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

Example 76 with TableScanOperator

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

the class SparkMapJoinOptimizer method getMapJoinConversionInfo.

/**
 *   This method returns the big table position in a map-join. If the given join
 *   cannot be converted to a map-join (This could happen for several reasons - one
 *   of them being presence of 2 or more big tables that cannot fit in-memory), it returns -1.
 *
 *   Otherwise, it returns an int value that is the index of the big table in the set
 *   MapJoinProcessor.bigTableCandidateSet
 *
 * @param joinOp
 * @param context
 * @return an array of 3 long values, first value is the position,
 *   second value is the connected map join size, and the third is big table data size.
 */
private long[] getMapJoinConversionInfo(JoinOperator joinOp, OptimizeSparkProcContext context) {
    Set<Integer> bigTableCandidateSet = MapJoinProcessor.getBigTableCandidates(joinOp.getConf().getConds());
    long maxSize = context.getConf().getLongVar(HiveConf.ConfVars.HIVECONVERTJOINNOCONDITIONALTASKTHRESHOLD);
    int bigTablePosition = -1;
    Statistics bigInputStat = null;
    long totalSize = 0;
    int pos = 0;
    // bigTableFound means we've encountered a table that's bigger than the
    // max. This table is either the big table or we cannot convert.
    boolean bigTableFound = false;
    boolean useTsStats = context.getConf().getBoolean(HiveConf.ConfVars.SPARK_USE_TS_STATS_FOR_MAPJOIN.varname, false);
    // If so, mark that branch as the big table branch.
    if (useTsStats) {
        LOG.debug("Checking map join optimization for operator {} using TS stats", joinOp);
        for (Operator<? extends OperatorDesc> parentOp : joinOp.getParentOperators()) {
            if (isBigTableBranch(parentOp)) {
                if (bigTablePosition < 0 && bigTableCandidateSet.contains(pos) && !containUnionWithoutRS(parentOp.getParentOperators().get(0))) {
                    LOG.debug("Found a big table branch with parent operator {} and position {}", parentOp, pos);
                    bigTablePosition = pos;
                    bigTableFound = true;
                    bigInputStat = new Statistics(0, Long.MAX_VALUE);
                } else {
                    // Either we've found multiple big table branches, or the current branch cannot
                    // be a big table branch. Disable mapjoin for these cases.
                    LOG.debug("Cannot enable map join optimization for operator {}", joinOp);
                    return new long[] { -1, 0, 0 };
                }
            }
            pos++;
        }
    }
    pos = 0;
    for (Operator<? extends OperatorDesc> parentOp : joinOp.getParentOperators()) {
        // Skip the potential big table identified above
        if (pos == bigTablePosition) {
            pos++;
            continue;
        }
        Statistics currInputStat = null;
        if (useTsStats) {
            // Not adding other stats (e.g., # of rows, col stats) since only data size is used here
            for (TableScanOperator root : OperatorUtils.findOperatorsUpstream(parentOp, TableScanOperator.class)) {
                if (currInputStat == null) {
                    currInputStat = root.getStatistics().clone();
                } else {
                    currInputStat.addBasicStats(root.getStatistics());
                }
            }
        } else {
            currInputStat = parentOp.getStatistics();
        }
        if (currInputStat == null) {
            LOG.warn("Couldn't get statistics from: " + parentOp);
            return new long[] { -1, 0, 0 };
        }
        // But, this is tricky to implement, and we'll leave it as a future work for now.
        if (containUnionWithoutRS(parentOp.getParentOperators().get(0))) {
            return new long[] { -1, 0, 0 };
        }
        long inputSize = currInputStat.getDataSize();
        if (bigInputStat == null || inputSize > bigInputStat.getDataSize()) {
            if (bigTableFound) {
                // on size and there's another one that's bigger.
                return new long[] { -1, 0, 0 };
            }
            if (inputSize > maxSize) {
                if (!bigTableCandidateSet.contains(pos)) {
                    // big for the map side.
                    return new long[] { -1, 0, 0 };
                }
                bigTableFound = true;
            }
            if (bigInputStat != null) {
                // we're replacing the current big table with a new one. Need
                // to count the current one as a map table then.
                totalSize += bigInputStat.getDataSize();
            }
            if (totalSize > maxSize) {
                // hence cannot convert.
                return new long[] { -1, 0, 0 };
            }
            if (bigTableCandidateSet.contains(pos)) {
                bigTablePosition = pos;
                bigInputStat = currInputStat;
            }
        } else {
            totalSize += currInputStat.getDataSize();
            if (totalSize > maxSize) {
                // cannot hold all map tables in memory. Cannot convert.
                return new long[] { -1, 0, 0 };
            }
        }
        pos++;
    }
    if (bigTablePosition == -1) {
        // No big table candidates.
        return new long[] { -1, 0, 0 };
    }
    // Final check, find size of already-calculated Mapjoin Operators in same work (spark-stage).
    // We need to factor this in to prevent overwhelming Spark executor-memory.
    long connectedMapJoinSize = getConnectedMapJoinSize(joinOp.getParentOperators().get(bigTablePosition), joinOp, context);
    if ((connectedMapJoinSize + totalSize) > maxSize) {
        return new long[] { -1, 0, 0 };
    }
    return new long[] { bigTablePosition, connectedMapJoinSize, totalSize };
}

