Skip to content

What's different in 2.0

akarnokd edited this page Oct 9, 2018 · 68 revisions

RxJava 2.0 has been completely rewritten from scratch on top of the Reactive-Streams specification. The specification itself has evolved out of RxJava 1.x and provides a common baseline for reactive systems and libraries.

Because Reactive-Streams has a different architecture, it mandates changes to some well known RxJava types. This wiki page attempts to summarize what has changed and describes how to rewrite 1.x code into 2.x code.

For technical details on how to write operators for 2.x, please visit the Writing Operators wiki page.

Contents

Maven address and base package

To allow having RxJava 1.x and RxJava 2.x side-by-side, RxJava 2.x is under the maven coordinates io.reactivex.rxjava2:rxjava:2.x.y and classes are accessible below io.reactivex.

Users switching from 1.x to 2.x have to re-organize their imports, but carefully.

Javadoc

The official Javadoc pages for 2.x is hosted at http://reactivex.io/RxJava/2.x/javadoc/

Nulls

RxJava 2.x no longer accepts null values and the following will yield NullPointerException immediately or as a signal to downstream:

Observable.just(null);

Single.just(null);

Observable.fromCallable(() -> null)
    .subscribe(System.out::println, Throwable::printStackTrace);

Observable.just(1).map(v -> null)
    .subscribe(System.out::println, Throwable::printStackTrace);

This means that Observable<Void> can no longer emit any values but only terminate normally or with an exception. API designers may instead choose to define Observable<Object> with no guarantee on what Object will be (which should be irrelevant anyway). For example, if one needs a signaller-like source, a shared enum can be defined and its solo instance onNext'd:

enum Irrelevant { INSTANCE; }

Observable<Object> source = Observable.create((ObservableEmitter<Object> emitter) -> {
   System.out.println("Side-effect 1");
   emitter.onNext(Irrelevant.INSTANCE);

   System.out.println("Side-effect 2");
   emitter.onNext(Irrelevant.INSTANCE);

   System.out.println("Side-effect 3");
   emitter.onNext(Irrelevant.INSTANCE);
});

source.subscribe(e -> { /* Ignored. */ }, Throwable::printStackTrace);

Observable and Flowable

A small regret about introducing backpressure in RxJava 0.x is that instead of having a separate base reactive class, the Observable itself was retrofitted. The main issue with backpressure is that many hot sources, such as UI events, can't be reasonably backpressured and cause unexpected MissingBackpressureException (i.e., beginners don't expect them).

We try to remedy this situation in 2.x by having io.reactivex.Observable non-backpressured and the new io.reactivex.Flowable be the backpressure-enabled base reactive class.

The good news is that operator names remain (mostly) the same. The bad news is that one should be careful when performing 'organize imports' as it may select the non-backpressured io.reactivex.Observable unintended.

Which type to use?

When architecting dataflows (as an end-consumer of RxJava) or deciding upon what type your 2.x compatible library should take and return, you can consider a few factors that should help you avoid problems down the line such as MissingBackpressureException or OutOfMemoryError.

When to use Observable

  • You have a flow of no more than 1000 elements at its longest: i.e., you have so few elements over time that there is practically no chance for OOME in your application.
  • You deal with GUI events such as mouse moves or touch events: these can rarely be backpressured reasonably and aren't that frequent. You may be able to handle an element frequency of 1000 Hz or less with Observable but consider using sampling/debouncing anyway.
  • Your flow is essentially synchronous but your platform doesn't support Java Streams or you miss features from it. Using Observable has lower overhead in general than Flowable. (You could also consider IxJava which is optimized for Iterable flows supporting Java 6+).

When to use Flowable

  • Dealing with 10k+ of elements that are generated in some fashion somewhere and thus the chain can tell the source to limit the amount it generates.
  • Reading (parsing) files from disk is inherently blocking and pull-based which works well with backpressure as you control, for example, how many lines you read from this for a specified request amount).
  • Reading from a database through JDBC is also blocking and pull-based and is controlled by you by calling ResultSet.next() for likely each downstream request.
  • Network (Streaming) IO where either the network helps or the protocol used supports requesting some logical amount.
  • Many blocking and/or pull-based data sources which may eventually get a non-blocking reactive API/driver in the future.

