Search in sources :

Example 1 with VectorMapOperator

use of org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator in project hive by apache.

the class SparkMapRecordHandler method init.

@Override
public <K, V> void init(JobConf job, OutputCollector<K, V> output, Reporter reporter) throws Exception {
    perfLogger.PerfLogBegin(CLASS_NAME, PerfLogger.SPARK_INIT_OPERATORS);
    super.init(job, output, reporter);
    isLogInfoEnabled = LOG.isInfoEnabled();
    try {
        jc = job;
        execContext = new ExecMapperContext(jc);
        // create map and fetch operators
        MapWork mrwork = Utilities.getMapWork(job);
        CompilationOpContext runtimeCtx = new CompilationOpContext();
        if (mrwork.getVectorMode()) {
            mo = new VectorMapOperator(runtimeCtx);
        } else {
            mo = new MapOperator(runtimeCtx);
        }
        mo.setConf(mrwork);
        // initialize map operator
        mo.initialize(jc, null);
        mo.setChildren(job);
        LOG.info(mo.dump(0));
        // initialize map local work
        localWork = mrwork.getMapRedLocalWork();
        execContext.setLocalWork(localWork);
        MapredContext.init(true, new JobConf(jc));
        MapredContext.get().setReporter(reporter);
        mo.passExecContext(execContext);
        mo.initializeLocalWork(jc);
        mo.initializeMapOperator(jc);
        OperatorUtils.setChildrenCollector(mo.getChildOperators(), output);
        mo.setReporter(rp);
        if (localWork == null) {
            return;
        }
        //The following code is for mapjoin
        //initialize all the dummy ops
        LOG.info("Initializing dummy operator");
        List<Operator<? extends OperatorDesc>> dummyOps = localWork.getDummyParentOp();
        for (Operator<? extends OperatorDesc> dummyOp : dummyOps) {
            dummyOp.setExecContext(execContext);
            dummyOp.initialize(jc, null);
        }
    } catch (Throwable e) {
        abort = true;
        if (e instanceof OutOfMemoryError) {
            // Don't create a new object if we are already out of memory
            throw (OutOfMemoryError) e;
        } else {
            throw new RuntimeException("Map operator initialization failed: " + e, e);
        }
    }
    perfLogger.PerfLogEnd(CLASS_NAME, PerfLogger.SPARK_INIT_OPERATORS);
}
Also used : Operator(org.apache.hadoop.hive.ql.exec.Operator) MapOperator(org.apache.hadoop.hive.ql.exec.MapOperator) AbstractMapOperator(org.apache.hadoop.hive.ql.exec.AbstractMapOperator) VectorMapOperator(org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator) ExecMapperContext(org.apache.hadoop.hive.ql.exec.mr.ExecMapperContext) VectorMapOperator(org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator) MapOperator(org.apache.hadoop.hive.ql.exec.MapOperator) AbstractMapOperator(org.apache.hadoop.hive.ql.exec.AbstractMapOperator) VectorMapOperator(org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator) MapWork(org.apache.hadoop.hive.ql.plan.MapWork) CompilationOpContext(org.apache.hadoop.hive.ql.CompilationOpContext) JobConf(org.apache.hadoop.mapred.JobConf) OperatorDesc(org.apache.hadoop.hive.ql.plan.OperatorDesc)

Example 2 with VectorMapOperator

use of org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator in project hive by apache.

the class ExecMapper method configure.

@Override
public void configure(JobConf job) {
    execContext = new ExecMapperContext(job);
    // Allocate the bean at the beginning -
    try {
        l4j.info("conf classpath = " + Arrays.asList(((URLClassLoader) job.getClassLoader()).getURLs()));
        l4j.info("thread classpath = " + Arrays.asList(((URLClassLoader) Thread.currentThread().getContextClassLoader()).getURLs()));
    } catch (Exception e) {
        l4j.info("cannot get classpath: " + e.getMessage());
    }
    setDone(false);
    try {
        jc = job;
        execContext.setJc(jc);
        // create map and fetch operators
        MapWork mrwork = Utilities.getMapWork(job);
        CompilationOpContext runtimeCtx = new CompilationOpContext();
        if (mrwork.getVectorMode()) {
            mo = new VectorMapOperator(runtimeCtx);
        } else {
            mo = new MapOperator(runtimeCtx);
        }
        mo.setConf(mrwork);
        // initialize map operator
        mo.initialize(job, null);
        mo.setChildren(job);
        l4j.info(mo.dump(0));
        // initialize map local work
        localWork = mrwork.getMapRedLocalWork();
        execContext.setLocalWork(localWork);
        MapredContext.init(true, new JobConf(jc));
        mo.passExecContext(execContext);
        mo.initializeLocalWork(jc);
        mo.initializeMapOperator(jc);
        if (localWork == null) {
            return;
        }
        //The following code is for mapjoin
        //initialize all the dummy ops
        l4j.info("Initializing dummy operator");
        List<Operator<? extends OperatorDesc>> dummyOps = localWork.getDummyParentOp();
        for (Operator<? extends OperatorDesc> dummyOp : dummyOps) {
            dummyOp.passExecContext(execContext);
            dummyOp.initialize(jc, null);
        }
    } catch (Throwable e) {
        abort = true;
        if (e instanceof OutOfMemoryError) {
            // Don't create a new object if we are already out of memory
            throw (OutOfMemoryError) e;
        } else {
            throw new RuntimeException("Map operator initialization failed", e);
        }
    }
}
Also used : 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) AbstractMapOperator(org.apache.hadoop.hive.ql.exec.AbstractMapOperator) VectorMapOperator(org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator) IOException(java.io.IOException) 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) MapWork(org.apache.hadoop.hive.ql.plan.MapWork) CompilationOpContext(org.apache.hadoop.hive.ql.CompilationOpContext) JobConf(org.apache.hadoop.mapred.JobConf) OperatorDesc(org.apache.hadoop.hive.ql.plan.OperatorDesc)

