Examples with DataSourceDescriptor org.apache.tez.dag.api.DataSourceDescriptor used on opensource projects

Example 11 with DataSourceDescriptor

use of org.apache.tez.dag.api.DataSourceDescriptor in project tez by apache.

the class TestMRRJobsDAGApi method testMRRSleepJobDagSubmitCore.

public State testMRRSleepJobDagSubmitCore(boolean dagViaRPC, boolean killDagWhileRunning, boolean closeSessionBeforeSubmit, TezClient reUseTezSession, boolean genSplitsInAM, Class<? extends InputInitializer> initializerClass, Map<String, LocalResource> additionalLocalResources) throws IOException, InterruptedException, TezException, ClassNotFoundException, YarnException {
    LOG.info("\n\n\nStarting testMRRSleepJobDagSubmit().");
    JobConf stage1Conf = new JobConf(mrrTezCluster.getConfig());
    JobConf stage2Conf = new JobConf(mrrTezCluster.getConfig());
    JobConf stage3Conf = new JobConf(mrrTezCluster.getConfig());
    stage1Conf.setLong(MRRSleepJob.MAP_SLEEP_TIME, 1);
    stage1Conf.setInt(MRRSleepJob.MAP_SLEEP_COUNT, 1);
    stage1Conf.setInt(MRJobConfig.NUM_MAPS, 1);
    stage1Conf.set(MRJobConfig.MAP_CLASS_ATTR, SleepMapper.class.getName());
    stage1Conf.set(MRJobConfig.MAP_OUTPUT_KEY_CLASS, IntWritable.class.getName());
    stage1Conf.set(MRJobConfig.MAP_OUTPUT_VALUE_CLASS, IntWritable.class.getName());
    stage1Conf.set(MRJobConfig.INPUT_FORMAT_CLASS_ATTR, SleepInputFormat.class.getName());
    stage1Conf.set(MRJobConfig.PARTITIONER_CLASS_ATTR, MRRSleepJobPartitioner.class.getName());
    stage2Conf.setLong(MRRSleepJob.REDUCE_SLEEP_TIME, 1);
    stage2Conf.setInt(MRRSleepJob.REDUCE_SLEEP_COUNT, 1);
    stage2Conf.setInt(MRJobConfig.NUM_REDUCES, 1);
    stage2Conf.set(MRJobConfig.REDUCE_CLASS_ATTR, ISleepReducer.class.getName());
    stage2Conf.set(MRJobConfig.MAP_OUTPUT_KEY_CLASS, IntWritable.class.getName());
    stage2Conf.set(MRJobConfig.MAP_OUTPUT_VALUE_CLASS, IntWritable.class.getName());
    stage2Conf.set(MRJobConfig.PARTITIONER_CLASS_ATTR, MRRSleepJobPartitioner.class.getName());
    stage3Conf.setLong(MRRSleepJob.REDUCE_SLEEP_TIME, 1);
    stage3Conf.setInt(MRRSleepJob.REDUCE_SLEEP_COUNT, 1);
    stage3Conf.setInt(MRJobConfig.NUM_REDUCES, 1);
    stage3Conf.set(MRJobConfig.REDUCE_CLASS_ATTR, SleepReducer.class.getName());
    stage3Conf.set(MRJobConfig.MAP_OUTPUT_KEY_CLASS, IntWritable.class.getName());
    stage3Conf.set(MRJobConfig.MAP_OUTPUT_VALUE_CLASS, IntWritable.class.getName());
    MRHelpers.translateMRConfToTez(stage1Conf);
    MRHelpers.translateMRConfToTez(stage2Conf);
    MRHelpers.translateMRConfToTez(stage3Conf);
    MRHelpers.configureMRApiUsage(stage1Conf);
    MRHelpers.configureMRApiUsage(stage2Conf);
    MRHelpers.configureMRApiUsage(stage3Conf);
    Path remoteStagingDir = remoteFs.makeQualified(new Path("/tmp", String.valueOf(new Random().nextInt(100000))));
    TezClientUtils.ensureStagingDirExists(conf, remoteStagingDir);
    UserPayload stage1Payload = TezUtils.createUserPayloadFromConf(stage1Conf);
    UserPayload stage2Payload = TezUtils.createUserPayloadFromConf(stage2Conf);
    UserPayload stage3Payload = TezUtils.createUserPayloadFromConf(stage3Conf);
