Search in sources :

Example 1 with JoinFlattenFunction

use of co.cask.cdap.etl.spark.function.JoinFlattenFunction in project cdap by caskdata.

the class SparkPipelineRunner method runPipeline.

public void runPipeline(PipelinePhase pipelinePhase, String sourcePluginType, JavaSparkExecutionContext sec, Map<String, Integer> stagePartitions, PipelinePluginContext pluginContext) throws Exception {
    MacroEvaluator macroEvaluator = new DefaultMacroEvaluator(sec.getWorkflowToken(), sec.getRuntimeArguments(), sec.getLogicalStartTime(), sec, sec.getNamespace());
    Map<String, SparkCollection<Object>> stageDataCollections = new HashMap<>();
    Map<String, SparkCollection<ErrorRecord<Object>>> stageErrorCollections = new HashMap<>();
    // should never happen, but removes warning
    if (pipelinePhase.getDag() == null) {
        throw new IllegalStateException("Pipeline phase has no connections.");
    }
    for (String stageName : pipelinePhase.getDag().getTopologicalOrder()) {
        StageInfo stageInfo = pipelinePhase.getStage(stageName);
        //noinspection ConstantConditions
        String pluginType = stageInfo.getPluginType();
        // don't want to do an additional filter for stages that can emit errors,
        // but aren't connected to an ErrorTransform
        boolean hasErrorOutput = false;
        Set<String> outputs = pipelinePhase.getStageOutputs(stageInfo.getName());
        for (String output : outputs) {
            //noinspection ConstantConditions
            if (ErrorTransform.PLUGIN_TYPE.equals(pipelinePhase.getStage(output).getPluginType())) {
                hasErrorOutput = true;
                break;
            }
        }
        SparkCollection<Object> stageData = null;
        Map<String, SparkCollection<Object>> inputDataCollections = new HashMap<>();
        Set<String> stageInputs = stageInfo.getInputs();
        for (String inputStageName : stageInputs) {
            inputDataCollections.put(inputStageName, stageDataCollections.get(inputStageName));
        }
        // initialize the stageRDD as the union of all input RDDs.
        if (!inputDataCollections.isEmpty()) {
            Iterator<SparkCollection<Object>> inputCollectionIter = inputDataCollections.values().iterator();
            stageData = inputCollectionIter.next();
            // don't union inputs records if we're joining or if we're processing errors
            while (!BatchJoiner.PLUGIN_TYPE.equals(pluginType) && !ErrorTransform.PLUGIN_TYPE.equals(pluginType) && inputCollectionIter.hasNext()) {
                stageData = stageData.union(inputCollectionIter.next());
            }
        }
        SparkCollection<ErrorRecord<Object>> stageErrors = null;
        PluginFunctionContext pluginFunctionContext = new PluginFunctionContext(stageInfo, sec);
        if (stageData == null) {
            // null in the other else-if conditions
            if (sourcePluginType.equals(pluginType)) {
                SparkCollection<Tuple2<Boolean, Object>> combinedData = getSource(stageInfo);
                if (hasErrorOutput) {
                    // need to cache, otherwise the stage can be computed twice, once for output and once for errors.
                    combinedData.cache();
                    stageErrors = combinedData.flatMap(stageInfo, Compat.convert(new OutputFilter<>()));
                }
                stageData = combinedData.flatMap(stageInfo, Compat.convert(new ErrorFilter<>()));
            } else {
                throw new IllegalStateException(String.format("Stage '%s' has no input and is not a source.", stageName));
            }
        } else if (BatchSink.PLUGIN_TYPE.equals(pluginType)) {
            stageData.store(stageInfo, Compat.convert(new BatchSinkFunction(pluginFunctionContext)));
        } else if (Transform.PLUGIN_TYPE.equals(pluginType)) {
            SparkCollection<Tuple2<Boolean, Object>> combinedData = stageData.transform(stageInfo);
            if (hasErrorOutput) {
                // need to cache, otherwise the stage can be computed twice, once for output and once for errors.
                combinedData.cache();
                stageErrors = combinedData.flatMap(stageInfo, Compat.convert(new OutputFilter<>()));
            }
            stageData = combinedData.flatMap(stageInfo, Compat.convert(new ErrorFilter<>()));
        } else if (ErrorTransform.PLUGIN_TYPE.equals(pluginType)) {
            // union all the errors coming into this stage
            SparkCollection<ErrorRecord<Object>> inputErrors = null;
            for (String inputStage : stageInputs) {
                SparkCollection<ErrorRecord<Object>> inputErrorsFromStage = stageErrorCollections.get(inputStage);
                if (inputErrorsFromStage == null) {
                    continue;
                }
                if (inputErrors == null) {
                    inputErrors = inputErrorsFromStage;
                } else {
                    inputErrors = inputErrors.union(inputErrorsFromStage);
                }
            }
            if (inputErrors != null) {
                SparkCollection<Tuple2<Boolean, Object>> combinedData = inputErrors.flatMap(stageInfo, Compat.convert(new ErrorTransformFunction<>(pluginFunctionContext)));
                if (hasErrorOutput) {
                    // need to cache, otherwise the stage can be computed twice, once for output and once for errors.
                    combinedData.cache();
                    stageErrors = combinedData.flatMap(stageInfo, Compat.convert(new OutputFilter<>()));
                }
                stageData = combinedData.flatMap(stageInfo, Compat.convert(new ErrorFilter<>()));
            }
        } else if (SparkCompute.PLUGIN_TYPE.equals(pluginType)) {
            SparkCompute<Object, Object> sparkCompute = pluginContext.newPluginInstance(stageName, macroEvaluator);
            stageData = stageData.compute(stageInfo, sparkCompute);
        } else if (SparkSink.PLUGIN_TYPE.equals(pluginType)) {
            SparkSink<Object> sparkSink = pluginContext.newPluginInstance(stageName, macroEvaluator);
            stageData.store(stageInfo, sparkSink);