Single

The 2.x Single reactive base type, which can emit a single onSuccess or onError has been redesigned from scratch. Its architecture now derives from the Reactive-Streams design. Its consumer type (rx.Single.SingleSubscriber<T>) has been changed from being a class that accepts rx.Subscription resources to be an interface io.reactivex.SingleObserver<T> that has only 3 methods:

interface SingleObserver<T> {
    void onSubscribe(Disposable d);
    void onSuccess(T value);
    void onError(Throwable error);
}

and follows the protocol onSubscribe (onSuccess | onError)?.

Completable

The Completable type remains largely the same. It was already designed along the Reactive-Streams style for 1.x so no user-level changes there.

Similar to the naming changes, rx.Completable.CompletableSubscriber has become io.reactivex.CompletableObserver with onSubscribe(Disposable):

interface CompletableObserver<T> {
    void onSubscribe(Disposable d);
    void onComplete();
    void onError(Throwable error);
}

and still follows the protocol onSubscribe (onComplete | onError)?.

Maybe

RxJava 2.0.0-RC2 introduced a new base reactive type called Maybe. Conceptually, it is a union of Single and Completable providing the means to capture an emission pattern where there could be 0 or 1 item or an error signalled by some reactive source.

The Maybe class is accompanied by MaybeSource as its base interface type, MaybeObserver<T> as its signal-receiving interface and follows the protocol onSubscribe (onSuccess | onError | onComplete)?. Because there could be at most 1 element emitted, the Maybe type has no notion of backpressure (because there is no buffer bloat possible as with unknown length Flowables or Observables.

This means that an invocation of onSubscribe(Disposable) is potentially followed by one of the other onXXX methods. Unlike Flowable, if there is only a single value to be signalled, only onSuccess is called and onComplete is not.

Working with this new base reactive type is practically the same as the others as it offers a modest subset of the Flowable operators that make sense with a 0 or 1 item sequence.

Maybe.just(1)
.map(v -> v + 1)
.filter(v -> v == 1)
.defaultIfEmpty(2)
.test()
.assertResult(2);

Base reactive interfaces

Following the style of extending the Reactive-Streams Publisher<T> in Flowable, the other base reactive classes now extend similar base interfaces (in package io.reactivex):

interface ObservableSource<T> {
    void subscribe(Observer<? super T> observer);
}

interface SingleSource<T> {
    void subscribe(SingleObserver<? super T> observer);
}

interface CompletableSource {
    void subscribe(CompletableObserver observer);
}

interface MaybeSource<T> {
    void subscribe(MaybeObserver<? super T> observer);
}

Therefore, many operators that required some reactive base type from the user now accept Publisher and XSource:

Flowable<R> flatMap(Function<? super T, ? extends Publisher<? extends R>> mapper);

Observable<R> flatMap(Function<? super T, ? extends ObservableSource<? extends R>> mapper);

By having Publisher as input this way, you can compose with other Reactive-Streams compliant libraries without the need to wrap them or convert them into Flowable first.

If an operator has to offer a reactive base type, however, the user will receive the full reactive class (as giving out an XSource is practically useless as it doesn't have operators on it):

Flowable<Flowable<Integer>> windows = source.window(5);

source.compose((Flowable<T> flowable) -> 
    flowable
    .subscribeOn(Schedulers.io())
    .observeOn(AndroidSchedulers.mainThread()));

Subjects and Processors

In the Reactive-Streams specification, the Subject-like behavior, namely being a consumer and supplier of events at the same time, is done by the org.reactivestreams.Processor interface. As with the Observable/Flowable split, the backpressure-aware, Reactive-Streams compliant implementations are based on the FlowableProcessor<T> class (which extends Flowable to give a rich set of instance operators). An important change regarding Subjects (and by extension, FlowableProcessor) that they no longer support T -> R like conversion (that is, input is of type T and the output is of type R). (We never had a use for it in 1.x and the original Subject<T, R> came from .NET where there is a Subject<T> overload because .NET allows the same class name with a different number of type arguments.)

The io.reactivex.subjects.AsyncSubject, io.reactivex.subjects.BehaviorSubject, io.reactivex.subjects.PublishSubject, io.reactivex.subjects.ReplaySubject and io.reactivex.subjects.UnicastSubject in 2.x don't support backpressure (as part of the 2.x Observable family).

The io.reactivex.processors.AsyncProcessor, io.reactivex.processors.BehaviorProcessor, io.reactivex.processors.PublishProcessor, io.reactivex.processors.ReplayProcessor and io.reactivex.processors.UnicastProcessor are backpressure-aware. The BehaviorProcessor and PublishProcessor don't coordinate requests (use Flowable.publish() for that) of their downstream subscribers and will signal them MissingBackpressureException if the downstream can't keep up. The other XProcessor types honor backpressure of their downstream subscribers but otherwise, when subscribed to a source (optional), they consume it in an unbounded manner (requesting Long.MAX_VALUE).