    DAG dag = DAG.create("testMRRSleepJobDagSubmit-" + random.nextInt(1000));
    Class<? extends InputInitializer> inputInitializerClazz = genSplitsInAM ? (initializerClass == null ? MRInputAMSplitGenerator.class : initializerClass) : null;
    LOG.info("Using initializer class: " + initializerClass);
    DataSourceDescriptor dsd;
    if (!genSplitsInAM) {
        dsd = MRInputHelpers.configureMRInputWithLegacySplitGeneration(stage1Conf, remoteStagingDir, true);
    } else {
        if (initializerClass == null) {
            dsd = MRInputLegacy.createConfigBuilder(stage1Conf, SleepInputFormat.class).build();
        } else {
            InputInitializerDescriptor iid = InputInitializerDescriptor.create(inputInitializerClazz.getName());
            dsd = MRInputLegacy.createConfigBuilder(stage1Conf, SleepInputFormat.class).setCustomInitializerDescriptor(iid).build();
        }
    }
    Vertex stage1Vertex = Vertex.create("map", ProcessorDescriptor.create(MapProcessor.class.getName()).setUserPayload(stage1Payload), dsd.getNumberOfShards(), Resource.newInstance(256, 1));
    stage1Vertex.addDataSource("MRInput", dsd);
    Vertex stage2Vertex = Vertex.create("ireduce", ProcessorDescriptor.create(ReduceProcessor.class.getName()).setUserPayload(stage2Payload), 1, Resource.newInstance(256, 1));
    Vertex stage3Vertex = Vertex.create("reduce", ProcessorDescriptor.create(ReduceProcessor.class.getName()).setUserPayload(stage3Payload), 1, Resource.newInstance(256, 1));
    stage3Conf.setBoolean(TezRuntimeConfiguration.TEZ_RUNTIME_CONVERT_USER_PAYLOAD_TO_HISTORY_TEXT, true);
    DataSinkDescriptor dataSinkDescriptor = MROutputLegacy.createConfigBuilder(stage3Conf, NullOutputFormat.class).build();
    Assert.assertFalse(dataSinkDescriptor.getOutputDescriptor().getHistoryText().isEmpty());
    stage3Vertex.addDataSink("MROutput", dataSinkDescriptor);
    // TODO env, resources
    dag.addVertex(stage1Vertex);
    dag.addVertex(stage2Vertex);
    dag.addVertex(stage3Vertex);
    Edge edge1 = Edge.create(stage1Vertex, stage2Vertex, EdgeProperty.create(DataMovementType.SCATTER_GATHER, DataSourceType.PERSISTED, SchedulingType.SEQUENTIAL, OutputDescriptor.create(OrderedPartitionedKVOutput.class.getName()).setUserPayload(stage2Payload), InputDescriptor.create(OrderedGroupedInputLegacy.class.getName()).setUserPayload(stage2Payload)));
    Edge edge2 = Edge.create(stage2Vertex, stage3Vertex, EdgeProperty.create(DataMovementType.SCATTER_GATHER, DataSourceType.PERSISTED, SchedulingType.SEQUENTIAL, OutputDescriptor.create(OrderedPartitionedKVOutput.class.getName()).setUserPayload(stage3Payload), InputDescriptor.create(OrderedGroupedInputLegacy.class.getName()).setUserPayload(stage3Payload)));
    dag.addEdge(edge1);
    dag.addEdge(edge2);
    TezConfiguration tezConf = new TezConfiguration(mrrTezCluster.getConfig());
    tezConf.set(TezConfiguration.TEZ_AM_STAGING_DIR, remoteStagingDir.toString());
    DAGClient dagClient = null;
    boolean reuseSession = reUseTezSession != null;
    TezClient tezSession = null;
    if (!dagViaRPC) {
        Preconditions.checkArgument(reuseSession == false);
    }
    if (!reuseSession) {
