Examples with UnsupportedSideInputReader org.apache.beam.runners.core.UnsupportedSideInputReader used on opensource projects

Example 1 with UnsupportedSideInputReader

use of org.apache.beam.runners.core.UnsupportedSideInputReader in project beam by apache.

the class SparkGroupAlsoByWindowViaWindowSet method groupAlsoByWindow.

public static <K, InputT, W extends BoundedWindow> JavaDStream<WindowedValue<KV<K, Iterable<InputT>>>> groupAlsoByWindow(JavaDStream<WindowedValue<KV<K, Iterable<WindowedValue<InputT>>>>> inputDStream, final Coder<K> keyCoder, final Coder<WindowedValue<InputT>> wvCoder, final WindowingStrategy<?, W> windowingStrategy, final SparkRuntimeContext runtimeContext, final List<Integer> sourceIds) {
    final IterableCoder<WindowedValue<InputT>> itrWvCoder = IterableCoder.of(wvCoder);
    final Coder<InputT> iCoder = ((FullWindowedValueCoder<InputT>) wvCoder).getValueCoder();
    final Coder<? extends BoundedWindow> wCoder = ((FullWindowedValueCoder<InputT>) wvCoder).getWindowCoder();
    final Coder<WindowedValue<KV<K, Iterable<InputT>>>> wvKvIterCoder = FullWindowedValueCoder.of(KvCoder.of(keyCoder, IterableCoder.of(iCoder)), wCoder);
    final TimerInternals.TimerDataCoder timerDataCoder = TimerInternals.TimerDataCoder.of(windowingStrategy.getWindowFn().windowCoder());
    long checkpointDurationMillis = runtimeContext.getPipelineOptions().as(SparkPipelineOptions.class).getCheckpointDurationMillis();
    // we have to switch to Scala API to avoid Optional in the Java API, see: SPARK-4819.
    // we also have a broader API for Scala (access to the actual key and entire iterator).
    // we use coders to convert objects in the PCollection to byte arrays, so they
    // can be transferred over the network for the shuffle and be in serialized form
    // for checkpointing.
    // for readability, we add comments with actual type next to byte[].
    // to shorten line length, we use:
    //---- WV: WindowedValue
    //---- Iterable: Itr
    //---- AccumT: A
    //---- InputT: I
    DStream<Tuple2<ByteArray, byte[]>> /*Itr<WV<I>>*/
    pairDStream = inputDStream.transformToPair(new Function<JavaRDD<WindowedValue<KV<K, Iterable<WindowedValue<InputT>>>>>, JavaPairRDD<ByteArray, byte[]>>() {

        // we use mapPartitions with the RDD API because its the only available API
        // that allows to preserve partitioning.
        @Override
        public JavaPairRDD<ByteArray, byte[]> call(JavaRDD<WindowedValue<KV<K, Iterable<WindowedValue<InputT>>>>> rdd) throws Exception {
            return rdd.mapPartitions(TranslationUtils.functionToFlatMapFunction(WindowingHelpers.<KV<K, Iterable<WindowedValue<InputT>>>>unwindowFunction()), true).mapPartitionsToPair(TranslationUtils.<K, Iterable<WindowedValue<InputT>>>toPairFlatMapFunction(), true).mapPartitionsToPair(TranslationUtils.pairFunctionToPairFlatMapFunction(CoderHelpers.toByteFunction(keyCoder, itrWvCoder)), true);
        }
    }).dstream();
    PairDStreamFunctions<ByteArray, byte[]> pairDStreamFunctions = DStream.toPairDStreamFunctions(pairDStream, JavaSparkContext$.MODULE$.<ByteArray>fakeClassTag(), JavaSparkContext$.MODULE$.<byte[]>fakeClassTag(), null);
    int defaultNumPartitions = pairDStreamFunctions.defaultPartitioner$default$1();
    Partitioner partitioner = pairDStreamFunctions.defaultPartitioner(defaultNumPartitions);
    // use updateStateByKey to scan through the state and update elements and timers.
    DStream<Tuple2<ByteArray, Tuple2<StateAndTimers, List<byte[]>>>> /*WV<KV<K, Itr<I>>>*/
    firedStream = pairDStreamFunctions.updateStateByKey(new SerializableFunction1<scala.collection.Iterator<Tuple3</*K*/
    ByteArray, Seq<byte[]>, Option<Tuple2<StateAndTimers, /*WV<KV<K, Itr<I>>>*/
    List<byte[]>>>>>, scala.collection.Iterator<Tuple2</*K*/
    ByteArray, Tuple2<StateAndTimers, /*WV<KV<K, Itr<I>>>*/
    List<byte[]>>>>>() {