TestSubject

The 1.x TestSubject has been dropped. Its functionality can be achieved via TestScheduler, PublishProcessor/PublishSubject and observeOn(testScheduler)/scheduler parameter.

TestScheduler scheduler = new TestScheduler();
PublishSubject<Integer> ps = PublishSubject.create();

TestObserver<Integer> ts = ps.delay(1000, TimeUnit.MILLISECONDS, scheduler)
.test();

ts.assertEmpty();

ps.onNext(1);

scheduler.advanceTimeBy(999, TimeUnit.MILLISECONDS);

ts.assertEmpty();

scheduler.advanceTimeBy(1, TimeUnit.MILLISECONDS);

ts.assertValue(1);

SerializedSubject

The SerializedSubject is no longer a public class. You have to use Subject.toSerialized() and FlowableProcessor.toSerialized() instead.

Other classes

The rx.observables.ConnectableObservable is now io.reactivex.observables.ConnectableObservable<T> and io.reactivex.flowables.ConnectableFlowable<T>.

GroupedObservable

The rx.observables.GroupedObservable is now io.reactivex.observables.GroupedObservable<T> and io.reactivex.flowables.GroupedFlowable<T>.

In 1.x, you could create an instance with GroupedObservable.from() which was used internally by 1.x. In 2.x, all use cases now extend GroupedObservable directly thus the factory methods are no longer available; the whole class is now abstract.

You can extend the class and add your own custom subscribeActual behavior to achieve something similar to the 1.x features:

class MyGroup<K, V> extends GroupedObservable<K, V> {
    final K key;

    final Subject<V> subject;

    public MyGroup(K key) {
        this.key = key;
        this.subject = PublishSubject.create();
    }

    @Override
    public T getKey() {
        return key;
    }

    @Override
    protected void subscribeActual(Observer<? super T> observer) {
        subject.subscribe(observer);
    }
}

(The same approach works with GroupedFlowable as well.)

Functional interfaces

Because both 1.x and 2.x is aimed at Java 6+, we can't use the Java 8 functional interfaces such as java.util.function.Function. Instead, we defined our own functional interfaces in 1.x and 2.x follows this tradition.

One notable difference is that all our functional interfaces now define throws Exception. This is a large convenience for consumers and mappers that otherwise throw and would need try-catch to transform or suppress a checked exception.

Flowable.just("file.txt")
.map(name -> Files.readLines(name))
.subscribe(lines -> System.out.println(lines.size()), Throwable::printStackTrace);

If the file doesn't exist or can't be read properly, the end consumer will print out IOException directly. Note also the Files.readLines(name) invoked without try-catch.

Actions

As the opportunity to reduce component count, 2.x doesn't define Action3-Action9 and ActionN (these were unused within RxJava itself anyway).

The remaining action interfaces were named according to the Java 8 functional types. The no argument Action0 is replaced by the io.reactivex.functions.Action for the operators and java.lang.Runnable for the Scheduler methods. Action1 has been renamed to Consumer and Action2 is called BiConsumer. ActionN is replaced by the Consumer<Object[]> type declaration.

Functions

We followed the naming convention of Java 8 by defining io.reactivex.functions.Function and io.reactivex.functions.BiFunction, plus renaming Func3 - Func9 into Function3 - Function9 respectively. The FuncN is replaced by the Function<Object[], R> type declaration.

In addition, operators requiring a predicate no longer use Func1<T, Boolean> but have a separate, primitive-returning type of Predicate<T> (allows better inlining due to no autoboxing).

Subscriber

The Reactive-Streams specification has its own Subscriber as an interface. This interface is lightweight and combines request management with cancellation into a single interface org.reactivestreams.Subscription instead of having rx.Producer and rx.Subscription separately. This allows creating stream consumers with less internal state than the quite heavy rx.Subscriber of 1.x.

Flowable.range(1, 10).subscribe(new Subscriber<Integer>() {
    @Override
    public void onSubscribe(Subscription s) {
        s.request(Long.MAX_VALUE);
    }

    @Override
    public void onNext(Integer t) {
        System.out.println(t);
    }

    @Override
    public void onError(Throwable t) {
        t.printStackTrace();
    }

    @Override
    public void onComplete() {
        System.out.println("Done");
    }
});

Due to the name conflict, replacing the package from rx to org.reactivestreams is not enough. In addition, org.reactivestreams.Subscriber has no notion of adding resources to it, cancelling it or requesting from the outside.

