Examples with PartitionDesc org.apache.hadoop.hive.ql.plan.PartitionDesc used on opensource projects

Example 81 with PartitionDesc

use of org.apache.hadoop.hive.ql.plan.PartitionDesc 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);
    }
}

Example 82 with PartitionDesc

use of org.apache.hadoop.hive.ql.plan.PartitionDesc in project hive by apache.

the class MapRecordProcessor method init.

@Override
void init(MRTaskReporter mrReporter, Map<String, LogicalInput> inputs, Map<String, LogicalOutput> outputs) throws Exception {
    perfLogger.perfLogBegin(CLASS_NAME, PerfLogger.TEZ_INIT_OPERATORS);
    super.init(mrReporter, inputs, outputs);
    checkAbortCondition();
    String key = processorContext.getTaskVertexName() + MAP_PLAN_KEY;
    cacheKeys.add(key);
    // create map and fetch operators
    if (!isInCompaction) {
        mapWork = cache.retrieve(key, () -> Utilities.getMapWork(jconf));
    } else {
        // During query-based compaction, we don't want to retrieve old MapWork from the cache, we want a new mapper
        // and new UDF validate_acid_sort_order instance for each bucket, otherwise validate_acid_sort_order will fail.
        mapWork = Utilities.getMapWork(jconf);
    }
    // TODO HIVE-14042. Cleanup may be required if exiting early.
    Utilities.setMapWork(jconf, mapWork);
    for (PartitionDesc part : mapWork.getAliasToPartnInfo().values()) {
        TableDesc tableDesc = part.getTableDesc();
        Utilities.copyJobSecretToTableProperties(tableDesc);
    }
    String prefixes = jconf.get(DagUtils.TEZ_MERGE_WORK_FILE_PREFIXES);
    if (prefixes != null) {
        mergeWorkList = new ArrayList<>();
        for (final String prefix : prefixes.split(",")) {
            if (prefix == null || prefix.isEmpty()) {
                continue;
            }
            key = processorContext.getTaskVertexName() + prefix;
            cacheKeys.add(key);
            checkAbortCondition();
            mergeWorkList.add((MapWork) cache.retrieve(key, () -> Utilities.getMergeWork(jconf, prefix)));
        }
    }
    MapredContext.init(true, new JobConf(jconf));
    ((TezContext) MapredContext.get()).setInputs(inputs);
    ((TezContext) MapredContext.get()).setTezProcessorContext(processorContext);
    // Update JobConf using MRInput, info like filename comes via this
    checkAbortCondition();
    legacyMRInput = getMRInput(inputs);
    if (legacyMRInput != null) {
        Configuration updatedConf = legacyMRInput.getConfigUpdates();
        if (updatedConf != null) {
            for (Entry<String, String> entry : updatedConf) {
                jconf.set(entry.getKey(), entry.getValue());
            }
        }
    }
    checkAbortCondition();
    createOutputMap();
    // Start all the Outputs.
    for (Entry<String, LogicalOutput> outputEntry : outputs.entrySet()) {
        LOG.debug("Starting Output: " + outputEntry.getKey());
        outputEntry.getValue().start();
        ((TezKVOutputCollector) outMap.get(outputEntry.getKey())).initialize();
    }
    checkAbortCondition();
    try {
        CompilationOpContext runtimeCtx = new CompilationOpContext();
        if (mapWork.getVectorMode()) {
            mapOp = new VectorMapOperator(runtimeCtx);
        } else {
            mapOp = new MapOperator(runtimeCtx);
        }
        // Not synchronizing creation of mapOp with an invocation. Check immediately
        // after creation in case abort has been set.
        // Relying on the regular flow to clean up the actual operator. i.e. If an exception is
        // thrown, an attempt will be made to cleanup the op.
        // If we are here - exit out via an exception. If we're in the middle of the opeartor.initialize
        // call further down, we rely upon op.abort().
        checkAbortCondition();
        mapOp.clearConnectedOperators();
        mapOp.setExecContext(execContext);
        boolean fromCache = false;
        if (mergeWorkList != null) {
            AbstractMapOperator mergeMapOp = null;
            for (BaseWork mergeWork : mergeWorkList) {
                // TODO HIVE-14042. What is mergeWork, and why is it not part of the regular operator chain.
                // The mergeMapOp.initialize call further down can block, and will not receive information
                // about an abort request.
                MapWork mergeMapWork = (MapWork) mergeWork;
                if (mergeMapWork.getVectorMode()) {
                    mergeMapOp = new VectorMapOperator(runtimeCtx);
                } else {
                    mergeMapOp = new MapOperator(runtimeCtx);
                }
                mergeMapOpList.add(mergeMapOp);
                // initialize the merge operators first.
                if (mergeMapOp != null) {
                    mergeMapOp.setConf(mergeMapWork);
                    LOG.info("Input name is {}", mergeMapWork.getName());
                    jconf.set(Utilities.INPUT_NAME, mergeMapWork.getName());
                    mergeMapOp.initialize(jconf, null);
                    // if there are no files/partitions to read, we need to skip trying to read
                    MultiMRInput multiMRInput = multiMRInputMap.get(mergeMapWork.getName());
                    boolean skipRead = false;
                    if (multiMRInput == null) {
                        LOG.info("Multi MR Input for work {} is null. Skipping read.", mergeMapWork.getName());
                        skipRead = true;
                    } else {
                        Collection<KeyValueReader> keyValueReaders = multiMRInput.getKeyValueReaders();
