本文主要研究一下FluxSink的机制

FluxSink

reactor-core-3.1.3.RELEASE-sources.jar!/reactor/core/publisher/FluxSink.java

/**
 * Wrapper API around a downstream Subscriber for emitting any number of
 * next signals followed by zero or one onError/onComplete.
 * <p>
 * @param <T> the value type
 */
public interface FluxSink<T> {

    /**
     * @see Subscriber#onComplete()
     */
    void complete();

    /**
     * Return the current subscriber {@link Context}.
     * <p>
     *   {@link Context} can be enriched via {@link Flux#subscriberContext(Function)}
     *   operator or directly by a child subscriber overriding
     *   {@link CoreSubscriber#currentContext()}
     *
     * @return the current subscriber {@link Context}.
     */
    Context currentContext();

    /**
     * @see Subscriber#onError(Throwable)
     * @param e the exception to signal, not null
     */
    void error(Throwable e);

    /**
     * Try emitting, might throw an unchecked exception.
     * @see Subscriber#onNext(Object)
     * @param t the value to emit, not null
     */
    FluxSink<T> next(T t);

    /**
     * The current outstanding request amount.
     * @return the current outstanding request amount
     */
    long requestedFromDownstream();

    /**
     * Returns true if the downstream cancelled the sequence.
     * @return true if the downstream cancelled the sequence
     */
    boolean isCancelled();

    /**
     * Attaches a {@link LongConsumer} to this {@link FluxSink} that will be notified of
     * any request to this sink.
     * <p>
     * For push/pull sinks created using {@link Flux#create(java.util.function.Consumer)}
     * or {@link Flux#create(java.util.function.Consumer, FluxSink.OverflowStrategy)},
     * the consumer
     * is invoked for every request to enable a hybrid backpressure-enabled push/pull model.
     * When bridging with asynchronous listener-based APIs, the {@code onRequest} callback
     * may be used to request more data from source if required and to manage backpressure
     * by delivering data to sink only when requests are pending.
     * <p>
     * For push-only sinks created using {@link Flux#push(java.util.function.Consumer)}
     * or {@link Flux#push(java.util.function.Consumer, FluxSink.OverflowStrategy)},
     * the consumer is invoked with an initial request of {@code Long.MAX_VALUE} when this method
     * is invoked.
     *
     * @param consumer the consumer to invoke on each request
     * @return {@link FluxSink} with a consumer that is notified of requests
     */
    FluxSink<T> onRequest(LongConsumer consumer);

    /**
     * Associates a disposable resource with this FluxSink
     * that will be disposed in case the downstream cancels the sequence
     * via {@link org.reactivestreams.Subscription#cancel()}.
     * @param d the disposable callback to use
     * @return the {@link FluxSink} with resource to be disposed on cancel signal
     */
    FluxSink<T> onCancel(Disposable d);

    /**
     * Associates a disposable resource with this FluxSink
     * that will be disposed on the first terminate signal which may be
     * a cancel, complete or error signal.
     * @param d the disposable callback to use
     * @return the {@link FluxSink} with resource to be disposed on first terminate signal
     */
    FluxSink<T> onDispose(Disposable d);

    /**
     * Enumeration for backpressure handling.
     */
    enum OverflowStrategy {
        /**
         * Completely ignore downstream backpressure requests.
         * <p>
         * This may yield {@link IllegalStateException} when queues get full downstream.
         */
        IGNORE,
        /**
         * Signal an {@link IllegalStateException} when the downstream can't keep up
         */
        ERROR,
        /**
         * Drop the incoming signal if the downstream is not ready to receive it.
         */
        DROP,
        /**
         * Downstream will get only the latest signals from upstream.
         */
        LATEST,
        /**
         * Buffer all signals if the downstream can't keep up.
         * <p>
         * Warning! This does unbounded buffering and may lead to {@link OutOfMemoryError}.
         */
        BUFFER
    }
}
注意OverflowStrategy.BUFFER使用的是一个无界队列,需要额外注意OOM问题

实例

    public static void main(String[] args) throws InterruptedException {
        final Flux<Integer> flux = Flux.<Integer> create(fluxSink -> {
            //NOTE sink:class reactor.core.publisher.FluxCreate$SerializedSink
            LOGGER.info("sink:{}",fluxSink.getClass());
            while (true) {
                LOGGER.info("sink next");
                fluxSink.next(ThreadLocalRandom.current().nextInt());
            }
        }, FluxSink.OverflowStrategy.BUFFER);

        //NOTE flux:class reactor.core.publisher.FluxCreate,prefetch:-1
        LOGGER.info("flux:{},prefetch:{}",flux.getClass(),flux.getPrefetch());

        flux.subscribe(e -> {
            LOGGER.info("subscribe:{}",e);
            try {
                TimeUnit.SECONDS.sleep(10);
            } catch (InterruptedException e1) {
                e1.printStackTrace();
            }
        });

        TimeUnit.MINUTES.sleep(20);
    }
这里create创建的是reactor.core.publisher.FluxCreate,而其sink是reactor.core.publisher.FluxCreate$SerializedSink