Example 77 with TableScanOperator

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

the class GenSparkSkewJoinProcessor method processSkewJoin.

@SuppressWarnings("unchecked")
public static void processSkewJoin(JoinOperator joinOp, Task<? extends Serializable> currTask, ReduceWork reduceWork, ParseContext parseCtx) throws SemanticException {
    SparkWork currentWork = ((SparkTask) currTask).getWork();
    if (currentWork.getChildren(reduceWork).size() > 0) {
        LOG.warn("Skip runtime skew join as the ReduceWork has child work and hasn't been split.");
        return;
    }
    List<Task<? extends Serializable>> children = currTask.getChildTasks();
    Path baseTmpDir = parseCtx.getContext().getMRTmpPath();
    JoinDesc joinDescriptor = joinOp.getConf();
    Map<Byte, List<ExprNodeDesc>> joinValues = joinDescriptor.getExprs();
    int numAliases = joinValues.size();
    Map<Byte, Path> bigKeysDirMap = new HashMap<Byte, Path>();
    Map<Byte, Map<Byte, Path>> smallKeysDirMap = new HashMap<Byte, Map<Byte, Path>>();
    Map<Byte, Path> skewJoinJobResultsDir = new HashMap<Byte, Path>();
    Byte[] tags = joinDescriptor.getTagOrder();
    // for each joining table, set dir for big key and small keys properly
    for (int i = 0; i < numAliases; i++) {
        Byte alias = tags[i];
        bigKeysDirMap.put(alias, GenMRSkewJoinProcessor.getBigKeysDir(baseTmpDir, alias));
        Map<Byte, Path> smallKeysMap = new HashMap<Byte, Path>();
        smallKeysDirMap.put(alias, smallKeysMap);
        for (Byte src2 : tags) {
            if (!src2.equals(alias)) {
                smallKeysMap.put(src2, GenMRSkewJoinProcessor.getSmallKeysDir(baseTmpDir, alias, src2));
            }
        }
        skewJoinJobResultsDir.put(alias, GenMRSkewJoinProcessor.getBigKeysSkewJoinResultDir(baseTmpDir, alias));
    }
    joinDescriptor.setHandleSkewJoin(true);
    joinDescriptor.setBigKeysDirMap(bigKeysDirMap);
    joinDescriptor.setSmallKeysDirMap(smallKeysDirMap);
    joinDescriptor.setSkewKeyDefinition(HiveConf.getIntVar(parseCtx.getConf(), HiveConf.ConfVars.HIVESKEWJOINKEY));
    // create proper table/column desc for spilled tables
    TableDesc keyTblDesc = (TableDesc) reduceWork.getKeyDesc().clone();
    List<String> joinKeys = Utilities.getColumnNames(keyTblDesc.getProperties());
    List<String> joinKeyTypes = Utilities.getColumnTypes(keyTblDesc.getProperties());
    Map<Byte, TableDesc> tableDescList = new HashMap<Byte, TableDesc>();
    Map<Byte, RowSchema> rowSchemaList = new HashMap<Byte, RowSchema>();
    Map<Byte, List<ExprNodeDesc>> newJoinValues = new HashMap<Byte, List<ExprNodeDesc>>();
    Map<Byte, List<ExprNodeDesc>> newJoinKeys = new HashMap<Byte, List<ExprNodeDesc>>();
    // used for create mapJoinDesc, should be in order
    List<TableDesc> newJoinValueTblDesc = new ArrayList<TableDesc>();
    for (int i = 0; i < tags.length; i++) {
        newJoinValueTblDesc.add(null);
    }
    for (int i = 0; i < numAliases; i++) {
        Byte alias = tags[i];
        List<ExprNodeDesc> valueCols = joinValues.get(alias);
        String colNames = "";
        String colTypes = "";
        int columnSize = valueCols.size();
        List<ExprNodeDesc> newValueExpr = new ArrayList<ExprNodeDesc>();