Example 3 with VectorMapOperator

use of org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator 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
    mapWork = (MapWork) cache.retrieve(key, new Callable<Object>() {

        @Override
        public Object call() {
            return Utilities.getMapWork(jconf);
        }
    });
    // TODO HIVE-14042. Cleanup may be required if exiting early.
    Utilities.setMapWork(jconf, mapWork);
    String prefixes = jconf.get(DagUtils.TEZ_MERGE_WORK_FILE_PREFIXES);
    if (prefixes != null) {
        mergeWorkList = new ArrayList<MapWork>();
        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, new Callable<Object>() {

                @Override
                public Object call() {
                    return 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()) {
        l4j.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);
                    l4j.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) {
                        l4j.info("Multi MR Input for work " + mergeMapWork.getName() + " is null. Skipping read.");
                        skipRead = true;
                    } else {
                        Collection<KeyValueReader> keyValueReaders = multiMRInput.getKeyValueReaders();
                        if ((keyValueReaders == null) || (keyValueReaders.isEmpty())) {
                            l4j.info("Key value readers are null or empty and hence skipping read. " + "KeyValueReaders = " + keyValueReaders);
                            skipRead = true;
                        }
                    }
                    if (skipRead) {
                        List<Operator<?>> children = new ArrayList<Operator<?>>();
                        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);
        l4j.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);
        l4j.info(mapOp.dump(0));
        // set memory available for operators
        long memoryAvailableToTask = processorContext.getTotalMemoryAvailableToTask();
        if (mapOp.getConf() != null) {
            mapOp.getConf().setMaxMemoryAvailable(memoryAvailableToTask);
            l4j.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 Callable<DynamicValueRegistryTez>() {

            @Override
            public DynamicValueRegistryTez call() {
                return 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() == false) {
            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) {
            l4j.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)

Aggregations

CompilationOpContext (org.apache.hadoop.hive.ql.CompilationOpContext)3 AbstractMapOperator (org.apache.hadoop.hive.ql.exec.AbstractMapOperator)3 MapOperator (org.apache.hadoop.hive.ql.exec.MapOperator)3 Operator (org.apache.hadoop.hive.ql.exec.Operator)3 VectorMapOperator (org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator)3 MapWork (org.apache.hadoop.hive.ql.plan.MapWork)3 JobConf (org.apache.hadoop.mapred.JobConf)3 ExecMapperContext (org.apache.hadoop.hive.ql.exec.mr.ExecMapperContext)2 OperatorDesc (org.apache.hadoop.hive.ql.plan.OperatorDesc)2 IOException (java.io.IOException)1 ArrayList (java.util.ArrayList)1 Configuration (org.apache.hadoop.conf.Configuration)1 DummyStoreOperator (org.apache.hadoop.hive.ql.exec.DummyStoreOperator)1 HashTableDummyOperator (org.apache.hadoop.hive.ql.exec.HashTableDummyOperator)1 TezDummyStoreOperator (org.apache.hadoop.hive.ql.exec.TezDummyStoreOperator)1 RegistryConfTez (org.apache.hadoop.hive.ql.exec.tez.DynamicValueRegistryTez.RegistryConfTez)1 TezKVOutputCollector (org.apache.hadoop.hive.ql.exec.tez.TezProcessor.TezKVOutputCollector)1 BaseWork (org.apache.hadoop.hive.ql.plan.BaseWork)1 MultiMRInput (org.apache.tez.mapreduce.input.MultiMRInput)1 LogicalOutput (org.apache.tez.runtime.api.LogicalOutput)1