Examples with MultimapSideInputHandler org.apache.beam.runners.fnexecution.state.StateRequestHandlers.MultimapSideInputHandler used on opensource projects

Example 1 with MultimapSideInputHandler

use of org.apache.beam.runners.fnexecution.state.StateRequestHandlers.MultimapSideInputHandler in project beam by apache.

the class BatchSideInputHandlerFactory method forMultimapSideInput.

@Override
public <K, V, W extends BoundedWindow> MultimapSideInputHandler<K, V, W> forMultimapSideInput(String transformId, String sideInputId, KvCoder<K, V> elementCoder, Coder<W> windowCoder) {
    PCollectionNode collectionNode = sideInputToCollection.get(SideInputId.newBuilder().setTransformId(transformId).setLocalName(sideInputId).build());
    checkArgument(collectionNode != null, "No side input for %s/%s", transformId, sideInputId);
    Coder<K> keyCoder = elementCoder.getKeyCoder();
    Map<Object, Map<Object, KV<K, List<V>>>> /* structural key */
    data = new HashMap<>();
    List<WindowedValue<KV<K, V>>> broadcastVariable = sideInputGetter.getSideInput(collectionNode.getId());
    for (WindowedValue<KV<K, V>> windowedValue : broadcastVariable) {
        K key = windowedValue.getValue().getKey();
        V value = windowedValue.getValue().getValue();
        for (BoundedWindow boundedWindow : windowedValue.getWindows()) {
            @SuppressWarnings("unchecked") W window = (W) boundedWindow;
            Object structuralW = windowCoder.structuralValue(window);
            Object structuralK = keyCoder.structuralValue(key);
            KV<K, List<V>> records = data.computeIfAbsent(structuralW, o -> new HashMap<>()).computeIfAbsent(structuralK, o -> KV.of(key, new ArrayList<>()));
            records.getValue().add(value);
        }
    }
    return new MultimapSideInputHandler<K, V, W>() {

        @Override
        public Iterable<V> get(K key, W window) {
            KV<K, List<V>> records = data.getOrDefault(windowCoder.structuralValue(window), Collections.emptyMap()).get(keyCoder.structuralValue(key));
            if (records == null) {
                return Collections.emptyList();
            }
            return Collections.unmodifiableList(records.getValue());
        }

        @Override
        public Coder<V> valueCoder() {
            return elementCoder.getValueCoder();
        }

        @Override
        public Iterable<K> get(W window) {
            Map<Object, KV<K, List<V>>> records = data.getOrDefault(windowCoder.structuralValue(window), Collections.emptyMap());
            return Iterables.unmodifiableIterable(FluentIterable.concat(records.values()).transform(kListKV -> kListKV.getKey()));
        }

        @Override
        public Coder<K> keyCoder() {
            return elementCoder.getKeyCoder();
        }
    };
}

Example 2 with MultimapSideInputHandler

use of org.apache.beam.runners.fnexecution.state.StateRequestHandlers.MultimapSideInputHandler in project beam by apache.

the class StreamingSideInputHandlerFactory method forMultimapSideInput.

@Override
public <K, V, W extends BoundedWindow> MultimapSideInputHandler<K, V, W> forMultimapSideInput(String transformId, String sideInputId, KvCoder<K, V> elementCoder, Coder<W> windowCoder) {
    PCollectionView collectionNode = sideInputToCollection.get(SideInputId.newBuilder().setTransformId(transformId).setLocalName(sideInputId).build());
    checkArgument(collectionNode != null, "No side input for %s/%s", transformId, sideInputId);
    return new MultimapSideInputHandler<K, V, W>() {

        @Override
        public Iterable<V> get(K key, W window) {
            Iterable<KV<K, V>> values = (Iterable<KV<K, V>>) runnerHandler.getIterable(collectionNode, window);
            Object structuralK = keyCoder().structuralValue(key);
            ArrayList<V> result = new ArrayList<>();
            // find values for the given key
            for (KV<K, V> kv : values) {
                if (structuralK.equals(keyCoder().structuralValue(kv.getKey()))) {
                    result.add(kv.getValue());
                }
            }
            return Collections.unmodifiableList(result);
        }