                        if ((keyValueReaders == null) || (keyValueReaders.isEmpty())) {
                            LOG.info("Key value readers are null or empty and hence skipping read. " + "KeyValueReaders = {}", keyValueReaders);
                            skipRead = true;
                        }
                    }
                    if (skipRead) {
                        List<Operator<?>> children = new ArrayList<>();
                        children.addAll(mergeMapOp.getConf().getAliasToWork().values());
                        // do the same thing as setChildren when there is nothing to read.
                        // the setChildren method initializes the object inspector needed by the operators
                        // based on path and partition information which we don't have in this case.
                        mergeMapOp.initEmptyInputChildren(children, jconf);
                    } else {
                        // the setChildren method initializes the object inspector needed by the operators
                        // based on path and partition information.
                        mergeMapOp.setChildren(jconf);
                    }
                    Operator<? extends OperatorDesc> finalOp = getFinalOp(mergeMapOp);
                    if (finalOp instanceof TezDummyStoreOperator) {
                        // we ensure that we don't try to read any data in case of skip read.
                        ((TezDummyStoreOperator) finalOp).setFetchDone(skipRead);
                        mapOp.setConnectedOperators(mergeMapWork.getTag(), (DummyStoreOperator) finalOp);
                    } else {
                        // found the plan is already connected which means this is derived from the cache.
                        fromCache = true;
                    }
                    mergeMapOp.passExecContext(new ExecMapperContext(jconf));
                    mergeMapOp.initializeLocalWork(jconf);
                }
            }
        }
        if (!fromCache) {
            // if not from cache, we still need to hook up the plans.
            ((TezContext) (MapredContext.get())).setDummyOpsMap(mapOp.getConnectedOperators());
        }
        // initialize map operator
        mapOp.setConf(mapWork);
        LOG.info("Main input name is " + mapWork.getName());
        jconf.set(Utilities.INPUT_NAME, mapWork.getName());
        mapOp.initialize(jconf, null);
        checkAbortCondition();
        mapOp.setChildren(jconf);
        mapOp.passExecContext(execContext);
        LOG.info(mapOp.dump(0));
        // set memory available for operators
        long memoryAvailableToTask = processorContext.getTotalMemoryAvailableToTask();
        if (mapOp.getConf() != null) {
            mapOp.getConf().setMaxMemoryAvailable(memoryAvailableToTask);
            LOG.info("Memory available for operators set to {}", LlapUtil.humanReadableByteCount(memoryAvailableToTask));
        }
        OperatorUtils.setMemoryAvailable(mapOp.getChildOperators(), memoryAvailableToTask);
        mapOp.initializeLocalWork(jconf);
        // Setup values registry
        checkAbortCondition();
        String valueRegistryKey = DynamicValue.DYNAMIC_VALUE_REGISTRY_CACHE_KEY;
        // On LLAP dynamic value registry might already be cached.
        final DynamicValueRegistryTez registryTez = dynamicValueCache.retrieve(valueRegistryKey, () -> new DynamicValueRegistryTez());
        dynamicValueCacheKeys.add(valueRegistryKey);
        RegistryConfTez registryConf = new RegistryConfTez(jconf, mapWork, processorContext, inputs);
        registryTez.init(registryConf);
        checkAbortCondition();
        initializeMapRecordSources();
        mapOp.initializeMapOperator(jconf);
        if ((mergeMapOpList != null) && !mergeMapOpList.isEmpty()) {
            for (AbstractMapOperator mergeMapOp : mergeMapOpList) {
                jconf.set(Utilities.INPUT_NAME, mergeMapOp.getConf().getName());
                // TODO HIVE-14042. abort handling: Handling of mergeMapOp
                mergeMapOp.initializeMapOperator(jconf);
            }
        }
        // Initialization isn't finished until all parents of all operators
        // are initialized. For broadcast joins that means initializing the
        // dummy parent operators as well.
        List<HashTableDummyOperator> dummyOps = mapWork.getDummyOps();
        jconf.set(Utilities.INPUT_NAME, mapWork.getName());
        if (dummyOps != null) {
            for (Operator<? extends OperatorDesc> dummyOp : dummyOps) {
                dummyOp.setExecContext(execContext);
                // TODO HIVE-14042. Handling of dummyOps, and propagating abort information to them
                dummyOp.initialize(jconf, null);
            }
        }
        OperatorUtils.setChildrenCollector(mapOp.getChildOperators(), outMap);
        mapOp.setReporter(reporter);
        MapredContext.get().setReporter(reporter);
    } catch (Throwable e) {
        setAborted(true);
        if (e instanceof OutOfMemoryError) {
            // Don't create a new object if we are already out of memory
            throw (OutOfMemoryError) e;
        } else if (e instanceof InterruptedException) {
            LOG.info("Hit an interrupt while initializing MapRecordProcessor. Message={}", e.getMessage());
            throw (InterruptedException) e;
        } else {
            throw new RuntimeException("Map operator initialization failed", e);
        }
    }
    perfLogger.perfLogEnd(CLASS_NAME, PerfLogger.TEZ_INIT_OPERATORS);
}