To bridge the gap we defined abstract classes DefaultSubscriber, ResourceSubscriber and DisposableSubscriber (plus their XObserver variants) for Flowable (and Observable) respectively that offers resource tracking support (of Disposables) just like rx.Subscriber and can be cancelled/disposed externally via dispose():

ResourceSubscriber<Integer> subscriber = new ResourceSubscriber<Integer>() {
    @Override
    public void onStart() {
        request(Long.MAX_VALUE);
    }

    @Override
    public void onNext(Integer t) {
        System.out.println(t);
    }

    @Override
    public void onError(Throwable t) {
        t.printStackTrace();
    }

    @Override
    public void onComplete() {
        System.out.println("Done");
    }
};

Flowable.range(1, 10).delay(1, TimeUnit.SECONDS).subscribe(subscriber);

subscriber.dispose();

Note also that due to Reactive-Streams compatibility, the method onCompleted has been renamed to onComplete without the trailing d.

Since 1.x Observable.subscribe(Subscriber) returned Subscription, users often added the Subscription to a CompositeSubscription for example:

CompositeSubscription composite = new CompositeSubscription();

composite.add(Observable.range(1, 5).subscribe(new TestSubscriber<Integer>()));

Due to the Reactive-Streams specification, Publisher.subscribe returns void and the pattern by itself no longer works in 2.0. To remedy this, the method E subscribeWith(E subscriber) has been added to each base reactive class which returns its input subscriber/observer as is. With the two examples before, the 2.x code can now look like this since ResourceSubscriber implements Disposable directly:

CompositeDisposable composite2 = new CompositeDisposable();

composite2.add(Flowable.range(1, 5).subscribeWith(subscriber));

Calling request from onSubscribe/onStart

Note that due to how request management works, calling request(n) from Subscriber.onSubscribe or ResourceSubscriber.onStart may trigger calls to onNext immediately before the request() call itself returns to the onSubscribe/onStart method of yours:

Flowable.range(1, 3).subscribe(new Subscriber<Integer>() {

    @Override
    public void onSubscribe(Subscription s) {
        System.out.println("OnSubscribe start");
        s.request(Long.MAX_VALUE);
        System.out.println("OnSubscribe end");
    }

    @Override
    public void onNext(Integer v) {
        System.out.println(v);
    }

    @Override
    public void onError(Throwable e) {
        e.printStackTrace();
    }

    @Override
    public void onComplete() {
        System.out.println("Done");
    }
});

This will print:

OnSubscribe start
1
2
3
Done
OnSubscribe end

The problem comes when one does some initialization in onSubscribe/onStart after calling request there and onNext may or may not see the effects of the initialization. To avoid this situation, make sure you call request after all initialization have been done in onSubscribe/onStart.

This behavior differs from 1.x where a request call went through a deferred logic that accumulated requests until an upstream Producer arrived at some time (This nature adds overhead to all operators and consumers in 1.x.) In 2.x, there is always a Subscription coming down first and 90% of the time there is no need to defer requesting.

Subscription

In RxJava 1.x, the interface rx.Subscription was responsible for stream and resource lifecycle management, namely unsubscribing a sequence and releasing general resources such as scheduled tasks. The Reactive-Streams specification took this name for specifying an interaction point between a source and a consumer: org.reactivestreams.Subscription allows requesting a positive amount from the upstream and allows cancelling the sequence.

To avoid the name clash, the 1.x rx.Subscription has been renamed into io.reactivex.Disposable (somewhat resembling .NET's own IDisposable).

Because Reactive-Streams base interface, org.reactivestreams.Publisher defines the subscribe() method as void, Flowable.subscribe(Subscriber) no longer returns any Subscription (or Disposable). The other base reactive types also follow this signature with their respective subscriber types.

The other overloads of subscribe now return Disposable in 2.x.

The original Subscription container types have been renamed and updated

  • CompositeSubscription to CompositeDisposable
  • SerialSubscription and MultipleAssignmentSubscription have been merged into SerialDisposable. The set() method disposes the old value and replace() method does not.
  • RefCountSubscription has been removed.

