Example #1
0
  /*
   * Helper function to create Vertex for given ReduceWork.
   */
  private Vertex createVertex(
      JobConf conf,
      ReduceWork reduceWork,
      LocalResource appJarLr,
      List<LocalResource> additionalLr,
      FileSystem fs,
      Path mrScratchDir,
      Context ctx)
      throws Exception {

    // set up operator plan
    Utilities.setReduceWork(conf, reduceWork, mrScratchDir, false);

    // create the directories FileSinkOperators need
    Utilities.createTmpDirs(conf, reduceWork);

    // Call once here, will be updated when we find edges
    MultiStageMRConfToTezTranslator.translateVertexConfToTez(conf, null);

    // create the vertex
    Vertex reducer =
        new Vertex(
            reduceWork.getName(),
            new ProcessorDescriptor(ReduceTezProcessor.class.getName())
                .setUserPayload(MRHelpers.createUserPayloadFromConf(conf)),
            reduceWork.getNumReduceTasks(),
            getContainerResource(conf));

    Map<String, String> environment = new HashMap<String, String>();

    MRHelpers.updateEnvironmentForMRTasks(conf, environment, false);
    reducer.setTaskEnvironment(environment);

    reducer.setJavaOpts(getContainerJavaOpts(conf));

    Map<String, LocalResource> localResources = new HashMap<String, LocalResource>();
    localResources.put(getBaseName(appJarLr), appJarLr);
    for (LocalResource lr : additionalLr) {
      localResources.put(getBaseName(lr), lr);
    }
    reducer.setTaskLocalResources(localResources);

    return reducer;
  }
Example #2
0
  /*
   * 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,
      TezWork tezWork)
      throws Exception {

    Path tezDir = getTezDir(mrScratchDir);

    // set up the operator plan
    Utilities.setMapWork(conf, mapWork, mrScratchDir, false);

    // create the directories FileSinkOperators need
    Utilities.createTmpDirs(conf, mapWork);

    // Tez ask us to call this even if there's no preceding vertex
    MultiStageMRConfToTezTranslator.translateVertexConfToTez(conf, null);

    // finally create the vertex
    Vertex map = null;

    // use tez to combine splits
    boolean useTezGroupedSplits = false;

    int numTasks = -1;
    Class amSplitGeneratorClass = null;
    InputSplitInfo inputSplitInfo = null;
    Class inputFormatClass = conf.getClass("mapred.input.format.class", InputFormat.class);

    boolean vertexHasCustomInput = false;
    if (tezWork != null) {
      for (BaseWork baseWork : tezWork.getParents(mapWork)) {
        if (tezWork.getEdgeType(baseWork, mapWork) == EdgeType.CUSTOM_EDGE) {
          vertexHasCustomInput = true;
        }
      }
    }
    if (vertexHasCustomInput) {
      useTezGroupedSplits = false;
      // grouping happens in execution phase. Setting the class to TezGroupedSplitsInputFormat
      // here would cause pre-mature grouping which would be incorrect.
      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 {
      // we'll set up tez to combine spits for us iff the input format
      // is HiveInputFormat
      if (inputFormatClass == HiveInputFormat.class) {
        useTezGroupedSplits = true;
        conf.setClass(
            "mapred.input.format.class", TezGroupedSplitsInputFormat.class, InputFormat.class);
      }
    }

    if (HiveConf.getBoolVar(conf, ConfVars.HIVE_AM_SPLIT_GENERATION)) {
      // if we're generating the splits in the AM, we just need to set
      // the correct plugin.
      amSplitGeneratorClass = MRInputAMSplitGenerator.class;
    } else {
      // client side split generation means we have to compute them now
      inputSplitInfo =
          MRHelpers.generateInputSplits(
              conf, new Path(tezDir, "split_" + mapWork.getName().replaceAll(" ", "_")));
      numTasks = inputSplitInfo.getNumTasks();
    }

    byte[] serializedConf = MRHelpers.createUserPayloadFromConf(conf);
    map =
        new Vertex(
            mapWork.getName(),
            new ProcessorDescriptor(MapTezProcessor.class.getName()).setUserPayload(serializedConf),
            numTasks,
            getContainerResource(conf));
    Map<String, String> environment = new HashMap<String, String>();
    MRHelpers.updateEnvironmentForMRTasks(conf, environment, true);
    map.setTaskEnvironment(environment);
    map.setJavaOpts(getContainerJavaOpts(conf));

    assert mapWork.getAliasToWork().keySet().size() == 1;

