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

Example 6 with DummyStoreOperator

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

the class ConvertJoinMapJoin method convertJoinSMBJoin.

// replaces the join operator with a new CommonJoinOperator, removes the
// parent reduce sinks
private void convertJoinSMBJoin(JoinOperator joinOp, OptimizeTezProcContext context, int mapJoinConversionPos, int numBuckets, boolean adjustParentsChildren) throws SemanticException {
    MapJoinDesc mapJoinDesc = null;
    if (adjustParentsChildren) {
        mapJoinDesc = MapJoinProcessor.getMapJoinDesc(context.conf, joinOp, joinOp.getConf().isLeftInputJoin(), joinOp.getConf().getBaseSrc(), joinOp.getConf().getMapAliases(), mapJoinConversionPos, true);
    } else {
        JoinDesc joinDesc = joinOp.getConf();
        // retain the original join desc in the map join.
        mapJoinDesc = new MapJoinDesc(MapJoinProcessor.getKeys(joinOp.getConf().isLeftInputJoin(), joinOp.getConf().getBaseSrc(), joinOp).getSecond(), null, joinDesc.getExprs(), null, null, joinDesc.getOutputColumnNames(), mapJoinConversionPos, joinDesc.getConds(), joinDesc.getFilters(), joinDesc.getNoOuterJoin(), null, joinDesc.getMemoryMonitorInfo(), joinDesc.getInMemoryDataSize());
        mapJoinDesc.setNullSafes(joinDesc.getNullSafes());
        mapJoinDesc.setFilterMap(joinDesc.getFilterMap());
        mapJoinDesc.setResidualFilterExprs(joinDesc.getResidualFilterExprs());
        // keep column expression map, explain plan uses this to display
        mapJoinDesc.setColumnExprMap(joinDesc.getColumnExprMap());
        mapJoinDesc.setReversedExprs(joinDesc.getReversedExprs());
        mapJoinDesc.resetOrder();
    }
    CommonMergeJoinOperator mergeJoinOp = (CommonMergeJoinOperator) OperatorFactory.get(joinOp.getCompilationOpContext(), new CommonMergeJoinDesc(numBuckets, mapJoinConversionPos, mapJoinDesc), joinOp.getSchema());
    int numReduceSinks = joinOp.getOpTraits().getNumReduceSinks();
    OpTraits opTraits = new OpTraits(joinOp.getOpTraits().getBucketColNames(), numBuckets, joinOp.getOpTraits().getSortCols(), numReduceSinks);
    mergeJoinOp.setOpTraits(opTraits);
    mergeJoinOp.setStatistics(joinOp.getStatistics());
    for (Operator<? extends OperatorDesc> parentOp : joinOp.getParentOperators()) {
        int pos = parentOp.getChildOperators().indexOf(joinOp);
        parentOp.getChildOperators().remove(pos);
        parentOp.getChildOperators().add(pos, mergeJoinOp);
    }
    for (Operator<? extends OperatorDesc> childOp : joinOp.getChildOperators()) {
        int pos = childOp.getParentOperators().indexOf(joinOp);
        childOp.getParentOperators().remove(pos);
        childOp.getParentOperators().add(pos, mergeJoinOp);
    }
    List<Operator<? extends OperatorDesc>> childOperators = mergeJoinOp.getChildOperators();
    List<Operator<? extends OperatorDesc>> parentOperators = mergeJoinOp.getParentOperators();
    childOperators.clear();
    parentOperators.clear();
    childOperators.addAll(joinOp.getChildOperators());
    parentOperators.addAll(joinOp.getParentOperators());
    mergeJoinOp.getConf().setGenJoinKeys(false);
    if (adjustParentsChildren) {
        mergeJoinOp.getConf().setGenJoinKeys(true);
        List<Operator<? extends OperatorDesc>> newParentOpList = new ArrayList<Operator<? extends OperatorDesc>>();
        for (Operator<? extends OperatorDesc> parentOp : mergeJoinOp.getParentOperators()) {
            for (Operator<? extends OperatorDesc> grandParentOp : parentOp.getParentOperators()) {
                grandParentOp.getChildOperators().remove(parentOp);
                grandParentOp.getChildOperators().add(mergeJoinOp);
                newParentOpList.add(grandParentOp);
            }
        }
        mergeJoinOp.getParentOperators().clear();
        mergeJoinOp.getParentOperators().addAll(newParentOpList);
        List<Operator<? extends OperatorDesc>> parentOps = new ArrayList<Operator<? extends OperatorDesc>>(mergeJoinOp.getParentOperators());
        for (Operator<? extends OperatorDesc> parentOp : parentOps) {
            int parentIndex = mergeJoinOp.getParentOperators().indexOf(parentOp);
            if (parentIndex == mapJoinConversionPos) {
                continue;
            }
            // insert the dummy store operator here
            DummyStoreOperator dummyStoreOp = new TezDummyStoreOperator(mergeJoinOp.getCompilationOpContext());
            dummyStoreOp.setParentOperators(new ArrayList<Operator<? extends OperatorDesc>>());
            dummyStoreOp.setChildOperators(new ArrayList<Operator<? extends OperatorDesc>>());
            dummyStoreOp.getChildOperators().add(mergeJoinOp);
            int index = parentOp.getChildOperators().indexOf(mergeJoinOp);
            parentOp.getChildOperators().remove(index);
            parentOp.getChildOperators().add(index, dummyStoreOp);
            dummyStoreOp.getParentOperators().add(parentOp);
            mergeJoinOp.getParentOperators().remove(parentIndex);
            mergeJoinOp.getParentOperators().add(parentIndex, dummyStoreOp);
        }
    }
    mergeJoinOp.cloneOriginalParentsList(mergeJoinOp.getParentOperators());
}

