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

Example 81 with OperatorDesc

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

the class Vectorizer method validateAndVectorizeOperatorTree.

private Operator<? extends OperatorDesc> validateAndVectorizeOperatorTree(Operator<? extends OperatorDesc> nonVecRootOperator, boolean isReduce, boolean isTezOrSpark, VectorTaskColumnInfo vectorTaskColumnInfo) throws VectorizerCannotVectorizeException {
    VectorizationContext taskVContext = new VectorizationContext("Task", vectorTaskColumnInfo.allColumnNames, vectorTaskColumnInfo.allTypeInfos, vectorTaskColumnInfo.allDataTypePhysicalVariations, hiveConf);
    List<Operator<? extends OperatorDesc>> currentParentList = newOperatorList();
    currentParentList.add(nonVecRootOperator);
    // Start with dummy vector operator as the parent of the parallel vector operator tree we are
    // creating
    Operator<? extends OperatorDesc> dummyVectorOperator = new DummyVectorOperator(taskVContext);
    List<Operator<? extends OperatorDesc>> currentVectorParentList = newOperatorList();
    currentVectorParentList.add(dummyVectorOperator);
    delayedFixups.clear();
    do {
        List<Operator<? extends OperatorDesc>> nextParentList = newOperatorList();
        List<Operator<? extends OperatorDesc>> nextVectorParentList = newOperatorList();
        final int count = currentParentList.size();
        for (int i = 0; i < count; i++) {
            Operator<? extends OperatorDesc> parent = currentParentList.get(i);
            List<Operator<? extends OperatorDesc>> childrenList = parent.getChildOperators();
            if (childrenList == null || childrenList.size() == 0) {
                continue;
            }
            Operator<? extends OperatorDesc> vectorParent = currentVectorParentList.get(i);
            /*
         * Vectorize this parent's children.  Plug them into vectorParent's children list.
         *
         * Add those children / vector children to nextParentList / nextVectorParentList.
         */
            doProcessChildren(parent, vectorParent, nextParentList, nextVectorParentList, isReduce, isTezOrSpark, vectorTaskColumnInfo);
        }
        currentParentList = nextParentList;
        currentVectorParentList = nextVectorParentList;
    } while (currentParentList.size() > 0);
    runDelayedFixups();
    return dummyVectorOperator;
}

Example 82 with OperatorDesc

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

the class SparkSMBJoinHintOptimizer method removeSmallTableReduceSink.

/**
 * In bucket mapjoin, there are ReduceSinks that mark a small table parent (Reduce Sink are removed from big-table).
 * In SMB join these are not expected for any parents, either from small or big tables.
 * @param mapJoinOp
 */
@SuppressWarnings("unchecked")
private void removeSmallTableReduceSink(MapJoinOperator mapJoinOp) {
    SMBJoinDesc smbJoinDesc = new SMBJoinDesc(mapJoinOp.getConf());
    List<Operator<? extends OperatorDesc>> parentOperators = mapJoinOp.getParentOperators();
    for (int i = 0; i < parentOperators.size(); i++) {
        Operator<? extends OperatorDesc> par = parentOperators.get(i);
        if (i != smbJoinDesc.getPosBigTable()) {
            if (par instanceof ReduceSinkOperator) {
                List<Operator<? extends OperatorDesc>> grandParents = par.getParentOperators();
                Preconditions.checkArgument(grandParents.size() == 1, "AssertionError: expect # of parents to be 1, but was " + grandParents.size());
                Operator<? extends OperatorDesc> grandParent = grandParents.get(0);
                grandParent.removeChild(par);
                grandParent.setChildOperators(Utilities.makeList(mapJoinOp));
                mapJoinOp.getParentOperators().set(i, grandParent);
            }
        }
    }
}

Example 83 with OperatorDesc

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

the class CommonJoinTaskDispatcher method processCurrentTask.

