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

Example 71 with Operator

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

the class MapJoinProcessor method genMapJoinLocalWork.

/**
   * Generate the MapRed Local Work for the given map-join operator
   *
   * @param newWork
   * @param mapJoinOp
   *          map-join operator for which local work needs to be generated.
   * @param bigTablePos
   * @throws SemanticException
   */
private static void genMapJoinLocalWork(MapredWork newWork, MapJoinOperator mapJoinOp, int bigTablePos) throws SemanticException {
    // keep the small table alias to avoid concurrent modification exception
    ArrayList<String> smallTableAliasList = new ArrayList<String>();
    // create a new  MapredLocalWork
    MapredLocalWork newLocalWork = new MapredLocalWork(new LinkedHashMap<String, Operator<? extends OperatorDesc>>(), new LinkedHashMap<String, FetchWork>());
    for (Map.Entry<String, Operator<? extends OperatorDesc>> entry : newWork.getMapWork().getAliasToWork().entrySet()) {
        String alias = entry.getKey();
        Operator<? extends OperatorDesc> op = entry.getValue();
        // if the table scan is for big table; then skip it
        // tracing down the operator tree from the table scan operator
        Operator<? extends OperatorDesc> parentOp = op;
        Operator<? extends OperatorDesc> childOp = op.getChildOperators().get(0);
        while ((childOp != null) && (!childOp.equals(mapJoinOp))) {
            parentOp = childOp;
            assert parentOp.getChildOperators().size() == 1;
            childOp = parentOp.getChildOperators().get(0);
        }
        if (childOp == null) {
            throw new SemanticException("Cannot find join op by tracing down the table scan operator tree");
        }
        // skip the big table pos
        int i = childOp.getParentOperators().indexOf(parentOp);
        if (i == bigTablePos) {
            continue;
        }
        // set alias to work and put into smallTableAliasList
        newLocalWork.getAliasToWork().put(alias, op);
        smallTableAliasList.add(alias);
        // get input path and remove this alias from pathToAlias
        // because this file will be fetched by fetch operator
        LinkedHashMap<Path, ArrayList<String>> pathToAliases = newWork.getMapWork().getPathToAliases();
        // keep record all the input path for this alias
        HashSet<Path> pathSet = new HashSet<>();
        HashSet<Path> emptyPath = new HashSet<>();
        for (Map.Entry<Path, ArrayList<String>> entry2 : pathToAliases.entrySet()) {
            Path path = entry2.getKey();
            ArrayList<String> list = entry2.getValue();
            if (list.contains(alias)) {
                // add to path set
                pathSet.add(path);
                //remove this alias from the alias list
                list.remove(alias);
                if (list.size() == 0) {
                    emptyPath.add(path);
                }
            }
        }
        //remove the path, with which no alias associates
        for (Path path : emptyPath) {
            newWork.getMapWork().removePathToAlias(path);
        }
        // create fetch work
        FetchWork fetchWork = null;
        List<Path> partDir = new ArrayList<Path>();
        List<PartitionDesc> partDesc = new ArrayList<PartitionDesc>();
        for (Path tablePath : pathSet) {
            PartitionDesc partitionDesc = newWork.getMapWork().getPathToPartitionInfo().get(tablePath);
            // create fetchwork for non partitioned table
            if (partitionDesc.getPartSpec() == null || partitionDesc.getPartSpec().size() == 0) {
                fetchWork = new FetchWork(tablePath, partitionDesc.getTableDesc());
                break;
            }
            // if table is partitioned,add partDir and partitionDesc
            partDir.add(tablePath);
            partDesc.add(partitionDesc);
        }
        // create fetchwork for partitioned table
        if (fetchWork == null) {
            TableDesc table = newWork.getMapWork().getAliasToPartnInfo().get(alias).getTableDesc();
            fetchWork = new FetchWork(partDir, partDesc, table);
        }
        // set alias to fetch work
        newLocalWork.getAliasToFetchWork().put(alias, fetchWork);
    }
    // remove small table ailias from aliasToWork;Avoid concurrent modification
    for (String alias : smallTableAliasList) {
        newWork.getMapWork().getAliasToWork().remove(alias);
    }
    // set up local work
    newWork.getMapWork().setMapRedLocalWork(newLocalWork);
    // remove reducer
    newWork.setReduceWork(null);
}

Example 72 with Operator

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

the class GenMRUnion1 method process.

