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

Example 1 with MapJoinDesc

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

the class MapJoinProcessor method convertSMBJoinToMapJoin.

/**
   * convert a sortmerge join to a a map-side join.
   *
   * @param opParseCtxMap
   * @param smbJoinOp
   *          join operator
   * @param joinTree
   *          qb join tree
   * @param bigTablePos
   *          position of the source to be read as part of map-reduce framework. All other sources
   *          are cached in memory
   * @param noCheckOuterJoin
   */
public static MapJoinOperator convertSMBJoinToMapJoin(HiveConf hconf, SMBMapJoinOperator smbJoinOp, int bigTablePos, boolean noCheckOuterJoin) throws SemanticException {
    // Create a new map join operator
    SMBJoinDesc smbJoinDesc = smbJoinOp.getConf();
    List<ExprNodeDesc> keyCols = smbJoinDesc.getKeys().get(Byte.valueOf((byte) 0));
    TableDesc keyTableDesc = PlanUtils.getMapJoinKeyTableDesc(hconf, PlanUtils.getFieldSchemasFromColumnList(keyCols, MAPJOINKEY_FIELDPREFIX));
    MapJoinDesc mapJoinDesc = new MapJoinDesc(smbJoinDesc.getKeys(), keyTableDesc, smbJoinDesc.getExprs(), smbJoinDesc.getValueTblDescs(), smbJoinDesc.getValueTblDescs(), smbJoinDesc.getOutputColumnNames(), bigTablePos, smbJoinDesc.getConds(), smbJoinDesc.getFilters(), smbJoinDesc.isNoOuterJoin(), smbJoinDesc.getDumpFilePrefix());
    mapJoinDesc.setStatistics(smbJoinDesc.getStatistics());
    RowSchema joinRS = smbJoinOp.getSchema();
    // The mapjoin has the same schema as the join operator
    MapJoinOperator mapJoinOp = (MapJoinOperator) OperatorFactory.getAndMakeChild(smbJoinOp.getCompilationOpContext(), mapJoinDesc, joinRS, new ArrayList<Operator<? extends OperatorDesc>>());
    // change the children of the original join operator to point to the map
    // join operator
    List<Operator<? extends OperatorDesc>> childOps = smbJoinOp.getChildOperators();
    for (Operator<? extends OperatorDesc> childOp : childOps) {
        childOp.replaceParent(smbJoinOp, mapJoinOp);
    }
    mapJoinOp.setChildOperators(childOps);
    smbJoinOp.setChildOperators(null);
    // change the parent of the original SMBjoin operator to point to the map
    // join operator
    List<Operator<? extends OperatorDesc>> parentOps = smbJoinOp.getParentOperators();
    for (Operator<? extends OperatorDesc> parentOp : parentOps) {
        parentOp.replaceChild(smbJoinOp, mapJoinOp);
    }
    mapJoinOp.setParentOperators(parentOps);
    smbJoinOp.setParentOperators(null);
    return mapJoinOp;
}

Example 2 with MapJoinDesc

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

the class AbstractBucketJoinProc method convertMapJoinToBucketMapJoin.

