聊聊FluxFlatMap的concurrency及prefetch參數

序

本文主要研究下FluxFlatMap的concurrency及prefetch參數

實例

@Test
    public void testConcurrencyAndPrefetch(){
        int concurrency = 3;
        int prefetch = 6;
        Flux.range(1,100)
                .log()
                .flatMap(i -> Flux.just(1,2,3,4,5,6,7,8,9,10).log(),
                        concurrency,prefetch)
                .subscribe();
    }

部分輸出

23:29:38.515 [main] DEBUG reactor.util.Loggers$LoggerFactory - Using Slf4j logging framework
23:29:38.534 [main] INFO reactor.Flux.Range.1 - | onSubscribe([Synchronous Fuseable] FluxRange.RangeSubscription)
23:29:38.537 [main] INFO reactor.Flux.Range.1 - | request(3)
23:29:38.537 [main] INFO reactor.Flux.Range.1 - | onNext(1)
23:29:38.538 [main] INFO reactor.Flux.Array.2 - | onSubscribe([Synchronous Fuseable] FluxArray.ArraySubscription)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | request(6)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | onNext(1)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | onNext(2)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | onNext(3)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | onNext(4)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | onNext(5)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | request(5)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | onNext(6)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | onNext(7)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | onNext(8)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | onNext(9)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | onNext(10)
23:29:38.539 [main] INFO reactor.Flux.Array.2 - | request(5)
23:29:38.540 [main] INFO reactor.Flux.Array.2 - | onComplete()
23:29:38.540 [main] INFO reactor.Flux.Range.1 - | request(1)
23:29:38.540 [main] INFO reactor.Flux.Range.1 - | onNext(2)
23:29:38.540 [main] INFO reactor.Flux.Array.3 - | onSubscribe([Synchronous Fuseable] FluxArray.ArraySubscription)
23:29:38.540 [main] INFO reactor.Flux.Array.3 - | request(6)
23:29:38.540 [main] INFO reactor.Flux.Array.3 - | onNext(1)
23:29:38.540 [main] INFO reactor.Flux.Array.3 - | onNext(2)

但看外內兩個flux的第一次request，能夠初步看到分別是concurrency及prefetch

源碼解析

Flux

reactor-core-3.1.5.RELEASE-sources.jar!/reactor/core/publisher/Flux.java

/**
     * Transform the elements emitted by this {@link Flux} asynchronously into Publishers,
     * then flatten these inner publishers into a single {@link Flux} through merging,
     * which allow them to interleave.
     * <p>
     * There are three dimensions to this operator that can be compared with
     * {@link #flatMapSequential(Function) flatMapSequential} and {@link #concatMap(Function) concatMap}:
     * <ul>
     *     <li><b>Generation of inners and subscription</b>: this operator is eagerly
     *     subscribing to its inners.</li>
     *     <li><b>Ordering of the flattened values</b>: this operator does not necessarily preserve
     *     original ordering, as inner element are flattened as they arrive.</li>
     *     <li><b>Interleaving</b>: this operator lets values from different inners interleave
     *     (similar to merging the inner sequences).</li>
     * </ul>
     * The concurrency argument allows to control how many {@link Publisher} can be
     * subscribed to and merged in parallel. The prefetch argument allows to give an
     * arbitrary prefetch size to the merged {@link Publisher}.
     *
     * <p>
     * <img class="marble" src="https://raw.githubusercontent.com/reactor/reactor-core/v3.1.3.RELEASE/src/docs/marble/flatmapc.png" alt="">
     *
     * @param mapper the {@link Function} to transform input sequence into N sequences {@link Publisher}
     * @param concurrency the maximum number of in-flight inner sequences
     * @param prefetch the maximum in-flight elements from each inner {@link Publisher} sequence
     * @param <V> the merged output sequence type
     *
     * @return a merged {@link Flux}
     */
    public final <V> Flux<V> flatMap(Function<? super T, ? extends Publisher<? extends V>> mapper, int
            concurrency, int prefetch) {
        return flatMap(mapper, false, concurrency, prefetch);
    }

    final <V> Flux<V> flatMap(Function<? super T, ? extends Publisher<? extends
            V>> mapper, boolean delayError, int concurrency, int prefetch) {
        return onAssembly(new FluxFlatMap<>(
                this,
                mapper,
                delayError,
                concurrency,
                Queues.get(concurrency),
                prefetch,
                Queues.get(prefetch)
        ));
    }

這裏使用的是FluxFlatMap

FluxFlatMap

reactor-core-3.1.5.RELEASE-sources.jar!/reactor/core/publisher/FluxFlatMap.java

FluxFlatMap(Flux<? extends T> source,
            Function<? super T, ? extends Publisher<? extends R>> mapper,
            boolean delayError,
            int maxConcurrency,
            Supplier<? extends Queue<R>> mainQueueSupplier,
            int prefetch,
            Supplier<? extends Queue<R>> innerQueueSupplier) {
        super(source);
        if (prefetch <= 0) {
            throw new IllegalArgumentException("prefetch > 0 required but it was " + prefetch);
        }
        if (maxConcurrency <= 0) {
            throw new IllegalArgumentException("maxConcurrency > 0 required but it was " + maxConcurrency);
        }
        this.mapper = Objects.requireNonNull(mapper, "mapper");
        this.delayError = delayError;
        this.prefetch = prefetch;
        this.maxConcurrency = maxConcurrency;
        this.mainQueueSupplier =
                Objects.requireNonNull(mainQueueSupplier, "mainQueueSupplier");
        this.innerQueueSupplier =
                Objects.requireNonNull(innerQueueSupplier, "innerQueueSupplier");
    }

