT
   *
   *           Sorting Key: grouping key if no DISTINCT grouping + distinct key
   *           if DISTINCT
   *
   *           Reducer: iterate/terminatePartial (mode = PARTIAL1)
   *
   *           STAGE 2
   *
   *           Partitioning Key: grouping key
   *
   *           Sorting Key: grouping key if no DISTINCT grouping + distinct key
   *           if DISTINCT
   *
   *           Reducer: merge/terminate (mode = FINAL)
   */
@SuppressWarnings("nls")
private Operator genGroupByPlan2MR(String dest, QB qb, Operator input) throws SemanticException {
    QBParseInfo parseInfo = qb.getParseInfo();
    ObjectPair<List<ASTNode>, List<Integer>> grpByExprsGroupingSets = getGroupByGroupingSetsForClause(parseInfo, dest);
    List<ASTNode> grpByExprs = grpByExprsGroupingSets.getFirst();
    List<Integer> groupingSets = grpByExprsGroupingSets.getSecond();
    // HIVE-3508 has been filed for this
    if (!groupingSets.isEmpty()) {
        throw new SemanticException(ErrorMsg.HIVE_GROUPING_SETS_AGGR_NOMAPAGGR.getMsg());
    }
    // ////// 1. Generate ReduceSinkOperator
    // There is a special case when we want the rows to be randomly distributed
    // to
    // reducers for load balancing problem. That happens when there is no
    // DISTINCT
    // operator. We set the numPartitionColumns to -1 for this purpose. This is
    // captured by WritableComparableHiveObject.hashCode() function.
    ReduceSinkOperator reduceSinkOperatorInfo = genGroupByPlanReduceSinkOperator(qb, dest, input, grpByExprs, (parseInfo.getDistinctFuncExprsForClause(dest).isEmpty() ? -1 : Integer.MAX_VALUE), false, -1, false, false);
    // ////// 2. Generate GroupbyOperator
    Map<String, GenericUDAFEvaluator> genericUDAFEvaluators = new LinkedHashMap<String, GenericUDAFEvaluator>();
    GroupByOperator groupByOperatorInfo = (GroupByOperator) genGroupByPlanGroupByOperator(parseInfo, dest, reduceSinkOperatorInfo, reduceSinkOperatorInfo, GroupByDesc.Mode.PARTIAL1, genericUDAFEvaluators);
    int numReducers = -1;
    if (grpByExprs.isEmpty()) {
        numReducers = 1;
    }
    // ////// 3. Generate ReduceSinkOperator2
    Operator reduceSinkOperatorInfo2 = genGroupByPlanReduceSinkOperator2MR(parseInfo, dest, groupByOperatorInfo, grpByExprs.size(), numReducers, false);
    // ////// 4. Generate GroupbyOperator2
    Operator groupByOperatorInfo2 = genGroupByPlanGroupByOperator2MR(parseInfo, dest, reduceSinkOperatorInfo2, GroupByDesc.Mode.FINAL, genericUDAFEvaluators, false);
    return groupByOperatorInfo2;
}

Example 9 with GenericUDAFEvaluator

use of org.apache.hadoop.hive.ql.udf.generic.GenericUDAFEvaluator in project hive by apache.

the class SemanticAnalyzer method genGroupByPlanMapGroupByOperator.