        TezConfiguration tempTezconf = new TezConfiguration(tezConf);
        if (!dagViaRPC) {
            tempTezconf.setBoolean(TezConfiguration.TEZ_AM_SESSION_MODE, false);
        } else {
            tempTezconf.setBoolean(TezConfiguration.TEZ_AM_SESSION_MODE, true);
        }
        tezSession = TezClient.create("testsession", tempTezconf);
        tezSession.start();
    } else {
        tezSession = reUseTezSession;
    }
    if (!dagViaRPC) {
        // TODO Use utility method post TEZ-205 to figure out AM arguments etc.
        dagClient = tezSession.submitDAG(dag);
    }
    if (dagViaRPC && closeSessionBeforeSubmit) {
        YarnClient yarnClient = YarnClient.createYarnClient();
        yarnClient.init(mrrTezCluster.getConfig());
        yarnClient.start();
        boolean sentKillSession = false;
        while (true) {
            Thread.sleep(500l);
            ApplicationReport appReport = yarnClient.getApplicationReport(tezSession.getAppMasterApplicationId());
            if (appReport == null) {
                continue;
            }
            YarnApplicationState appState = appReport.getYarnApplicationState();
            if (!sentKillSession) {
                if (appState == YarnApplicationState.RUNNING) {
                    tezSession.stop();
                    sentKillSession = true;
                }
            } else {
                if (appState == YarnApplicationState.FINISHED || appState == YarnApplicationState.KILLED || appState == YarnApplicationState.FAILED) {
                    LOG.info("Application completed after sending session shutdown" + ", yarnApplicationState=" + appState + ", finalAppStatus=" + appReport.getFinalApplicationStatus());
                    Assert.assertEquals(YarnApplicationState.FINISHED, appState);
                    Assert.assertEquals(FinalApplicationStatus.SUCCEEDED, appReport.getFinalApplicationStatus());
                    break;
                }
            }
        }
        yarnClient.stop();
        return null;
    }
    if (dagViaRPC) {
        LOG.info("Submitting dag to tez session with appId=" + tezSession.getAppMasterApplicationId() + " and Dag Name=" + dag.getName());
        if (additionalLocalResources != null) {
            tezSession.addAppMasterLocalFiles(additionalLocalResources);
        }
        dagClient = tezSession.submitDAG(dag);
        Assert.assertEquals(TezAppMasterStatus.RUNNING, tezSession.getAppMasterStatus());
    }
    DAGStatus dagStatus = dagClient.getDAGStatus(null);
    while (!dagStatus.isCompleted()) {
        LOG.info("Waiting for job to complete. Sleeping for 500ms." + " Current state: " + dagStatus.getState());
        Thread.sleep(500l);
        if (killDagWhileRunning && dagStatus.getState() == DAGStatus.State.RUNNING) {
            LOG.info("Killing running dag/session");
            if (dagViaRPC) {
                tezSession.stop();
            } else {
                dagClient.tryKillDAG();
            }
        }
        dagStatus = dagClient.getDAGStatus(null);
    }
    if (!reuseSession) {
        tezSession.stop();
    }
    return dagStatus.getState();
}

