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

Example 1 with InputSplitInfoMem

use of org.apache.tez.mapreduce.hadoop.InputSplitInfoMem in project tez by apache.

the class MRInputAMSplitGenerator method initialize.

@Override
public List<Event> initialize() throws Exception {
    StopWatch sw = new StopWatch().start();
    MRInputUserPayloadProto userPayloadProto = MRInputHelpers.parseMRInputPayload(getContext().getInputUserPayload());
    sw.stop();
    if (LOG.isDebugEnabled()) {
        LOG.debug("Time to parse MRInput payload into prot: " + sw.now(TimeUnit.MILLISECONDS));
    }
    sw.reset().start();
    Configuration conf = TezUtils.createConfFromByteString(userPayloadProto.getConfigurationBytes());
    sendSerializedEvents = conf.getBoolean(MRJobConfig.MR_TEZ_INPUT_INITIALIZER_SERIALIZE_EVENT_PAYLOAD, MRJobConfig.MR_TEZ_INPUT_INITIALIZER_SERIALIZE_EVENT_PAYLOAD_DEFAULT);
    sw.stop();
    if (LOG.isDebugEnabled()) {
        LOG.debug("Emitting serialized splits: " + sendSerializedEvents + " for input " + getContext().getInputName());
        LOG.debug("Time converting ByteString to configuration: " + sw.now(TimeUnit.MILLISECONDS));
    }
    sw.reset().start();
    int totalResource = getContext().getTotalAvailableResource().getMemory();
    int taskResource = getContext().getVertexTaskResource().getMemory();
    float waves = conf.getFloat(TezSplitGrouper.TEZ_GROUPING_SPLIT_WAVES, TezSplitGrouper.TEZ_GROUPING_SPLIT_WAVES_DEFAULT);
    int numTasks = (int) ((totalResource * waves) / taskResource);
    boolean groupSplits = userPayloadProto.getGroupingEnabled();
    boolean sortSplits = userPayloadProto.getSortSplitsEnabled();
    LOG.info("Input " + getContext().getInputName() + " asking for " + numTasks + " tasks. Headroom: " + totalResource + ". Task Resource: " + taskResource + ". waves: " + waves + ". groupingEnabled: " + groupSplits + ". SortSplitsEnabled: " + sortSplits);
    // Read all credentials into the credentials instance stored in JobConf.
    JobConf jobConf = new JobConf(conf);
    jobConf.getCredentials().mergeAll(UserGroupInformation.getCurrentUser().getCredentials());
    InputSplitInfoMem inputSplitInfo = null;
    inputSplitInfo = MRInputHelpers.generateInputSplitsToMem(jobConf, groupSplits, sortSplits, groupSplits ? numTasks : 0);
    sw.stop();
    if (LOG.isDebugEnabled()) {
        LOG.debug("Time to create splits to mem: " + sw.now(TimeUnit.MILLISECONDS));
    }
    List<Event> events = Lists.newArrayListWithCapacity(inputSplitInfo.getNumTasks() + 1);
    InputConfigureVertexTasksEvent configureVertexEvent = InputConfigureVertexTasksEvent.create(inputSplitInfo.getNumTasks(), VertexLocationHint.create(inputSplitInfo.getTaskLocationHints()), InputSpecUpdate.getDefaultSinglePhysicalInputSpecUpdate());
    events.add(configureVertexEvent);
    if (sendSerializedEvents) {
        MRSplitsProto splitsProto = inputSplitInfo.getSplitsProto();
        int count = 0;
        for (MRSplitProto mrSplit : splitsProto.getSplitsList()) {
            // Unnecessary array copy, can be avoided by using ByteBuffer instead of a raw array.
            InputDataInformationEvent diEvent = InputDataInformationEvent.createWithSerializedPayload(count++, mrSplit.toByteString().asReadOnlyByteBuffer());
            events.add(diEvent);
        }
    } else {
        int count = 0;
        if (inputSplitInfo.holdsNewFormatSplits()) {
            for (org.apache.hadoop.mapreduce.InputSplit split : inputSplitInfo.getNewFormatSplits()) {
                InputDataInformationEvent diEvent = InputDataInformationEvent.createWithObjectPayload(count++, split);
                events.add(diEvent);
            }
        } else {
            for (org.apache.hadoop.mapred.InputSplit split : inputSplitInfo.getOldFormatSplits()) {
                InputDataInformationEvent diEvent = InputDataInformationEvent.createWithObjectPayload(count++, split);
                events.add(diEvent);
            }
        }
    }
    return events;
}

Example 2 with InputSplitInfoMem

use of org.apache.tez.mapreduce.hadoop.InputSplitInfoMem in project hive by apache.

the class HiveSplitGenerator method initialize.