/*
   * Convert mapjoin to a bucketed mapjoin.
   * The operator tree is not changed, but the mapjoin descriptor in the big table is
   * enhanced to keep the big table bucket -> small table buckets mapping.
   */
protected void convertMapJoinToBucketMapJoin(MapJoinOperator mapJoinOp, BucketJoinProcCtx context) throws SemanticException {
    MapJoinDesc desc = mapJoinOp.getConf();
    Map<String, Map<String, List<String>>> aliasBucketFileNameMapping = new LinkedHashMap<String, Map<String, List<String>>>();
    Map<String, List<Integer>> tblAliasToNumberOfBucketsInEachPartition = context.getTblAliasToNumberOfBucketsInEachPartition();
    Map<String, List<List<String>>> tblAliasToBucketedFilePathsInEachPartition = context.getTblAliasToBucketedFilePathsInEachPartition();
    Map<Partition, List<String>> bigTblPartsToBucketFileNames = context.getBigTblPartsToBucketFileNames();
    Map<Partition, Integer> bigTblPartsToBucketNumber = context.getBigTblPartsToBucketNumber();
    List<String> joinAliases = context.getJoinAliases();
    String baseBigAlias = context.getBaseBigAlias();
    // sort bucket names for the big table
    for (List<String> partBucketNames : bigTblPartsToBucketFileNames.values()) {
        Collections.sort(partBucketNames);
    }
    // in the big table to bucket file names in small tables.
    for (int j = 0; j < joinAliases.size(); j++) {
        String alias = joinAliases.get(j);
        if (alias.equals(baseBigAlias)) {
            continue;
        }
        for (List<String> names : tblAliasToBucketedFilePathsInEachPartition.get(alias)) {
            Collections.sort(names);
        }
        List<Integer> smallTblBucketNums = tblAliasToNumberOfBucketsInEachPartition.get(alias);
        List<List<String>> smallTblFilesList = tblAliasToBucketedFilePathsInEachPartition.get(alias);
        Map<String, List<String>> mappingBigTableBucketFileNameToSmallTableBucketFileNames = new LinkedHashMap<String, List<String>>();
        aliasBucketFileNameMapping.put(alias, mappingBigTableBucketFileNameToSmallTableBucketFileNames);
        // for each bucket file in big table, get the corresponding bucket file
        // name in the small table.
        // more than 1 partition in the big table, do the mapping for each partition
        Iterator<Entry<Partition, List<String>>> bigTblPartToBucketNames = bigTblPartsToBucketFileNames.entrySet().iterator();
        Iterator<Entry<Partition, Integer>> bigTblPartToBucketNum = bigTblPartsToBucketNumber.entrySet().iterator();
        while (bigTblPartToBucketNames.hasNext()) {
            assert bigTblPartToBucketNum.hasNext();
            int bigTblBucketNum = bigTblPartToBucketNum.next().getValue();
            List<String> bigTblBucketNameList = bigTblPartToBucketNames.next().getValue();
            fillMappingBigTableBucketFileNameToSmallTableBucketFileNames(smallTblBucketNums, smallTblFilesList, mappingBigTableBucketFileNameToSmallTableBucketFileNames, bigTblBucketNum, bigTblBucketNameList, desc.getBigTableBucketNumMapping());
        }
    }
    desc.setAliasBucketFileNameMapping(aliasBucketFileNameMapping);
    desc.setBigTableAlias(baseBigAlias);
    boolean bigTablePartitioned = context.isBigTablePartitioned();
    if (bigTablePartitioned) {
        desc.setBigTablePartSpecToFileMapping(convert(bigTblPartsToBucketFileNames));
    }
    Map<Integer, Set<String>> posToAliasMap = mapJoinOp.getPosToAliasMap();
    Map<String, String> aliasToNewAliasMap = context.getAliasToNewAliasMap();
    if (aliasToNewAliasMap != null && posToAliasMap != null) {
        for (Map.Entry<String, String> entry : aliasToNewAliasMap.entrySet()) {
            for (Set<String> aliases : posToAliasMap.values()) {
                if (aliases.remove(entry.getKey())) {
                    aliases.add(entry.getValue());
                }
            }
        }
    }
    // successfully convert to bucket map join
    desc.setBucketMapJoin(true);
}

Example 3 with MapJoinDesc

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

the class GenMRSkewJoinProcessor method processSkewJoin.