Example 12 with DataSourceDescriptor

use of org.apache.tez.dag.api.DataSourceDescriptor in project tez by apache.

the class CartesianProduct method createDAG.

private DAG createDAG(TezConfiguration tezConf) throws IOException {
    InputDescriptor inputDescriptor = InputDescriptor.create(FakeInput.class.getName());
    InputInitializerDescriptor inputInitializerDescriptor = InputInitializerDescriptor.create(FakeInputInitializer.class.getName());
    DataSourceDescriptor dataSourceDescriptor = DataSourceDescriptor.create(inputDescriptor, inputInitializerDescriptor, null);
    Vertex v1 = Vertex.create(VERTEX1, ProcessorDescriptor.create(TokenProcessor.class.getName()));
    v1.addDataSource(INPUT, dataSourceDescriptor);
    Vertex v2 = Vertex.create(VERTEX2, ProcessorDescriptor.create(TokenProcessor.class.getName()));
    v2.addDataSource(INPUT, dataSourceDescriptor);
    OutputDescriptor outputDescriptor = OutputDescriptor.create(FakeOutput.class.getName());
    OutputCommitterDescriptor outputCommitterDescriptor = OutputCommitterDescriptor.create(FakeOutputCommitter.class.getName());
    DataSinkDescriptor dataSinkDescriptor = DataSinkDescriptor.create(outputDescriptor, outputCommitterDescriptor, null);
    CartesianProductConfig cartesianProductConfig = new CartesianProductConfig(Arrays.asList(sourceVertices));
    UserPayload userPayload = cartesianProductConfig.toUserPayload(tezConf);
    Vertex v3 = Vertex.create(VERTEX3, ProcessorDescriptor.create(JoinProcessor.class.getName()));
    v3.addDataSink(OUTPUT, dataSinkDescriptor);
    v3.setVertexManagerPlugin(VertexManagerPluginDescriptor.create(CartesianProductVertexManager.class.getName()).setUserPayload(userPayload));
    EdgeManagerPluginDescriptor edgeManagerDescriptor = EdgeManagerPluginDescriptor.create(CartesianProductEdgeManager.class.getName());
    edgeManagerDescriptor.setUserPayload(userPayload);
    UnorderedPartitionedKVEdgeConfig edgeConf = UnorderedPartitionedKVEdgeConfig.newBuilder(Text.class.getName(), IntWritable.class.getName(), RoundRobinPartitioner.class.getName()).build();
    EdgeProperty edgeProperty = edgeConf.createDefaultCustomEdgeProperty(edgeManagerDescriptor);
    return DAG.create("CrossProduct").addVertex(v1).addVertex(v2).addVertex(v3).addEdge(Edge.create(v1, v3, edgeProperty)).addEdge(Edge.create(v2, v3, edgeProperty));
}