Backpressure

The Reactive-Streams specification mandates operators supporting backpressure, specifically via the guarantee that they don't overflow their consumers when those don't request. Operators of the new Flowable base reactive type now consider downstream request amounts properly, however, this doesn't mean MissingBackpressureException is gone. The exception is still there but this time, the operator that can't signal more onNext will signal this exception instead (allowing better identification of who is not properly backpressured).

As an alternative, the 2.x Observable doesn't do backpressure at all and is available as a choice to switch over.

Reactive-Streams compliance

updated in 2.0.7

The Flowable-based sources and operators are, as of 2.0.7, fully Reactive-Streams version 1.0.0 specification compliant.

Before 2.0.7, the operator strict() had to be applied in order to achieve the same level of compliance. In 2.0.7, the operator strict() returns this, is deprecated and will be removed completely in 2.1.0.

As one of the primary goals of RxJava 2, the design focuses on performance and in order enable it, RxJava 2.0.7 adds a custom io.reactivex.FlowableSubscriber interface (extends org.reactivestreams.Subscriber) but adds no new methods to it. The new interface is constrained to RxJava 2 and represents a consumer to Flowable that is able to work in a mode that relaxes the Reactive-Streams version 1.0.0 specification in rules §1.3, §2.3, §2.12 and §3.9:

  • §1.3 relaxation: onSubscribe may run concurrently with onNext in case the FlowableSubscriber calls request() from inside onSubscribe and it is the resposibility of FlowableSubscriber to ensure thread-safety between the remaining instructions in onSubscribe and onNext.
  • §2.3 relaxation: calling Subscription.cancel and Subscription.request from FlowableSubscriber.onComplete() or FlowableSubscriber.onError() is considered a no-operation.
  • §2.12 relaxation: if the same FlowableSubscriber instance is subscribed to multiple sources, it must ensure its onXXX methods remain thread safe.
  • §3.9 relaxation: issuing a non-positive request() will not stop the current stream but signal an error via RxJavaPlugins.onError.

From a user's perspective, if one was using the the subscribe methods other than Flowable.subscribe(Subscriber<? super T>), there is no need to do anything regarding this change and there is no extra penalty for it.

If one was using Flowable.subscribe(Subscriber<? super T>) with the built-in RxJava Subscriber implementations such as DisposableSubscriber, TestSubscriber and ResourceSubscriber, there is a small runtime overhead (one instanceof check) associated when the code is not recompiled against 2.0.7.

If a custom class implementing Subscriber was employed before, subscribing it to a Flowable adds an internal wrapper that ensures observing the Flowable is 100% compliant with the specification at the cost of some per-item overhead.

In order to help lift these extra overheads, a new method Flowable.subscribe(FlowableSubscriber<? super T>) has been added which exposes the original behavior from before 2.0.7. It is recommended that new custom consumer implementations extend FlowableSubscriber instead of just Subscriber.

Runtime hooks

The 2.x redesigned the RxJavaPlugins class which now supports changing the hooks at runtime. Tests that want to override the schedulers and the lifecycle of the base reactive types can do it on a case-by-case basis through callback functions.

The class-based RxJavaObservableHook and friends are now gone and RxJavaHooks functionality is incorporated into RxJavaPlugins.

Error handling

One important design requirement for 2.x is that no Throwable errors should be swallowed. This means errors that can't be emitted because the downstream's lifecycle already reached its terminal state or the downstream cancelled a sequence which was about to emit an error.

Such errors are routed to the RxJavaPlugins.onError handler. This handler can be overridden with the method RxJavaPlugins.setErrorHandler(Consumer<Throwable>). Without a specific handler, RxJava defaults to printing the Throwable's stacktrace to the console and calls the current thread's uncaught exception handler.

On desktop Java, this latter handler does nothing on an ExecutorService backed Scheduler and the application can keep running. However, Android is more strict and terminates the application in such uncaught exception cases.

If this behavior is desirable can be debated, but in any case, if you want to avoid such calls to the uncaught exception handler, the final application that uses RxJava 2 (directly or transitively) should set a no-op handler:

// If Java 8 lambdas are supported
RxJavaPlugins.setErrorHandler(e -> { });

// If no Retrolambda or Jack 
RxJavaPlugins.setErrorHandler(Functions.<Throwable>emptyConsumer());

It is not advised intermediate libraries change the error handler outside their own testing environment.

Unfortunately, RxJava can't tell which of these out-of-lifecycle, undeliverable exceptions should or shouldn't crash your app. Identifying the source and reason for these exceptions can be tiresome, especially if they originate from a source and get routed to RxJavaPlugins.onError somewhere lower the chain.