/**
   * Create tasks for processing skew joins. The idea is (HIVE-964) to use
   * separated jobs and map-joins to handle skew joins.
   * <p>
   * <ul>
   * <li>
   * Number of mr jobs to handle skew keys is the number of table minus 1 (we
   * can stream the last table, so big keys in the last table will not be a
   * problem).
   * <li>
   * At runtime in Join, we output big keys in one table into one corresponding
   * directories, and all same keys in other tables into different dirs(one for
   * each table). The directories will look like:
   * <ul>
   * <li>
   * dir-T1-bigkeys(containing big keys in T1), dir-T2-keys(containing keys
   * which is big in T1),dir-T3-keys(containing keys which is big in T1), ...
   * <li>
   * dir-T1-keys(containing keys which is big in T2), dir-T2-bigkeys(containing
   * big keys in T2),dir-T3-keys(containing keys which is big in T2), ...
   * <li>
   * dir-T1-keys(containing keys which is big in T3), dir-T2-keys(containing big
   * keys in T3),dir-T3-bigkeys(containing keys which is big in T3), ... .....
   * </ul>
   * </ul>
   * For each table, we launch one mapjoin job, taking the directory containing
   * big keys in this table and corresponding dirs in other tables as input.
   * (Actally one job for one row in the above.)
   *
   * <p>
   * For more discussions, please check
   * https://issues.apache.org/jira/browse/HIVE-964.
   *
   */
@SuppressWarnings("unchecked")
public static void processSkewJoin(JoinOperator joinOp, Task<? extends Serializable> currTask, ParseContext parseCtx) throws SemanticException {
    // now does not work with outer joins
    if (!GenMRSkewJoinProcessor.skewJoinEnabled(parseCtx.getConf(), joinOp)) {
        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 (int i = 0; i < numAliases; i++) {
        Byte alias = tags[i];
        bigKeysDirMap.put(alias, 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, getSmallKeysDir(baseTmpDir, alias, src2));
            }
        }
        skewJoinJobResultsDir.put(alias, getBigKeysSkewJoinResultDir(baseTmpDir, alias));
    }
    joinDescriptor.setHandleSkewJoin(true);
    joinDescriptor.setBigKeysDirMap(bigKeysDirMap);
    joinDescriptor.setSmallKeysDirMap(smallKeysDirMap);
    joinDescriptor.setSkewKeyDefinition(HiveConf.getIntVar(parseCtx.getConf(), HiveConf.ConfVars.HIVESKEWJOINKEY));
    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>>();
    MapredWork currPlan = (MapredWork) currTask.getWork();
    TableDesc keyTblDesc = (TableDesc) currPlan.getReduceWork().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 (Byte tag : tags) {
        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));
            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));
        }
        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.valueOf((byte) i), Utilities.getTableDesc(valueColNames, valueColTypes));
    }
    joinDescriptor.setSkewKeysValuesTables(tableDescList);
    joinDescriptor.setKeyTableDesc(keyTblDesc);
    for (int i = 0; i < numAliases - 1; i++) {
        Byte src = tags[i];
        MapWork newPlan = PlanUtils.getMapRedWork().getMapWork();
        // This code has been only added for testing
        boolean mapperCannotSpanPartns = parseCtx.getConf().getBoolVar(HiveConf.ConfVars.HIVE_MAPPER_CANNOT_SPAN_MULTIPLE_PARTITIONS);
        newPlan.setMapperCannotSpanPartns(mapperCannotSpanPartns);
        MapredWork clonePlan = SerializationUtilities.clonePlan(currPlan);
        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).setTableDesc(tableDescList.get((byte) k));
            parentOps[k] = ts;
        }
        Operator<? extends OperatorDesc> tblScan_op = parentOps[i];
        ArrayList<String> aliases = new ArrayList<String>();
        String alias = src.toString().intern();
        aliases.add(alias);
        Path bigKeyDirPath = bigKeysDirMap.get(src);
        newPlan.addPathToAlias(bigKeyDirPath, aliases);
        newPlan.getAliasToWork().put(alias, tblScan_op);
        PartitionDesc part = new PartitionDesc(tableDescList.get(src), null);
        newPlan.addPathToPartitionInfo(bigKeyDirPath, part);
        newPlan.getAliasToPartnInfo().put(alias, part);
        Operator<? extends OperatorDesc> reducer = clonePlan.getReduceWork().getReducer();
        assert reducer instanceof JoinOperator;
        JoinOperator cloneJoinOp = (JoinOperator) reducer;
        String dumpFilePrefix = "mapfile" + PlanUtils.getCountForMapJoinDumpFilePrefix();
        MapJoinDesc mapJoinDescriptor = new MapJoinDesc(newJoinKeys, keyTblDesc, newJoinValues, newJoinValueTblDesc, newJoinValueTblDesc, joinDescriptor.getOutputColumnNames(), i, joinDescriptor.getConds(), joinDescriptor.getFilters(), joinDescriptor.getNoOuterJoin(), dumpFilePrefix);
        mapJoinDescriptor.setTagOrder(tags);
        mapJoinDescriptor.setHandleSkewJoin(false);
        mapJoinDescriptor.setNullSafes(joinDescriptor.getNullSafes());
        MapredLocalWork localPlan = new MapredLocalWork(new LinkedHashMap<String, Operator<? extends OperatorDesc>>(), new LinkedHashMap<String, FetchWork>());
        Map<Byte, Path> smallTblDirs = smallKeysDirMap.get(src);
        for (int j = 0; j < numAliases; j++) {
            if (j == i) {
                continue;
            }
            Byte small_alias = tags[j];
            Operator<? extends OperatorDesc> tblScan_op2 = parentOps[j];
            localPlan.getAliasToWork().put(small_alias.toString(), tblScan_op2);
            Path tblDir = smallTblDirs.get(small_alias);
            localPlan.getAliasToFetchWork().put(small_alias.toString(), new FetchWork(tblDir, tableDescList.get(small_alias)));
        }
        newPlan.setMapRedLocalWork(localPlan);
        // construct a map join and set it as the child operator of tblScan_op
        MapJoinOperator mapJoinOp = (MapJoinOperator) OperatorFactory.getAndMakeChild(joinOp.getCompilationOpContext(), mapJoinDescriptor, (RowSchema) null, parentOps);
        // change the children of the original join operator to point to the map
        // join operator
        List<Operator<? extends OperatorDesc>> childOps = cloneJoinOp.getChildOperators();
        for (Operator<? extends OperatorDesc> childOp : childOps) {
            childOp.replaceParent(cloneJoinOp, mapJoinOp);
        }
        mapJoinOp.setChildOperators(childOps);
        HiveConf jc = new HiveConf(parseCtx.getConf(), GenMRSkewJoinProcessor.class);
        newPlan.setNumMapTasks(HiveConf.getIntVar(jc, HiveConf.ConfVars.HIVESKEWJOINMAPJOINNUMMAPTASK));
        newPlan.setMinSplitSize(HiveConf.getLongVar(jc, HiveConf.ConfVars.HIVESKEWJOINMAPJOINMINSPLIT));
        newPlan.setInputformat(HiveInputFormat.class.getName());
        MapredWork w = new MapredWork();
        w.setMapWork(newPlan);
        Task<? extends Serializable> skewJoinMapJoinTask = TaskFactory.get(w, jc);
        skewJoinMapJoinTask.setFetchSource(currTask.isFetchSource());
        bigKeysDirToTaskMap.put(bigKeyDirPath, skewJoinMapJoinTask);
        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);
    }
    ConditionalResolverSkewJoinCtx context = new ConditionalResolverSkewJoinCtx(bigKeysDirToTaskMap, children);
    ConditionalWork cndWork = new ConditionalWork(listWorks);
    ConditionalTask cndTsk = (ConditionalTask) TaskFactory.get(cndWork, parseCtx.getConf());
    cndTsk.setListTasks(listTasks);
    cndTsk.setResolver(new ConditionalResolverSkewJoin());
    cndTsk.setResolverCtx(context);
    currTask.setChildTasks(new ArrayList<Task<? extends Serializable>>());
    currTask.addDependentTask(cndTsk);
    return;
}