/**
   * Generate the map-side GroupByOperator for the Query Block
   * (qb.getParseInfo().getXXX(dest)). The new GroupByOperator will be a child
   * of the inputOperatorInfo.
   *
   * @param mode
   *          The mode of the aggregation (HASH)
   * @param genericUDAFEvaluators
   *          If not null, this function will store the mapping from Aggregation
   *          StringTree to the genericUDAFEvaluator in this parameter, so it
   *          can be used in the next-stage GroupBy aggregations.
   * @return the new GroupByOperator
   */
@SuppressWarnings("nls")
private Operator genGroupByPlanMapGroupByOperator(QB qb, String dest, List<ASTNode> grpByExprs, Operator inputOperatorInfo, GroupByDesc.Mode mode, Map<String, GenericUDAFEvaluator> genericUDAFEvaluators, List<Integer> groupingSetKeys, boolean groupingSetsPresent) throws SemanticException {
    RowResolver groupByInputRowResolver = opParseCtx.get(inputOperatorInfo).getRowResolver();
    QBParseInfo parseInfo = qb.getParseInfo();
    RowResolver groupByOutputRowResolver = new RowResolver();
    groupByOutputRowResolver.setIsExprResolver(true);
    ArrayList<ExprNodeDesc> groupByKeys = new ArrayList<ExprNodeDesc>();
    ArrayList<String> outputColumnNames = new ArrayList<String>();
    ArrayList<AggregationDesc> aggregations = new ArrayList<AggregationDesc>();
    Map<String, ExprNodeDesc> colExprMap = new HashMap<String, ExprNodeDesc>();
    for (int i = 0; i < grpByExprs.size(); ++i) {
        ASTNode grpbyExpr = grpByExprs.get(i);
        ExprNodeDesc grpByExprNode = genExprNodeDesc(grpbyExpr, groupByInputRowResolver);
        if ((grpByExprNode instanceof ExprNodeColumnDesc) && ExprNodeDescUtils.indexOf(grpByExprNode, groupByKeys) >= 0) {
            // Skip duplicated grouping keys, it happens when define column alias.
            grpByExprs.remove(i--);
            continue;
        }
        groupByKeys.add(grpByExprNode);
        String field = getColumnInternalName(i);
        outputColumnNames.add(field);
        groupByOutputRowResolver.putExpression(grpbyExpr, new ColumnInfo(field, grpByExprNode.getTypeInfo(), "", false));
        colExprMap.put(field, groupByKeys.get(groupByKeys.size() - 1));
    }
    // The grouping set key is present after the grouping keys, before the distinct keys
    int groupingSetsPosition = -1;
    // for the grouping set (corresponding to the rollup).
    if (groupingSetsPresent) {
        groupingSetsPosition = groupByKeys.size();
        createNewGroupingKey(groupByKeys, outputColumnNames, groupByOutputRowResolver, colExprMap);
    }
    // If there is a distinctFuncExp, add all parameters to the reduceKeys.
    if (!parseInfo.getDistinctFuncExprsForClause(dest).isEmpty()) {
        List<ASTNode> list = parseInfo.getDistinctFuncExprsForClause(dest);
        for (ASTNode value : list) {
            // 0 is function name
            for (int i = 1; i < value.getChildCount(); i++) {
                ASTNode parameter = (ASTNode) value.getChild(i);
                if (groupByOutputRowResolver.getExpression(parameter) == null) {
                    ExprNodeDesc distExprNode = genExprNodeDesc(parameter, groupByInputRowResolver);
                    groupByKeys.add(distExprNode);
                    String field = getColumnInternalName(groupByKeys.size() - 1);
                    outputColumnNames.add(field);
                    groupByOutputRowResolver.putExpression(parameter, new ColumnInfo(field, distExprNode.getTypeInfo(), "", false));
                    colExprMap.put(field, groupByKeys.get(groupByKeys.size() - 1));
                }
            }
        }
    }
    // For each aggregation
    HashMap<String, ASTNode> aggregationTrees = parseInfo.getAggregationExprsForClause(dest);
    assert (aggregationTrees != null);
    boolean containsDistinctAggr = false;
    for (Map.Entry<String, ASTNode> entry : aggregationTrees.entrySet()) {
        ASTNode value = entry.getValue();
        String aggName = unescapeIdentifier(value.getChild(0).getText());
        ArrayList<ExprNodeDesc> aggParameters = new ArrayList<ExprNodeDesc>();
        // 0 is the function name
        for (int i = 1; i < value.getChildCount(); i++) {
            ASTNode paraExpr = (ASTNode) value.getChild(i);
            ExprNodeDesc paraExprNode = genExprNodeDesc(paraExpr, groupByInputRowResolver);
            aggParameters.add(paraExprNode);
        }
        boolean isDistinct = value.getType() == HiveParser.TOK_FUNCTIONDI;
        containsDistinctAggr = containsDistinctAggr || isDistinct;
        boolean isAllColumns = value.getType() == HiveParser.TOK_FUNCTIONSTAR;
        Mode amode = groupByDescModeToUDAFMode(mode, isDistinct);
        GenericUDAFEvaluator genericUDAFEvaluator = getGenericUDAFEvaluator(aggName, aggParameters, value, isDistinct, isAllColumns);
        assert (genericUDAFEvaluator != null);
        GenericUDAFInfo udaf = getGenericUDAFInfo(genericUDAFEvaluator, amode, aggParameters);
        aggregations.add(new AggregationDesc(aggName.toLowerCase(), udaf.genericUDAFEvaluator, udaf.convertedParameters, isDistinct, amode));
        String field = getColumnInternalName(groupByKeys.size() + aggregations.size() - 1);
        outputColumnNames.add(field);
        if (groupByOutputRowResolver.getExpression(value) == null) {
            groupByOutputRowResolver.putExpression(value, new ColumnInfo(field, udaf.returnType, "", false));
        }
        // GroupByOperators
        if (genericUDAFEvaluators != null) {
            genericUDAFEvaluators.put(entry.getKey(), genericUDAFEvaluator);
        }
    }
    float groupByMemoryUsage = HiveConf.getFloatVar(conf, HiveConf.ConfVars.HIVEMAPAGGRHASHMEMORY);
    float memoryThreshold = HiveConf.getFloatVar(conf, HiveConf.ConfVars.HIVEMAPAGGRMEMORYTHRESHOLD);
    Operator op = putOpInsertMap(OperatorFactory.getAndMakeChild(new GroupByDesc(mode, outputColumnNames, groupByKeys, aggregations, false, groupByMemoryUsage, memoryThreshold, groupingSetKeys, groupingSetsPresent, groupingSetsPosition, containsDistinctAggr), new RowSchema(groupByOutputRowResolver.getColumnInfos()), inputOperatorInfo), groupByOutputRowResolver);
    op.setColumnExprMap(colExprMap);
    return op;
}