Therefore, 2.0.6 introduces specific exception wrappers to help distinguish and track down what was happening the time of the error:

  • OnErrorNotImplementedException: reintroduced to detect when the user forgot to add error handling to subscribe().
  • ProtocolViolationException: indicates a bug in an operator
  • UndeliverableException: wraps the original exception that can't be delivered due to lifecycle restrictions on a Subscriber/Observer. It is automatically applied by RxJavaPlugins.onError with intact stacktrace that may help find which exact operator rerouted the original error.

If an undeliverable exception is an instance/descendant of NullPointerException, IllegalStateException (UndeliverableException and ProtocolViolationException extend this), IllegalArgumentException, CompositeException, MissingBackpressureException or OnErrorNotImplementedException, the UndeliverableException wrapping doesn't happen.

In addition, some 3rd party libraries/code throw when they get interrupted by a cancel/dispose call which leads to an undeliverable exception most of the time. Internal changes in 2.0.6 now consistently cancel or dispose a Subscription/Disposable before cancelling/disposing a task or worker (which causes the interrupt on the target thread).

// in some library
try {
   doSomethingBlockingly()
} catch (InterruptedException ex) {
   // check if the interrupt is due to cancellation
   // if so, no need to signal the InterruptedException
   if (!disposable.isDisposed()) {
      observer.onError(ex);
   }
}

If the library/code already did this, the undeliverable InterruptedExceptions should stop now. If this pattern was not employed before, we encourage updating the code/library in question.

If one decides to add a non-empty global error consumer, here is an example that manages the typical undeliverable exceptions based on whether they represent a likely bug or an ignorable application/network state:

RxJavaPlugins.setErrorHandler(e -> {
    if (e instanceof UndeliverableException) {
        e = e.getCause();
    }
    if ((e instanceof IOException) || (e instanceof SocketException)) {
        // fine, irrelevant network problem or API that throws on cancellation
        return;
    }
    if (e instanceof InterruptedException) {
        // fine, some blocking code was interrupted by a dispose call
        return;
    }
    if ((e instanceof NullPointerException) || (e instanceof IllegalArgumentException)) {
        // that's likely a bug in the application
        Thread.currentThread().getUncaughtExceptionHandler()
            .handleException(Thread.currentThread(), e);
        return;
    }
    if (e instanceof IllegalStateException) {
        // that's a bug in RxJava or in a custom operator
        Thread.currentThread().getUncaughtExceptionHandler()
            .handleException(Thread.currentThread(), e);
        return;
    }
    Log.warning("Undeliverable exception received, not sure what to do", e);
});

Schedulers

The 2.x API still supports the main default scheduler types: computation, io, newThread and trampoline, accessible through io.reactivex.schedulers.Schedulers utility class.

The immediate scheduler is not present in 2.x. It was frequently misused and didn't implement the Scheduler specification correctly anyway; it contained blocking sleep for delayed action and didn't support recursive scheduling at all. Use Schedulers.trampoline() instead.

The Schedulers.test() has been removed as well to avoid the conceptional difference with the rest of the default schedulers. Those return a "global" scheduler instance whereas test() returned always a new instance of the TestScheduler. Test developers are now encouraged to simply new TestScheduler() in their code.

The io.reactivex.Scheduler abstract base class now supports scheduling tasks directly without the need to create and then destroy a Worker (which is often forgotten):

public abstract class Scheduler {
    
    public Disposable scheduleDirect(Runnable task) { ... }
 
    public Disposable scheduleDirect(Runnable task, long delay, TimeUnit unit) { ... }

    public Disposable scheduleDirectPeriodically(Runnable task, long initialDelay, 
        long period, TimeUnit unit) { ... }

    public long now(TimeUnit unit) { ... }

    // ... rest is the same: lifecycle methods, worker creation
}

The main purpose is to avoid the tracking overhead of the Workers for typically one-shot tasks. The methods have a default implementation that reuses createWorker properly but can be overridden with more efficient implementations if necessary.

The method that returns the scheduler's own notion of current time, now() has been changed to accept a TimeUnit to indicate the unit of measure.

Entering the reactive world

One of the design flaws of RxJava 1.x was the exposure of the rx.Observable.create() method that while powerful, not the typical operator you want to use to enter the reactive world. Unfortunately, so many depend on it that we couldn't remove or rename it.

