Examples with SparkWork org.apache.hadoop.hive.ql.plan.SparkWork used on opensource projects

Example 1 with SparkWork

use of org.apache.hadoop.hive.ql.plan.SparkWork in project hive by apache.

the class GenMapRedUtils method addStatsTask.

/**
   * Add the StatsTask as a dependent task of the MoveTask
   * because StatsTask will change the Table/Partition metadata. For atomicity, we
   * should not change it before the data is actually there done by MoveTask.
   *
   * @param nd
   *          the FileSinkOperator whose results are taken care of by the MoveTask.
   * @param mvTask
   *          The MoveTask that moves the FileSinkOperator's results.
   * @param currTask
   *          The MapRedTask that the FileSinkOperator belongs to.
   * @param hconf
   *          HiveConf
   */
public static void addStatsTask(FileSinkOperator nd, MoveTask mvTask, Task<? extends Serializable> currTask, HiveConf hconf) {
    MoveWork mvWork = mvTask.getWork();
    StatsWork statsWork = null;
    if (mvWork.getLoadTableWork() != null) {
        statsWork = new StatsWork(mvWork.getLoadTableWork());
    } else if (mvWork.getLoadFileWork() != null) {
        statsWork = new StatsWork(mvWork.getLoadFileWork());
    }
    assert statsWork != null : "Error when generating StatsTask";
    statsWork.setSourceTask(currTask);
    statsWork.setStatsReliable(hconf.getBoolVar(ConfVars.HIVE_STATS_RELIABLE));
    statsWork.setStatsTmpDir(nd.getConf().getStatsTmpDir());
    if (currTask.getWork() instanceof MapredWork) {
        MapredWork mrWork = (MapredWork) currTask.getWork();
        mrWork.getMapWork().setGatheringStats(true);
        if (mrWork.getReduceWork() != null) {
            mrWork.getReduceWork().setGatheringStats(true);
        }
    } else if (currTask.getWork() instanceof SparkWork) {
        SparkWork work = (SparkWork) currTask.getWork();
        for (BaseWork w : work.getAllWork()) {
            w.setGatheringStats(true);
        }
    } else {
        // must be TezWork
        TezWork work = (TezWork) currTask.getWork();
        for (BaseWork w : work.getAllWork()) {
            w.setGatheringStats(true);
        }
    }
    // AggKey in StatsWork is used for stats aggregation while StatsAggPrefix
    // in FileSinkDesc is used for stats publishing. They should be consistent.
    statsWork.setAggKey(nd.getConf().getStatsAggPrefix());
    Task<? extends Serializable> statsTask = TaskFactory.get(statsWork, hconf);
    // subscribe feeds from the MoveTask so that MoveTask can forward the list
    // of dynamic partition list to the StatsTask
    mvTask.addDependentTask(statsTask);
    statsTask.subscribeFeed(mvTask);
}

Example 2 with SparkWork

use of org.apache.hadoop.hive.ql.plan.SparkWork in project hive by apache.

the class SparkProcessAnalyzeTable method process.