Example 10 with GenericUDAFEvaluator

use of org.apache.hadoop.hive.ql.udf.generic.GenericUDAFEvaluator in project hive by apache.

the class SemanticAnalyzer method genGroupByPlanGroupByOperator.

/**
   * Generate the GroupByOperator for the Query Block (parseInfo.getXXX(dest)).
   * The new GroupByOperator will be a child of the reduceSinkOperatorInfo.
   *
   * @param mode
   *          The mode of the aggregation (PARTIAL1 or COMPLETE)
   * @param genericUDAFEvaluators
   *          If not null, this function will store the mapping from Aggregation
   *          StringTree to the genericUDAFEvaluator in this parameter, so it
   *          can be used in the next-stage GroupBy aggregations.
   * @return the new GroupByOperator
   */
@SuppressWarnings("nls")
private Operator genGroupByPlanGroupByOperator(QBParseInfo parseInfo, String dest, Operator input, ReduceSinkOperator rs, GroupByDesc.Mode mode, Map<String, GenericUDAFEvaluator> genericUDAFEvaluators) throws SemanticException {
    RowResolver groupByInputRowResolver = opParseCtx.get(input).getRowResolver();
    RowResolver groupByOutputRowResolver = new RowResolver();
    groupByOutputRowResolver.setIsExprResolver(true);
    ArrayList<ExprNodeDesc> groupByKeys = new ArrayList<ExprNodeDesc>();
    ArrayList<AggregationDesc> aggregations = new ArrayList<AggregationDesc>();
    ArrayList<String> outputColumnNames = new ArrayList<String>();
    Map<String, ExprNodeDesc> colExprMap = new HashMap<String, ExprNodeDesc>();
    List<ASTNode> grpByExprs = getGroupByForClause(parseInfo, dest);
    for (int i = 0; i < grpByExprs.size(); ++i) {
        ASTNode grpbyExpr = grpByExprs.get(i);
        ColumnInfo exprInfo = groupByInputRowResolver.getExpression(grpbyExpr);
        if (exprInfo == null) {
            throw new SemanticException(ErrorMsg.INVALID_COLUMN.getMsg(grpbyExpr));
        }
        groupByKeys.add(new ExprNodeColumnDesc(exprInfo.getType(), exprInfo.getInternalName(), "", false));
        String field = getColumnInternalName(i);
        outputColumnNames.add(field);
        ColumnInfo oColInfo = new ColumnInfo(field, exprInfo.getType(), null, false);
        groupByOutputRowResolver.putExpression(grpbyExpr, oColInfo);
        addAlternateGByKeyMappings(grpbyExpr, oColInfo, input, groupByOutputRowResolver);
        colExprMap.put(field, groupByKeys.get(groupByKeys.size() - 1));
    }
    // For each aggregation
    HashMap<String, ASTNode> aggregationTrees = parseInfo.getAggregationExprsForClause(dest);
    assert (aggregationTrees != null);
    // get the last colName for the reduce KEY
    // it represents the column name corresponding to distinct aggr, if any
    String lastKeyColName = null;
    List<String> inputKeyCols = rs.getConf().getOutputKeyColumnNames();
    if (inputKeyCols.size() > 0) {
        lastKeyColName = inputKeyCols.get(inputKeyCols.size() - 1);
    }
    List<ExprNodeDesc> reduceValues = rs.getConf().getValueCols();
    int numDistinctUDFs = 0;
    for (Map.Entry<String, ASTNode> entry : aggregationTrees.entrySet()) {
        ASTNode value = entry.getValue();
        // This is the GenericUDAF name