Since 2.x is a fresh start, we won't make that mistake again. Each reactive base type Flowable, Observable, Single, Maybe and Completable feature a safe create operator that does the right thing regarding backpressure (for Flowable) and cancellation (all):

Flowable.create((FlowableEmitter<Integer> emitter) -> {
    emitter.onNext(1);
    emitter.onNext(2);
    emitter.onComplete();
}, BackpressureStrategy.BUFFER);

Practically, the 1.x fromEmitter (formerly fromAsync) has been renamed to Flowable.create. The other base reactive types have similar create methods (minus the backpressure strategy).

Leaving the reactive world

Apart from subscribing to the base types with their respective consumers (Subscriber, Observer, SingleObserver, MaybeObserver and CompletableObserver) and functional-interface based consumers (such as subscribe(Consumer<T>, Consumer<Throwable>, Action)), the formerly separate 1.x BlockingObservable (and similar classes for the others) has been integrated with the main reactive type. Now you can directly block for some results by invoking a blockingX operation directly:

List<Integer> list = Flowable.range(1, 100).toList().blockingGet(); // toList() returns Single

Integer i = Flowable.range(100, 100).blockingLast();

(The reason for this is twofold: performance and ease of use of the library as a synchronous Java 8 Streams-like processor.)

Another significant difference between rx.Subscriber (and co) and org.reactivestreams.Subscriber (and co) is that in 2.x, your Subscribers and Observers are not allowed to throw anything but fatal exceptions (see Exceptions.throwIfFatal()). (The Reactive-Streams specification allows throwing NullPointerException if the onSubscribe, onNext or onError receives a null value, but RxJava doesn't let nulls in any way.) This means the following code is no longer legal:

Subscriber<Integer> subscriber = new Subscriber<Integer>() {
    @Override
    public void onSubscribe(Subscription s) {
        s.request(Long.MAX_VALUE);
    }

    public void onNext(Integer t) {
        if (t == 1) {
            throw new IllegalArgumentException();
        }
    }

    public void onError(Throwable e) {
        if (e instanceof IllegalArgumentException) {
            throw new UnsupportedOperationException();
        }
    }

    public void onComplete() {
        throw new NoSuchElementException();
    }
};

Flowable.just(1).subscribe(subscriber);

The same applies to Observer, SingleObserver, MaybeObserver and CompletableObserver.

Since many of the existing code targeting 1.x do such things, the method safeSubscribe has been introduced that does handle these non-conforming consumers.

Alternatively, you can use the subscribe(Consumer<T>, Consumer<Throwable>, Action) (and similar) methods to provide a callback/lambda that can throw:

Flowable.just(1)
.subscribe(
    subscriber::onNext, 
    subscriber::onError, 
    subscriber::onComplete, 
    subscriber::onSubscribe
);

Testing

Testing RxJava 2.x works the same way as it does in 1.x. Flowable can be tested with io.reactivex.subscribers.TestSubscriber whereas the non-backpressured Observable, Single, Maybe and Completable can be tested with io.reactivex.observers.TestObserver.

test() "operator"

To support our internal testing, all base reactive types now feature test() methods (which is a huge convenience for us) returning TestSubscriber or TestObserver:

TestSubscriber<Integer> ts = Flowable.range(1, 5).test();

TestObserver<Integer> to = Observable.range(1, 5).test();

TestObserver<Integer> tso = Single.just(1).test();

TestObserver<Integer> tmo = Maybe.just(1).test();

TestObserver<Integer> tco = Completable.complete().test();

The second convenience is that most TestSubscriber/TestObserver methods return the instance itself allowing chaining the various assertX methods. The third convenience is that you can now fluently test your sources without the need to create or introduce TestSubscriber/TestObserver instance in your code:

Flowable.range(1, 5)
.test()
.assertResult(1, 2, 3, 4, 5)
;

Notable new assert methods

  • assertResult(T... items): asserts if subscribed, received exactly the given items in the given order followed by onComplete and no errors
  • assertFailure(Class<? extends Throwable> clazz, T... items): asserts if subscribed, received exactly the given items in the given order followed by a Throwable error of wich clazz.isInstance() returns true.
  • assertFailureAndMessage(Class<? extends Throwable> clazz, String message, T... items): same as assertFailure plus validates the getMessage() contains the specified message
  • awaitDone(long time, TimeUnit unit) awaits a terminal event (blockingly) and cancels the sequence if the timeout elapsed.
  • assertOf(Consumer<TestSubscriber<T>> consumer) compose some assertions into the fluent chain (used internally for fusion test as operator fusion is not part of the public API right now).

