Search in sources :

Example 1 with TezKVOutputCollector

use of org.apache.hadoop.hive.ql.exec.tez.TezProcessor.TezKVOutputCollector in project hive by apache.

the class RecordProcessor method createOutputMap.

protected void createOutputMap() {
    Preconditions.checkState(outMap == null, "Outputs should only be setup once");
    outMap = new HashMap<>();
    for (Entry<String, LogicalOutput> entry : outputs.entrySet()) {
        TezKVOutputCollector collector = new TezKVOutputCollector(entry.getValue());
        outMap.put(entry.getKey(), collector);
    }
}
Also used : TezKVOutputCollector(org.apache.hadoop.hive.ql.exec.tez.TezProcessor.TezKVOutputCollector) LogicalOutput(org.apache.tez.runtime.api.LogicalOutput)

Example 2 with TezKVOutputCollector

use of org.apache.hadoop.hive.ql.exec.tez.TezProcessor.TezKVOutputCollector 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);
}
Also used : AbstractMapOperator(org.apache.hadoop.hive.ql.exec.AbstractMapOperator) HashTableDummyOperator(org.apache.hadoop.hive.ql.exec.HashTableDummyOperator) TezDummyStoreOperator(org.apache.hadoop.hive.ql.exec.TezDummyStoreOperator) MapOperator(org.apache.hadoop.hive.ql.exec.MapOperator) VectorMapOperator(org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator) Operator(org.apache.hadoop.hive.ql.exec.Operator) DummyStoreOperator(org.apache.hadoop.hive.ql.exec.DummyStoreOperator) AbstractMapOperator(org.apache.hadoop.hive.ql.exec.AbstractMapOperator) Configuration(org.apache.hadoop.conf.Configuration) KeyValueReader(org.apache.tez.runtime.library.api.KeyValueReader) ArrayList(java.util.ArrayList) VectorMapOperator(org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator) TezDummyStoreOperator(org.apache.hadoop.hive.ql.exec.TezDummyStoreOperator) MapOperator(org.apache.hadoop.hive.ql.exec.MapOperator) VectorMapOperator(org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator) AbstractMapOperator(org.apache.hadoop.hive.ql.exec.AbstractMapOperator) JobConf(org.apache.hadoop.mapred.JobConf) BaseWork(org.apache.hadoop.hive.ql.plan.BaseWork) MultiMRInput(org.apache.tez.mapreduce.input.MultiMRInput) ExecMapperContext(org.apache.hadoop.hive.ql.exec.mr.ExecMapperContext) LogicalOutput(org.apache.tez.runtime.api.LogicalOutput) HashTableDummyOperator(org.apache.hadoop.hive.ql.exec.HashTableDummyOperator) TezKVOutputCollector(org.apache.hadoop.hive.ql.exec.tez.TezProcessor.TezKVOutputCollector) MapWork(org.apache.hadoop.hive.ql.plan.MapWork) CompilationOpContext(org.apache.hadoop.hive.ql.CompilationOpContext) RegistryConfTez(org.apache.hadoop.hive.ql.exec.tez.DynamicValueRegistryTez.RegistryConfTez) PartitionDesc(org.apache.hadoop.hive.ql.plan.PartitionDesc) TableDesc(org.apache.hadoop.hive.ql.plan.TableDesc)

Aggregations

TezKVOutputCollector (org.apache.hadoop.hive.ql.exec.tez.TezProcessor.TezKVOutputCollector)2 LogicalOutput (org.apache.tez.runtime.api.LogicalOutput)2 ArrayList (java.util.ArrayList)1 Configuration (org.apache.hadoop.conf.Configuration)1 CompilationOpContext (org.apache.hadoop.hive.ql.CompilationOpContext)1 AbstractMapOperator (org.apache.hadoop.hive.ql.exec.AbstractMapOperator)1 DummyStoreOperator (org.apache.hadoop.hive.ql.exec.DummyStoreOperator)1 HashTableDummyOperator (org.apache.hadoop.hive.ql.exec.HashTableDummyOperator)1 MapOperator (org.apache.hadoop.hive.ql.exec.MapOperator)1 Operator (org.apache.hadoop.hive.ql.exec.Operator)1 TezDummyStoreOperator (org.apache.hadoop.hive.ql.exec.TezDummyStoreOperator)1 ExecMapperContext (org.apache.hadoop.hive.ql.exec.mr.ExecMapperContext)1 RegistryConfTez (org.apache.hadoop.hive.ql.exec.tez.DynamicValueRegistryTez.RegistryConfTez)1 VectorMapOperator (org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator)1 BaseWork (org.apache.hadoop.hive.ql.plan.BaseWork)1 MapWork (org.apache.hadoop.hive.ql.plan.MapWork)1 PartitionDesc (org.apache.hadoop.hive.ql.plan.PartitionDesc)1 TableDesc (org.apache.hadoop.hive.ql.plan.TableDesc)1 JobConf (org.apache.hadoop.mapred.JobConf)1 MultiMRInput (org.apache.tez.mapreduce.input.MultiMRInput)1