Examples with OuterJoinFlattenFunction co.cask.cdap.etl.spark.function.OuterJoinFlattenFunction used on opensource projects

Example 1 with OuterJoinFlattenFunction

use of co.cask.cdap.etl.spark.function.OuterJoinFlattenFunction 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);
    }
}

Example 2 with OuterJoinFlattenFunction

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

the class SparkPipelineRunner method runPipeline.

public void runPipeline(PipelinePhase pipelinePhase, String sourcePluginType, JavaSparkExecutionContext sec, Map<String, Integer> stagePartitions, PluginContext pluginContext, Map<String, StageStatisticsCollector> collectors) throws Exception {
    MacroEvaluator macroEvaluator = new DefaultMacroEvaluator(new BasicArguments(sec), sec.getLogicalStartTime(), sec, sec.getNamespace());
    Map<String, EmittedRecords> emittedRecords = 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()) {
        StageSpec stageSpec = pipelinePhase.getStage(stageName);
        // noinspection ConstantConditions
        String pluginType = stageSpec.getPluginType();
        EmittedRecords.Builder emittedBuilder = EmittedRecords.builder();
        // don't want to do an additional filter for stages that can emit errors,
        // but aren't connected to an ErrorTransform
        // similarly, don't want to do an additional filter for alerts when the stage isn't connected to
        // an AlertPublisher
        boolean hasErrorOutput = false;
        boolean hasAlertOutput = false;
        Set<String> outputs = pipelinePhase.getStageOutputs(stageSpec.getName());
        for (String output : outputs) {
            String outputPluginType = pipelinePhase.getStage(output).getPluginType();
            // noinspection ConstantConditions
            if (ErrorTransform.PLUGIN_TYPE.equals(outputPluginType)) {
                hasErrorOutput = true;
            } else if (AlertPublisher.PLUGIN_TYPE.equals(outputPluginType)) {
                hasAlertOutput = true;
            }
        }
        SparkCollection<Object> stageData = null;
        Map<String, SparkCollection<Object>> inputDataCollections = new HashMap<>();
        Set<String> stageInputs = pipelinePhase.getStageInputs(stageName);
        for (String inputStageName : stageInputs) {
            StageSpec inputStageSpec = pipelinePhase.getStage(inputStageName);
            if (inputStageSpec == null) {
                // means the input to this stage is in a separate phase. For example, it is an action.
                continue;
            }
            String port = null;
            // not errors or alerts or output port records
            if (!Constants.Connector.PLUGIN_TYPE.equals(inputStageSpec.getPluginType()) && !Constants.Connector.PLUGIN_TYPE.equals(pluginType)) {
                port = inputStageSpec.getOutputPorts().get(stageName).getPort();
            }
            SparkCollection<Object> inputRecords = port == null ? emittedRecords.get(inputStageName).outputRecords : emittedRecords.get(inputStageName).outputPortRecords.get(port);
            inputDataCollections.put(inputStageName, inputRecords);
        }
        // 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());
            }
        }
        boolean isConnectorSource = Constants.Connector.PLUGIN_TYPE.equals(pluginType) && pipelinePhase.getSources().contains(stageName);
        boolean isConnectorSink = Constants.Connector.PLUGIN_TYPE.equals(pluginType) && pipelinePhase.getSinks().contains(stageName);
        StageStatisticsCollector collector = collectors.get(stageName) == null ? new NoopStageStatisticsCollector() : collectors.get(stageName);
        PluginFunctionContext pluginFunctionContext = new PluginFunctionContext(stageSpec, sec, collector);
        if (stageData == null) {
            // null in the other else-if conditions
            if (sourcePluginType.equals(pluginType) || isConnectorSource) {
                SparkCollection<RecordInfo<Object>> combinedData = getSource(stageSpec, collector);
                emittedBuilder = addEmitted(emittedBuilder, pipelinePhase, stageSpec, combinedData, hasErrorOutput, hasAlertOutput);
            } else {