        List<ExprNodeDesc> newKeyExpr = new ArrayList<ExprNodeDesc>();
        ArrayList<ColumnInfo> columnInfos = new ArrayList<ColumnInfo>();
        boolean first = true;
        for (int k = 0; k < columnSize; k++) {
            TypeInfo type = valueCols.get(k).getTypeInfo();
            // any name, it does not matter.
            String newColName = i + "_VALUE_" + k;
            ColumnInfo columnInfo = new ColumnInfo(newColName, type, alias.toString(), false);
            columnInfos.add(columnInfo);
            newValueExpr.add(new ExprNodeColumnDesc(columnInfo.getType(), columnInfo.getInternalName(), columnInfo.getTabAlias(), false));
            if (!first) {
                colNames = colNames + ",";
                colTypes = colTypes + ",";
            }
            first = false;
            colNames = colNames + newColName;
            colTypes = colTypes + valueCols.get(k).getTypeString();
        }
        // we are putting join keys at last part of the spilled table
        for (int k = 0; k < joinKeys.size(); k++) {
            if (!first) {
                colNames = colNames + ",";
                colTypes = colTypes + ",";
            }
            first = false;
            colNames = colNames + joinKeys.get(k);
            colTypes = colTypes + joinKeyTypes.get(k);
            ColumnInfo columnInfo = new ColumnInfo(joinKeys.get(k), TypeInfoFactory.getPrimitiveTypeInfo(joinKeyTypes.get(k)), alias.toString(), false);
            columnInfos.add(columnInfo);
            newKeyExpr.add(new ExprNodeColumnDesc(columnInfo.getType(), columnInfo.getInternalName(), columnInfo.getTabAlias(), false));
        }
        newJoinValues.put(alias, newValueExpr);
        newJoinKeys.put(alias, newKeyExpr);
        tableDescList.put(alias, Utilities.getTableDesc(colNames, colTypes));
        rowSchemaList.put(alias, new RowSchema(columnInfos));
        // construct value table Desc
        String valueColNames = "";
        String valueColTypes = "";
        first = true;
        for (int k = 0; k < columnSize; k++) {
            // any name, it does not matter.
            String newColName = i + "_VALUE_" + k;
            if (!first) {
                valueColNames = valueColNames + ",";
                valueColTypes = valueColTypes + ",";
            }
            valueColNames = valueColNames + newColName;
            valueColTypes = valueColTypes + valueCols.get(k).getTypeString();
            first = false;
        }
        newJoinValueTblDesc.set((byte) i, Utilities.getTableDesc(valueColNames, valueColTypes));
    }
    joinDescriptor.setSkewKeysValuesTables(tableDescList);
    joinDescriptor.setKeyTableDesc(keyTblDesc);
    // create N-1 map join tasks
    HashMap<Path, Task<? extends Serializable>> bigKeysDirToTaskMap = new HashMap<Path, Task<? extends Serializable>>();
    List<Serializable> listWorks = new ArrayList<Serializable>();
    List<Task<? extends Serializable>> listTasks = new ArrayList<Task<? extends Serializable>>();
    for (int i = 0; i < numAliases - 1; i++) {
        Byte src = tags[i];
        HiveConf hiveConf = new HiveConf(parseCtx.getConf(), GenSparkSkewJoinProcessor.class);
        SparkWork sparkWork = new SparkWork(parseCtx.getConf().getVar(HiveConf.ConfVars.HIVEQUERYID));