@SuppressWarnings("unchecked")
@Override
public Object process(Node nd, Stack<Node> stack, NodeProcessorCtx procContext, Object... nodeOutputs) throws SemanticException {
    GenSparkProcContext context = (GenSparkProcContext) procContext;
    TableScanOperator tableScan = (TableScanOperator) nd;
    ParseContext parseContext = context.parseContext;
    @SuppressWarnings("rawtypes") Class<? extends InputFormat> inputFormat = tableScan.getConf().getTableMetadata().getInputFormatClass();
    if (parseContext.getQueryProperties().isAnalyzeCommand()) {
        Preconditions.checkArgument(tableScan.getChildOperators() == null || tableScan.getChildOperators().size() == 0, "AssertionError: expected tableScan.getChildOperators() to be null, " + "or tableScan.getChildOperators().size() to be 0");
        String alias = null;
        for (String a : parseContext.getTopOps().keySet()) {
            if (tableScan == parseContext.getTopOps().get(a)) {
                alias = a;
            }
        }
        Preconditions.checkArgument(alias != null, "AssertionError: expected alias to be not null");
        SparkWork sparkWork = context.currentTask.getWork();
        boolean partialScan = parseContext.getQueryProperties().isPartialScanAnalyzeCommand();
        boolean noScan = parseContext.getQueryProperties().isNoScanAnalyzeCommand();
        if (inputFormat.equals(OrcInputFormat.class) && (noScan || partialScan)) {
            // ANALYZE TABLE T [PARTITION (...)] COMPUTE STATISTICS partialscan;
            // ANALYZE TABLE T [PARTITION (...)] COMPUTE STATISTICS noscan;
            // There will not be any Spark job above this task
            StatsNoJobWork snjWork = new StatsNoJobWork(tableScan.getConf().getTableMetadata().getTableSpec());
            snjWork.setStatsReliable(parseContext.getConf().getBoolVar(HiveConf.ConfVars.HIVE_STATS_RELIABLE));
            Task<StatsNoJobWork> snjTask = TaskFactory.get(snjWork, parseContext.getConf());
            snjTask.setParentTasks(null);
            context.rootTasks.remove(context.currentTask);
            context.rootTasks.add(snjTask);
            return true;
        } else {
            // ANALYZE TABLE T [PARTITION (...)] COMPUTE STATISTICS;
            // The plan consists of a simple SparkTask followed by a StatsTask.
            // The Spark task is just a simple TableScanOperator
            StatsWork statsWork = new StatsWork(tableScan.getConf().getTableMetadata().getTableSpec());
            statsWork.setAggKey(tableScan.getConf().getStatsAggPrefix());
            statsWork.setStatsTmpDir(tableScan.getConf().getTmpStatsDir());
            statsWork.setSourceTask(context.currentTask);
            statsWork.setStatsReliable(parseContext.getConf().getBoolVar(HiveConf.ConfVars.HIVE_STATS_RELIABLE));
            Task<StatsWork> statsTask = TaskFactory.get(statsWork, parseContext.getConf());
            context.currentTask.addDependentTask(statsTask);
            // The plan consists of a StatsTask only.
            if (parseContext.getQueryProperties().isNoScanAnalyzeCommand()) {
                statsTask.setParentTasks(null);
                statsWork.setNoScanAnalyzeCommand(true);
                context.rootTasks.remove(context.currentTask);
                context.rootTasks.add(statsTask);
            }
            // ANALYZE TABLE T [PARTITION (...)] COMPUTE STATISTICS partialscan;
            if (parseContext.getQueryProperties().isPartialScanAnalyzeCommand()) {
                handlePartialScanCommand(tableScan, parseContext, statsWork, context, statsTask);
            }
            // NOTE: here we should use the new partition predicate pushdown API to get a list of pruned list,
            // and pass it to setTaskPlan as the last parameter
            Set<Partition> confirmedPartns = GenMapRedUtils.getConfirmedPartitionsForScan(tableScan);
            PrunedPartitionList partitions = null;
            if (confirmedPartns.size() > 0) {
                Table source = tableScan.getConf().getTableMetadata();
                List<String> partCols = GenMapRedUtils.getPartitionColumns(tableScan);
                partitions = new PrunedPartitionList(source, confirmedPartns, partCols, false);
            }
            MapWork w = utils.createMapWork(context, tableScan, sparkWork, partitions);
            w.setGatheringStats(true);
            return true;
        }
    }
    return null;
}

Example 3 with SparkWork

use of org.apache.hadoop.hive.ql.plan.SparkWork in project hive by apache.

the class GenMapRedUtils method createMRWorkForMergingFiles.

/**
   * @param fsInput The FileSink operator.
   * @param ctx The MR processing context.
   * @param finalName the final destination path the merge job should output.
   * @param dependencyTask
   * @param mvTasks
   * @param conf
   * @param currTask
   * @throws SemanticException