        String aggName = unescapeIdentifier(value.getChild(0).getText());
        boolean isDistinct = value.getType() == HiveParser.TOK_FUNCTIONDI;
        boolean isAllColumns = value.getType() == HiveParser.TOK_FUNCTIONSTAR;
        // Convert children to aggParameters
        ArrayList<ExprNodeDesc> aggParameters = new ArrayList<ExprNodeDesc>();
        // 0 is the function name
        for (int i = 1; i < value.getChildCount(); i++) {
            ASTNode paraExpr = (ASTNode) value.getChild(i);
            ColumnInfo paraExprInfo = groupByInputRowResolver.getExpression(paraExpr);
            if (paraExprInfo == null) {
                throw new SemanticException(ErrorMsg.INVALID_COLUMN.getMsg(paraExpr));
            }
            String paraExpression = paraExprInfo.getInternalName();
            assert (paraExpression != null);
            if (isDistinct && lastKeyColName != null) {
                // if aggr is distinct, the parameter is name is constructed as
                // KEY.lastKeyColName:<tag>._colx
                paraExpression = Utilities.ReduceField.KEY.name() + "." + lastKeyColName + ":" + numDistinctUDFs + "." + getColumnInternalName(i - 1);
            }
            ExprNodeDesc expr = new ExprNodeColumnDesc(paraExprInfo.getType(), paraExpression, paraExprInfo.getTabAlias(), paraExprInfo.getIsVirtualCol());
            ExprNodeDesc reduceValue = isConstantParameterInAggregationParameters(paraExprInfo.getInternalName(), reduceValues);
            if (reduceValue != null) {
                // this parameter is a constant
                expr = reduceValue;
            }
            aggParameters.add(expr);
        }
        if (isDistinct) {
            numDistinctUDFs++;
        }
        Mode amode = groupByDescModeToUDAFMode(mode, isDistinct);
        GenericUDAFEvaluator genericUDAFEvaluator = getGenericUDAFEvaluator(aggName, aggParameters, value, isDistinct, isAllColumns);
        assert (genericUDAFEvaluator != null);
        GenericUDAFInfo udaf = getGenericUDAFInfo(genericUDAFEvaluator, amode, aggParameters);
        aggregations.add(new AggregationDesc(aggName.toLowerCase(), udaf.genericUDAFEvaluator, udaf.convertedParameters, isDistinct, amode));
        String field = getColumnInternalName(groupByKeys.size() + aggregations.size() - 1);
        outputColumnNames.add(field);
        groupByOutputRowResolver.putExpression(value, new ColumnInfo(field, udaf.returnType, "", false));
        // GroupByOperators
        if (genericUDAFEvaluators != null) {
            genericUDAFEvaluators.put(entry.getKey(), genericUDAFEvaluator);
        }
    }
    float groupByMemoryUsage = HiveConf.getFloatVar(conf, HiveConf.ConfVars.HIVEMAPAGGRHASHMEMORY);
    float memoryThreshold = HiveConf.getFloatVar(conf, HiveConf.ConfVars.HIVEMAPAGGRMEMORYTHRESHOLD);
    Operator op = putOpInsertMap(OperatorFactory.getAndMakeChild(new GroupByDesc(mode, outputColumnNames, groupByKeys, aggregations, false, groupByMemoryUsage, memoryThreshold, null, false, -1, numDistinctUDFs > 0), new RowSchema(groupByOutputRowResolver.getColumnInfos()), input), groupByOutputRowResolver);
    op.setColumnExprMap(colExprMap);
    return op;
}