Example 4 with MapJoinDesc

use of org.apache.hadoop.hive.ql.plan.MapJoinDesc 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;
}

Example 5 with MapJoinDesc

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

the class GenSparkSkewJoinProcessor method setMemUsage.

private static void setMemUsage(MapJoinOperator mapJoinOp, Task<? extends Serializable> task, ParseContext parseContext) {
    MapJoinResolver.LocalMapJoinProcCtx context = new MapJoinResolver.LocalMapJoinProcCtx(task, parseContext);
    try {
        new LocalMapJoinProcFactory.LocalMapJoinProcessor().hasGroupBy(mapJoinOp, context);
    } catch (Exception e) {
        LOG.warn("Error setting memory usage.", e);
        return;
    }
    MapJoinDesc mapJoinDesc = mapJoinOp.getConf();
    HiveConf conf = context.getParseCtx().getConf();
    float hashtableMemoryUsage;
    if (context.isFollowedByGroupBy()) {
        hashtableMemoryUsage = conf.getFloatVar(HiveConf.ConfVars.HIVEHASHTABLEFOLLOWBYGBYMAXMEMORYUSAGE);
    } else {
        hashtableMemoryUsage = conf.getFloatVar(HiveConf.ConfVars.HIVEHASHTABLEMAXMEMORYUSAGE);
    }
    mapJoinDesc.setHashTableMemoryUsage(hashtableMemoryUsage);
}