Examples with InputSplitInfo org.apache.tez.mapreduce.hadoop.InputSplitInfo used on opensource projects

Example 1 with InputSplitInfo

use of org.apache.tez.mapreduce.hadoop.InputSplitInfo 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;
}