    @Override
    public void subscribe(CoreSubscriber<? super R> actual) {

        if (trySubscribeScalarMap(source, actual, mapper, false)) {
            return;
        }

        source.subscribe(new FlatMapMain<>(actual,
                mapper,
                delayError,
                maxConcurrency,
                mainQueueSupplier,
                prefetch, innerQueueSupplier));
    }

這裏能夠看到subscribe的時候使用了FlatMapMain

FlatMapMain

static final class FlatMapMain<T, R> extends FlatMapTracker<FlatMapInner<R>>
            implements InnerOperator<T, R> {

        FlatMapMain(CoreSubscriber<? super R> actual,
                Function<? super T, ? extends Publisher<? extends R>> mapper,
                boolean delayError,
                int maxConcurrency,
                Supplier<? extends Queue<R>> mainQueueSupplier,
                int prefetch,
                Supplier<? extends Queue<R>> innerQueueSupplier) {
            this.actual = actual;
            this.mapper = mapper;
            this.delayError = delayError;
            this.maxConcurrency = maxConcurrency;
            this.mainQueueSupplier = mainQueueSupplier;
            this.prefetch = prefetch;
            this.innerQueueSupplier = innerQueueSupplier;
            this.limit = Operators.unboundedOrLimit(maxConcurrency);
        }

        @Override
        public void request(long n) {
            if (Operators.validate(n)) {
                Operators.addCap(REQUESTED, this, n);
                drain();
            }
        }

        @Override
        public void onSubscribe(Subscription s) {
            if (Operators.validate(this.s, s)) {
                this.s = s;

                actual.onSubscribe(this);
                s.request(Operators.unboundedOrPrefetch(maxConcurrency));
            }
        }

        @SuppressWarnings("unchecked")
        @Override
        public void onNext(T t) {
            if (done) {
                Operators.onNextDropped(t, actual.currentContext());
                return;
            }

            Publisher<? extends R> p;

            try {
                p = Objects.requireNonNull(mapper.apply(t),
                "The mapper returned a null Publisher");
            }
            catch (Throwable e) {
                onError(Operators.onOperatorError(s, e, t, actual.currentContext()));
                return;
            }

            if (p instanceof Callable) {
                R v;
                try {
                    v = ((Callable<R>) p).call();
                }
                catch (Throwable e) {
                    if (!delayError || !Exceptions.addThrowable(ERROR, this, e)) {
                        onError(Operators.onOperatorError(s, e, t, actual.currentContext()));
                    }
                    return;
                }
                tryEmitScalar(v);
            }
            else {
                FlatMapInner<R> inner = new FlatMapInner<>(this, prefetch);
                if (add(inner)) {

                    p.subscribe(inner);
                }
            }

        }

        //...
}

這個能夠理解爲對外層flux的操做，能夠看到onSubscribe的時候，其內部request的大小爲Operators.unboundedOrPrefetch(maxConcurrency)，也就是第一個參數concurrency

在onNext操做裏頭，對裏頭的flux使用了FlatMapInner

FlatMapInner

static final class FlatMapInner<R>
            implements InnerConsumer<R>, Subscription {

        FlatMapInner(FlatMapMain<?, R> parent, int prefetch) {
            this.parent = parent;
            this.prefetch = prefetch;
//            this.limit = prefetch >> 2;
            this.limit = Operators.unboundedOrLimit(prefetch);
        }

        @Override
        public void onSubscribe(Subscription s) {
            if (Operators.setOnce(S, this, s)) {
                if (s instanceof Fuseable.QueueSubscription) {
                    @SuppressWarnings("unchecked") Fuseable.QueueSubscription<R> f =
                            (Fuseable.QueueSubscription<R>) s;
                    int m = f.requestFusion(Fuseable.ANY | Fuseable.THREAD_BARRIER);
                    if (m == Fuseable.SYNC) {
                        sourceMode = Fuseable.SYNC;
                        queue = f;
                        done = true;
                        parent.drain();
                        return;
                    }
                    if (m == Fuseable.ASYNC) {
                        sourceMode = Fuseable.ASYNC;
                        queue = f;
                    }
                    // NONE is just fall-through as the queue will be created on demand
                }
                s.request(Operators.unboundedOrPrefetch(prefetch));
            }
        }        

        @Override
        public void request(long n) {
            long p = produced + n;
            if (p >= limit) {
                produced = 0L;
                s.request(p);
            }
            else {
                produced = p;
            }
        }        
}

subscribe的時候，request的數量爲Operators.unboundedOrPrefetch(prefetch)
這裏能夠看到這裏對prefetch進行右移2操做，至關於除以4，做爲limit，limit是個判斷，用來對inner的flux的request數量進行限制

小結

flatMap的兩個參數concurrency及prefetch，分別是做用於外頭及裏頭的兩個flux，第一次request都是使用該值，後續的話，其內部會對request的數量進行判斷和調整。

doc

• webflux-concurrency-model