        @Override
        public scala.collection.Iterator<Tuple2</*K*/
        ByteArray, Tuple2<StateAndTimers, /*WV<KV<K, Itr<I>>>*/
        List<byte[]>>>> apply(final scala.collection.Iterator<Tuple3</*K*/
        ByteArray, Seq<byte[]>, Option<Tuple2<StateAndTimers, /*WV<KV<K, Itr<I>>>*/
        List<byte[]>>>>> iter) {
            //--- ACTUAL STATEFUL OPERATION:
            //
            // Input Iterator: the partition (~bundle) of a cogrouping of the input
            // and the previous state (if exists).
            //
            // Output Iterator: the output key, and the updated state.
            //
            // possible input scenarios for (K, Seq, Option<S>):
            // (1) Option<S>.isEmpty: new data with no previous state.
            // (2) Seq.isEmpty: no new data, but evaluating previous state (timer-like behaviour).
            // (3) Seq.nonEmpty && Option<S>.isDefined: new data with previous state.
            final SystemReduceFn<K, InputT, Iterable<InputT>, Iterable<InputT>, W> reduceFn = SystemReduceFn.buffering(((FullWindowedValueCoder<InputT>) wvCoder).getValueCoder());
            final OutputWindowedValueHolder<K, InputT> outputHolder = new OutputWindowedValueHolder<>();
            // use in memory Aggregators since Spark Accumulators are not resilient
            // in stateful operators, once done with this partition.
            final MetricsContainerImpl cellProvider = new MetricsContainerImpl("cellProvider");
            final CounterCell droppedDueToClosedWindow = cellProvider.getCounter(MetricName.named(SparkGroupAlsoByWindowViaWindowSet.class, GroupAlsoByWindowsAggregators.DROPPED_DUE_TO_CLOSED_WINDOW_COUNTER));
            final CounterCell droppedDueToLateness = cellProvider.getCounter(MetricName.named(SparkGroupAlsoByWindowViaWindowSet.class, GroupAlsoByWindowsAggregators.DROPPED_DUE_TO_LATENESS_COUNTER));
            AbstractIterator<Tuple2<ByteArray, Tuple2<StateAndTimers, List<byte[]>>>> /*WV<KV<K, Itr<I>>>*/
            outIter = new AbstractIterator<Tuple2</*K*/
            ByteArray, Tuple2<StateAndTimers, /*WV<KV<K, Itr<I>>>*/
            List<byte[]>>>>() {