        @Override
        public Iterable<K> get(W window) {
            Iterable<KV<K, V>> values = (Iterable<KV<K, V>>) runnerHandler.getIterable(collectionNode, window);
            Map<Object, K> result = new HashMap<>();
            // find all keys
            for (KV<K, V> kv : values) {
                result.putIfAbsent(keyCoder().structuralValue(kv.getKey()), kv.getKey());
            }
            return Collections.unmodifiableCollection(result.values());
        }

        @Override
        public Coder<K> keyCoder() {
            return elementCoder.getKeyCoder();
        }

        @Override
        public Coder<V> valueCoder() {
            return elementCoder.getValueCoder();
        }
    };
}

Example 3 with MultimapSideInputHandler

use of org.apache.beam.runners.fnexecution.state.StateRequestHandlers.MultimapSideInputHandler in project beam by apache.

the class RemoteExecutionTest method testExecutionWithSideInput.

@Test
public void testExecutionWithSideInput() throws Exception {
    launchSdkHarness(PipelineOptionsFactory.create());
    Pipeline p = Pipeline.create();
    addExperiment(p.getOptions().as(ExperimentalOptions.class), "beam_fn_api");
    // TODO(BEAM-10097): Remove experiment once all portable runners support this view type
    addExperiment(p.getOptions().as(ExperimentalOptions.class), "use_runner_v2");
    PCollection<String> input = p.apply("impulse", Impulse.create()).apply("create", ParDo.of(new DoFn<byte[], String>() {

        @ProcessElement
        public void process(ProcessContext ctxt) {
            ctxt.output("zero");
            ctxt.output("one");
            ctxt.output("two");
        }
    })).setCoder(StringUtf8Coder.of());
    PCollectionView<Iterable<String>> iterableView = input.apply("createIterableSideInput", View.asIterable());
    PCollectionView<Map<String, Iterable<String>>> multimapView = input.apply(WithKeys.of("key")).apply("createMultimapSideInput", View.asMultimap());
    input.apply("readSideInput", ParDo.of(new DoFn<String, KV<String, String>>() {

        @ProcessElement
        public void processElement(ProcessContext context) {
            for (String value : context.sideInput(iterableView)) {
                context.output(KV.of(context.element(), value));
            }
            for (Map.Entry<String, Iterable<String>> entry : context.sideInput(multimapView).entrySet()) {
                for (String value : entry.getValue()) {
                    context.output(KV.of(context.element(), entry.getKey() + ":" + value));
                }
            }
        }
    }).withSideInputs(iterableView, multimapView)).setCoder(KvCoder.of(StringUtf8Coder.of(), StringUtf8Coder.of())).apply("gbk", GroupByKey.create());
    RunnerApi.Pipeline pipelineProto = PipelineTranslation.toProto(p);
    FusedPipeline fused = GreedyPipelineFuser.fuse(pipelineProto);
    Optional<ExecutableStage> optionalStage = Iterables.tryFind(fused.getFusedStages(), (ExecutableStage stage) -> !stage.getSideInputs().isEmpty());
    checkState(optionalStage.isPresent(), "Expected a stage with side inputs.");
    ExecutableStage stage = optionalStage.get();
    ExecutableProcessBundleDescriptor descriptor = ProcessBundleDescriptors.fromExecutableStage("test_stage", stage, dataServer.getApiServiceDescriptor(), stateServer.getApiServiceDescriptor());
    BundleProcessor processor = controlClient.getProcessor(descriptor.getProcessBundleDescriptor(), descriptor.getRemoteInputDestinations(), stateDelegator);
    Map<String, Coder> remoteOutputCoders = descriptor.getRemoteOutputCoders();
    Map<String, Collection<WindowedValue<?>>> outputValues = new HashMap<>();
    Map<String, RemoteOutputReceiver<?>> outputReceivers = new HashMap<>();
    for (Entry<String, Coder> remoteOutputCoder : remoteOutputCoders.entrySet()) {
        List<WindowedValue<?>> outputContents = Collections.synchronizedList(new ArrayList<>());
        outputValues.put(remoteOutputCoder.getKey(), outputContents);
        outputReceivers.put(remoteOutputCoder.getKey(), RemoteOutputReceiver.of((Coder<WindowedValue<?>>) remoteOutputCoder.getValue(), outputContents::add));
    }
    StateRequestHandler stateRequestHandler = StateRequestHandlers.forSideInputHandlerFactory(descriptor.getSideInputSpecs(), new SideInputHandlerFactory() {