/**
   * Union Operator encountered . Currently, the algorithm is pretty simple: If
   * all the sub-queries are map-only, don't do anything. Otherwise, insert a
   * FileSink on top of all the sub-queries.
   *
   * This can be optimized later on.
   *
   * @param nd
   *          the file sink operator encountered
   * @param opProcCtx
   *          context
   */
@Override
public Object process(Node nd, Stack<Node> stack, NodeProcessorCtx opProcCtx, Object... nodeOutputs) throws SemanticException {
    UnionOperator union = (UnionOperator) nd;
    GenMRProcContext ctx = (GenMRProcContext) opProcCtx;
    ParseContext parseCtx = ctx.getParseCtx();
    UnionProcContext uCtx = parseCtx.getUCtx();
    // Map-only subqueries can be optimized in future to not write to a file in
    // future
    Map<Operator<? extends OperatorDesc>, GenMapRedCtx> mapCurrCtx = ctx.getMapCurrCtx();
    if (union.getConf().isAllInputsInSameReducer()) {
        // All inputs of this UnionOperator are in the same Reducer.
        // We do not need to break the operator tree.
        mapCurrCtx.put((Operator<? extends OperatorDesc>) nd, new GenMapRedCtx(ctx.getCurrTask(), ctx.getCurrAliasId()));
        return null;
    }
    UnionParseContext uPrsCtx = uCtx.getUnionParseContext(union);
    ctx.setCurrUnionOp(union);
    // map-reduce job
    if (uPrsCtx.allMapOnlySubQ()) {
        return processMapOnlyUnion(union, stack, ctx, uCtx);
    }
    assert uPrsCtx != null;
    Task<? extends Serializable> currTask = ctx.getCurrTask();
    int pos = UnionProcFactory.getPositionParent(union, stack);
    Task<? extends Serializable> uTask = null;
    MapredWork uPlan = null;
    // union is encountered for the first time
    GenMRUnionCtx uCtxTask = ctx.getUnionTask(union);
    if (uCtxTask == null) {
        uPlan = GenMapRedUtils.getMapRedWork(parseCtx);
        uTask = TaskFactory.get(uPlan, parseCtx.getConf());
        uCtxTask = new GenMRUnionCtx(uTask);
        ctx.setUnionTask(union, uCtxTask);
    } else {
        uTask = uCtxTask.getUTask();
    }
    // Copy into the current union task plan if
    if (uPrsCtx.getMapOnlySubq(pos) && uPrsCtx.getRootTask(pos)) {
        processSubQueryUnionMerge(ctx, uCtxTask, union, stack);
        if (ctx.getRootTasks().contains(currTask)) {
            ctx.getRootTasks().remove(currTask);
        }
    } else // If it a map-reduce job, create a temporary file
    {
        // is the current task a root task
        if (shouldBeRootTask(currTask) && !ctx.getRootTasks().contains(currTask) && (currTask.getParentTasks() == null || currTask.getParentTasks().isEmpty())) {
            ctx.getRootTasks().add(currTask);
        }
        processSubQueryUnionCreateIntermediate(union.getParentOperators().get(pos), union, uTask, ctx, uCtxTask);
        // the currAliasId and CurrTopOp is not valid any more
        ctx.setCurrAliasId(null);
        ctx.setCurrTopOp(null);
        ctx.getOpTaskMap().put(null, uTask);
    }
    ctx.setCurrTask(uTask);
    mapCurrCtx.put((Operator<? extends OperatorDesc>) nd, new GenMapRedCtx(ctx.getCurrTask(), null));
    return true;
}

Example 73 with Operator

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

the class GenMapRedUtils method initUnionPlan.