Example 7 with DummyStoreOperator

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

the class MergeJoinProc method process.

@Override
public Object process(Node nd, Stack<Node> stack, NodeProcessorCtx procCtx, Object... nodeOutputs) throws SemanticException {
    GenTezProcContext context = (GenTezProcContext) procCtx;
    CommonMergeJoinOperator mergeJoinOp = (CommonMergeJoinOperator) nd;
    if (stack.size() < 2) {
        // safety check for L53 to get parentOp, although it is very unlikely that
        // stack size is less than 2, i.e., there is only one MergeJoinOperator in the stack.
        context.currentMergeJoinOperator = mergeJoinOp;
        return null;
    }
    TezWork tezWork = context.currentTask.getWork();
    @SuppressWarnings("unchecked") Operator<? extends OperatorDesc> parentOp = (Operator<? extends OperatorDesc>) ((stack.get(stack.size() - 2)));
    // we need to set the merge work that has been created as part of the dummy store walk. If a
    // merge work already exists for this merge join operator, add the dummy store work to the
    // merge work. Else create a merge work, add above work to the merge work
    MergeJoinWork mergeWork = null;
    if (context.opMergeJoinWorkMap.containsKey(mergeJoinOp)) {
        // we already have the merge work corresponding to this merge join operator
        mergeWork = context.opMergeJoinWorkMap.get(mergeJoinOp);
    } else {
        mergeWork = new MergeJoinWork();
        tezWork.add(mergeWork);
        context.opMergeJoinWorkMap.put(mergeJoinOp, mergeWork);
    }
    if (!(stack.get(stack.size() - 2) instanceof DummyStoreOperator)) {
        /* this may happen in one of the following case:
      TS[0], FIL[26], SEL[2], DUMMY_STORE[30], MERGEJOIN[29]]
                                              /                              
      TS[3], FIL[27], SEL[5], ---------------
      */
        context.currentMergeJoinOperator = mergeJoinOp;
        mergeWork.setTag(mergeJoinOp.getTagForOperator(parentOp));
        return null;
    }
    // Guaranteed to be just 1 because each DummyStoreOperator can be part of only one work.
    BaseWork parentWork = context.childToWorkMap.get(parentOp).get(0);
    mergeWork.addMergedWork(null, parentWork, context.leafOperatorToFollowingWork);
    mergeWork.setMergeJoinOperator(mergeJoinOp);
    tezWork.setVertexType(mergeWork, VertexType.MULTI_INPUT_UNINITIALIZED_EDGES);
    for (BaseWork grandParentWork : tezWork.getParents(parentWork)) {
        TezEdgeProperty edgeProp = tezWork.getEdgeProperty(grandParentWork, parentWork);
        tezWork.disconnect(grandParentWork, parentWork);
        tezWork.connect(grandParentWork, mergeWork, edgeProp);
    }
    for (BaseWork childWork : tezWork.getChildren(parentWork)) {
        TezEdgeProperty edgeProp = tezWork.getEdgeProperty(parentWork, childWork);
        tezWork.disconnect(parentWork, childWork);
        tezWork.connect(mergeWork, childWork, edgeProp);
    }
    tezWork.remove(parentWork);
    DummyStoreOperator dummyOp = (DummyStoreOperator) (stack.get(stack.size() - 2));
    parentWork.setTag(mergeJoinOp.getTagForOperator(dummyOp));
    mergeJoinOp.getParentOperators().remove(dummyOp);
    dummyOp.getChildOperators().clear();
    return true;
}