Example 83 with PartitionDesc

use of org.apache.hadoop.hive.ql.plan.PartitionDesc in project hive by apache.

the class SymbolicInputFormat method rework.

public void rework(HiveConf job, MapredWork work) throws IOException {
    Map<Path, PartitionDesc> pathToParts = work.getMapWork().getPathToPartitionInfo();
    List<Path> toRemovePaths = new ArrayList<>();
    Map<Path, PartitionDesc> toAddPathToPart = new HashMap<>();
    Map<Path, List<String>> pathToAliases = work.getMapWork().getPathToAliases();
    for (Map.Entry<Path, PartitionDesc> pathPartEntry : pathToParts.entrySet()) {
        Path path = pathPartEntry.getKey();
        PartitionDesc partDesc = pathPartEntry.getValue();
        // this path points to a symlink path
        if (partDesc.getInputFileFormatClass().equals(SymlinkTextInputFormat.class)) {
            // change to TextInputFormat
            partDesc.setInputFileFormatClass(TextInputFormat.class);
            FileSystem fileSystem = path.getFileSystem(job);
            FileStatus fStatus = fileSystem.getFileStatus(path);
            FileStatus[] symlinks = null;
            if (!fStatus.isDir()) {
                symlinks = new FileStatus[] { fStatus };
            } else {
                symlinks = fileSystem.listStatus(path, FileUtils.HIDDEN_FILES_PATH_FILTER);
            }
            toRemovePaths.add(path);
            List<String> aliases = pathToAliases.remove(path);
            for (FileStatus symlink : symlinks) {
                BufferedReader reader = null;
                try {
                    reader = new BufferedReader(new InputStreamReader(fileSystem.open(symlink.getPath())));
                    partDesc.setInputFileFormatClass(TextInputFormat.class);
                    String line;
                    while ((line = reader.readLine()) != null) {
                        // no check for the line? How to check?
                        // if the line is invalid for any reason, the job will fail.
                        FileStatus[] matches = fileSystem.globStatus(new Path(line));
                        for (FileStatus fileStatus : matches) {
                            Path schemaLessPath = Path.getPathWithoutSchemeAndAuthority(fileStatus.getPath());
                            StringInternUtils.internUriStringsInPath(schemaLessPath);
                            toAddPathToPart.put(schemaLessPath, partDesc);
                            pathToAliases.put(schemaLessPath, aliases);
                        }
                    }
                } finally {
                    org.apache.hadoop.io.IOUtils.closeStream(reader);
                }
            }
        }
    }
    for (Entry<Path, PartitionDesc> toAdd : toAddPathToPart.entrySet()) {
        work.getMapWork().addPathToPartitionInfo(toAdd.getKey(), toAdd.getValue());
    }
    for (Path toRemove : toRemovePaths) {
        work.getMapWork().removePathToPartitionInfo(toRemove);
    }
}