        @Override
        public <V, W extends BoundedWindow> IterableSideInputHandler<V, W> forIterableSideInput(String pTransformId, String sideInputId, Coder<V> elementCoder, Coder<W> windowCoder) {
            return new IterableSideInputHandler<V, W>() {

                @Override
                public Iterable<V> get(W window) {
                    return (Iterable) Arrays.asList("A", "B", "C");
                }

                @Override
                public Coder<V> elementCoder() {
                    return elementCoder;
                }
            };
        }

        @Override
        public <K, V, W extends BoundedWindow> MultimapSideInputHandler<K, V, W> forMultimapSideInput(String pTransformId, String sideInputId, KvCoder<K, V> elementCoder, Coder<W> windowCoder) {
            return new MultimapSideInputHandler<K, V, W>() {

                @Override
                public Iterable<K> get(W window) {
                    return (Iterable) Arrays.asList("key1", "key2");
                }

                @Override
                public Iterable<V> get(K key, W window) {
                    if ("key1".equals(key)) {
                        return (Iterable) Arrays.asList("H", "I", "J");
                    } else if ("key2".equals(key)) {
                        return (Iterable) Arrays.asList("M", "N", "O");
                    }
                    return Collections.emptyList();
                }

                @Override
                public Coder<K> keyCoder() {
                    return elementCoder.getKeyCoder();
                }

                @Override
                public Coder<V> valueCoder() {
                    return elementCoder.getValueCoder();
                }
            };
        }
    });
    BundleProgressHandler progressHandler = BundleProgressHandler.ignored();
    try (RemoteBundle bundle = processor.newBundle(outputReceivers, stateRequestHandler, progressHandler)) {
        Iterables.getOnlyElement(bundle.getInputReceivers().values()).accept(valueInGlobalWindow("X"));
        Iterables.getOnlyElement(bundle.getInputReceivers().values()).accept(valueInGlobalWindow("Y"));
    }
    for (Collection<WindowedValue<?>> windowedValues : outputValues.values()) {
        assertThat(windowedValues, containsInAnyOrder(valueInGlobalWindow(KV.of("X", "A")), valueInGlobalWindow(KV.of("X", "B")), valueInGlobalWindow(KV.of("X", "C")), valueInGlobalWindow(KV.of("X", "key1:H")), valueInGlobalWindow(KV.of("X", "key1:I")), valueInGlobalWindow(KV.of("X", "key1:J")), valueInGlobalWindow(KV.of("X", "key2:M")), valueInGlobalWindow(KV.of("X", "key2:N")), valueInGlobalWindow(KV.of("X", "key2:O")), valueInGlobalWindow(KV.of("Y", "A")), valueInGlobalWindow(KV.of("Y", "B")), valueInGlobalWindow(KV.of("Y", "C")), valueInGlobalWindow(KV.of("Y", "key1:H")), valueInGlobalWindow(KV.of("Y", "key1:I")), valueInGlobalWindow(KV.of("Y", "key1:J")), valueInGlobalWindow(KV.of("Y", "key2:M")), valueInGlobalWindow(KV.of("Y", "key2:N")), valueInGlobalWindow(KV.of("Y", "key2:O"))));
    }
}

Example 4 with MultimapSideInputHandler

use of org.apache.beam.runners.fnexecution.state.StateRequestHandlers.MultimapSideInputHandler in project beam by apache.

the class RemoteExecutionTest method testExecutionWithSideInputCaching.

@Test
public void testExecutionWithSideInputCaching() throws Exception {
    Pipeline p = Pipeline.create();
    addExperiment(p.getOptions().as(ExperimentalOptions.class), "beam_fn_api");
    // TODO(BEAM-10097): Remove experiment once all portable runners support this view type
    addExperiment(p.getOptions().as(ExperimentalOptions.class), "use_runner_v2");
    launchSdkHarness(p.getOptions());
    PCollection<String> input = p.apply("impulse", Impulse.create()).apply("create", ParDo.of(new DoFn<byte[], String>() {