   * create a Map-only merge job using CombineHiveInputFormat for all partitions with
   * following operators:
   *          MR job J0:
   *          ...
   *          |
   *          v
   *          FileSinkOperator_1 (fsInput)
   *          |
   *          v
   *          Merge job J1:
   *          |
   *          v
   *          TableScan (using CombineHiveInputFormat) (tsMerge)
   *          |
   *          v
   *          FileSinkOperator (fsMerge)
   *
   *          Here the pathToPartitionInfo & pathToAlias will remain the same, which means the paths
   *          do
   *          not contain the dynamic partitions (their parent). So after the dynamic partitions are
   *          created (after the first job finished before the moveTask or ConditionalTask start),
   *          we need to change the pathToPartitionInfo & pathToAlias to include the dynamic
   *          partition
   *          directories.
   *
   */
public static void createMRWorkForMergingFiles(FileSinkOperator fsInput, Path finalName, DependencyCollectionTask dependencyTask, List<Task<MoveWork>> mvTasks, HiveConf conf, Task<? extends Serializable> currTask) throws SemanticException {
    //
    // 1. create the operator tree
    //
    FileSinkDesc fsInputDesc = fsInput.getConf();
    // Create a TableScan operator
    RowSchema inputRS = fsInput.getSchema();
    TableScanOperator tsMerge = GenMapRedUtils.createTemporaryTableScanOperator(fsInput.getCompilationOpContext(), inputRS);
    // Create a FileSink operator
    TableDesc ts = (TableDesc) fsInputDesc.getTableInfo().clone();
    FileSinkDesc fsOutputDesc = new FileSinkDesc(finalName, ts, conf.getBoolVar(ConfVars.COMPRESSRESULT));
    FileSinkOperator fsOutput = (FileSinkOperator) OperatorFactory.getAndMakeChild(fsOutputDesc, inputRS, tsMerge);
    // If the input FileSinkOperator is a dynamic partition enabled, the tsMerge input schema
    // needs to include the partition column, and the fsOutput should have
    // a DynamicPartitionCtx to indicate that it needs to dynamically partitioned.
    DynamicPartitionCtx dpCtx = fsInputDesc.getDynPartCtx();
    if (dpCtx != null && dpCtx.getNumDPCols() > 0) {
        // adding DP ColumnInfo to the RowSchema signature
        ArrayList<ColumnInfo> signature = inputRS.getSignature();
        String tblAlias = fsInputDesc.getTableInfo().getTableName();
        for (String dpCol : dpCtx.getDPColNames()) {
            ColumnInfo colInfo = new ColumnInfo(dpCol, // all partition column type should be string
            TypeInfoFactory.stringTypeInfo, tblAlias, // partition column is virtual column
            true);
            signature.add(colInfo);
        }
        inputRS.setSignature(signature);
        // create another DynamicPartitionCtx, which has a different input-to-DP column mapping
        DynamicPartitionCtx dpCtx2 = new DynamicPartitionCtx(dpCtx);
        fsOutputDesc.setDynPartCtx(dpCtx2);
        // update the FileSinkOperator to include partition columns
        usePartitionColumns(fsInputDesc.getTableInfo().getProperties(), dpCtx.getDPColNames());
    } else {
        // non-partitioned table
        fsInputDesc.getTableInfo().getProperties().remove(org.apache.hadoop.hive.metastore.api.hive_metastoreConstants.META_TABLE_PARTITION_COLUMNS);
    }
    //
    // 2. Constructing a conditional task consisting of a move task and a map reduce task
    //
    MoveWork dummyMv = new MoveWork(null, null, null, new LoadFileDesc(fsInputDesc.getFinalDirName(), finalName, true, null, null), false);
    MapWork cplan;
    Serializable work;
    if ((conf.getBoolVar(ConfVars.HIVEMERGERCFILEBLOCKLEVEL) && fsInputDesc.getTableInfo().getInputFileFormatClass().equals(RCFileInputFormat.class)) || (conf.getBoolVar(ConfVars.HIVEMERGEORCFILESTRIPELEVEL) && fsInputDesc.getTableInfo().getInputFileFormatClass().equals(OrcInputFormat.class))) {
        cplan = GenMapRedUtils.createMergeTask(fsInputDesc, finalName, dpCtx != null && dpCtx.getNumDPCols() > 0, fsInput.getCompilationOpContext());
        if (conf.getVar(ConfVars.HIVE_EXECUTION_ENGINE).equals("tez")) {
            work = new TezWork(conf.getVar(HiveConf.ConfVars.HIVEQUERYID), conf);
            cplan.setName("File Merge");
            ((TezWork) work).add(cplan);
        } else if (conf.getVar(ConfVars.HIVE_EXECUTION_ENGINE).equals("spark")) {
            work = new SparkWork(conf.getVar(HiveConf.ConfVars.HIVEQUERYID));
            cplan.setName("Spark Merge File Work");
            ((SparkWork) work).add(cplan);
        } else {
            work = cplan;
        }
    } else {
        cplan = createMRWorkForMergingFiles(conf, tsMerge, fsInputDesc);
        if (conf.getVar(ConfVars.HIVE_EXECUTION_ENGINE).equals("tez")) {
            work = new TezWork(conf.getVar(HiveConf.ConfVars.HIVEQUERYID), conf);
            cplan.setName("File Merge");
            ((TezWork) work).add(cplan);
        } else if (conf.getVar(ConfVars.HIVE_EXECUTION_ENGINE).equals("spark")) {
            work = new SparkWork(conf.getVar(HiveConf.ConfVars.HIVEQUERYID));
            cplan.setName("Spark Merge File Work");
            ((SparkWork) work).add(cplan);
        } else {
            work = new MapredWork();
            ((MapredWork) work).setMapWork(cplan);
        }
    }
    // use CombineHiveInputFormat for map-only merging
    cplan.setInputformat("org.apache.hadoop.hive.ql.io.CombineHiveInputFormat");
    // NOTE: we should gather stats in MR1 rather than MR2 at merge job since we don't
    // know if merge MR2 will be triggered at execution time
    Task<MoveWork> mvTask = GenMapRedUtils.findMoveTask(mvTasks, fsOutput);
    ConditionalTask cndTsk = GenMapRedUtils.createCondTask(conf, currTask, dummyMv, work, fsInputDesc.getFinalDirName(), finalName, mvTask, dependencyTask);
    // keep the dynamic partition context in conditional task resolver context
    ConditionalResolverMergeFilesCtx mrCtx = (ConditionalResolverMergeFilesCtx) cndTsk.getResolverCtx();
    mrCtx.setDPCtx(fsInputDesc.getDynPartCtx());
    mrCtx.setLbCtx(fsInputDesc.getLbCtx());
}