Example 84 with PartitionDesc

use of org.apache.hadoop.hive.ql.plan.PartitionDesc in project hive by apache.

the class AvroGenericRecordReader method getSchema.

/**
 * Attempt to retrieve the reader schema.  We have a couple opportunities
 * to provide this, depending on whether or not we're just selecting data
 * or running with a MR job.
 * @return  Reader schema for the Avro object, or null if it has not been provided.
 * @throws AvroSerdeException
 */
private Schema getSchema(JobConf job, FileSplit split) throws AvroSerdeException, IOException {
    // Inside of a MR job, we can pull out the actual properties
    if (AvroSerdeUtils.insideMRJob(job)) {
        MapWork mapWork = Utilities.getMapWork(job);
        // that matches our input split.
        for (Map.Entry<Path, PartitionDesc> pathsAndParts : mapWork.getPathToPartitionInfo().entrySet()) {
            Path partitionPath = pathsAndParts.getKey();
            if (pathIsInPartition(split.getPath(), partitionPath)) {
                LOG.info("Matching partition {} with input split {}", partitionPath, split);
                Properties props = pathsAndParts.getValue().getProperties();
                if (props.containsKey(AvroTableProperties.SCHEMA_LITERAL.getPropName()) || props.containsKey(AvroTableProperties.SCHEMA_URL.getPropName())) {
                    return AvroSerdeUtils.determineSchemaOrThrowException(job, props);
                } else {
                    // If it's not in this property, it won't be in any others
                    return null;
                }
            }
        }
        LOG.info("Unable to match filesplit {} with a partition.", split);
    }
    // In "select * from table" situations (non-MR), we can add things to the job
    // It's safe to add this to the job since it's not *actually* a mapred job.
    // Here the global state is confined to just this process.
    String s = job.get(AvroTableProperties.AVRO_SERDE_SCHEMA.getPropName());
    if (StringUtils.isNotBlank(s)) {
        LOG.info("Found the avro schema in the job");
        LOG.debug("Avro schema: {}", s);
        return AvroSerdeUtils.getSchemaFor(s);
    }
    // No more places to get the schema from. Give up.  May have to re-encode later.
    return null;
}

Example 85 with PartitionDesc

use of org.apache.hadoop.hive.ql.plan.PartitionDesc in project hive by apache.

the class GenMRSkewJoinProcessor method processSkewJoin.