    String alias = mapWork.getAliasToWork().keySet().iterator().next();

    byte[] mrInput = null;
    if (useTezGroupedSplits) {
      mrInput =
          MRHelpers.createMRInputPayloadWithGrouping(
              serializedConf, HiveInputFormat.class.getName());
    } else {
      mrInput = MRHelpers.createMRInputPayload(serializedConf, null);
    }
    map.addInput(
        alias,
        new InputDescriptor(MRInputLegacy.class.getName()).setUserPayload(mrInput),
        amSplitGeneratorClass);

    Map<String, LocalResource> localResources = new HashMap<String, LocalResource>();
    localResources.put(getBaseName(appJarLr), appJarLr);
    for (LocalResource lr : additionalLr) {
      localResources.put(getBaseName(lr), lr);
    }

    if (inputSplitInfo != null) {
      // only relevant for client-side split generation
      map.setTaskLocationsHint(inputSplitInfo.getTaskLocationHints());
      MRHelpers.updateLocalResourcesForInputSplits(
          FileSystem.get(conf), inputSplitInfo, localResources);
    }

    map.setTaskLocalResources(localResources);
    return map;
  }
Example #3
0
  /** Execute a query plan using Hadoop. */
  @SuppressWarnings({"deprecation", "unchecked"})
  @Override
  public int execute(DriverContext driverContext) {

    IOPrepareCache ioPrepareCache = IOPrepareCache.get();
    ioPrepareCache.clear();

    boolean success = true;

    Context ctx = driverContext.getCtx();
    boolean ctxCreated = false;
    Path emptyScratchDir;

    MapWork mWork = work.getMapWork();
    ReduceWork rWork = work.getReduceWork();

    try {
      if (ctx == null) {
        ctx = new Context(job);
        ctxCreated = true;
      }

      emptyScratchDir = ctx.getMRTmpPath();
      FileSystem fs = emptyScratchDir.getFileSystem(job);
      fs.mkdirs(emptyScratchDir);
    } catch (IOException e) {
      e.printStackTrace();
      console.printError(
          "Error launching map-reduce job",
          "\n" + org.apache.hadoop.util.StringUtils.stringifyException(e));
      return 5;
    }

    HiveFileFormatUtils.prepareJobOutput(job);
    // See the javadoc on HiveOutputFormatImpl and HadoopShims.prepareJobOutput()
    job.setOutputFormat(HiveOutputFormatImpl.class);

    job.setMapperClass(ExecMapper.class);

    job.setMapOutputKeyClass(HiveKey.class);
    job.setMapOutputValueClass(BytesWritable.class);

    try {
      String partitioner = HiveConf.getVar(job, ConfVars.HIVEPARTITIONER);
      job.setPartitionerClass(JavaUtils.loadClass(partitioner));
    } catch (ClassNotFoundException e) {
      throw new RuntimeException(e.getMessage(), e);
    }

    if (mWork.getNumMapTasks() != null) {
      job.setNumMapTasks(mWork.getNumMapTasks().intValue());
    }

    if (mWork.getMaxSplitSize() != null) {
      HiveConf.setLongVar(
          job, HiveConf.ConfVars.MAPREDMAXSPLITSIZE, mWork.getMaxSplitSize().longValue());
    }

    if (mWork.getMinSplitSize() != null) {
      HiveConf.setLongVar(
          job, HiveConf.ConfVars.MAPREDMINSPLITSIZE, mWork.getMinSplitSize().longValue());
    }

    if (mWork.getMinSplitSizePerNode() != null) {
      HiveConf.setLongVar(
          job,
          HiveConf.ConfVars.MAPREDMINSPLITSIZEPERNODE,
          mWork.getMinSplitSizePerNode().longValue());
    }

    if (mWork.getMinSplitSizePerRack() != null) {
      HiveConf.setLongVar(
          job,
          HiveConf.ConfVars.MAPREDMINSPLITSIZEPERRACK,
          mWork.getMinSplitSizePerRack().longValue());
    }

    job.setNumReduceTasks(rWork != null ? rWork.getNumReduceTasks().intValue() : 0);
    job.setReducerClass(ExecReducer.class);

    // set input format information if necessary
    setInputAttributes(job);

    // Turn on speculative execution for reducers
    boolean useSpeculativeExecReducers =
        HiveConf.getBoolVar(job, HiveConf.ConfVars.HIVESPECULATIVEEXECREDUCERS);
    HiveConf.setBoolVar(
        job, HiveConf.ConfVars.HADOOPSPECULATIVEEXECREDUCERS, useSpeculativeExecReducers);

    String inpFormat = HiveConf.getVar(job, HiveConf.ConfVars.HIVEINPUTFORMAT);

    if (mWork.isUseBucketizedHiveInputFormat()) {
      inpFormat = BucketizedHiveInputFormat.class.getName();
    }