@Override
public Task<? extends Serializable> processCurrentTask(MapRedTask currTask, ConditionalTask conditionalTask, Context context) throws SemanticException {
    // whether it contains common join op; if contains, return this common join op
    JoinOperator joinOp = getJoinOp(currTask);
    if (joinOp == null || joinOp.getConf().isFixedAsSorted()) {
        return null;
    }
    currTask.setTaskTag(Task.COMMON_JOIN);
    MapWork currWork = currTask.getWork().getMapWork();
    // create conditional work list and task list
    List<Serializable> listWorks = new ArrayList<Serializable>();
    List<Task<? extends Serializable>> listTasks = new ArrayList<Task<? extends Serializable>>();
    // create task to aliases mapping and alias to input file mapping for resolver
    // Must be deterministic order map for consistent q-test output across Java versions
    HashMap<Task<? extends Serializable>, Set<String>> taskToAliases = new LinkedHashMap<Task<? extends Serializable>, Set<String>>();
    HashMap<Path, ArrayList<String>> pathToAliases = currWork.getPathToAliases();
    Map<String, Operator<? extends OperatorDesc>> aliasToWork = currWork.getAliasToWork();
    // start to generate multiple map join tasks
    JoinDesc joinDesc = joinOp.getConf();
    if (aliasToSize == null) {
        aliasToSize = new HashMap<String, Long>();
    }
    try {
        long aliasTotalKnownInputSize = getTotalKnownInputSize(context, currWork, pathToAliases, aliasToSize);
        Set<Integer> bigTableCandidates = MapJoinProcessor.getBigTableCandidates(joinDesc.getConds());
        // no table could be the big table; there is no need to convert
        if (bigTableCandidates.isEmpty()) {
            return null;
        }
        // if any of bigTableCandidates is from multi-sourced, bigTableCandidates should
        // only contain multi-sourced because multi-sourced cannot be hashed or direct readable
        bigTableCandidates = multiInsertBigTableCheck(joinOp, bigTableCandidates);
        Configuration conf = context.getConf();
        // If sizes of at least n-1 tables in a n-way join is known, and their sum is smaller than
        // the threshold size, convert the join into map-join and don't create a conditional task
        boolean convertJoinMapJoin = HiveConf.getBoolVar(conf, HiveConf.ConfVars.HIVECONVERTJOINNOCONDITIONALTASK);
        int bigTablePosition = -1;
        if (convertJoinMapJoin) {
            // This is the threshold that the user has specified to fit in mapjoin
            long mapJoinSize = HiveConf.getLongVar(conf, HiveConf.ConfVars.HIVECONVERTJOINNOCONDITIONALTASKTHRESHOLD);
            Long bigTableSize = null;
            Set<String> aliases = aliasToWork.keySet();
            for (int tablePosition : bigTableCandidates) {
                Operator<?> parent = joinOp.getParentOperators().get(tablePosition);
                Set<String> participants = GenMapRedUtils.findAliases(currWork, parent);
                long sumOfOthers = Utilities.sumOfExcept(aliasToSize, aliases, participants);
                if (sumOfOthers < 0 || sumOfOthers > mapJoinSize) {
                    // some small alias is not known or too big
                    continue;
                }
                if (bigTableSize == null && bigTablePosition >= 0 && tablePosition < bigTablePosition) {
                    // prefer right most alias
                    continue;
                }
                long aliasSize = Utilities.sumOf(aliasToSize, participants);
                if (bigTableSize == null || bigTableSize < 0 || (aliasSize >= 0 && aliasSize >= bigTableSize)) {
                    bigTablePosition = tablePosition;
                    bigTableSize = aliasSize;
                }
            }
        }
        currWork.setLeftInputJoin(joinOp.getConf().isLeftInputJoin());
        currWork.setBaseSrc(joinOp.getConf().getBaseSrc());
        currWork.setMapAliases(joinOp.getConf().getMapAliases());
        if (bigTablePosition >= 0) {
            // create map join task and set big table as bigTablePosition
            MapRedTask newTask = convertTaskToMapJoinTask(currTask.getWork(), bigTablePosition);
            newTask.setTaskTag(Task.MAPJOIN_ONLY_NOBACKUP);
            newTask.setFetchSource(currTask.isFetchSource());
            replaceTask(currTask, newTask);
            // joined with multiple small tables on different keys
            if ((newTask.getChildTasks() != null) && (newTask.getChildTasks().size() == 1)) {
                mergeMapJoinTaskIntoItsChildMapRedTask(newTask, conf);
            }
            return newTask;
        }
        long ThresholdOfSmallTblSizeSum = HiveConf.getLongVar(conf, HiveConf.ConfVars.HIVESMALLTABLESFILESIZE);
        for (int pos = 0; pos < joinOp.getNumParent(); pos++) {
            // this table cannot be big table
            if (!bigTableCandidates.contains(pos)) {
                continue;
            }
            Operator<?> startOp = joinOp.getParentOperators().get(pos);
            Set<String> aliases = GenMapRedUtils.findAliases(currWork, startOp);
            long aliasKnownSize = Utilities.sumOf(aliasToSize, aliases);
            if (cannotConvert(aliasKnownSize, aliasTotalKnownInputSize, ThresholdOfSmallTblSizeSum)) {
                continue;
            }
            MapredWork newWork = SerializationUtilities.clonePlan(currTask.getWork());
            // create map join task and set big table as i
            MapRedTask newTask = convertTaskToMapJoinTask(newWork, pos);
            // add into conditional task
            listWorks.add(newTask.getWork());
            listTasks.add(newTask);
            newTask.setTaskTag(Task.CONVERTED_MAPJOIN);
            newTask.setFetchSource(currTask.isFetchSource());
            // set up backup task
            newTask.setBackupTask(currTask);
            newTask.setBackupChildrenTasks(currTask.getChildTasks());
            // put the mapping task to aliases
            taskToAliases.put(newTask, aliases);
        }
    } catch (Exception e) {
        throw new SemanticException("Generate Map Join Task Error: " + e.getMessage(), e);
    }
    if (listTasks.isEmpty()) {
        return currTask;
    }
    // insert current common join task to conditional task
    listWorks.add(currTask.getWork());
    listTasks.add(currTask);
    // clear JoinTree and OP Parse Context
    currWork.setLeftInputJoin(false);
    currWork.setBaseSrc(null);
    currWork.setMapAliases(null);
    // create conditional task and insert conditional task into task tree
    ConditionalWork cndWork = new ConditionalWork(listWorks);
    ConditionalTask cndTsk = (ConditionalTask) TaskFactory.get(cndWork);
    cndTsk.setListTasks(listTasks);
    // set resolver and resolver context
    cndTsk.setResolver(new ConditionalResolverCommonJoin());
    ConditionalResolverCommonJoinCtx resolverCtx = new ConditionalResolverCommonJoinCtx();
    resolverCtx.setPathToAliases(pathToAliases);
    resolverCtx.setAliasToKnownSize(aliasToSize);
    resolverCtx.setTaskToAliases(taskToAliases);
    resolverCtx.setCommonJoinTask(currTask);
    resolverCtx.setLocalTmpDir(context.getLocalScratchDir(false));
    resolverCtx.setHdfsTmpDir(context.getMRScratchDir());
    cndTsk.setResolverCtx(resolverCtx);
    // replace the current task with the new generated conditional task
    replaceTaskWithConditionalTask(currTask, cndTsk);
    return cndTsk;
}

Example 84 with OperatorDesc

use of org.apache.hadoop.hive.ql.plan.OperatorDesc 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 85 with OperatorDesc

use of org.apache.hadoop.hive.ql.plan.OperatorDesc 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);
}