/**
 * Create tasks for processing skew joins. The idea is (HIVE-964) to use
 * separated jobs and map-joins to handle skew joins.
 * <p>
 * <ul>
 * <li>
 * Number of mr jobs to handle skew keys is the number of table minus 1 (we
 * can stream the last table, so big keys in the last table will not be a
 * problem).
 * <li>
 * At runtime in Join, we output big keys in one table into one corresponding
 * directories, and all same keys in other tables into different dirs(one for
 * each table). The directories will look like:
 * <ul>
 * <li>
 * dir-T1-bigkeys(containing big keys in T1), dir-T2-keys(containing keys
 * which is big in T1),dir-T3-keys(containing keys which is big in T1), ...
 * <li>
 * dir-T1-keys(containing keys which is big in T2), dir-T2-bigkeys(containing
 * big keys in T2),dir-T3-keys(containing keys which is big in T2), ...
 * <li>
 * dir-T1-keys(containing keys which is big in T3), dir-T2-keys(containing big
 * keys in T3),dir-T3-bigkeys(containing keys which is big in T3), ... .....
 * </ul>
 * </ul>
 * For each table, we launch one mapjoin job, taking the directory containing
 * big keys in this table and corresponding dirs in other tables as input.
 * (Actually one job for one row in the above.)
 *
 * <p>
 * For more discussions, please check
 * https://issues.apache.org/jira/browse/HIVE-964.
 */
@SuppressWarnings("unchecked")
public static void processSkewJoin(JoinOperator joinOp, Task<?> currTask, ParseContext parseCtx) throws SemanticException {
    // now does not work with outer joins
    if (!GenMRSkewJoinProcessor.skewJoinEnabled(parseCtx.getConf(), joinOp)) {
        return;
    }
    List<Task<?>> children = currTask.getChildTasks();
    Path baseTmpDir = parseCtx.getContext().getMRTmpPath();
    JoinDesc joinDescriptor = joinOp.getConf();
    Map<Byte, List<ExprNodeDesc>> joinValues = joinDescriptor.getExprs();
    int numAliases = joinValues.size();
    Map<Byte, Path> bigKeysDirMap = new HashMap<Byte, Path>();
    Map<Byte, Map<Byte, Path>> smallKeysDirMap = new HashMap<Byte, Map<Byte, Path>>();
    Map<Byte, Path> skewJoinJobResultsDir = new HashMap<Byte, Path>();
    Byte[] tags = joinDescriptor.getTagOrder();
    for (int i = 0; i < numAliases; i++) {
        Byte alias = tags[i];
        bigKeysDirMap.put(alias, getBigKeysDir(baseTmpDir, alias));
        Map<Byte, Path> smallKeysMap = new HashMap<Byte, Path>();
        smallKeysDirMap.put(alias, smallKeysMap);
        for (Byte src2 : tags) {
            if (!src2.equals(alias)) {
                smallKeysMap.put(src2, getSmallKeysDir(baseTmpDir, alias, src2));
            }
        }
        skewJoinJobResultsDir.put(alias, getBigKeysSkewJoinResultDir(baseTmpDir, alias));
    }
    joinDescriptor.setHandleSkewJoin(true);
    joinDescriptor.setBigKeysDirMap(bigKeysDirMap);
    joinDescriptor.setSmallKeysDirMap(smallKeysDirMap);
    joinDescriptor.setSkewKeyDefinition(HiveConf.getIntVar(parseCtx.getConf(), HiveConf.ConfVars.HIVESKEWJOINKEY));
    HashMap<Path, Task<?>> bigKeysDirToTaskMap = new HashMap<Path, Task<?>>();
    List<Serializable> listWorks = new ArrayList<Serializable>();
    List<Task<?>> listTasks = new ArrayList<Task<?>>();
    MapredWork currPlan = (MapredWork) currTask.getWork();
    TableDesc keyTblDesc = (TableDesc) currPlan.getReduceWork().getKeyDesc().clone();
    List<String> joinKeys = Utilities.getColumnNames(keyTblDesc.getProperties());
    List<String> joinKeyTypes = Utilities.getColumnTypes(keyTblDesc.getProperties());
    Map<Byte, TableDesc> tableDescList = new HashMap<Byte, TableDesc>();
    Map<Byte, RowSchema> rowSchemaList = new HashMap<Byte, RowSchema>();
    Map<Byte, List<ExprNodeDesc>> newJoinValues = new HashMap<Byte, List<ExprNodeDesc>>();
    Map<Byte, List<ExprNodeDesc>> newJoinKeys = new HashMap<Byte, List<ExprNodeDesc>>();
    // used for create mapJoinDesc, should be in order
    List<TableDesc> newJoinValueTblDesc = new ArrayList<TableDesc>();
    for (Byte tag : tags) {
        newJoinValueTblDesc.add(null);
    }
    for (int i = 0; i < numAliases; i++) {
        Byte alias = tags[i];