Example 8 with DummyStoreOperator

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

the class SparkCompiler method generateTaskTreeHelper.

private void generateTaskTreeHelper(GenSparkProcContext procCtx, List<Node> topNodes) throws SemanticException {
    // create a walker which walks the tree in a DFS manner while maintaining
    // the operator stack. The dispatcher generates the plan from the operator tree
    Map<Rule, NodeProcessor> opRules = new LinkedHashMap<Rule, NodeProcessor>();
    GenSparkWork genSparkWork = new GenSparkWork(GenSparkUtils.getUtils());
    opRules.put(new RuleRegExp("Split Work - ReduceSink", ReduceSinkOperator.getOperatorName() + "%"), genSparkWork);
    opRules.put(new RuleRegExp("Split Work - SparkPartitionPruningSink", SparkPartitionPruningSinkOperator.getOperatorName() + "%"), genSparkWork);
    opRules.put(new TypeRule(MapJoinOperator.class), new SparkReduceSinkMapJoinProc());
    opRules.put(new RuleRegExp("Split Work + Move/Merge - FileSink", FileSinkOperator.getOperatorName() + "%"), new CompositeProcessor(new SparkFileSinkProcessor(), genSparkWork));
    opRules.put(new RuleRegExp("Handle Analyze Command", TableScanOperator.getOperatorName() + "%"), new SparkProcessAnalyzeTable(GenSparkUtils.getUtils()));
    opRules.put(new RuleRegExp("Remember union", UnionOperator.getOperatorName() + "%"), new NodeProcessor() {

        @Override
        public Object process(Node n, Stack<Node> s, NodeProcessorCtx procCtx, Object... os) throws SemanticException {
            GenSparkProcContext context = (GenSparkProcContext) procCtx;
            UnionOperator union = (UnionOperator) n;
            // simply need to remember that we've seen a union.
            context.currentUnionOperators.add(union);
            return null;
        }
    });
    /**
     *  SMB join case:   (Big)   (Small)  (Small)
     *                     TS       TS       TS
     *                      \       |       /
     *                       \      DS     DS
     *                         \   |    /
     *                         SMBJoinOP
     *
     * Some of the other processors are expecting only one traversal beyond SMBJoinOp.
     * We need to traverse from the big-table path only, and stop traversing on the
     * small-table path once we reach SMBJoinOp.
     * Also add some SMB join information to the context, so we can properly annotate
     * the MapWork later on.
     */
    opRules.put(new TypeRule(SMBMapJoinOperator.class), new NodeProcessor() {

        @Override
        public Object process(Node currNode, Stack<Node> stack, NodeProcessorCtx procCtx, Object... os) throws SemanticException {
            GenSparkProcContext context = (GenSparkProcContext) procCtx;
            SMBMapJoinOperator currSmbNode = (SMBMapJoinOperator) currNode;
            SparkSMBMapJoinInfo smbMapJoinCtx = context.smbMapJoinCtxMap.get(currSmbNode);
            if (smbMapJoinCtx == null) {
                smbMapJoinCtx = new SparkSMBMapJoinInfo();
                context.smbMapJoinCtxMap.put(currSmbNode, smbMapJoinCtx);
            }
            for (Node stackNode : stack) {
                if (stackNode instanceof DummyStoreOperator) {
                    // If coming from small-table side, do some book-keeping, and skip traversal.
                    smbMapJoinCtx.smallTableRootOps.add(context.currentRootOperator);
                    return true;
                }
            }
            // If coming from big-table side, do some book-keeping, and continue traversal
            smbMapJoinCtx.bigTableRootOp = context.currentRootOperator;
            return false;
        }
    });
    // The dispatcher fires the processor corresponding to the closest matching
    // rule and passes the context along
    Dispatcher disp = new DefaultRuleDispatcher(null, opRules, procCtx);
    GraphWalker ogw = new GenSparkWorkWalker(disp, procCtx);
    ogw.startWalking(topNodes, null);
}