Example 4 with SparkWork

use of org.apache.hadoop.hive.ql.plan.SparkWork in project hive by apache.

the class GenMapRedUtils method isMergeRequired.

/**
   * Returns true iff the fsOp requires a merge
   * @param mvTasks
   * @param hconf
   * @param fsOp
   * @param currTask
   * @param isInsertTable
   * @return
   */
public static boolean isMergeRequired(List<Task<MoveWork>> mvTasks, HiveConf hconf, FileSinkOperator fsOp, Task<? extends Serializable> currTask, boolean isInsertTable) {
    // Has the user enabled merging of files for map-only jobs or for all jobs
    if ((mvTasks != null) && (!mvTasks.isEmpty())) {
        // no need of merging if the move is to a local file system
        MoveTask mvTask = (MoveTask) GenMapRedUtils.findMoveTask(mvTasks, fsOp);
        if (mvTask != null && isInsertTable && hconf.getBoolVar(ConfVars.HIVESTATSAUTOGATHER) && !fsOp.getConf().isMaterialization()) {
            // mark the MapredWork and FileSinkOperator for gathering stats
            fsOp.getConf().setGatherStats(true);
            fsOp.getConf().setStatsReliable(hconf.getBoolVar(ConfVars.HIVE_STATS_RELIABLE));
            if (!mvTask.hasFollowingStatsTask()) {
                GenMapRedUtils.addStatsTask(fsOp, mvTask, currTask, hconf);
            }
        }
        if ((mvTask != null) && !mvTask.isLocal() && fsOp.getConf().canBeMerged()) {
            if (currTask.getWork() instanceof TezWork) {
                // config
                return hconf.getBoolVar(ConfVars.HIVEMERGETEZFILES);
            } else if (currTask.getWork() instanceof SparkWork) {
                // spark has its own config for merging
                return hconf.getBoolVar(ConfVars.HIVEMERGESPARKFILES);
            }
            if (fsOp.getConf().isLinkedFileSink()) {
                // possibly by a big margin. So, merge aggresively.
                if (hconf.getBoolVar(ConfVars.HIVEMERGEMAPFILES) || hconf.getBoolVar(ConfVars.HIVEMERGEMAPREDFILES)) {
                    return true;
                }
            } else {
                // or for a map-reduce job
                if (currTask.getWork() instanceof MapredWork) {
                    ReduceWork reduceWork = ((MapredWork) currTask.getWork()).getReduceWork();
                    boolean mergeMapOnly = hconf.getBoolVar(ConfVars.HIVEMERGEMAPFILES) && reduceWork == null;
                    boolean mergeMapRed = hconf.getBoolVar(ConfVars.HIVEMERGEMAPREDFILES) && reduceWork != null;
                    if (mergeMapOnly || mergeMapRed) {
                        return true;
                    }
                } else {
                    return false;
                }
            }
        }
    }
    return false;
}

Example 5 with SparkWork

use of org.apache.hadoop.hive.ql.plan.SparkWork in project hive by apache.

the class SparkReduceSinkMapJoinProc method process.