        Task<? extends Serializable> skewJoinMapJoinTask = TaskFactory.get(sparkWork);
        skewJoinMapJoinTask.setFetchSource(currTask.isFetchSource());
        // create N TableScans
        Operator<? extends OperatorDesc>[] parentOps = new TableScanOperator[tags.length];
        for (int k = 0; k < tags.length; k++) {
            Operator<? extends OperatorDesc> ts = GenMapRedUtils.createTemporaryTableScanOperator(joinOp.getCompilationOpContext(), rowSchemaList.get((byte) k));
            ((TableScanOperator) ts).setTableDescSkewJoin(tableDescList.get((byte) k));
            parentOps[k] = ts;
        }
        // create the MapJoinOperator
        String dumpFilePrefix = "mapfile" + PlanUtils.getCountForMapJoinDumpFilePrefix();
        MapJoinDesc mapJoinDescriptor = new MapJoinDesc(newJoinKeys, keyTblDesc, newJoinValues, newJoinValueTblDesc, newJoinValueTblDesc, joinDescriptor.getOutputColumnNames(), i, joinDescriptor.getConds(), joinDescriptor.getFilters(), joinDescriptor.getNoOuterJoin(), dumpFilePrefix, joinDescriptor.getMemoryMonitorInfo(), joinDescriptor.getInMemoryDataSize());
        mapJoinDescriptor.setTagOrder(tags);
        mapJoinDescriptor.setHandleSkewJoin(false);
        mapJoinDescriptor.setNullSafes(joinDescriptor.getNullSafes());
        mapJoinDescriptor.setColumnExprMap(joinDescriptor.getColumnExprMap());
        // temporarily, mark it as child of all the TS
        MapJoinOperator mapJoinOp = (MapJoinOperator) OperatorFactory.getAndMakeChild(joinOp.getCompilationOpContext(), mapJoinDescriptor, null, parentOps);
        // clone the original join operator, and replace it with the MJ
        // this makes sure MJ has the same downstream operator plan as the original join
        List<Operator<?>> reducerList = new ArrayList<Operator<?>>();
        reducerList.add(reduceWork.getReducer());
        Operator<? extends OperatorDesc> reducer = SerializationUtilities.cloneOperatorTree(reducerList).get(0);
        Preconditions.checkArgument(reducer instanceof JoinOperator, "Reducer should be join operator, but actually is " + reducer.getName());
        JoinOperator cloneJoinOp = (JoinOperator) reducer;
        List<Operator<? extends OperatorDesc>> childOps = cloneJoinOp.getChildOperators();
        for (Operator<? extends OperatorDesc> childOp : childOps) {
            childOp.replaceParent(cloneJoinOp, mapJoinOp);
        }
        mapJoinOp.setChildOperators(childOps);
        // set memory usage for the MJ operator
        setMemUsage(mapJoinOp, skewJoinMapJoinTask, parseCtx);
        // create N MapWorks and add them to the SparkWork
        MapWork bigMapWork = null;
        Map<Byte, Path> smallTblDirs = smallKeysDirMap.get(src);
        for (int j = 0; j < tags.length; j++) {
            MapWork mapWork = PlanUtils.getMapRedWork().getMapWork();
            sparkWork.add(mapWork);
            // This code has been only added for testing
            boolean mapperCannotSpanPartns = parseCtx.getConf().getBoolVar(HiveConf.ConfVars.HIVE_MAPPER_CANNOT_SPAN_MULTIPLE_PARTITIONS);
            mapWork.setMapperCannotSpanPartns(mapperCannotSpanPartns);
            Operator<? extends OperatorDesc> tableScan = parentOps[j];
            String alias = tags[j].toString();
