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

Example 11 with BaseWork

use of org.apache.hadoop.hive.ql.plan.BaseWork 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 12 with BaseWork

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

the class SparkSkewJoinProcFactory method splitTask.

/**
 * If the join is not in a leaf ReduceWork, the spark task has to be split into 2 tasks.
 */
private static void splitTask(SparkTask currentTask, ReduceWork reduceWork, ParseContext parseContext) throws SemanticException {
    SparkWork currentWork = currentTask.getWork();
    Set<Operator<?>> reduceSinkSet = OperatorUtils.getOp(reduceWork, ReduceSinkOperator.class);
    if (currentWork.getChildren(reduceWork).size() == 1 && canSplit(currentWork) && reduceSinkSet.size() == 1) {
        ReduceSinkOperator reduceSink = (ReduceSinkOperator) reduceSinkSet.iterator().next();
        BaseWork childWork = currentWork.getChildren(reduceWork).get(0);
        SparkEdgeProperty originEdge = currentWork.getEdgeProperty(reduceWork, childWork);
        // disconnect the reduce work from its child. this should produce two isolated sub graphs
        currentWork.disconnect(reduceWork, childWork);
        // move works following the current reduce work into a new spark work
        SparkWork newWork = new SparkWork(parseContext.getConf().getVar(HiveConf.ConfVars.HIVEQUERYID));
        newWork.add(childWork);
        copyWorkGraph(currentWork, newWork, childWork);
        // remove them from current spark work
        for (BaseWork baseWork : newWork.getAllWorkUnsorted()) {
            currentWork.remove(baseWork);
            currentWork.getCloneToWork().remove(baseWork);
        }
        // create TS to read intermediate data
        Context baseCtx = parseContext.getContext();
        Path taskTmpDir = baseCtx.getMRTmpPath();
        Operator<? extends OperatorDesc> rsParent = reduceSink.getParentOperators().get(0);
        TableDesc tableDesc = PlanUtils.getIntermediateFileTableDesc(PlanUtils.getFieldSchemasFromRowSchema(rsParent.getSchema(), "temporarycol"));
        // this will insert FS and TS between the RS and its parent
        TableScanOperator tableScanOp = GenMapRedUtils.createTemporaryFile(rsParent, reduceSink, taskTmpDir, tableDesc, parseContext);
        // create new MapWork
        MapWork mapWork = PlanUtils.getMapRedWork().getMapWork();
        mapWork.setName("Map " + GenSparkUtils.getUtils().getNextSeqNumber());
        newWork.add(mapWork);
        newWork.connect(mapWork, childWork, originEdge);
        // setup the new map work
        String streamDesc = taskTmpDir.toUri().toString();
        if (GenMapRedUtils.needsTagging((ReduceWork) childWork)) {
            Operator<? extends OperatorDesc> childReducer = ((ReduceWork) childWork).getReducer();
            String id = null;
            if (childReducer instanceof JoinOperator) {
                if (parseContext.getJoinOps().contains(childReducer)) {
                    id = ((JoinOperator) childReducer).getConf().getId();
                }
            } else if (childReducer instanceof MapJoinOperator) {
                if (parseContext.getMapJoinOps().contains(childReducer)) {
                    id = ((MapJoinOperator) childReducer).getConf().getId();
                }
            } else if (childReducer instanceof SMBMapJoinOperator) {
                if (parseContext.getSmbMapJoinOps().contains(childReducer)) {
                    id = ((SMBMapJoinOperator) childReducer).getConf().getId();
                }
            }
            if (id != null) {
                streamDesc = id + ":$INTNAME";
            } else {
                streamDesc = "$INTNAME";
            }
            String origStreamDesc = streamDesc;
            int pos = 0;
            while (mapWork.getAliasToWork().get(streamDesc) != null) {
                streamDesc = origStreamDesc.concat(String.valueOf(++pos));
            }
        }
        GenMapRedUtils.setTaskPlan(taskTmpDir, streamDesc, tableScanOp, mapWork, false, tableDesc);
        // insert the new task between current task and its child
        @SuppressWarnings("unchecked") Task<? extends Serializable> newTask = TaskFactory.get(newWork);
        List<Task<? extends Serializable>> childTasks = currentTask.getChildTasks();
        // must have at most one child
        if (childTasks != null && childTasks.size() > 0) {
            Task<? extends Serializable> childTask = childTasks.get(0);
            currentTask.removeDependentTask(childTask);
            newTask.addDependentTask(childTask);
        }
        currentTask.addDependentTask(newTask);
        newTask.setFetchSource(currentTask.isFetchSource());
    }
}

Example 13 with BaseWork

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

the class SparkSkewJoinProcFactory method copyWorkGraph.