/* (non-Javadoc)
   * This processor addresses the RS-MJ case that occurs in spark on the small/hash
   * table side of things. The work that RS will be a part of must be connected
   * to the MJ work via be a broadcast edge.
   * We should not walk down the tree when we encounter this pattern because:
   * the type of work (map work or reduce work) needs to be determined
   * on the basis of the big table side because it may be a mapwork (no need for shuffle)
   * or reduce work.
   */
@SuppressWarnings("unchecked")
@Override
public Object process(Node nd, Stack<Node> stack, NodeProcessorCtx procContext, Object... nodeOutputs) throws SemanticException {
    GenSparkProcContext context = (GenSparkProcContext) procContext;
    if (!nd.getClass().equals(MapJoinOperator.class)) {
        return null;
    }
    MapJoinOperator mapJoinOp = (MapJoinOperator) nd;
    if (stack.size() < 2 || !(stack.get(stack.size() - 2) instanceof ReduceSinkOperator)) {
        context.currentMapJoinOperators.add(mapJoinOp);
        return null;
    }
    context.preceedingWork = null;
    context.currentRootOperator = null;
    ReduceSinkOperator parentRS = (ReduceSinkOperator) stack.get(stack.size() - 2);
    // remove the tag for in-memory side of mapjoin
    parentRS.getConf().setSkipTag(true);
    parentRS.setSkipTag(true);
    // remember the original parent list before we start modifying it.
    if (!context.mapJoinParentMap.containsKey(mapJoinOp)) {
        List<Operator<?>> parents = new ArrayList<Operator<?>>(mapJoinOp.getParentOperators());
        context.mapJoinParentMap.put(mapJoinOp, parents);
    }
    List<BaseWork> mapJoinWork;
    /*
     *  If there was a pre-existing work generated for the big-table mapjoin side,
     *  we need to hook the work generated for the RS (associated with the RS-MJ pattern)
     *  with the pre-existing work.
     *
     *  Otherwise, we need to associate that the mapjoin op
     *  to be linked to the RS work (associated with the RS-MJ pattern).
     *
     */
    mapJoinWork = context.mapJoinWorkMap.get(mapJoinOp);
    int workMapSize = context.childToWorkMap.get(parentRS).size();
    Preconditions.checkArgument(workMapSize == 1, "AssertionError: expected context.childToWorkMap.get(parentRS).size() to be 1, but was " + workMapSize);
    BaseWork parentWork = context.childToWorkMap.get(parentRS).get(0);
    // set the link between mapjoin and parent vertex
    int pos = context.mapJoinParentMap.get(mapJoinOp).indexOf(parentRS);
    if (pos == -1) {
        throw new SemanticException("Cannot find position of parent in mapjoin");
    }
    LOG.debug("Mapjoin " + mapJoinOp + ", pos: " + pos + " --> " + parentWork.getName());
    mapJoinOp.getConf().getParentToInput().put(pos, parentWork.getName());
    SparkEdgeProperty edgeProp = new SparkEdgeProperty(SparkEdgeProperty.SHUFFLE_NONE);
    if (mapJoinWork != null) {
        for (BaseWork myWork : mapJoinWork) {
            // link the work with the work associated with the reduce sink that triggered this rule
            SparkWork sparkWork = context.currentTask.getWork();
            LOG.debug("connecting " + parentWork.getName() + " with " + myWork.getName());
            sparkWork.connect(parentWork, myWork, edgeProp);
        }
    }
    // remember in case we need to connect additional work later
    Map<BaseWork, SparkEdgeProperty> linkWorkMap = null;
    if (context.linkOpWithWorkMap.containsKey(mapJoinOp)) {
        linkWorkMap = context.linkOpWithWorkMap.get(mapJoinOp);
    } else {
        linkWorkMap = new HashMap<BaseWork, SparkEdgeProperty>();
    }
    linkWorkMap.put(parentWork, edgeProp);
    context.linkOpWithWorkMap.put(mapJoinOp, linkWorkMap);
    List<ReduceSinkOperator> reduceSinks = context.linkWorkWithReduceSinkMap.get(parentWork);
    if (reduceSinks == null) {
        reduceSinks = new ArrayList<ReduceSinkOperator>();
    }
    reduceSinks.add(parentRS);
    context.linkWorkWithReduceSinkMap.put(parentWork, reduceSinks);
    // create the dummy operators