            ArrayList<String> aliases = new ArrayList<String>();
            aliases.add(alias);
            Path path;
            if (j == i) {
                path = bigKeysDirMap.get(tags[j]);
                bigKeysDirToTaskMap.put(path, skewJoinMapJoinTask);
                bigMapWork = mapWork;
            } else {
                path = smallTblDirs.get(tags[j]);
            }
            mapWork.addPathToAlias(path, aliases);
            mapWork.getAliasToWork().put(alias, tableScan);
            PartitionDesc partitionDesc = new PartitionDesc(tableDescList.get(tags[j]), null);
            mapWork.addPathToPartitionInfo(path, partitionDesc);
            mapWork.getAliasToPartnInfo().put(alias, partitionDesc);
            mapWork.setName("Map " + GenSparkUtils.getUtils().getNextSeqNumber());
        }
        // connect all small dir map work to the big dir map work
        Preconditions.checkArgument(bigMapWork != null, "Haven't identified big dir MapWork");
        // these 2 flags are intended only for the big-key map work
        bigMapWork.setNumMapTasks(HiveConf.getIntVar(hiveConf, HiveConf.ConfVars.HIVESKEWJOINMAPJOINNUMMAPTASK));
        bigMapWork.setMinSplitSize(HiveConf.getLongVar(hiveConf, HiveConf.ConfVars.HIVESKEWJOINMAPJOINMINSPLIT));
        // use HiveInputFormat so that we can control the number of map tasks
        bigMapWork.setInputformat(HiveInputFormat.class.getName());
        for (BaseWork work : sparkWork.getRoots()) {
            Preconditions.checkArgument(work instanceof MapWork, "All root work should be MapWork, but got " + work.getClass().getSimpleName());
            if (work != bigMapWork) {
                sparkWork.connect(work, bigMapWork, new SparkEdgeProperty(SparkEdgeProperty.SHUFFLE_NONE));
            }
        }
        // insert SparkHashTableSink and Dummy operators
        for (int j = 0; j < tags.length; j++) {
            if (j != i) {
                insertSHTS(tags[j], (TableScanOperator) parentOps[j], bigMapWork);
            }
        }
        listWorks.add(skewJoinMapJoinTask.getWork());
        listTasks.add(skewJoinMapJoinTask);
    }
    if (children != null) {
        for (Task<? extends Serializable> tsk : listTasks) {
            for (Task<? extends Serializable> oldChild : children) {
                tsk.addDependentTask(oldChild);
            }
        }
        currTask.setChildTasks(new ArrayList<Task<? extends Serializable>>());
        for (Task<? extends Serializable> oldChild : children) {
            oldChild.getParentTasks().remove(currTask);
        }
        listTasks.addAll(children);
        for (Task<? extends Serializable> oldChild : children) {
            listWorks.add(oldChild.getWork());
        }
    }
    ConditionalResolverSkewJoin.ConditionalResolverSkewJoinCtx context = new ConditionalResolverSkewJoin.ConditionalResolverSkewJoinCtx(bigKeysDirToTaskMap, children);
    ConditionalWork cndWork = new ConditionalWork(listWorks);
    ConditionalTask cndTsk = (ConditionalTask) TaskFactory.get(cndWork);
    cndTsk.setListTasks(listTasks);
    cndTsk.setResolver(new ConditionalResolverSkewJoin());
    cndTsk.setResolverCtx(context);
    currTask.setChildTasks(new ArrayList<Task<? extends Serializable>>());
    currTask.addDependentTask(cndTsk);
}