        @ProcessElement
        public void process(ProcessContext ctxt) {
            ctxt.output("zero");
            ctxt.output("one");
            ctxt.output("two");
        }
    })).setCoder(StringUtf8Coder.of());
    PCollectionView<Iterable<String>> iterableView = input.apply("createIterableSideInput", View.asIterable());
    PCollectionView<Map<String, Iterable<String>>> multimapView = input.apply(WithKeys.of("key")).apply("createMultimapSideInput", View.asMultimap());
    input.apply("readSideInput", ParDo.of(new DoFn<String, KV<String, String>>() {

        @ProcessElement
        public void processElement(ProcessContext context) {
            for (String value : context.sideInput(iterableView)) {
                context.output(KV.of(context.element(), value));
            }
            for (Map.Entry<String, Iterable<String>> entry : context.sideInput(multimapView).entrySet()) {
                for (String value : entry.getValue()) {
                    context.output(KV.of(context.element(), entry.getKey() + ":" + value));
                }
            }
        }
    }).withSideInputs(iterableView, multimapView)).setCoder(KvCoder.of(StringUtf8Coder.of(), StringUtf8Coder.of())).apply("gbk", GroupByKey.create());
    RunnerApi.Pipeline pipelineProto = PipelineTranslation.toProto(p);
    FusedPipeline fused = GreedyPipelineFuser.fuse(pipelineProto);
    Optional<ExecutableStage> optionalStage = Iterables.tryFind(fused.getFusedStages(), (ExecutableStage stage) -> !stage.getSideInputs().isEmpty());
    checkState(optionalStage.isPresent(), "Expected a stage with side inputs.");
    ExecutableStage stage = optionalStage.get();
    ExecutableProcessBundleDescriptor descriptor = ProcessBundleDescriptors.fromExecutableStage("test_stage", stage, dataServer.getApiServiceDescriptor(), stateServer.getApiServiceDescriptor());
    BundleProcessor processor = controlClient.getProcessor(descriptor.getProcessBundleDescriptor(), descriptor.getRemoteInputDestinations(), stateDelegator);
    Map<String, Coder> remoteOutputCoders = descriptor.getRemoteOutputCoders();
    Map<String, Collection<WindowedValue<?>>> outputValues = new HashMap<>();
    Map<String, RemoteOutputReceiver<?>> outputReceivers = new HashMap<>();
    for (Entry<String, Coder> remoteOutputCoder : remoteOutputCoders.entrySet()) {
        List<WindowedValue<?>> outputContents = Collections.synchronizedList(new ArrayList<>());
        outputValues.put(remoteOutputCoder.getKey(), outputContents);
        outputReceivers.put(remoteOutputCoder.getKey(), RemoteOutputReceiver.of((Coder<WindowedValue<?>>) remoteOutputCoder.getValue(), outputContents::add));
    }
    StoringStateRequestHandler stateRequestHandler = new StoringStateRequestHandler(StateRequestHandlers.forSideInputHandlerFactory(descriptor.getSideInputSpecs(), new SideInputHandlerFactory() {

        @Override
        public <V, W extends BoundedWindow> IterableSideInputHandler<V, W> forIterableSideInput(String pTransformId, String sideInputId, Coder<V> elementCoder, Coder<W> windowCoder) {
            return new IterableSideInputHandler<V, W>() {

                @Override
                public Iterable<V> get(W window) {
                    return (Iterable) Arrays.asList("A", "B", "C");
                }

                @Override
                public Coder<V> elementCoder() {
                    return elementCoder;
                }
            };
        }

        @Override
        public <K, V, W extends BoundedWindow> MultimapSideInputHandler<K, V, W> forMultimapSideInput(String pTransformId, String sideInputId, KvCoder<K, V> elementCoder, Coder<W> windowCoder) {
            return new MultimapSideInputHandler<K, V, W>() {

                @Override
                public Iterable<K> get(W window) {
                    return (Iterable) Arrays.asList("key1", "key2");
                }

                @Override
                public Iterable<V> get(K key, W window) {
                    if ("key1".equals(key)) {
                        return (Iterable) Arrays.asList("H", "I", "J");
                    } else if ("key2".equals(key)) {
                        return (Iterable) Arrays.asList("M", "N", "O");
                    }
                    return Collections.emptyList();
                }

                @Override
                public Coder<K> keyCoder() {
                    return elementCoder.getKeyCoder();
                }