        } else if (BatchAggregator.PLUGIN_TYPE.equals(pluginType)) {
            Integer partitions = stagePartitions.get(stageName);
            SparkCollection<Tuple2<Boolean, Object>> combinedData = stageData.aggregate(stageInfo, partitions);
            if (hasErrorOutput) {
                // need to cache, otherwise the stage can be computed twice, once for output and once for errors.
                combinedData.cache();
                stageErrors = combinedData.flatMap(stageInfo, Compat.convert(new OutputFilter<>()));
            }
            stageData = combinedData.flatMap(stageInfo, Compat.convert(new ErrorFilter<>()));
        } else if (BatchJoiner.PLUGIN_TYPE.equals(pluginType)) {
            BatchJoiner<Object, Object, Object> joiner = pluginContext.newPluginInstance(stageName, macroEvaluator);
            BatchJoinerRuntimeContext joinerRuntimeContext = pluginFunctionContext.createBatchRuntimeContext();
            joiner.initialize(joinerRuntimeContext);
            Map<String, SparkPairCollection<Object, Object>> preJoinStreams = new HashMap<>();
            for (Map.Entry<String, SparkCollection<Object>> inputStreamEntry : inputDataCollections.entrySet()) {
                String inputStage = inputStreamEntry.getKey();
                SparkCollection<Object> inputStream = inputStreamEntry.getValue();
                preJoinStreams.put(inputStage, addJoinKey(stageInfo, inputStage, inputStream));
            }
            Set<String> remainingInputs = new HashSet<>();
            remainingInputs.addAll(inputDataCollections.keySet());
            Integer numPartitions = stagePartitions.get(stageName);
            SparkPairCollection<Object, List<JoinElement<Object>>> joinedInputs = null;
            // inner join on required inputs
            for (final String inputStageName : joiner.getJoinConfig().getRequiredInputs()) {
                SparkPairCollection<Object, Object> preJoinCollection = preJoinStreams.get(inputStageName);
                if (joinedInputs == null) {
                    joinedInputs = preJoinCollection.mapValues(new InitialJoinFunction<>(inputStageName));
                } else {
                    JoinFlattenFunction<Object> joinFlattenFunction = new JoinFlattenFunction<>(inputStageName);
                    joinedInputs = numPartitions == null ? joinedInputs.join(preJoinCollection).mapValues(joinFlattenFunction) : joinedInputs.join(preJoinCollection, numPartitions).mapValues(joinFlattenFunction);
                }
                remainingInputs.remove(inputStageName);
            }
            // outer join on non-required inputs
            boolean isFullOuter = joinedInputs == null;
            for (final String inputStageName : remainingInputs) {
                SparkPairCollection<Object, Object> preJoinStream = preJoinStreams.get(inputStageName);
                if (joinedInputs == null) {
                    joinedInputs = preJoinStream.mapValues(new InitialJoinFunction<>(inputStageName));
                } else {
                    if (isFullOuter) {
                        OuterJoinFlattenFunction<Object> flattenFunction = new OuterJoinFlattenFunction<>(inputStageName);
                        joinedInputs = numPartitions == null ? joinedInputs.fullOuterJoin(preJoinStream).mapValues(flattenFunction) : joinedInputs.fullOuterJoin(preJoinStream, numPartitions).mapValues(flattenFunction);
                    } else {
                        LeftJoinFlattenFunction<Object> flattenFunction = new LeftJoinFlattenFunction<>(inputStageName);
                        joinedInputs = numPartitions == null ? joinedInputs.leftOuterJoin(preJoinStream).mapValues(flattenFunction) : joinedInputs.leftOuterJoin(preJoinStream, numPartitions).mapValues(flattenFunction);
                    }
                }
            }
            // should never happen, but removes warnings
            if (joinedInputs == null) {
                throw new IllegalStateException("There are no inputs into join stage " + stageName);
            }
            stageData = mergeJoinResults(stageInfo, joinedInputs).cache();
        } else if (Windower.PLUGIN_TYPE.equals(pluginType)) {
            Windower windower = pluginContext.newPluginInstance(stageName, macroEvaluator);
            stageData = stageData.window(stageInfo, windower);
        } else {
            throw new IllegalStateException(String.format("Stage %s is of unsupported plugin type %s.", stageName, pluginType));
        }
        if (shouldCache(pipelinePhase, stageInfo)) {
            stageData = stageData.cache();
            if (stageErrors != null) {
                stageErrors = stageErrors.cache();
            }
        }
        stageDataCollections.put(stageName, stageData);
        stageErrorCollections.put(stageName, stageErrors);
    }
}
Also used : OutputFilter(co.cask.cdap.etl.spark.function.OutputFilter) MacroEvaluator(co.cask.cdap.api.macro.MacroEvaluator) DefaultMacroEvaluator(co.cask.cdap.etl.common.DefaultMacroEvaluator) HashSet(java.util.HashSet) Set(java.util.Set) HashMap(java.util.HashMap) StageInfo(co.cask.cdap.etl.planner.StageInfo) PluginFunctionContext(co.cask.cdap.etl.spark.function.PluginFunctionContext) DefaultMacroEvaluator(co.cask.cdap.etl.common.DefaultMacroEvaluator) BatchJoinerRuntimeContext(co.cask.cdap.etl.api.batch.BatchJoinerRuntimeContext) Windower(co.cask.cdap.etl.api.streaming.Windower) ErrorFilter(co.cask.cdap.etl.spark.function.ErrorFilter) JoinFlattenFunction(co.cask.cdap.etl.spark.function.JoinFlattenFunction) OuterJoinFlattenFunction(co.cask.cdap.etl.spark.function.OuterJoinFlattenFunction) LeftJoinFlattenFunction(co.cask.cdap.etl.spark.function.LeftJoinFlattenFunction) BatchJoiner(co.cask.cdap.etl.api.batch.BatchJoiner) JoinElement(co.cask.cdap.etl.api.JoinElement) BatchSinkFunction(co.cask.cdap.etl.spark.function.BatchSinkFunction) SparkSink(co.cask.cdap.etl.api.batch.SparkSink) OuterJoinFlattenFunction(co.cask.cdap.etl.spark.function.OuterJoinFlattenFunction) LeftJoinFlattenFunction(co.cask.cdap.etl.spark.function.LeftJoinFlattenFunction) Tuple2(scala.Tuple2) HashMap(java.util.HashMap) Map(java.util.Map) ErrorRecord(co.cask.cdap.etl.api.ErrorRecord)