    LOG.info("Using " + inpFormat);

    try {
      job.setInputFormat(JavaUtils.loadClass(inpFormat));
    } catch (ClassNotFoundException e) {
      throw new RuntimeException(e.getMessage(), e);
    }

    // No-Op - we don't really write anything here ..
    job.setOutputKeyClass(Text.class);
    job.setOutputValueClass(Text.class);

    // Transfer HIVEAUXJARS and HIVEADDEDJARS to "tmpjars" so hadoop understands
    // it
    String auxJars = HiveConf.getVar(job, HiveConf.ConfVars.HIVEAUXJARS);
    String addedJars = HiveConf.getVar(job, HiveConf.ConfVars.HIVEADDEDJARS);
    if (StringUtils.isNotBlank(auxJars) || StringUtils.isNotBlank(addedJars)) {
      String allJars =
          StringUtils.isNotBlank(auxJars)
              ? (StringUtils.isNotBlank(addedJars) ? addedJars + "," + auxJars : auxJars)
              : addedJars;
      LOG.info("adding libjars: " + allJars);
      initializeFiles("tmpjars", allJars);
    }

    // Transfer HIVEADDEDFILES to "tmpfiles" so hadoop understands it
    String addedFiles = HiveConf.getVar(job, HiveConf.ConfVars.HIVEADDEDFILES);
    if (StringUtils.isNotBlank(addedFiles)) {
      initializeFiles("tmpfiles", addedFiles);
    }
    int returnVal = 0;
    boolean noName = StringUtils.isEmpty(HiveConf.getVar(job, HiveConf.ConfVars.HADOOPJOBNAME));

    if (noName) {
      // This is for a special case to ensure unit tests pass
      HiveConf.setVar(job, HiveConf.ConfVars.HADOOPJOBNAME, "JOB" + Utilities.randGen.nextInt());
    }
    String addedArchives = HiveConf.getVar(job, HiveConf.ConfVars.HIVEADDEDARCHIVES);
    // Transfer HIVEADDEDARCHIVES to "tmparchives" so hadoop understands it
    if (StringUtils.isNotBlank(addedArchives)) {
      initializeFiles("tmparchives", addedArchives);
    }

    try {
      MapredLocalWork localwork = mWork.getMapRedLocalWork();
      if (localwork != null && localwork.hasStagedAlias()) {
        if (!ShimLoader.getHadoopShims().isLocalMode(job)) {
          Path localPath = localwork.getTmpPath();
          Path hdfsPath = mWork.getTmpHDFSPath();

          FileSystem hdfs = hdfsPath.getFileSystem(job);
          FileSystem localFS = localPath.getFileSystem(job);
          FileStatus[] hashtableFiles = localFS.listStatus(localPath);
          int fileNumber = hashtableFiles.length;
          String[] fileNames = new String[fileNumber];

          for (int i = 0; i < fileNumber; i++) {
            fileNames[i] = hashtableFiles[i].getPath().getName();
          }

          // package and compress all the hashtable files to an archive file
          String stageId = this.getId();
          String archiveFileName = Utilities.generateTarFileName(stageId);
          localwork.setStageID(stageId);

          CompressionUtils.tar(localPath.toUri().getPath(), fileNames, archiveFileName);
          Path archivePath = Utilities.generateTarPath(localPath, stageId);
          LOG.info("Archive " + hashtableFiles.length + " hash table files to " + archivePath);

          // upload archive file to hdfs
          Path hdfsFilePath = Utilities.generateTarPath(hdfsPath, stageId);
          short replication = (short) job.getInt("mapred.submit.replication", 10);
          hdfs.copyFromLocalFile(archivePath, hdfsFilePath);
          hdfs.setReplication(hdfsFilePath, replication);
          LOG.info("Upload 1 archive file  from" + archivePath + " to: " + hdfsFilePath);

          // add the archive file to distributed cache
          DistributedCache.createSymlink(job);
          DistributedCache.addCacheArchive(hdfsFilePath.toUri(), job);
          LOG.info(
              "Add 1 archive file to distributed cache. Archive file: " + hdfsFilePath.toUri());
        }
      }
      work.configureJobConf(job);
      List<Path> inputPaths = Utilities.getInputPaths(job, mWork, emptyScratchDir, ctx, false);
      Utilities.setInputPaths(job, inputPaths);