Example 78 with TableScanOperator

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

the class GenSparkSkewJoinProcessor method insertSHTS.

/**
 * Insert SparkHashTableSink and HashTableDummy between small dir TS and MJ.
 */
@SuppressWarnings("unchecked")
private static void insertSHTS(byte tag, TableScanOperator tableScan, MapWork bigMapWork) {
    Preconditions.checkArgument(tableScan.getChildOperators().size() == 1 && tableScan.getChildOperators().get(0) instanceof MapJoinOperator);
    HashTableDummyDesc desc = new HashTableDummyDesc();
    HashTableDummyOperator dummyOp = (HashTableDummyOperator) OperatorFactory.get(tableScan.getCompilationOpContext(), desc);
    dummyOp.getConf().setTbl(tableScan.getTableDescSkewJoin());
    MapJoinOperator mapJoinOp = (MapJoinOperator) tableScan.getChildOperators().get(0);
    mapJoinOp.replaceParent(tableScan, dummyOp);
    List<Operator<? extends OperatorDesc>> mapJoinChildren = new ArrayList<Operator<? extends OperatorDesc>>();
    mapJoinChildren.add(mapJoinOp);
    dummyOp.setChildOperators(mapJoinChildren);
    bigMapWork.addDummyOp(dummyOp);
    MapJoinDesc mjDesc = mapJoinOp.getConf();
    // mapjoin should not be affected by join reordering
    mjDesc.resetOrder();
    SparkHashTableSinkDesc hashTableSinkDesc = new SparkHashTableSinkDesc(mjDesc);
    SparkHashTableSinkOperator hashTableSinkOp = (SparkHashTableSinkOperator) OperatorFactory.get(tableScan.getCompilationOpContext(), hashTableSinkDesc);
    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);
    }
    tableScan.replaceChild(mapJoinOp, hashTableSinkOp);
    List<Operator<? extends OperatorDesc>> tableScanParents = new ArrayList<Operator<? extends OperatorDesc>>();
    tableScanParents.add(tableScan);
    hashTableSinkOp.setParentOperators(tableScanParents);
    hashTableSinkOp.getConf().setTag(tag);
}

Example 79 with TableScanOperator

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

the class SamplingOptimizer method resolve.

public PhysicalContext resolve(PhysicalContext pctx) throws SemanticException {
    for (Task<?> task : pctx.getRootTasks()) {
        if (!(task instanceof MapRedTask) || !((MapRedTask) task).getWork().isFinalMapRed()) {
            // this could be replaced by bucketing on RS + bucketed fetcher for next MR
            continue;
        }
        MapredWork mrWork = ((MapRedTask) task).getWork();
        MapWork mapWork = mrWork.getMapWork();
        ReduceWork reduceWork = mrWork.getReduceWork();
        if (reduceWork == null || reduceWork.getNumReduceTasks() != 1 || mapWork.getAliasToWork().size() != 1 || mapWork.getSamplingType() > 0 || reduceWork.getReducer() == null) {
            continue;
        }
        // GROUPBY operator in reducer may not be processed in parallel. Skip optimizing.
        if (OperatorUtils.findSingleOperator(reduceWork.getReducer(), GroupByOperator.class) != null) {
            continue;
        }
        Operator<?> operator = mapWork.getAliasToWork().values().iterator().next();
        if (!(operator instanceof TableScanOperator)) {
            continue;
        }
        TableScanOperator tsop = (TableScanOperator) operator;
        Table tbl = tsop.getConf().getTableMetadata();
        if (tbl == null) {
            continue;
        }
        if (AcidUtils.isInsertOnlyTable(tbl.getParameters())) {
            // sampler will limit the input to the the correct directories, but we don't care about MR.
            continue;
        }
        ReduceSinkOperator child = OperatorUtils.findSingleOperator(operator, ReduceSinkOperator.class);
        if (child == null || child.getConf().getNumReducers() != 1 || !child.getConf().getPartitionCols().isEmpty()) {
            continue;
        }
        child.getConf().setNumReducers(-1);
        reduceWork.setNumReduceTasks(-1);
        mapWork.setSamplingType(MapWork.SAMPLING_ON_START);
    }
    return pctx;
}