/**
   * Initialize the current union plan.
   *
   * @param op
   *          the reduce sink operator encountered
   * @param opProcCtx
   *          processing context
   */
public static void initUnionPlan(ReduceSinkOperator op, UnionOperator currUnionOp, GenMRProcContext opProcCtx, Task<? extends Serializable> unionTask) throws SemanticException {
    Operator<? extends OperatorDesc> reducer = op.getChildOperators().get(0);
    MapredWork plan = (MapredWork) unionTask.getWork();
    HashMap<Operator<? extends OperatorDesc>, Task<? extends Serializable>> opTaskMap = opProcCtx.getOpTaskMap();
    opTaskMap.put(reducer, unionTask);
    plan.setReduceWork(new ReduceWork());
    plan.getReduceWork().setReducer(reducer);
    plan.getReduceWork().setReducer(reducer);
    ReduceSinkDesc desc = op.getConf();
    plan.getReduceWork().setNumReduceTasks(desc.getNumReducers());
    if (needsTagging(plan.getReduceWork())) {
        plan.getReduceWork().setNeedsTagging(true);
    }
    initUnionPlan(opProcCtx, currUnionOp, unionTask, false);
}

Example 74 with Operator

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

the class GenMROperator method process.

/**
   * Reduce Scan encountered.
   *
   * @param nd
   *          the reduce sink operator encountered
   * @param procCtx
   *          context
   */
public Object process(Node nd, Stack<Node> stack, NodeProcessorCtx procCtx, Object... nodeOutputs) throws SemanticException {
    GenMRProcContext ctx = (GenMRProcContext) procCtx;
    Map<Operator<? extends OperatorDesc>, GenMapRedCtx> mapCurrCtx = ctx.getMapCurrCtx();
    GenMapRedCtx mapredCtx = mapCurrCtx.get(stack.get(stack.size() - 2));
    mapCurrCtx.put((Operator<? extends OperatorDesc>) nd, new GenMapRedCtx(mapredCtx.getCurrTask(), mapredCtx.getCurrAliasId()));
    return true;
}

Example 75 with Operator

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

the class GenMRTableScan1 method process.

