Examples with RegistryConfTez org.apache.hadoop.hive.ql.exec.tez.DynamicValueRegistryTez.RegistryConfTez used on opensource projects

Example 1 with RegistryConfTez

use of org.apache.hadoop.hive.ql.exec.tez.DynamicValueRegistryTez.RegistryConfTez in project hive by apache.

the class ReduceRecordProcessor 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);
    MapredContext.init(false, new JobConf(jconf));
    List<LogicalInput> shuffleInputs = getShuffleInputs(inputs);
    // TODO HIVE-14042. Move to using a loop and a timed wait once TEZ-3302 is fixed.
    checkAbortCondition();
    if (shuffleInputs != null) {
        l4j.info("Waiting for ShuffleInputs to become ready");
        processorContext.waitForAllInputsReady(new ArrayList<Input>(shuffleInputs));
    }
    connectOps.clear();
    ReduceWork redWork = reduceWork;
    l4j.info("Main work is " + reduceWork.getName());
    List<HashTableDummyOperator> workOps = reduceWork.getDummyOps();
    HashSet<HashTableDummyOperator> dummyOps = workOps == null ? null : new HashSet<>(workOps);
    tagToReducerMap.put(redWork.getTag(), redWork);
    if (mergeWorkList != null) {
        for (BaseWork mergeWork : mergeWorkList) {
            if (l4j.isDebugEnabled()) {
                l4j.debug("Additional work " + mergeWork.getName());
            }
            workOps = mergeWork.getDummyOps();
            if (workOps != null) {
                if (dummyOps == null) {
                    dummyOps = new HashSet<>(workOps);
                } else {
                    dummyOps.addAll(workOps);
                }
            }
            ReduceWork mergeReduceWork = (ReduceWork) mergeWork;
            reducer = mergeReduceWork.getReducer();
            // Check immediately after reducer is assigned, in cae the abort came in during
            checkAbortCondition();
            DummyStoreOperator dummyStoreOp = getJoinParentOp(reducer);
            connectOps.put(mergeReduceWork.getTag(), dummyStoreOp);
            tagToReducerMap.put(mergeReduceWork.getTag(), mergeReduceWork);
        }
        ((TezContext) MapredContext.get()).setDummyOpsMap(connectOps);
    }
    checkAbortCondition();
    bigTablePosition = (byte) reduceWork.getTag();
    ObjectInspector[] mainWorkOIs = null;
    ((TezContext) MapredContext.get()).setInputs(inputs);
    ((TezContext) MapredContext.get()).setTezProcessorContext(processorContext);
    int numTags = reduceWork.getTagToValueDesc().size();
    reducer = reduceWork.getReducer();
    // Check immediately after reducer is assigned, in cae the abort came in during
    checkAbortCondition();
    // set memory available for operators
    long memoryAvailableToTask = processorContext.getTotalMemoryAvailableToTask();
    if (reducer.getConf() != null) {
        reducer.getConf().setMaxMemoryAvailable(memoryAvailableToTask);
        l4j.info("Memory available for operators set to {}", LlapUtil.humanReadableByteCount(memoryAvailableToTask));
    }
    OperatorUtils.setMemoryAvailable(reducer.getChildOperators(), memoryAvailableToTask);
    // Setup values registry
    String valueRegistryKey = DynamicValue.DYNAMIC_VALUE_REGISTRY_CACHE_KEY;
    DynamicValueRegistryTez registryTez = dynamicValueCache.retrieve(valueRegistryKey, new Callable<DynamicValueRegistryTez>() {

        @Override
        public DynamicValueRegistryTez call() {
            return new DynamicValueRegistryTez();
        }
    });
    dynamicValueCacheKeys.add(valueRegistryKey);
    RegistryConfTez registryConf = new RegistryConfTez(jconf, reduceWork, processorContext, inputs);
    registryTez.init(registryConf);
    checkAbortCondition();
    if (numTags > 1) {
        sources = new ReduceRecordSource[numTags];
        mainWorkOIs = new ObjectInspector[numTags];
        initializeMultipleSources(reduceWork, numTags, mainWorkOIs, sources);
        ((TezContext) MapredContext.get()).setRecordSources(sources);
        reducer.initialize(jconf, mainWorkOIs);
    } else {
        numTags = tagToReducerMap.keySet().size();
        sources = new ReduceRecordSource[numTags];
        mainWorkOIs = new ObjectInspector[numTags];
        for (int i : tagToReducerMap.keySet()) {
            redWork = tagToReducerMap.get(i);
            reducer = redWork.getReducer();
            // Check immediately after reducer is assigned, in cae the abort came in during
            checkAbortCondition();
            initializeSourceForTag(redWork, i, mainWorkOIs, sources, redWork.getTagToValueDesc().get(0), redWork.getTagToInput().get(0));
            reducer.initializeLocalWork(jconf);
        }
        reducer = reduceWork.getReducer();
        // Check immediately after reducer is assigned, in cae the abort came in during
        checkAbortCondition();
        ((TezContext) MapredContext.get()).setRecordSources(sources);
        reducer.initialize(jconf, new ObjectInspector[] { mainWorkOIs[bigTablePosition] });
        for (int i : tagToReducerMap.keySet()) {
            if (i == bigTablePosition) {
                continue;
            }
            redWork = tagToReducerMap.get(i);
            reducer = redWork.getReducer();
            // Check immediately after reducer is assigned, in cae the abort came in during
            checkAbortCondition();
            reducer.initialize(jconf, new ObjectInspector[] { mainWorkOIs[i] });
        }
    }
    checkAbortCondition();
    reducer = reduceWork.getReducer();
    // initialize reduce operator tree
    try {
        l4j.info(reducer.dump(0));
        // dummy parent operators as well.
        if (dummyOps != null) {
            for (HashTableDummyOperator dummyOp : dummyOps) {
                // TODO HIVE-14042. Propagating abort to dummyOps.
                dummyOp.initialize(jconf, null);
                checkAbortCondition();
            }
        }
        // set output collector for any reduce sink operators in the pipeline.
        List<Operator<?>> children = new LinkedList<Operator<?>>();
        children.add(reducer);
        if (dummyOps != null) {
            children.addAll(dummyOps);
        }
        createOutputMap();
        OperatorUtils.setChildrenCollector(children, outMap);
        checkAbortCondition();
        reducer.setReporter(reporter);
        MapredContext.get().setReporter(reporter);
    } catch (Throwable e) {
        super.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 ReduceRecordProcessor. Message={}", e.getMessage());
            throw (InterruptedException) e;
        } else {
            throw new RuntimeException("Reduce operator initialization failed", e);
        }
    }
    perfLogger.PerfLogEnd(CLASS_NAME, PerfLogger.TEZ_INIT_OPERATORS);
}

Example 2 with RegistryConfTez

use of org.apache.hadoop.hive.ql.exec.tez.DynamicValueRegistryTez.RegistryConfTez 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);
}