Example 80 with TableScanOperator

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

the class TezCompiler method removeSemiJoinCyclesDueToMapsideJoins.

private static void removeSemiJoinCyclesDueToMapsideJoins(OptimizeTezProcContext procCtx) throws SemanticException {
    if (!procCtx.conf.getBoolVar(ConfVars.TEZ_DYNAMIC_SEMIJOIN_REDUCTION) || procCtx.parseContext.getRsToSemiJoinBranchInfo().size() == 0) {
        return;
    }
    Map<Rule, NodeProcessor> opRules = new LinkedHashMap<Rule, NodeProcessor>();
    opRules.put(new RuleRegExp("R1", MapJoinOperator.getOperatorName() + "%" + MapJoinOperator.getOperatorName() + "%"), new SemiJoinCycleRemovalDueToMapsideJoins());
    opRules.put(new RuleRegExp("R2", MapJoinOperator.getOperatorName() + "%" + CommonMergeJoinOperator.getOperatorName() + "%"), new SemiJoinCycleRemovalDueToMapsideJoins());
    opRules.put(new RuleRegExp("R3", CommonMergeJoinOperator.getOperatorName() + "%" + MapJoinOperator.getOperatorName() + "%"), new SemiJoinCycleRemovalDueToMapsideJoins());
    opRules.put(new RuleRegExp("R4", CommonMergeJoinOperator.getOperatorName() + "%" + CommonMergeJoinOperator.getOperatorName() + "%"), new SemiJoinCycleRemovalDueToMapsideJoins());
    SemiJoinCycleRemovalDueTOMapsideJoinContext ctx = new SemiJoinCycleRemovalDueTOMapsideJoinContext();
    Dispatcher disp = new DefaultRuleDispatcher(null, opRules, ctx);
    List<Node> topNodes = new ArrayList<Node>();
    topNodes.addAll(procCtx.parseContext.getTopOps().values());
    GraphWalker ogw = new PreOrderOnceWalker(disp);
    ogw.startWalking(topNodes, null);
    // process the list
    ParseContext pCtx = procCtx.parseContext;
    for (Operator<?> parentJoin : ctx.childParentMap.keySet()) {
        Operator<?> childJoin = ctx.childParentMap.get(parentJoin);
        if (parentJoin.getChildOperators().size() == 1) {
            continue;
        }
        for (Operator<?> child : parentJoin.getChildOperators()) {
            if (!(child instanceof SelectOperator)) {
                continue;
            }
            while (child.getChildOperators().size() > 0) {
                child = child.getChildOperators().get(0);
            }
            if (!(child instanceof ReduceSinkOperator)) {
                continue;
            }
            ReduceSinkOperator rs = ((ReduceSinkOperator) child);
            SemiJoinBranchInfo sjInfo = pCtx.getRsToSemiJoinBranchInfo().get(rs);
            if (sjInfo == null) {
                continue;
            }
            TableScanOperator ts = sjInfo.getTsOp();
            // cycle with childJoin.
            for (Operator<?> parent : childJoin.getParentOperators()) {
                if (parent == parentJoin) {
                    continue;
                }
                assert parent instanceof ReduceSinkOperator;
                while (parent.getParentOperators().size() > 0) {
                    parent = parent.getParentOperators().get(0);
                }
                if (parent == ts) {
                    // We have a cycle!
                    if (sjInfo.getIsHint()) {
                        throw new SemanticException("Removing hinted semijoin as it is creating cycles with mapside joins " + rs + " : " + ts);
                    }
                    if (LOG.isDebugEnabled()) {
                        LOG.debug("Semijoin cycle due to mapjoin. Removing semijoin " + OperatorUtils.getOpNamePretty(rs) + " - " + OperatorUtils.getOpNamePretty(ts));
                    }
                    GenTezUtils.removeBranch(rs);
                    GenTezUtils.removeSemiJoinOperator(pCtx, rs, ts);
                }
            }
        }
    }
}