      Utilities.setMapRedWork(job, work, ctx.getMRTmpPath());

      if (mWork.getSamplingType() > 0 && rWork != null && job.getNumReduceTasks() > 1) {
        try {
          handleSampling(ctx, mWork, job);
          job.setPartitionerClass(HiveTotalOrderPartitioner.class);
        } catch (IllegalStateException e) {
          console.printInfo("Not enough sampling data.. Rolling back to single reducer task");
          rWork.setNumReduceTasks(1);
          job.setNumReduceTasks(1);
        } catch (Exception e) {
          LOG.error("Sampling error", e);
          console.printError(
              e.toString(), "\n" + org.apache.hadoop.util.StringUtils.stringifyException(e));
          rWork.setNumReduceTasks(1);
          job.setNumReduceTasks(1);
        }
      }

      // remove the pwd from conf file so that job tracker doesn't show this
      // logs
      String pwd = HiveConf.getVar(job, HiveConf.ConfVars.METASTOREPWD);
      if (pwd != null) {
        HiveConf.setVar(job, HiveConf.ConfVars.METASTOREPWD, "HIVE");
      }
      JobClient jc = new JobClient(job);
      // make this client wait if job tracker is not behaving well.
      Throttle.checkJobTracker(job, LOG);

      if (mWork.isGatheringStats() || (rWork != null && rWork.isGatheringStats())) {
        // initialize stats publishing table
        StatsPublisher statsPublisher;
        StatsFactory factory = StatsFactory.newFactory(job);
        if (factory != null) {
          statsPublisher = factory.getStatsPublisher();
          List<String> statsTmpDir = Utilities.getStatsTmpDirs(mWork, job);
          if (rWork != null) {
            statsTmpDir.addAll(Utilities.getStatsTmpDirs(rWork, job));
          }
          StatsCollectionContext sc = new StatsCollectionContext(job);
          sc.setStatsTmpDirs(statsTmpDir);
          if (!statsPublisher.init(sc)) { // creating stats table if not exists
            if (HiveConf.getBoolVar(job, HiveConf.ConfVars.HIVE_STATS_RELIABLE)) {
              throw new HiveException(
                  ErrorMsg.STATSPUBLISHER_INITIALIZATION_ERROR.getErrorCodedMsg());
            }
          }
        }
      }

      Utilities.createTmpDirs(job, mWork);
      Utilities.createTmpDirs(job, rWork);

      SessionState ss = SessionState.get();
      if (HiveConf.getVar(job, HiveConf.ConfVars.HIVE_EXECUTION_ENGINE).equals("tez")
          && ss != null) {
        TezSessionState session = ss.getTezSession();
        TezSessionPoolManager.getInstance().close(session, true);
      }

      // Finally SUBMIT the JOB!
      rj = jc.submitJob(job);
      // replace it back
      if (pwd != null) {
        HiveConf.setVar(job, HiveConf.ConfVars.METASTOREPWD, pwd);
      }

      returnVal = jobExecHelper.progress(rj, jc, ctx.getHiveTxnManager());
      success = (returnVal == 0);
    } catch (Exception e) {
      e.printStackTrace();
      String mesg = " with exception '" + Utilities.getNameMessage(e) + "'";
      if (rj != null) {
        mesg = "Ended Job = " + rj.getJobID() + mesg;
      } else {
        mesg = "Job Submission failed" + mesg;
      }

      // Has to use full name to make sure it does not conflict with
      // org.apache.commons.lang.StringUtils
      console.printError(mesg, "\n" + org.apache.hadoop.util.StringUtils.stringifyException(e));

      success = false;
      returnVal = 1;
    } finally {
      Utilities.clearWork(job);
      try {
        if (ctxCreated) {
          ctx.clear();
        }

        if (rj != null) {
          if (returnVal != 0) {
            rj.killJob();
          }
          jobID = rj.getID().toString();
        }
      } catch (Exception e) {
        LOG.warn("Failed while cleaning up ", e);
      } finally {
        HadoopJobExecHelper.runningJobs.remove(rj);
      }
    }

    // get the list of Dynamic partition paths
    try {
      if (rj != null) {
        if (mWork.getAliasToWork() != null) {
          for (Operator<? extends OperatorDesc> op : mWork.getAliasToWork().values()) {
            op.jobClose(job, success);
          }
        }
        if (rWork != null) {
          rWork.getReducer().jobClose(job, success);
        }
      }
    } catch (Exception e) {
      // jobClose needs to execute successfully otherwise fail task
      if (success) {
        success = false;
        returnVal = 3;
        String mesg = "Job Commit failed with exception '" + Utilities.getNameMessage(e) + "'";
        console.printError(mesg, "\n" + org.apache.hadoop.util.StringUtils.stringifyException(e));
      }
    }

    return (returnVal);
  }