                @Override
                protected Tuple2</*K*/
                ByteArray, Tuple2<StateAndTimers, /*WV<KV<K, Itr<I>>>*/
                List<byte[]>>> computeNext() {
                    // (possibly) previous-state and (possibly) new data.
                    while (iter.hasNext()) {
                        // for each element in the partition:
                        Tuple3<ByteArray, Seq<byte[]>, Option<Tuple2<StateAndTimers, List<byte[]>>>> next = iter.next();
                        ByteArray encodedKey = next._1();
                        K key = CoderHelpers.fromByteArray(encodedKey.getValue(), keyCoder);
                        Seq<byte[]> seq = next._2();
                        Option<Tuple2<StateAndTimers, List<byte[]>>> prevStateAndTimersOpt = next._3();
                        SparkStateInternals<K> stateInternals;
                        SparkTimerInternals timerInternals = SparkTimerInternals.forStreamFromSources(sourceIds, GlobalWatermarkHolder.get());
                        // get state(internals) per key.
                        if (prevStateAndTimersOpt.isEmpty()) {
                            // no previous state.
                            stateInternals = SparkStateInternals.forKey(key);
                        } else {
                            // with pre-existing state.
                            StateAndTimers prevStateAndTimers = prevStateAndTimersOpt.get()._1();
                            stateInternals = SparkStateInternals.forKeyAndState(key, prevStateAndTimers.getState());
                            Collection<byte[]> serTimers = prevStateAndTimers.getTimers();
                            timerInternals.addTimers(SparkTimerInternals.deserializeTimers(serTimers, timerDataCoder));
                        }
                        ReduceFnRunner<K, InputT, Iterable<InputT>, W> reduceFnRunner = new ReduceFnRunner<>(key, windowingStrategy, ExecutableTriggerStateMachine.create(TriggerStateMachines.stateMachineForTrigger(TriggerTranslation.toProto(windowingStrategy.getTrigger()))), stateInternals, timerInternals, outputHolder, new UnsupportedSideInputReader("GroupAlsoByWindow"), reduceFn, runtimeContext.getPipelineOptions());
                        // clear before potential use.
                        outputHolder.clear();
                        if (!seq.isEmpty()) {
                            // new input for key.
                            try {
                                Iterable<WindowedValue<InputT>> elementsIterable = CoderHelpers.fromByteArray(seq.head(), itrWvCoder);
                                Iterable<WindowedValue<InputT>> validElements = LateDataUtils.dropExpiredWindows(key, elementsIterable, timerInternals, windowingStrategy, droppedDueToLateness);
                                reduceFnRunner.processElements(validElements);
                            } catch (Exception e) {
                                throw new RuntimeException("Failed to process element with ReduceFnRunner", e);
                            }
                        } else if (stateInternals.getState().isEmpty()) {
                            // no input and no state -> GC evict now.
                            continue;
                        }
                        try {
                            // advance the watermark to HWM to fire by timers.
                            timerInternals.advanceWatermark();
                            // call on timers that are ready.
                            reduceFnRunner.onTimers(timerInternals.getTimersReadyToProcess());
                        } catch (Exception e) {
                            throw new RuntimeException("Failed to process ReduceFnRunner onTimer.", e);
                        }
                        // this is mostly symbolic since actual persist is done by emitting output.
                        reduceFnRunner.persist();
                        // obtain output, if fired.
                        List<WindowedValue<KV<K, Iterable<InputT>>>> outputs = outputHolder.get();
                        if (!outputs.isEmpty() || !stateInternals.getState().isEmpty()) {
                            StateAndTimers updated = new StateAndTimers(stateInternals.getState(), SparkTimerInternals.serializeTimers(timerInternals.getTimers(), timerDataCoder));
                            // persist Spark's state by outputting.
                            List<byte[]> serOutput = CoderHelpers.toByteArrays(outputs, wvKvIterCoder);
                            return new Tuple2<>(encodedKey, new Tuple2<>(updated, serOutput));
                        }
                    // an empty state with no output, can be evicted completely - do nothing.
                    }
                    return endOfData();
                }
            };
            // log if there's something to log.
            long lateDropped = droppedDueToLateness.getCumulative();
            if (lateDropped > 0) {
                LOG.info(String.format("Dropped %d elements due to lateness.", lateDropped));
                droppedDueToLateness.inc(-droppedDueToLateness.getCumulative());
            }
            long closedWindowDropped = droppedDueToClosedWindow.getCumulative();
            if (closedWindowDropped > 0) {
                LOG.info(String.format("Dropped %d elements due to closed window.", closedWindowDropped));
                droppedDueToClosedWindow.inc(-droppedDueToClosedWindow.getCumulative());
            }
            return scala.collection.JavaConversions.asScalaIterator(outIter);
        }
    }, partitioner, true, JavaSparkContext$.MODULE$.<Tuple2<StateAndTimers, List<byte[]>>>fakeClassTag());
    if (checkpointDurationMillis > 0) {
        firedStream.checkpoint(new Duration(checkpointDurationMillis));
    }
    // go back to Java now.
    JavaPairDStream<ByteArray, Tuple2<StateAndTimers, List<byte[]>>> /*WV<KV<K, Itr<I>>>*/
    javaFiredStream = JavaPairDStream.fromPairDStream(firedStream, JavaSparkContext$.MODULE$.<ByteArray>fakeClassTag(), JavaSparkContext$.MODULE$.<Tuple2<StateAndTimers, List<byte[]>>>fakeClassTag());
    // filter state-only output (nothing to fire) and remove the state from the output.
    return javaFiredStream.filter(new Function<Tuple2</*K*/
    ByteArray, Tuple2<StateAndTimers, /*WV<KV<K, Itr<I>>>*/
    List<byte[]>>>, Boolean>() {