Example 13 with DataSourceDescriptor

use of org.apache.tez.dag.api.DataSourceDescriptor in project tez by apache.

the class UnionExample method createDAG.

private DAG createDAG(FileSystem fs, TezConfiguration tezConf, Map<String, LocalResource> localResources, Path stagingDir, String inputPath, String outputPath) throws IOException {
    DAG dag = DAG.create("UnionExample");
    int numMaps = -1;
    Configuration inputConf = new Configuration(tezConf);
    inputConf.setBoolean("mapred.mapper.new-api", false);
    inputConf.set("mapred.input.format.class", TextInputFormat.class.getName());
    inputConf.set(FileInputFormat.INPUT_DIR, inputPath);
    MRInput.MRInputConfigBuilder configurer = MRInput.createConfigBuilder(inputConf, null);
    DataSourceDescriptor dataSource = configurer.generateSplitsInAM(false).build();
    Vertex mapVertex1 = Vertex.create("map1", ProcessorDescriptor.create(TokenProcessor.class.getName()), numMaps).addDataSource("MRInput", dataSource);
    Vertex mapVertex2 = Vertex.create("map2", ProcessorDescriptor.create(TokenProcessor.class.getName()), numMaps).addDataSource("MRInput", dataSource);
    Vertex mapVertex3 = Vertex.create("map3", ProcessorDescriptor.create(TokenProcessor.class.getName()), numMaps).addDataSource("MRInput", dataSource);
    Vertex checkerVertex = Vertex.create("checker", ProcessorDescriptor.create(UnionProcessor.class.getName()), 1);
    Configuration outputConf = new Configuration(tezConf);
    outputConf.setBoolean("mapred.reducer.new-api", false);
    outputConf.set("mapred.output.format.class", TextOutputFormat.class.getName());
    outputConf.set(FileOutputFormat.OUTDIR, outputPath);
    DataSinkDescriptor od = MROutput.createConfigBuilder(outputConf, null).build();
    checkerVertex.addDataSink("union", od);
    Configuration allPartsConf = new Configuration(tezConf);
    DataSinkDescriptor od2 = MROutput.createConfigBuilder(allPartsConf, TextOutputFormat.class, outputPath + "-all-parts").build();
    checkerVertex.addDataSink("all-parts", od2);
    Configuration partsConf = new Configuration(tezConf);
    DataSinkDescriptor od1 = MROutput.createConfigBuilder(partsConf, TextOutputFormat.class, outputPath + "-parts").build();
    VertexGroup unionVertex = dag.createVertexGroup("union", mapVertex1, mapVertex2);
    unionVertex.addDataSink("parts", od1);
    OrderedPartitionedKVEdgeConfig edgeConf = OrderedPartitionedKVEdgeConfig.newBuilder(Text.class.getName(), IntWritable.class.getName(), HashPartitioner.class.getName()).build();
    dag.addVertex(mapVertex1).addVertex(mapVertex2).addVertex(mapVertex3).addVertex(checkerVertex).addEdge(Edge.create(mapVertex3, checkerVertex, edgeConf.createDefaultEdgeProperty())).addEdge(GroupInputEdge.create(unionVertex, checkerVertex, edgeConf.createDefaultEdgeProperty(), InputDescriptor.create(ConcatenatedMergedKeyValuesInput.class.getName())));
    return dag;
}

Example 14 with DataSourceDescriptor

use of org.apache.tez.dag.api.DataSourceDescriptor in project hive by apache.

the class DagUtils method createVertex.