Flux.subscribe

reactor-core-3.1.3.RELEASE-sources.jar!/reactor/core/publisher/Flux.java

    /**
     * Subscribe {@link Consumer} to this {@link Flux} that will respectively consume all the
     * elements in the sequence, handle errors, react to completion, and request upon subscription.
     * It will let the provided {@link Subscription subscriptionConsumer}
     * request the adequate amount of data, or request unbounded demand
     * {@code Long.MAX_VALUE} if no such consumer is provided.
     * <p>
     * For a passive version that observe and forward incoming data see {@link #doOnNext(java.util.function.Consumer)},
     * {@link #doOnError(java.util.function.Consumer)}, {@link #doOnComplete(Runnable)}
     * and {@link #doOnSubscribe(Consumer)}.
     * <p>For a version that gives you more control over backpressure and the request, see
     * {@link #subscribe(Subscriber)} with a {@link BaseSubscriber}.
     * <p>
     * Keep in mind that since the sequence can be asynchronous, this will immediately
     * return control to the calling thread. This can give the impression the consumer is
     * not invoked when executing in a main thread or a unit test for instance.
     *
     * <p>
     * <img class="marble" src="https://raw.githubusercontent.com/reactor/reactor-core/v3.1.3.RELEASE/src/docs/marble/subscribecomplete.png" alt="">
     *
     * @param consumer the consumer to invoke on each value
     * @param errorConsumer the consumer to invoke on error signal
     * @param completeConsumer the consumer to invoke on complete signal
     * @param subscriptionConsumer the consumer to invoke on subscribe signal, to be used
     * for the initial {@link Subscription#request(long) request}, or null for max request
     *
     * @return a new {@link Disposable} that can be used to cancel the underlying {@link Subscription}
     */
    public final Disposable subscribe(
            @Nullable Consumer<? super T> consumer,
            @Nullable Consumer<? super Throwable> errorConsumer,
            @Nullable Runnable completeConsumer,
            @Nullable Consumer<? super Subscription> subscriptionConsumer) {
        return subscribeWith(new LambdaSubscriber<>(consumer, errorConsumer,
                completeConsumer,
                subscriptionConsumer));
    }

    @Override
    public final void subscribe(Subscriber<? super T> actual) {
        onLastAssembly(this).subscribe(Operators.toCoreSubscriber(actual));
    }
创建的是LambdaSubscriber,最后调用FluxCreate.subscribe

FluxCreate.subscribe

reactor-core-3.1.3.RELEASE-sources.jar!/reactor/core/publisher/FluxCreate.java

    public void subscribe(CoreSubscriber<? super T> actual) {
        BaseSink<T> sink = createSink(actual, backpressure);

        actual.onSubscribe(sink);
        try {
            source.accept(
                    createMode == CreateMode.PUSH_PULL ? new SerializedSink<>(sink) :
                            sink);
        }
        catch (Throwable ex) {
            Exceptions.throwIfFatal(ex);
            sink.error(Operators.onOperatorError(ex, actual.currentContext()));
        }
    }
    static <T> BaseSink<T> createSink(CoreSubscriber<? super T> t,
            OverflowStrategy backpressure) {
        switch (backpressure) {
            case IGNORE: {
                return new IgnoreSink<>(t);
            }
            case ERROR: {
                return new ErrorAsyncSink<>(t);
            }
            case DROP: {
                return new DropAsyncSink<>(t);
            }
            case LATEST: {
                return new LatestAsyncSink<>(t);
            }
            default: {
                return new BufferAsyncSink<>(t, Queues.SMALL_BUFFER_SIZE);
            }
        }
    }    
先创建sink,这里创建的是BufferAsyncSink,然后调用LambdaSubscriber.onSubscribe
然后再调用source.accept,也就是调用fluxSink的lambda方法产生数据,开启stream模式

LambdaSubscriber.onSubscribe

reactor-core-3.1.3.RELEASE-sources.jar!/reactor/core/publisher/LambdaSubscriber.java

    public final void onSubscribe(Subscription s) {
        if (Operators.validate(subscription, s)) {
            this.subscription = s;
            if (subscriptionConsumer != null) {
                try {
                    subscriptionConsumer.accept(s);
                }
                catch (Throwable t) {
                    Exceptions.throwIfFatal(t);
                    s.cancel();
                    onError(t);
                }
            }
            else {
                s.request(Long.MAX_VALUE);
            }
        }
    }
这里又调用了BufferAsyncSink的request(Long.MAX_VALUE),实际是调用BaseSink的request
        public final void request(long n) {
            if (Operators.validate(n)) {
                Operators.addCap(REQUESTED, this, n);

                LongConsumer consumer = requestConsumer;
                if (n > 0 && consumer != null && !isCancelled()) {
                    consumer.accept(n);
                }
                onRequestedFromDownstream();
            }
        }
这里的onRequestedFromDownstream调用了BufferAsyncSink的onRequestedFromDownstream
        @Override
        void onRequestedFromDownstream() {
            drain();
        }
调用的是BufferAsyncSink的drain