                throw new IllegalStateException(String.format("Stage '%s' has no input and is not a source.", stageName));
            }
        } else if (BatchSink.PLUGIN_TYPE.equals(pluginType) || isConnectorSink) {
            stageData.store(stageSpec, Compat.convert(new BatchSinkFunction(pluginFunctionContext)));
        } else if (Transform.PLUGIN_TYPE.equals(pluginType)) {
            SparkCollection<RecordInfo<Object>> combinedData = stageData.transform(stageSpec, collector);
            emittedBuilder = addEmitted(emittedBuilder, pipelinePhase, stageSpec, combinedData, hasErrorOutput, hasAlertOutput);
        } else if (SplitterTransform.PLUGIN_TYPE.equals(pluginType)) {
            SparkCollection<RecordInfo<Object>> combinedData = stageData.multiOutputTransform(stageSpec, collector);
            emittedBuilder = addEmitted(emittedBuilder, pipelinePhase, stageSpec, combinedData, hasErrorOutput, hasAlertOutput);
        } 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 = emittedRecords.get(inputStage).errorRecords;
                if (inputErrorsFromStage == null) {
                    continue;
                }
                if (inputErrors == null) {
                    inputErrors = inputErrorsFromStage;
                } else {
                    inputErrors = inputErrors.union(inputErrorsFromStage);
                }
            }
            if (inputErrors != null) {
                SparkCollection<RecordInfo<Object>> combinedData = inputErrors.flatMap(stageSpec, Compat.convert(new ErrorTransformFunction<>(pluginFunctionContext)));
                emittedBuilder = addEmitted(emittedBuilder, pipelinePhase, stageSpec, combinedData, hasErrorOutput, hasAlertOutput);
            }
        } else if (SparkCompute.PLUGIN_TYPE.equals(pluginType)) {
            SparkCompute<Object, Object> sparkCompute = pluginContext.newPluginInstance(stageName, macroEvaluator);
            emittedBuilder = emittedBuilder.setOutput(stageData.compute(stageSpec, sparkCompute));
        } else if (SparkSink.PLUGIN_TYPE.equals(pluginType)) {
            SparkSink<Object> sparkSink = pluginContext.newPluginInstance(stageName, macroEvaluator);
            stageData.store(stageSpec, sparkSink);
        } else if (BatchAggregator.PLUGIN_TYPE.equals(pluginType)) {
            Integer partitions = stagePartitions.get(stageName);
            SparkCollection<RecordInfo<Object>> combinedData = stageData.aggregate(stageSpec, partitions, collector);
            emittedBuilder = addEmitted(emittedBuilder, pipelinePhase, stageSpec, combinedData, hasErrorOutput, hasAlertOutput);
        } 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(stageSpec, inputStage, inputStream, collector));
            }
            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);
            }
            emittedBuilder = emittedBuilder.setOutput(mergeJoinResults(stageSpec, joinedInputs, collector).cache());
        } else if (Windower.PLUGIN_TYPE.equals(pluginType)) {
            Windower windower = pluginContext.newPluginInstance(stageName, macroEvaluator);
            emittedBuilder = emittedBuilder.setOutput(stageData.window(stageSpec, windower));
        } else if (AlertPublisher.PLUGIN_TYPE.equals(pluginType)) {
            // union all the alerts coming into this stage
            SparkCollection<Alert> inputAlerts = null;
            for (String inputStage : stageInputs) {
                SparkCollection<Alert> inputErrorsFromStage = emittedRecords.get(inputStage).alertRecords;
                if (inputErrorsFromStage == null) {
                    continue;
                }
                if (inputAlerts == null) {
                    inputAlerts = inputErrorsFromStage;
                } else {
                    inputAlerts = inputAlerts.union(inputErrorsFromStage);
                }
            }
            if (inputAlerts != null) {
                inputAlerts.publishAlerts(stageSpec, collector);
            }
        } else {
            throw new IllegalStateException(String.format("Stage %s is of unsupported plugin type %s.", stageName, pluginType));
        }
        emittedRecords.put(stageName, emittedBuilder.build());
    }
}