    List<Operator<?>> dummyOperators = new ArrayList<Operator<?>>();
    // create an new operator: HashTableDummyOperator, which share the table desc
    HashTableDummyDesc desc = new HashTableDummyDesc();
    HashTableDummyOperator dummyOp = (HashTableDummyOperator) OperatorFactory.get(mapJoinOp.getCompilationOpContext(), desc);
    TableDesc tbl;
    // need to create the correct table descriptor for key/value
    RowSchema rowSchema = parentRS.getParentOperators().get(0).getSchema();
    tbl = PlanUtils.getReduceValueTableDesc(PlanUtils.getFieldSchemasFromRowSchema(rowSchema, ""));
    dummyOp.getConf().setTbl(tbl);
    Map<Byte, List<ExprNodeDesc>> keyExprMap = mapJoinOp.getConf().getKeys();
    List<ExprNodeDesc> keyCols = keyExprMap.get(Byte.valueOf((byte) 0));
    StringBuilder keyOrder = new StringBuilder();
    StringBuilder keyNullOrder = new StringBuilder();
    for (int i = 0; i < keyCols.size(); i++) {
        keyOrder.append("+");
        keyNullOrder.append("a");
    }
    TableDesc keyTableDesc = PlanUtils.getReduceKeyTableDesc(PlanUtils.getFieldSchemasFromColumnList(keyCols, "mapjoinkey"), keyOrder.toString(), keyNullOrder.toString());
    mapJoinOp.getConf().setKeyTableDesc(keyTableDesc);
    // let the dummy op be the parent of mapjoin op
    mapJoinOp.replaceParent(parentRS, dummyOp);
    List<Operator<? extends OperatorDesc>> dummyChildren = new ArrayList<Operator<? extends OperatorDesc>>();
    dummyChildren.add(mapJoinOp);
    dummyOp.setChildOperators(dummyChildren);
    dummyOperators.add(dummyOp);
    // cut the operator tree so as to not retain connections from the parent RS downstream
    List<Operator<? extends OperatorDesc>> childOperators = parentRS.getChildOperators();
    int childIndex = childOperators.indexOf(mapJoinOp);
    childOperators.remove(childIndex);
    // at task startup
    if (mapJoinWork != null) {
        for (BaseWork myWork : mapJoinWork) {
            myWork.addDummyOp(dummyOp);
        }
    }
    if (context.linkChildOpWithDummyOp.containsKey(mapJoinOp)) {
        for (Operator<?> op : context.linkChildOpWithDummyOp.get(mapJoinOp)) {
            dummyOperators.add(op);
        }
    }
    context.linkChildOpWithDummyOp.put(mapJoinOp, dummyOperators);
    // replace ReduceSinkOp with HashTableSinkOp for the RSops which are parents of MJop
    MapJoinDesc mjDesc = mapJoinOp.getConf();
    HiveConf conf = context.conf;
    // Unlike in MR, we may call this method multiple times, for each
    // small table HTS. But, since it's idempotent, it should be OK.
    mjDesc.resetOrder();
    float hashtableMemoryUsage;
    if (hasGroupBy(mapJoinOp, context)) {
        hashtableMemoryUsage = conf.getFloatVar(HiveConf.ConfVars.HIVEHASHTABLEFOLLOWBYGBYMAXMEMORYUSAGE);
    } else {
        hashtableMemoryUsage = conf.getFloatVar(HiveConf.ConfVars.HIVEHASHTABLEMAXMEMORYUSAGE);
    }
    mjDesc.setHashTableMemoryUsage(hashtableMemoryUsage);
    SparkHashTableSinkDesc hashTableSinkDesc = new SparkHashTableSinkDesc(mjDesc);
    SparkHashTableSinkOperator hashTableSinkOp = (SparkHashTableSinkOperator) OperatorFactory.get(mapJoinOp.getCompilationOpContext(), hashTableSinkDesc);
    byte tag = (byte) pos;
    int[] valueIndex = mjDesc.getValueIndex(tag);
    if (valueIndex != null) {
        List<ExprNodeDesc> newValues = new ArrayList<ExprNodeDesc>();
        List<ExprNodeDesc> values = hashTableSinkDesc.getExprs().get(tag);
        for (int index = 0; index < values.size(); index++) {
            if (valueIndex[index] < 0) {
                newValues.add(values.get(index));
            }
        }
        hashTableSinkDesc.getExprs().put(tag, newValues);
    }
    //get all parents of reduce sink
    List<Operator<? extends OperatorDesc>> rsParentOps = parentRS.getParentOperators();
    for (Operator<? extends OperatorDesc> parent : rsParentOps) {
        parent.replaceChild(parentRS, hashTableSinkOp);
    }
    hashTableSinkOp.setParentOperators(rsParentOps);
    hashTableSinkOp.getConf().setTag(tag);
    return true;
}