/*
   * Helper function to create Vertex from MapWork.
   */
private Vertex createVertex(JobConf conf, MapWork mapWork, LocalResource appJarLr, List<LocalResource> additionalLr, FileSystem fs, Path mrScratchDir, Context ctx, VertexType vertexType) throws Exception {
    Path tezDir = getTezDir(mrScratchDir);
    // set up the operator plan
    Utilities.cacheMapWork(conf, mapWork, mrScratchDir);
    // create the directories FileSinkOperators need
    Utilities.createTmpDirs(conf, mapWork);
    // finally create the vertex
    Vertex map = null;
    // use tez to combine splits
    boolean groupSplitsInInputInitializer;
    DataSourceDescriptor dataSource;
    int numTasks = -1;
    @SuppressWarnings("rawtypes") Class inputFormatClass = conf.getClass("mapred.input.format.class", InputFormat.class);
    boolean vertexHasCustomInput = VertexType.isCustomInputType(vertexType);
    LOG.info("Vertex has custom input? " + vertexHasCustomInput);
    if (vertexHasCustomInput) {
        groupSplitsInInputInitializer = false;
        // grouping happens in execution phase. The input payload should not enable grouping here,
        // it will be enabled in the CustomVertex.
        inputFormatClass = HiveInputFormat.class;
        conf.setClass("mapred.input.format.class", HiveInputFormat.class, InputFormat.class);
        // mapreduce.tez.input.initializer.serialize.event.payload should be set to false when using
        // this plug-in to avoid getting a serialized event at run-time.
        conf.setBoolean("mapreduce.tez.input.initializer.serialize.event.payload", false);
    } else {
        // is HiveInputFormat
        if (inputFormatClass == HiveInputFormat.class) {
            groupSplitsInInputInitializer = true;
        } else {
            groupSplitsInInputInitializer = false;
        }
    }
    if (mapWork instanceof MergeFileWork) {
        Path outputPath = ((MergeFileWork) mapWork).getOutputDir();
        // prepare the tmp output directory. The output tmp directory should
        // exist before jobClose (before renaming after job completion)
        Path tempOutPath = Utilities.toTempPath(outputPath);
        try {
            FileSystem tmpOutFS = tempOutPath.getFileSystem(conf);
            if (!tmpOutFS.exists(tempOutPath)) {
                tmpOutFS.mkdirs(tempOutPath);
            }
        } catch (IOException e) {
            throw new RuntimeException("Can't make path " + outputPath + " : " + e.getMessage(), e);
        }
    }
    // remember mapping of plan to input
    conf.set(Utilities.INPUT_NAME, mapWork.getName());
    if (HiveConf.getBoolVar(conf, ConfVars.HIVE_AM_SPLIT_GENERATION)) {
        // set up the operator plan. (before setting up splits on the AM)
        Utilities.setMapWork(conf, mapWork, mrScratchDir, false);
        // the correct plugin.
        if (groupSplitsInInputInitializer) {
            // Not setting a payload, since the MRInput payload is the same and can be accessed.
            InputInitializerDescriptor descriptor = InputInitializerDescriptor.create(HiveSplitGenerator.class.getName());
            dataSource = MRInputLegacy.createConfigBuilder(conf, inputFormatClass).groupSplits(true).setCustomInitializerDescriptor(descriptor).build();
        } else {
            // Not HiveInputFormat, or a custom VertexManager will take care of grouping splits
            if (vertexHasCustomInput) {
                dataSource = MultiMRInput.createConfigBuilder(conf, inputFormatClass).groupSplits(false).build();
            } else {
                dataSource = MRInputLegacy.createConfigBuilder(conf, inputFormatClass).groupSplits(false).build();
            }
        }
    } else {
        // Setup client side split generation.
        // we need to set this, because with HS2 and client side split
        // generation we end up not finding the map work. This is
        // because of thread local madness (tez split generation is
        // multi-threaded - HS2 plan cache uses thread locals). Setting
        // VECTOR_MODE/USE_VECTORIZED_INPUT_FILE_FORMAT causes the split gen code to use the conf instead
        // of the map work.
        conf.setBoolean(Utilities.VECTOR_MODE, mapWork.getVectorMode());
        conf.setBoolean(Utilities.USE_VECTORIZED_INPUT_FILE_FORMAT, mapWork.getUseVectorizedInputFileFormat());
        dataSource = MRInputHelpers.configureMRInputWithLegacySplitGeneration(conf, new Path(tezDir, "split_" + mapWork.getName().replaceAll(" ", "_")), true);
        numTasks = dataSource.getNumberOfShards();
        // set up the operator plan. (after generating splits - that changes configs)
        Utilities.setMapWork(conf, mapWork, mrScratchDir, false);
    }
    UserPayload serializedConf = TezUtils.createUserPayloadFromConf(conf);
    String procClassName = MapTezProcessor.class.getName();
    if (mapWork instanceof MergeFileWork) {
        procClassName = MergeFileTezProcessor.class.getName();
    }
    VertexExecutionContext executionContext = createVertexExecutionContext(mapWork);
    map = Vertex.create(mapWork.getName(), ProcessorDescriptor.create(procClassName).setUserPayload(serializedConf), numTasks, getContainerResource(conf));
    map.setTaskEnvironment(getContainerEnvironment(conf, true));
    map.setExecutionContext(executionContext);
    map.setTaskLaunchCmdOpts(getContainerJavaOpts(conf));
    assert mapWork.getAliasToWork().keySet().size() == 1;
    // Add the actual source input
    String alias = mapWork.getAliasToWork().keySet().iterator().next();
    map.addDataSource(alias, dataSource);
    Map<String, LocalResource> localResources = new HashMap<String, LocalResource>();
    localResources.put(getBaseName(appJarLr), appJarLr);
    for (LocalResource lr : additionalLr) {
        localResources.put(getBaseName(lr), lr);
    }
    map.addTaskLocalFiles(localResources);
    return map;
}