        @Override
        public Boolean call(Tuple2</*K*/
        ByteArray, Tuple2<StateAndTimers, /*WV<KV<K, Itr<I>>>*/
        List<byte[]>>> t2) throws Exception {
            // filter output if defined.
            return !t2._2()._2().isEmpty();
        }
    }).flatMap(new FlatMapFunction<Tuple2</*K*/
    ByteArray, Tuple2<StateAndTimers, /*WV<KV<K, Itr<I>>>*/
    List<byte[]>>>, WindowedValue<KV<K, Iterable<InputT>>>>() {

        @Override
        public Iterable<WindowedValue<KV<K, Iterable<InputT>>>> call(Tuple2</*K*/
        ByteArray, Tuple2<StateAndTimers, /*WV<KV<K, Itr<I>>>*/
        List<byte[]>>> t2) throws Exception {
            // return in serialized form.
            return CoderHelpers.fromByteArrays(t2._2()._2(), wvKvIterCoder);
        }
    });
}

Example 2 with UnsupportedSideInputReader

use of org.apache.beam.runners.core.UnsupportedSideInputReader in project beam by apache.

the class SparkGroupAlsoByWindowViaOutputBufferFn method call.

@Override
public Iterable<WindowedValue<KV<K, Iterable<InputT>>>> call(WindowedValue<KV<K, Iterable<WindowedValue<InputT>>>> windowedValue) throws Exception {
    K key = windowedValue.getValue().getKey();
    Iterable<WindowedValue<InputT>> values = windowedValue.getValue().getValue();
    //------ based on GroupAlsoByWindowsViaOutputBufferDoFn ------//
    // Used with Batch, we know that all the data is available for this key. We can't use the
    // timer manager from the context because it doesn't exist. So we create one and emulate the
    // watermark, knowing that we have all data and it is in timestamp order.
    InMemoryTimerInternals timerInternals = new InMemoryTimerInternals();
    timerInternals.advanceProcessingTime(Instant.now());
    timerInternals.advanceSynchronizedProcessingTime(Instant.now());
    StateInternals stateInternals = stateInternalsFactory.stateInternalsForKey(key);
    GABWOutputWindowedValue<K, InputT> outputter = new GABWOutputWindowedValue<>();
    ReduceFnRunner<K, InputT, Iterable<InputT>, W> reduceFnRunner = new ReduceFnRunner<>(key, windowingStrategy, ExecutableTriggerStateMachine.create(TriggerStateMachines.stateMachineForTrigger(TriggerTranslation.toProto(windowingStrategy.getTrigger()))), stateInternals, timerInternals, outputter, new UnsupportedSideInputReader("GroupAlsoByWindow"), reduceFn, runtimeContext.getPipelineOptions());
    // Process the grouped values.
    reduceFnRunner.processElements(values);
    // Finish any pending windows by advancing the input watermark to infinity.
    timerInternals.advanceInputWatermark(BoundedWindow.TIMESTAMP_MAX_VALUE);
    // Finally, advance the processing time to infinity to fire any timers.
    timerInternals.advanceProcessingTime(BoundedWindow.TIMESTAMP_MAX_VALUE);
    timerInternals.advanceSynchronizedProcessingTime(BoundedWindow.TIMESTAMP_MAX_VALUE);
    fireEligibleTimers(timerInternals, reduceFnRunner);
    reduceFnRunner.persist();
    return outputter.getOutputs();
}