/**
 * Copy a sub-graph from originWork to newWork.
 */
private static void copyWorkGraph(SparkWork originWork, SparkWork newWork, BaseWork baseWork) {
    for (BaseWork child : originWork.getChildren(baseWork)) {
        if (!newWork.contains(child)) {
            newWork.add(child);
            SparkEdgeProperty edgeProperty = originWork.getEdgeProperty(baseWork, child);
            newWork.connect(baseWork, child, edgeProperty);
            copyWorkGraph(originWork, newWork, child);
        }
    }
    for (BaseWork parent : originWork.getParents(baseWork)) {
        if (!newWork.contains(parent)) {
            newWork.add(parent);
            SparkEdgeProperty edgeProperty = originWork.getEdgeProperty(parent, baseWork);
            newWork.connect(parent, baseWork, edgeProperty);
            copyWorkGraph(originWork, newWork, parent);
        }
    }
}

Example 14 with BaseWork

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

the class GenSparkWorkWalker method walk.

/**
 * Walk the given operator.
 *
 * @param nd operator being walked
 */
@Override
protected void walk(Node nd) throws SemanticException {
    List<? extends Node> children = nd.getChildren();
    // maintain the stack of operators encountered
    opStack.push(nd);
    Boolean skip = dispatchAndReturn(nd, opStack);
    // save some positional state
    Operator<? extends OperatorDesc> currentRoot = ctx.currentRootOperator;
    Operator<? extends OperatorDesc> parentOfRoot = ctx.parentOfRoot;
    BaseWork preceedingWork = ctx.preceedingWork;
    if (skip == null || !skip) {
        // move all the children to the front of queue
        for (Node ch : children) {
            // and restore the state before walking each child
            ctx.currentRootOperator = currentRoot;
            ctx.parentOfRoot = parentOfRoot;
            ctx.preceedingWork = preceedingWork;
            walk(ch);
        }
    }
    // done with this operator
    opStack.pop();
}

Example 15 with BaseWork

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

the class SparkCompiler method generateTaskTree.

/**
 * TODO: need to turn on rules that's commented out and add more if necessary.
 */
@Override
protected void generateTaskTree(List<Task<? extends Serializable>> rootTasks, ParseContext pCtx, List<Task<MoveWork>> mvTask, Set<ReadEntity> inputs, Set<WriteEntity> outputs) throws SemanticException {
    PERF_LOGGER.PerfLogBegin(CLASS_NAME, PerfLogger.SPARK_GENERATE_TASK_TREE);
    GenSparkUtils utils = GenSparkUtils.getUtils();
    utils.resetSequenceNumber();
    ParseContext tempParseContext = getParseContext(pCtx, rootTasks);
    GenSparkProcContext procCtx = new GenSparkProcContext(conf, tempParseContext, mvTask, rootTasks, inputs, outputs, pCtx.getTopOps());
    // -------------------------------- First Pass ---------------------------------- //
    // Identify SparkPartitionPruningSinkOperators, and break OP tree if necessary
    Map<Rule, NodeProcessor> opRules = new LinkedHashMap<Rule, NodeProcessor>();
    opRules.put(new RuleRegExp("Clone OP tree for PartitionPruningSink", SparkPartitionPruningSinkOperator.getOperatorName() + "%"), new SplitOpTreeForDPP());
    Dispatcher disp = new DefaultRuleDispatcher(null, opRules, procCtx);
    GraphWalker ogw = new GenSparkWorkWalker(disp, procCtx);
    List<Node> topNodes = new ArrayList<Node>();
    topNodes.addAll(pCtx.getTopOps().values());
    ogw.startWalking(topNodes, null);
    // -------------------------------- Second Pass ---------------------------------- //
    // Process operator tree in two steps: first we process the extra op trees generated
    // in the first pass. Then we process the main op tree, and the result task will depend
    // on the task generated in the first pass.
    topNodes.clear();
    topNodes.addAll(procCtx.topOps.values());
    generateTaskTreeHelper(procCtx, topNodes);
    // the partitions used.
    if (!procCtx.clonedPruningTableScanSet.isEmpty()) {
        SparkTask pruningTask = SparkUtilities.createSparkTask(conf);
        SparkTask mainTask = procCtx.currentTask;
        pruningTask.addDependentTask(procCtx.currentTask);
        procCtx.rootTasks.remove(procCtx.currentTask);
        procCtx.rootTasks.add(pruningTask);
        procCtx.currentTask = pruningTask;
        topNodes.clear();
        topNodes.addAll(procCtx.clonedPruningTableScanSet);
        generateTaskTreeHelper(procCtx, topNodes);
        procCtx.currentTask = mainTask;
    }
    // we need to clone some operator plans and remove union operators still
    for (BaseWork w : procCtx.workWithUnionOperators) {
        GenSparkUtils.getUtils().removeUnionOperators(procCtx, w);
    }
    // we need to fill MapWork with 'local' work and bucket information for SMB Join.
    GenSparkUtils.getUtils().annotateMapWork(procCtx);
    // finally make sure the file sink operators are set up right
    for (FileSinkOperator fileSink : procCtx.fileSinkSet) {
        GenSparkUtils.getUtils().processFileSink(procCtx, fileSink);
    }
    // Process partition pruning sinks
    for (Operator<?> prunerSink : procCtx.pruningSinkSet) {
        utils.processPartitionPruningSink(procCtx, (SparkPartitionPruningSinkOperator) prunerSink);
    }
    PERF_LOGGER.PerfLogEnd(CLASS_NAME, PerfLogger.SPARK_GENERATE_TASK_TREE);
}