        List<ExprNodeDesc> valueCols = joinValues.get(alias);
        String colNames = "";
        String colTypes = "";
        int columnSize = valueCols.size();
        List<ExprNodeDesc> newValueExpr = new ArrayList<ExprNodeDesc>();
        List<ExprNodeDesc> newKeyExpr = new ArrayList<ExprNodeDesc>();
        ArrayList<ColumnInfo> columnInfos = new ArrayList<ColumnInfo>();
        boolean first = true;
        for (int k = 0; k < columnSize; k++) {
            TypeInfo type = valueCols.get(k).getTypeInfo();
            // any name, it does not matter.
            String newColName = i + "_VALUE_" + k;
            ColumnInfo columnInfo = new ColumnInfo(newColName, type, alias.toString(), false);
            columnInfos.add(columnInfo);
            newValueExpr.add(new ExprNodeColumnDesc(columnInfo));
            if (!first) {
                colNames = colNames + ",";
                colTypes = colTypes + ",";
            }
            first = false;
            colNames = colNames + newColName;
            colTypes = colTypes + valueCols.get(k).getTypeString();
        }
        // we are putting join keys at last part of the spilled table
        for (int k = 0; k < joinKeys.size(); k++) {
            if (!first) {
                colNames = colNames + ",";
                colTypes = colTypes + ",";
            }
            first = false;
            colNames = colNames + joinKeys.get(k);
            colTypes = colTypes + joinKeyTypes.get(k);
            ColumnInfo columnInfo = new ColumnInfo(joinKeys.get(k), TypeInfoFactory.getPrimitiveTypeInfo(joinKeyTypes.get(k)), alias.toString(), false);
            columnInfos.add(columnInfo);
            newKeyExpr.add(new ExprNodeColumnDesc(columnInfo));
        }
        newJoinValues.put(alias, newValueExpr);
        newJoinKeys.put(alias, newKeyExpr);
        tableDescList.put(alias, Utilities.getTableDesc(colNames, colTypes));
        rowSchemaList.put(alias, new RowSchema(columnInfos));
        // construct value table Desc
        String valueColNames = "";
        String valueColTypes = "";
        first = true;
        for (int k = 0; k < columnSize; k++) {
            // any name, it does not matter.
            String newColName = i + "_VALUE_" + k;
            if (!first) {
                valueColNames = valueColNames + ",";
                valueColTypes = valueColTypes + ",";
            }
            valueColNames = valueColNames + newColName;
            valueColTypes = valueColTypes + valueCols.get(k).getTypeString();
            first = false;
        }
        newJoinValueTblDesc.set(Byte.valueOf((byte) i), Utilities.getTableDesc(valueColNames, valueColTypes));
    }
    joinDescriptor.setSkewKeysValuesTables(tableDescList);
    joinDescriptor.setKeyTableDesc(keyTblDesc);
    for (int i = 0; i < numAliases - 1; i++) {
        Byte src = tags[i];
        MapWork newPlan = PlanUtils.getMapRedWork().getMapWork();
        // This code has been only added for testing
        boolean mapperCannotSpanPartns = parseCtx.getConf().getBoolVar(HiveConf.ConfVars.HIVE_MAPPER_CANNOT_SPAN_MULTIPLE_PARTITIONS);
        newPlan.setMapperCannotSpanPartns(mapperCannotSpanPartns);
        MapredWork clonePlan = SerializationUtilities.clonePlan(currPlan);
        Operator<? extends OperatorDesc>[] parentOps = new TableScanOperator[tags.length];
        for (int k = 0; k < tags.length; k++) {
            Operator<? extends OperatorDesc> ts = GenMapRedUtils.createTemporaryTableScanOperator(joinOp.getCompilationOpContext(), rowSchemaList.get((byte) k));
            ((TableScanOperator) ts).setTableDescSkewJoin(tableDescList.get((byte) k));
            parentOps[k] = ts;
        }
        Operator<? extends OperatorDesc> tblScan_op = parentOps[i];
        ArrayList<String> aliases = new ArrayList<String>();
        String alias = src.toString().intern();
        aliases.add(alias);
        Path bigKeyDirPath = bigKeysDirMap.get(src);
        newPlan.addPathToAlias(bigKeyDirPath, aliases);
        newPlan.getAliasToWork().put(alias, tblScan_op);