Example 3 with UnsupportedSideInputReader

use of org.apache.beam.runners.core.UnsupportedSideInputReader in project beam by apache.

the class GroupAlsoByWindowViaOutputBufferFn method call.

@Override
public Iterator<WindowedValue<KV<K, Iterable<InputT>>>> call(K key, Iterator<WindowedValue<KV<K, InputT>>> iterator) throws Exception {
    // we have to materialize the Iterator because ReduceFnRunner.processElements expects
    // to have all elements to merge the windows between each other.
    // possible OOM even though the spark framework spills to disk if a given group is too large to
    // fit in memory.
    ArrayList<WindowedValue<InputT>> values = new ArrayList<>();
    while (iterator.hasNext()) {
        WindowedValue<KV<K, InputT>> wv = iterator.next();
        values.add(wv.withValue(wv.getValue().getValue()));
    }
    // ------ based on GroupAlsoByWindowsViaOutputBufferDoFn ------//
    // Used with Batch, we know that all the data is available for this key. We can't use the
    // timer manager from the context because it doesn't exist. So we create one and emulate the
    // watermark, knowing that we have all data and it is in timestamp order.
    InMemoryTimerInternals timerInternals = new InMemoryTimerInternals();
    timerInternals.advanceProcessingTime(Instant.now());
    timerInternals.advanceSynchronizedProcessingTime(Instant.now());
    StateInternals stateInternals = stateInternalsFactory.stateInternalsForKey(key);
    GABWOutputWindowedValue<K, InputT> outputter = new GABWOutputWindowedValue<>();
    ReduceFnRunner<K, InputT, Iterable<InputT>, W> reduceFnRunner = new ReduceFnRunner<>(key, windowingStrategy, ExecutableTriggerStateMachine.create(TriggerStateMachines.stateMachineForTrigger(TriggerTranslation.toProto(windowingStrategy.getTrigger()))), stateInternals, timerInternals, outputter, new UnsupportedSideInputReader("GroupAlsoByWindow"), reduceFn, options.get());
    // Process the grouped values.
    reduceFnRunner.processElements(values);
    // Finish any pending windows by advancing the input watermark to infinity.
    timerInternals.advanceInputWatermark(BoundedWindow.TIMESTAMP_MAX_VALUE);
    // Finally, advance the processing time to infinity to fire any timers.
    timerInternals.advanceProcessingTime(BoundedWindow.TIMESTAMP_MAX_VALUE);
    timerInternals.advanceSynchronizedProcessingTime(BoundedWindow.TIMESTAMP_MAX_VALUE);
    fireEligibleTimers(timerInternals, reduceFnRunner);
    reduceFnRunner.persist();
    return outputter.getOutputs().iterator();
}

Example 4 with UnsupportedSideInputReader

use of org.apache.beam.runners.core.UnsupportedSideInputReader in project beam by apache.

the class SparkGroupAlsoByWindowViaOutputBufferFn method call.