@SuppressWarnings("unchecked")
@Override
public List<Event> initialize() throws Exception {
    if (getContext() != null) {
        // called from Tez AM.
        prepare(getContext());
    }
    // Setup the map work for this thread. Pruning modified the work instance to potentially remove
    // partitions. The same work instance must be used when generating splits.
    Utilities.setMapWork(jobConf, work);
    try {
        boolean sendSerializedEvents = conf.getBoolean("mapreduce.tez.input.initializer.serialize.event.payload", true);
        // perform dynamic partition pruning
        if (pruner != null) {
            pruner.initialize(getContext(), work, jobConf);
            pruner.prune();
        }
        InputSplitInfoMem inputSplitInfo = null;
        boolean generateConsistentSplits = HiveConf.getBoolVar(conf, HiveConf.ConfVars.HIVE_TEZ_GENERATE_CONSISTENT_SPLITS);
        LOG.info("GenerateConsistentSplitsInHive=" + generateConsistentSplits);
        String realInputFormatName = conf.get("mapred.input.format.class");
        boolean groupingEnabled = userPayloadProto.getGroupingEnabled();
        if (groupingEnabled) {
            // Need to instantiate the realInputFormat
            InputFormat<?, ?> inputFormat = (InputFormat<?, ?>) ReflectionUtils.newInstance(JavaUtils.loadClass(realInputFormatName), jobConf);
            int totalResource = 0;
            int taskResource = 0;
            int availableSlots = 0;
            // FIXME. Do the right thing Luke.
            if (getContext() == null) {
                // for now, totalResource = taskResource for llap
                availableSlots = 1;
            }
            if (getContext() != null) {
                totalResource = getContext().getTotalAvailableResource().getMemory();
                taskResource = getContext().getVertexTaskResource().getMemory();
                availableSlots = totalResource / taskResource;
            }
            if (HiveConf.getLongVar(conf, HiveConf.ConfVars.MAPREDMINSPLITSIZE, 1) <= 1) {
                // broken configuration from mapred-default.xml
                final long blockSize = conf.getLongBytes(DFSConfigKeys.DFS_BLOCK_SIZE_KEY, DFSConfigKeys.DFS_BLOCK_SIZE_DEFAULT);
                final long minGrouping = conf.getLong(TezMapReduceSplitsGrouper.TEZ_GROUPING_SPLIT_MIN_SIZE, TezMapReduceSplitsGrouper.TEZ_GROUPING_SPLIT_MIN_SIZE_DEFAULT);
                final long preferredSplitSize = Math.min(blockSize / 2, minGrouping);
                HiveConf.setLongVar(jobConf, HiveConf.ConfVars.MAPREDMINSPLITSIZE, preferredSplitSize);
                LOG.info("The preferred split size is " + preferredSplitSize);
            }
            float waves;
            // Create the un-grouped splits
            if (numSplits.isPresent()) {
                waves = numSplits.get().floatValue() / availableSlots;
            } else {
                waves = conf.getFloat(TezMapReduceSplitsGrouper.TEZ_GROUPING_SPLIT_WAVES, TezMapReduceSplitsGrouper.TEZ_GROUPING_SPLIT_WAVES_DEFAULT);
            }
            InputSplit[] splits;
            if (generateSingleSplit && conf.get(HiveConf.ConfVars.HIVETEZINPUTFORMAT.varname).equals(HiveInputFormat.class.getName())) {
                MapWork mapWork = Utilities.getMapWork(jobConf);
                List<Path> paths = Utilities.getInputPathsTez(jobConf, mapWork);
                FileSystem fs = paths.get(0).getFileSystem(jobConf);
                FileStatus[] fileStatuses = fs.listStatus(paths.get(0));
                if (fileStatuses.length == 0) {
                    // generate single split typically happens when reading data out of order by queries.
                    // if order by query returns no rows, no files will exists in input path
                    splits = new InputSplit[0];
                } else {
                    // if files exists in input path then it has to be 1 as this code path gets triggered only
                    // of order by queries which is expected to write only one file (written by one reducer)
                    Preconditions.checkState(paths.size() == 1 && fileStatuses.length == 1 && mapWork.getAliasToPartnInfo().size() == 1, "Requested to generate single split. Paths and fileStatuses are expected to be 1. " + "Got paths: " + paths.size() + " fileStatuses: " + fileStatuses.length);
                    splits = new InputSplit[1];
                    FileStatus fileStatus = fileStatuses[0];
                    BlockLocation[] locations = fs.getFileBlockLocations(fileStatus, 0, fileStatus.getLen());
                    Set<String> hostsSet = new HashSet<>();
                    for (BlockLocation location : locations) {
                        hostsSet.addAll(Lists.newArrayList(location.getHosts()));
                    }
                    String[] hosts = hostsSet.toArray(new String[0]);
                    FileSplit fileSplit = new FileSplit(fileStatus.getPath(), 0, fileStatus.getLen(), hosts);