                @Override
                public Coder<V> valueCoder() {
                    return elementCoder.getValueCoder();
                }
            };
        }
    }));
    String transformId = Iterables.get(stage.getSideInputs(), 0).transform().getId();
    stateRequestHandler.addCacheToken(BeamFnApi.ProcessBundleRequest.CacheToken.newBuilder().setSideInput(BeamFnApi.ProcessBundleRequest.CacheToken.SideInput.newBuilder().setSideInputId(iterableView.getTagInternal().getId()).setTransformId(transformId).build()).setToken(ByteString.copyFromUtf8("IterableSideInputToken")).build());
    stateRequestHandler.addCacheToken(BeamFnApi.ProcessBundleRequest.CacheToken.newBuilder().setSideInput(BeamFnApi.ProcessBundleRequest.CacheToken.SideInput.newBuilder().setSideInputId(multimapView.getTagInternal().getId()).setTransformId(transformId).build()).setToken(ByteString.copyFromUtf8("MulitmapSideInputToken")).build());
    BundleProgressHandler progressHandler = BundleProgressHandler.ignored();
    try (RemoteBundle bundle = processor.newBundle(outputReceivers, stateRequestHandler, progressHandler)) {
        Iterables.getOnlyElement(bundle.getInputReceivers().values()).accept(valueInGlobalWindow("X"));
    }
    try (RemoteBundle bundle = processor.newBundle(outputReceivers, stateRequestHandler, progressHandler)) {
        Iterables.getOnlyElement(bundle.getInputReceivers().values()).accept(valueInGlobalWindow("Y"));
    }
    for (Collection<WindowedValue<?>> windowedValues : outputValues.values()) {
        assertThat(windowedValues, containsInAnyOrder(valueInGlobalWindow(KV.of("X", "A")), valueInGlobalWindow(KV.of("X", "B")), valueInGlobalWindow(KV.of("X", "C")), valueInGlobalWindow(KV.of("X", "key1:H")), valueInGlobalWindow(KV.of("X", "key1:I")), valueInGlobalWindow(KV.of("X", "key1:J")), valueInGlobalWindow(KV.of("X", "key2:M")), valueInGlobalWindow(KV.of("X", "key2:N")), valueInGlobalWindow(KV.of("X", "key2:O")), valueInGlobalWindow(KV.of("Y", "A")), valueInGlobalWindow(KV.of("Y", "B")), valueInGlobalWindow(KV.of("Y", "C")), valueInGlobalWindow(KV.of("Y", "key1:H")), valueInGlobalWindow(KV.of("Y", "key1:I")), valueInGlobalWindow(KV.of("Y", "key1:J")), valueInGlobalWindow(KV.of("Y", "key2:M")), valueInGlobalWindow(KV.of("Y", "key2:N")), valueInGlobalWindow(KV.of("Y", "key2:O"))));
    }
    // Expect the following requests for the first bundle:
    // * one to read iterable side input
    // * one to read keys from multimap side input
    // * one to read key1 iterable from multimap side input
    // * one to read key2 iterable from multimap side input
    assertEquals(4, stateRequestHandler.receivedRequests.size());
    assertEquals(stateRequestHandler.receivedRequests.get(0).getStateKey().getIterableSideInput(), BeamFnApi.StateKey.IterableSideInput.newBuilder().setSideInputId(iterableView.getTagInternal().getId()).setTransformId(transformId).build());
    assertEquals(stateRequestHandler.receivedRequests.get(1).getStateKey().getMultimapKeysSideInput(), BeamFnApi.StateKey.MultimapKeysSideInput.newBuilder().setSideInputId(multimapView.getTagInternal().getId()).setTransformId(transformId).build());
    assertEquals(stateRequestHandler.receivedRequests.get(2).getStateKey().getMultimapSideInput(), BeamFnApi.StateKey.MultimapSideInput.newBuilder().setSideInputId(multimapView.getTagInternal().getId()).setTransformId(transformId).setKey(encode("key1")).build());
    assertEquals(stateRequestHandler.receivedRequests.get(3).getStateKey().getMultimapSideInput(), BeamFnApi.StateKey.MultimapSideInput.newBuilder().setSideInputId(multimapView.getTagInternal().getId()).setTransformId(transformId).setKey(encode("key2")).build());
}