@Override
public Iterator<WindowedValue<KV<K, Iterable<InputT>>>> call(KV<K, Iterable<WindowedValue<InputT>>> kv) throws Exception {
    K key = kv.getKey();
    Iterable<WindowedValue<InputT>> values = kv.getValue();
    // ------ based on GroupAlsoByWindowsViaOutputBufferDoFn ------//
    // Used with Batch, we know that all the data is available for this key. We can't use the
    // timer manager from the context because it doesn't exist. So we create one and emulate the
    // watermark, knowing that we have all data and it is in timestamp order.
    InMemoryTimerInternals timerInternals = new InMemoryTimerInternals();
    timerInternals.advanceProcessingTime(Instant.now());
    timerInternals.advanceSynchronizedProcessingTime(Instant.now());
    StateInternals stateInternals = stateInternalsFactory.stateInternalsForKey(key);
    GABWOutputWindowedValue<K, InputT> outputter = new GABWOutputWindowedValue<>();
    ReduceFnRunner<K, InputT, Iterable<InputT>, W> reduceFnRunner = new ReduceFnRunner<>(key, windowingStrategy, ExecutableTriggerStateMachine.create(TriggerStateMachines.stateMachineForTrigger(TriggerTranslation.toProto(windowingStrategy.getTrigger()))), stateInternals, timerInternals, outputter, new UnsupportedSideInputReader("GroupAlsoByWindow"), reduceFn, options.get());
    // Process the grouped values.
    reduceFnRunner.processElements(values);
    // Finish any pending windows by advancing the input watermark to infinity.
    timerInternals.advanceInputWatermark(BoundedWindow.TIMESTAMP_MAX_VALUE);
    // Finally, advance the processing time to infinity to fire any timers.
    timerInternals.advanceProcessingTime(BoundedWindow.TIMESTAMP_MAX_VALUE);
    timerInternals.advanceSynchronizedProcessingTime(BoundedWindow.TIMESTAMP_MAX_VALUE);
    fireEligibleTimers(timerInternals, reduceFnRunner);
    reduceFnRunner.persist();
    return outputter.getOutputs().iterator();
}

Example 5 with UnsupportedSideInputReader

use of org.apache.beam.runners.core.UnsupportedSideInputReader in project beam by apache.

the class GroupByWindowFunction method flatMap.

@Override
public void flatMap(KV<K, Iterable<WindowedValue<V>>> kIteratorKV, RecordCollector<WindowedValue<KV<K, Iterable<V>>>> collector) {
    try {
        K key = kIteratorKV.getKey();
        Iterable<WindowedValue<V>> values = kIteratorKV.getValue();
        InMemoryTimerInternals timerInternals = new InMemoryTimerInternals();
        timerInternals.advanceProcessingTime(Instant.now());
        timerInternals.advanceSynchronizedProcessingTime(Instant.now());
        StateInternals stateInternals = InMemoryStateInternals.forKey(key);
        GABWOutputWindowedValue<K, V> outputter = new GABWOutputWindowedValue<>();
        ReduceFnRunner<K, V, Iterable<V>, W> reduceFnRunner = new ReduceFnRunner<>(key, windowingStrategy, ExecutableTriggerStateMachine.create(TriggerStateMachines.stateMachineForTrigger(TriggerTranslation.toProto(windowingStrategy.getTrigger()))), stateInternals, timerInternals, outputter, new UnsupportedSideInputReader("GroupAlsoByWindow"), reduceFn, null);
        // Process the grouped values.
        reduceFnRunner.processElements(values);
        // Finish any pending windows by advancing the input watermark to infinity.
        timerInternals.advanceInputWatermark(BoundedWindow.TIMESTAMP_MAX_VALUE);
        // Finally, advance the processing time to infinity to fire any timers.
        timerInternals.advanceProcessingTime(BoundedWindow.TIMESTAMP_MAX_VALUE);
        timerInternals.advanceSynchronizedProcessingTime(BoundedWindow.TIMESTAMP_MAX_VALUE);
        fireEligibleTimers(timerInternals, reduceFnRunner);
        reduceFnRunner.persist();
        Iterator<WindowedValue<KV<K, Iterable<V>>>> outputs = outputter.getOutputs().iterator();
        while (outputs.hasNext()) {
            collector.collect(outputs.next());
        }
    } catch (Exception e) {
        LOG.info(e.getMessage());
    }
}