Aggregations

MacroEvaluator (co.cask.cdap.api.macro.MacroEvaluator)1 ErrorRecord (co.cask.cdap.etl.api.ErrorRecord)1 JoinElement (co.cask.cdap.etl.api.JoinElement)1 BatchJoiner (co.cask.cdap.etl.api.batch.BatchJoiner)1 BatchJoinerRuntimeContext (co.cask.cdap.etl.api.batch.BatchJoinerRuntimeContext)1 SparkSink (co.cask.cdap.etl.api.batch.SparkSink)1 Windower (co.cask.cdap.etl.api.streaming.Windower)1 DefaultMacroEvaluator (co.cask.cdap.etl.common.DefaultMacroEvaluator)1 StageInfo (co.cask.cdap.etl.planner.StageInfo)1 BatchSinkFunction (co.cask.cdap.etl.spark.function.BatchSinkFunction)1 ErrorFilter (co.cask.cdap.etl.spark.function.ErrorFilter)1 JoinFlattenFunction (co.cask.cdap.etl.spark.function.JoinFlattenFunction)1 LeftJoinFlattenFunction (co.cask.cdap.etl.spark.function.LeftJoinFlattenFunction)1 OuterJoinFlattenFunction (co.cask.cdap.etl.spark.function.OuterJoinFlattenFunction)1 OutputFilter (co.cask.cdap.etl.spark.function.OutputFilter)1 PluginFunctionContext (co.cask.cdap.etl.spark.function.PluginFunctionContext)1 HashMap (java.util.HashMap)1 HashSet (java.util.HashSet)1 Map (java.util.Map)1 Set (java.util.Set)1