Example 15 with DataSourceDescriptor

use of org.apache.tez.dag.api.DataSourceDescriptor in project hive by apache.

the class DagUtils method createVertex.

/*
   * Helper function to create Vertex from MapWork.
   */
private Vertex createVertex(JobConf conf, MapWork mapWork, FileSystem fs, Path mrScratchDir, Context ctx, VertexType vertexType, Map<String, LocalResource> localResources) throws Exception {
    // set up the operator plan
    Utilities.cacheMapWork(conf, mapWork, mrScratchDir);
    // create the directories FileSinkOperators need
    Utilities.createTmpDirs(conf, mapWork);
    // finally create the vertex
    Vertex map = null;
    // use tez to combine splits
    boolean groupSplitsInInputInitializer;
    DataSourceDescriptor dataSource;
    int numTasks = -1;
    @SuppressWarnings("rawtypes") Class inputFormatClass = conf.getClass("mapred.input.format.class", InputFormat.class);
    boolean vertexHasCustomInput = VertexType.isCustomInputType(vertexType);
    LOG.info("Vertex has custom input? " + vertexHasCustomInput);
    if (vertexHasCustomInput) {
        groupSplitsInInputInitializer = false;
        // grouping happens in execution phase. The input payload should not enable grouping here,
        // it will be enabled in the CustomVertex.
        inputFormatClass = HiveInputFormat.class;
        conf.setClass("mapred.input.format.class", HiveInputFormat.class, InputFormat.class);
        // mapreduce.tez.input.initializer.serialize.event.payload should be set to false when using
        // this plug-in to avoid getting a serialized event at run-time.
        conf.setBoolean("mapreduce.tez.input.initializer.serialize.event.payload", false);
    } else {
        // is HiveInputFormat
        if (inputFormatClass == HiveInputFormat.class) {
            groupSplitsInInputInitializer = true;
        } else {
            groupSplitsInInputInitializer = false;
        }
    }
    if (mapWork instanceof MergeFileWork) {
        Path outputPath = ((MergeFileWork) mapWork).getOutputDir();
        // prepare the tmp output directory. The output tmp directory should
        // exist before jobClose (before renaming after job completion)
        Path tempOutPath = Utilities.toTempPath(outputPath);
        try {
            FileSystem tmpOutFS = tempOutPath.getFileSystem(conf);
            if (!tmpOutFS.exists(tempOutPath)) {
                tmpOutFS.mkdirs(tempOutPath);
            }
        } catch (IOException e) {
            throw new RuntimeException("Can't make path " + outputPath + " : " + e.getMessage(), e);
        }
    }
    // remember mapping of plan to input
    conf.set(Utilities.INPUT_NAME, mapWork.getName());
    if (HiveConf.getBoolVar(conf, ConfVars.HIVE_AM_SPLIT_GENERATION)) {
        // set up the operator plan. (before setting up splits on the AM)
        Utilities.setMapWork(conf, mapWork, mrScratchDir, false);
        // the correct plugin.
        if (groupSplitsInInputInitializer) {
            // Not setting a payload, since the MRInput payload is the same and can be accessed.
            InputInitializerDescriptor descriptor = InputInitializerDescriptor.create(HiveSplitGenerator.class.getName());
            dataSource = MRInputLegacy.createConfigBuilder(conf, inputFormatClass).groupSplits(true).setCustomInitializerDescriptor(descriptor).build();
        } else {
            // Not HiveInputFormat, or a custom VertexManager will take care of grouping splits
            if (vertexHasCustomInput && vertexType == VertexType.MULTI_INPUT_UNINITIALIZED_EDGES) {
                // SMB Join.
                dataSource = MultiMRInput.createConfigBuilder(conf, inputFormatClass).groupSplits(false).build();
            } else {
                dataSource = MRInputLegacy.createConfigBuilder(conf, inputFormatClass).groupSplits(false).build();
            }
        }
    } else {
        // Setup client side split generation.
        // we need to set this, because with HS2 and client side split
        // generation we end up not finding the map work. This is
        // because of thread local madness (tez split generation is
        // multi-threaded - HS2 plan cache uses thread locals). Setting
        // VECTOR_MODE/USE_VECTORIZED_INPUT_FILE_FORMAT causes the split gen code to use the conf instead
        // of the map work.
        conf.setBoolean(Utilities.VECTOR_MODE, mapWork.getVectorMode());
        conf.setBoolean(Utilities.USE_VECTORIZED_INPUT_FILE_FORMAT, mapWork.getUseVectorizedInputFileFormat());
        InputSplitInfo inputSplitInfo = MRInputHelpers.generateInputSplitsToMem(conf, false, 0);
        InputInitializerDescriptor descriptor = InputInitializerDescriptor.create(MRInputSplitDistributor.class.getName());
        InputDescriptor inputDescriptor = InputDescriptor.create(MRInputLegacy.class.getName()).setUserPayload(UserPayload.create(MRRuntimeProtos.MRInputUserPayloadProto.newBuilder().setConfigurationBytes(TezUtils.createByteStringFromConf(conf)).setSplits(inputSplitInfo.getSplitsProto()).build().toByteString().asReadOnlyByteBuffer()));
        dataSource = DataSourceDescriptor.create(inputDescriptor, descriptor, null);
        numTasks = inputSplitInfo.getNumTasks();
        // set up the operator plan. (after generating splits - that changes configs)
        Utilities.setMapWork(conf, mapWork, mrScratchDir, false);
    }
    UserPayload serializedConf = TezUtils.createUserPayloadFromConf(conf);
    String procClassName = MapTezProcessor.class.getName();
    if (mapWork instanceof MergeFileWork) {
        procClassName = MergeFileTezProcessor.class.getName();
    }
    VertexExecutionContext executionContext = createVertexExecutionContext(mapWork);
    map = Vertex.create(mapWork.getName(), ProcessorDescriptor.create(procClassName).setUserPayload(serializedConf), numTasks, getContainerResource(conf));
    map.setTaskEnvironment(getContainerEnvironment(conf, true));
    map.setExecutionContext(executionContext);
    map.setTaskLaunchCmdOpts(getContainerJavaOpts(conf));
    assert mapWork.getAliasToWork().keySet().size() == 1;
    // Add the actual source input
    String alias = mapWork.getAliasToWork().keySet().iterator().next();
    map.addDataSource(alias, dataSource);
    map.addTaskLocalFiles(localResources);
    return map;
}