        PartitionDesc part = new PartitionDesc(tableDescList.get(src), null);
        newPlan.addPathToPartitionInfo(bigKeyDirPath, part);
        newPlan.getAliasToPartnInfo().put(alias, part);
        Operator<? extends OperatorDesc> reducer = clonePlan.getReduceWork().getReducer();
        assert reducer instanceof JoinOperator;
        JoinOperator cloneJoinOp = (JoinOperator) reducer;
        String dumpFilePrefix = "mapfile" + PlanUtils.getCountForMapJoinDumpFilePrefix();
        MapJoinDesc mapJoinDescriptor = new MapJoinDesc(newJoinKeys, keyTblDesc, newJoinValues, newJoinValueTblDesc, newJoinValueTblDesc, joinDescriptor.getOutputColumnNames(), i, joinDescriptor.getConds(), joinDescriptor.getFilters(), joinDescriptor.getNoOuterJoin(), dumpFilePrefix, joinDescriptor.getMemoryMonitorInfo(), joinDescriptor.getInMemoryDataSize());
        mapJoinDescriptor.setTagOrder(tags);
        mapJoinDescriptor.setHandleSkewJoin(false);
        mapJoinDescriptor.setNullSafes(joinDescriptor.getNullSafes());
        mapJoinDescriptor.setColumnExprMap(joinDescriptor.getColumnExprMap());
        MapredLocalWork localPlan = new MapredLocalWork(new LinkedHashMap<String, Operator<? extends OperatorDesc>>(), new LinkedHashMap<String, FetchWork>());
        Map<Byte, Path> smallTblDirs = smallKeysDirMap.get(src);
        for (int j = 0; j < numAliases; j++) {
            if (j == i) {
                continue;
            }
            Byte small_alias = tags[j];
            Operator<? extends OperatorDesc> tblScan_op2 = parentOps[j];
            localPlan.getAliasToWork().put(small_alias.toString(), tblScan_op2);
            Path tblDir = smallTblDirs.get(small_alias);
            localPlan.getAliasToFetchWork().put(small_alias.toString(), new FetchWork(tblDir, tableDescList.get(small_alias)));
        }
        newPlan.setMapRedLocalWork(localPlan);
        // construct a map join and set it as the child operator of tblScan_op
        MapJoinOperator mapJoinOp = (MapJoinOperator) OperatorFactory.getAndMakeChild(joinOp.getCompilationOpContext(), mapJoinDescriptor, (RowSchema) null, parentOps);
        // change the children of the original join operator to point to the map
        // join operator
        List<Operator<? extends OperatorDesc>> childOps = cloneJoinOp.getChildOperators();
        for (Operator<? extends OperatorDesc> childOp : childOps) {
            childOp.replaceParent(cloneJoinOp, mapJoinOp);
        }
        mapJoinOp.setChildOperators(childOps);
        HiveConf jc = new HiveConf(parseCtx.getConf(), GenMRSkewJoinProcessor.class);
        newPlan.setNumMapTasks(HiveConf.getIntVar(jc, HiveConf.ConfVars.HIVESKEWJOINMAPJOINNUMMAPTASK));
        newPlan.setMinSplitSize(HiveConf.getLongVar(jc, HiveConf.ConfVars.HIVESKEWJOINMAPJOINMINSPLIT));
        newPlan.setInputformat(HiveInputFormat.class.getName());
        MapredWork w = new MapredWork();
        w.setMapWork(newPlan);
        Task<?> skewJoinMapJoinTask = TaskFactory.get(w);
        skewJoinMapJoinTask.setFetchSource(currTask.isFetchSource());
        bigKeysDirToTaskMap.put(bigKeyDirPath, skewJoinMapJoinTask);
        listWorks.add(skewJoinMapJoinTask.getWork());
        listTasks.add(skewJoinMapJoinTask);
    }
    if (children != null) {
        for (Task<?> tsk : listTasks) {
            for (Task<?> oldChild : children) {
                tsk.addDependentTask(oldChild);
            }
        }
        currTask.setChildTasks(new ArrayList<Task<?>>());
        for (Task<?> oldChild : children) {
            oldChild.getParentTasks().remove(currTask);
        }
        listTasks.addAll(children);
    }
    ConditionalResolverSkewJoinCtx context = new ConditionalResolverSkewJoinCtx(bigKeysDirToTaskMap, children);
    ConditionalWork cndWork = new ConditionalWork(listWorks);
    ConditionalTask cndTsk = (ConditionalTask) TaskFactory.get(cndWork);
    cndTsk.setListTasks(listTasks);
    cndTsk.setResolver(new ConditionalResolverSkewJoin());
    cndTsk.setResolverCtx(context);
    currTask.setChildTasks(new ArrayList<Task<?>>());
    currTask.addDependentTask(cndTsk);
    return;
}