/**
   * Table Sink encountered.
   * @param nd
   *          the table sink operator encountered
   * @param opProcCtx
   *          context
   */
@Override
public Object process(Node nd, Stack<Node> stack, NodeProcessorCtx opProcCtx, Object... nodeOutputs) throws SemanticException {
    TableScanOperator op = (TableScanOperator) nd;
    GenMRProcContext ctx = (GenMRProcContext) opProcCtx;
    ParseContext parseCtx = ctx.getParseCtx();
    Class<? extends InputFormat> inputFormat = op.getConf().getTableMetadata().getInputFormatClass();
    Map<Operator<? extends OperatorDesc>, GenMapRedCtx> mapCurrCtx = ctx.getMapCurrCtx();
    // create a dummy MapReduce task
    MapredWork currWork = GenMapRedUtils.getMapRedWork(parseCtx);
    MapRedTask currTask = (MapRedTask) TaskFactory.get(currWork, parseCtx.getConf());
    ctx.setCurrTask(currTask);
    ctx.setCurrTopOp(op);
    for (String alias : parseCtx.getTopOps().keySet()) {
        Operator<? extends OperatorDesc> currOp = parseCtx.getTopOps().get(alias);
        if (currOp == op) {
            String currAliasId = alias;
            ctx.setCurrAliasId(currAliasId);
            mapCurrCtx.put(op, new GenMapRedCtx(currTask, currAliasId));
            if (parseCtx.getQueryProperties().isAnalyzeCommand()) {
                boolean partialScan = parseCtx.getQueryProperties().isPartialScanAnalyzeCommand();
                boolean noScan = parseCtx.getQueryProperties().isNoScanAnalyzeCommand();
                if (OrcInputFormat.class.isAssignableFrom(inputFormat) || MapredParquetInputFormat.class.isAssignableFrom(inputFormat)) {
                    // For ORC and Parquet, all the following statements are the same
                    // ANALYZE TABLE T [PARTITION (...)] COMPUTE STATISTICS
                    // ANALYZE TABLE T [PARTITION (...)] COMPUTE STATISTICS partialscan;
                    // ANALYZE TABLE T [PARTITION (...)] COMPUTE STATISTICS noscan;
                    // There will not be any MR or Tez job above this task
                    StatsNoJobWork snjWork = new StatsNoJobWork(op.getConf().getTableMetadata().getTableSpec());
                    snjWork.setStatsReliable(parseCtx.getConf().getBoolVar(HiveConf.ConfVars.HIVE_STATS_RELIABLE));
                    // If partition is specified, get pruned partition list
                    Set<Partition> confirmedParts = GenMapRedUtils.getConfirmedPartitionsForScan(op);
                    if (confirmedParts.size() > 0) {
                        Table source = op.getConf().getTableMetadata();
                        List<String> partCols = GenMapRedUtils.getPartitionColumns(op);
                        PrunedPartitionList partList = new PrunedPartitionList(source, confirmedParts, partCols, false);
                        snjWork.setPrunedPartitionList(partList);
                    }
                    Task<StatsNoJobWork> snjTask = TaskFactory.get(snjWork, parseCtx.getConf());
                    ctx.setCurrTask(snjTask);
                    ctx.setCurrTopOp(null);
                    ctx.getRootTasks().clear();
                    ctx.getRootTasks().add(snjTask);
                } else {
                    // ANALYZE TABLE T [PARTITION (...)] COMPUTE STATISTICS;
                    // The plan consists of a simple MapRedTask followed by a StatsTask.
                    // The MR task is just a simple TableScanOperator
                    StatsWork statsWork = new StatsWork(op.getConf().getTableMetadata().getTableSpec());
                    statsWork.setAggKey(op.getConf().getStatsAggPrefix());
                    statsWork.setStatsTmpDir(op.getConf().getTmpStatsDir());
                    statsWork.setSourceTask(currTask);
                    statsWork.setStatsReliable(parseCtx.getConf().getBoolVar(HiveConf.ConfVars.HIVE_STATS_RELIABLE));
                    Task<StatsWork> statsTask = TaskFactory.get(statsWork, parseCtx.getConf());
                    currTask.addDependentTask(statsTask);
                    if (!ctx.getRootTasks().contains(currTask)) {
                        ctx.getRootTasks().add(currTask);
                    }
                    // The plan consists of a StatsTask only.
                    if (noScan) {
                        statsTask.setParentTasks(null);
                        statsWork.setNoScanAnalyzeCommand(true);
                        ctx.getRootTasks().remove(currTask);
                        ctx.getRootTasks().add(statsTask);
                    }
                    // ANALYZE TABLE T [PARTITION (...)] COMPUTE STATISTICS partialscan;
                    if (partialScan) {
                        handlePartialScanCommand(op, ctx, parseCtx, currTask, statsWork, statsTask);
                    }
                    currWork.getMapWork().setGatheringStats(true);
                    if (currWork.getReduceWork() != null) {
                        currWork.getReduceWork().setGatheringStats(true);
                    }
                    // 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(op);
                    if (confirmedPartns.size() > 0) {
                        Table source = op.getConf().getTableMetadata();
                        List<String> partCols = GenMapRedUtils.getPartitionColumns(op);
                        PrunedPartitionList partList = new PrunedPartitionList(source, confirmedPartns, partCols, false);
                        GenMapRedUtils.setTaskPlan(currAliasId, op, currTask, false, ctx, partList);
                    } else {
                        // non-partitioned table
                        GenMapRedUtils.setTaskPlan(currAliasId, op, currTask, false, ctx);
                    }
                }
            }
            return true;
        }
    }
    assert false;
    return null;
}