                    String alias = mapWork.getAliases().get(0);
                    PartitionDesc partDesc = mapWork.getAliasToPartnInfo().get(alias);
                    String partIF = partDesc.getInputFileFormatClassName();
                    splits[0] = new HiveInputFormat.HiveInputSplit(fileSplit, partIF);
                }
            } else {
                // Raw splits
                splits = inputFormat.getSplits(jobConf, numSplits.orElse(Math.multiplyExact(availableSlots, (int) waves)));
            }
            // Sort the splits, so that subsequent grouping is consistent.
            Arrays.sort(splits, new InputSplitComparator());
            LOG.info("Number of input splits: " + splits.length + ". " + availableSlots + " available slots, " + waves + " waves. Input format is: " + realInputFormatName);
            // increment/set input counters
            InputInitializerContext inputInitializerContext = getContext();
            TezCounters tezCounters = null;
            String counterName;
            String groupName = null;
            String vertexName = null;
            if (inputInitializerContext != null) {
                try {
                    tezCounters = new TezCounters();
                    groupName = HiveInputCounters.class.getName();
                    vertexName = jobConf.get(Operator.CONTEXT_NAME_KEY, "");
                    counterName = Utilities.getVertexCounterName(HiveInputCounters.RAW_INPUT_SPLITS.name(), vertexName);
                    tezCounters.findCounter(groupName, counterName).increment(splits.length);
                    final List<Path> paths = Utilities.getInputPathsTez(jobConf, work);
                    counterName = Utilities.getVertexCounterName(HiveInputCounters.INPUT_DIRECTORIES.name(), vertexName);
                    tezCounters.findCounter(groupName, counterName).increment(paths.size());
                    final Set<String> files = new HashSet<>();
                    for (InputSplit inputSplit : splits) {
                        if (inputSplit instanceof FileSplit) {
                            final FileSplit fileSplit = (FileSplit) inputSplit;
                            final Path path = fileSplit.getPath();
                            // The assumption here is the path is a file. Only case this is different is ACID deltas.
                            // The isFile check is avoided here for performance reasons.
                            final String fileStr = path.toString();
                            if (!files.contains(fileStr)) {
                                files.add(fileStr);
                            }
                        }
                    }
                    counterName = Utilities.getVertexCounterName(HiveInputCounters.INPUT_FILES.name(), vertexName);
                    tezCounters.findCounter(groupName, counterName).increment(files.size());
                } catch (Exception e) {
                    LOG.warn("Caught exception while trying to update Tez counters", e);
                }
            }
            if (work.getIncludedBuckets() != null) {
                splits = pruneBuckets(work, splits);
            }
            Multimap<Integer, InputSplit> groupedSplits = splitGrouper.generateGroupedSplits(jobConf, conf, splits, waves, availableSlots, splitLocationProvider);
            // And finally return them in a flat array
            InputSplit[] flatSplits = groupedSplits.values().toArray(new InputSplit[0]);
            LOG.info("Number of split groups: " + flatSplits.length);
            if (inputInitializerContext != null) {
                try {
                    counterName = Utilities.getVertexCounterName(HiveInputCounters.GROUPED_INPUT_SPLITS.name(), vertexName);
                    tezCounters.findCounter(groupName, counterName).setValue(flatSplits.length);
                    LOG.debug("Published tez counters: {}", tezCounters);
                    inputInitializerContext.addCounters(tezCounters);
                } catch (Exception e) {
                    LOG.warn("Caught exception while trying to update Tez counters", e);
                }
            }
            List<TaskLocationHint> locationHints = splitGrouper.createTaskLocationHints(flatSplits, generateConsistentSplits);
            inputSplitInfo = new InputSplitInfoMem(flatSplits, locationHints, flatSplits.length, null, jobConf);
        } else {
            // If this is used in the future - make sure to disable grouping in the payload, if it isn't already disabled
            throw new RuntimeException("HiveInputFormat does not support non-grouped splits, InputFormatName is: " + realInputFormatName);
        // inputSplitInfo = MRInputHelpers.generateInputSplitsToMem(jobConf, false, 0);
        }
        return createEventList(sendSerializedEvents, inputSplitInfo);
    } finally {
        Utilities.clearWork(jobConf);
    }
}