BufferAsyncSink.drain

        void drain() {
            if (WIP.getAndIncrement(this) != 0) {
                return;
            }

            int missed = 1;
            final Subscriber<? super T> a = actual;
            final Queue<T> q = queue;

            for (; ; ) {
                long r = requested;
                long e = 0L;

                while (e != r) {
                    if (isCancelled()) {
                        q.clear();
                        return;
                    }

                    boolean d = done;

                    T o = q.poll();

                    boolean empty = o == null;

                    if (d && empty) {
                        Throwable ex = error;
                        if (ex != null) {
                            super.error(ex);
                        }
                        else {
                            super.complete();
                        }
                        return;
                    }

                    if (empty) {
                        break;
                    }

                    a.onNext(o);

                    e++;
                }

                if (e == r) {
                    if (isCancelled()) {
                        q.clear();
                        return;
                    }

                    boolean d = done;

                    boolean empty = q.isEmpty();

                    if (d && empty) {
                        Throwable ex = error;
                        if (ex != null) {
                            super.error(ex);
                        }
                        else {
                            super.complete();
                        }
                        return;
                    }
                }

                if (e != 0) {
                    Operators.produced(REQUESTED, this, e);
                }

                missed = WIP.addAndGet(this, -missed);
                if (missed == 0) {
                    break;
                }
            }
        }
这里的queue是创建BufferAsyncSink指定的,默认是Queues.SMALL_BUFFER_SIZE(Math.max(16,Integer.parseInt(System.getProperty("reactor.bufferSize.small", "256"))))
而这里的onNext则是同步调用LambdaSubscriber的consumer

FluxCreate.subscribe#source.accept

source.accept(
                    createMode == CreateMode.PUSH_PULL ? new SerializedSink<>(sink) :
                            sink);
CreateMode.PUSH_PULL这里对sink包装为SerializedSink,然后调用Flux.create自定义的lambda consumer
fluxSink -> {
            //NOTE sink:class reactor.core.publisher.FluxCreate$SerializedSink
            LOGGER.info("sink:{}",fluxSink.getClass());
            while (true) {
                LOGGER.info("sink next");
                fluxSink.next(ThreadLocalRandom.current().nextInt());
            }
        }
之后就开启数据推送

SerializedSink.next

reactor-core-3.1.3.RELEASE-sources.jar!/reactor/core/publisher/FluxCreate.java#SerializedSink.next

        public FluxSink<T> next(T t) {
            Objects.requireNonNull(t, "t is null in sink.next(t)");
            if (sink.isCancelled() || done) {
                Operators.onNextDropped(t, sink.currentContext());
                return this;
            }
            if (WIP.get(this) == 0 && WIP.compareAndSet(this, 0, 1)) {
                try {
                    sink.next(t);
                }
                catch (Throwable ex) {
                    Operators.onOperatorError(sink, ex, t, sink.currentContext());
                }
                if (WIP.decrementAndGet(this) == 0) {
                    return this;
                }
            }
            else {
                Queue<T> q = queue;
                synchronized (this) {
                    q.offer(t);
                }
                if (WIP.getAndIncrement(this) != 0) {
                    return this;
                }
            }
            drainLoop();
            return this;
        }
这里调用BufferAsyncSink.next,然后drainLoop之后才返回

BufferAsyncSink.next

        public FluxSink<T> next(T t) {
            queue.offer(t);
            drain();
            return this;
        }
这里将数据放入queue中,然后调用drain取数据,同步调用LambdaSubscriber的onNext

reactor-core-3.1.3.RELEASE-sources.jar!/reactor/core/publisher/LambdaSubscriber.java

    @Override
    public final void onNext(T x) {
        try {
            if (consumer != null) {
                consumer.accept(x);
            }
        }
        catch (Throwable t) {
            Exceptions.throwIfFatal(t);
            this.subscription.cancel();
            onError(t);
        }
    }
即同步调用自定义的subscribe方法,实例中除了log还会sleep,这里是同步阻塞的
这里调用完之后,fluxSink这里的next方法返回,然后继续循环
fluxSink -> {
            //NOTE sink:class reactor.core.publisher.FluxCreate$SerializedSink
            LOGGER.info("sink:{}",fluxSink.getClass());
            while (true) {
                LOGGER.info("sink next");
                fluxSink.next(ThreadLocalRandom.current().nextInt());
            }
        }

小结

fluxSink这里看是无限循环next产生数据,实则不用担心,如果subscribe与fluxSink都是同一个线程的话(本实例都是在main线程),它们是同步阻塞调用的。

subscribe的时候调用LambdaSubscriber.onSubscribe,request(N)请求数据,然后再调用source.accept,也就是调用fluxSink的lambda方法产生数据,开启stream模式

这里的fluxSink.next里头阻塞调用了subscribe的consumer,返回之后才继续循环。

至于BUFFER模式OOM的问题,可以思考下如何产生。


codecraft
11.9k 声望2k 粉丝

当一个代码的工匠回首往事时,不因虚度年华而悔恨,也不因碌碌无为而羞愧,这样,当他老的时候,可以很自豪告诉世人,我曾经将代码注入生命去打造互联网的浪潮之巅,那是个很疯狂的时代,我在一波波的浪潮上留下...


引用和评论

0 条评论