One of the benefits is that changing Flowable to Observable here the test code part doesn't have to change at all due to the implicit type change of the TestSubscriber to TestObserver.

cancel and request upfront

The test() method on TestObserver has a test(boolean cancel) overload which cancels/disposes the TestSubscriber/TestObserver before it even gets subscribed:

PublishSubject<Integer> pp = PublishSubject.create();

// nobody subscribed yet
assertFalse(pp.hasSubscribers());

pp.test(true);

// nobody remained subscribed
assertFalse(pp.hasSubscribers());

TestSubscriber has the test(long initialRequest) and test(long initialRequest, boolean cancel) overloads to specify the initial request amount and whether the TestSubscriber should be also immediately cancelled. If the initialRequest is given, the TestSubscriber offers the requestMore(long) method to keep requesting in a fluent manner:

Flowable.range(1, 5)
.test(0)
.assertValues()
.requestMore(1)
.assertValues(1)
.requestMore(2)
.assertValues(1, 2, 3)
.requestMore(2)
.assertResult(1, 2, 3, 4, 5);

or alternatively the TestSubscriber instance has to be captured to gain access to its request() method:

PublishProcessor<Integer> pp = PublishProcessor.create();

TestSubscriber<Integer> ts = pp.test(0L);

ts.request(1);

pp.onNext(1);
pp.onNext(2);

ts.assertFailure(MissingBackpressureException.class, 1);

Testing an async source

Given an asynchronous source, fluent blocking for a terminal event is now possible:

Flowable.just(1)
.subscribeOn(Schedulers.single())
.test()
.awaitDone(5, TimeUnit.SECONDS)
.assertResult(1);

Mockito & TestSubscriber

Those who are using Mockito and mocked Observer in 1.x has to mock the Subscriber.onSubscribe method to issue an initial request, otherwise, the sequence will hang or fail with hot sources:

@SuppressWarnings("unchecked")
public static <T> Subscriber<T> mockSubscriber() {
    Subscriber<T> w = mock(Subscriber.class);

    Mockito.doAnswer(new Answer<Object>() {
        @Override
        public Object answer(InvocationOnMock a) throws Throwable {
            Subscription s = a.getArgumentAt(0, Subscription.class);
            s.request(Long.MAX_VALUE);
            return null;
        }
    }).when(w).onSubscribe((Subscription)any());

    return w;
}

Operator differences

Most operators are still there in 2.x and practically all of them have the same behavior as they had in 1.x. The following subsections list each base reactive type and the difference between 1.x and 2.x.

Generally, many operators gained overloads that now allow specifying the internal buffer size or prefetch amount they should run their upstream (or inner sources).

Some operator overloads have been renamed with a postfix, such as fromArray, fromIterable etc. The reason for this is that when the library is compiled with Java 8, the javac often can't disambiguate between functional interface types.

Operators marked as @Beta or @Experimental in 1.x are promoted to standard.

1.x Observable to 2.x Flowable

Factory methods:

1.x 2.x
amb added amb(ObservableSource...) overload, 2-9 argument versions dropped
RxRingBuffer.SIZE bufferSize()
combineLatest added varargs overload, added overloads with bufferSize argument, combineLatest(List) dropped
concat added overload with prefetch argument, 5-9 source overloads dropped, use concatArray instead
N/A added concatArray and concatArrayDelayError
N/A added concatArrayEager and concatArrayEagerDelayError
concatDelayError added overloads with option to delay till the current ends or till the very end
concatEagerDelayError added overloads with option to delay till the current ends or till the very end
create(SyncOnSubscribe) replaced with generate + overloads (distinct interfaces, you can implement them all at once)
create(AsnycOnSubscribe) not present
create(OnSubscribe) repurposed with safe create(FlowableOnSubscribe, BackpressureStrategy), raw support via unsafeCreate()
from disambiguated into fromArray, fromIterable, fromFuture
N/A added fromPublisher
fromAsync renamed to create()
N/A added intervalRange()
limit dropped, use take
merge added overloads with prefetch
mergeDelayError added overloads with prefetch
sequenceEqual added overload with bufferSize
switchOnNext added overload with prefetch
switchOnNextDelayError added overload with prefetch
timer deprecated overloads dropped
zip added overloads with bufferSize and delayErrors capabilities, disambiguated to zipArray and zipIterable

Instance methods:

1.x 2.x
all RC3 returns Single<Boolean> now
any RC3 returns Single<Boolean> now
asObservable renamed to hide(), hides all identities now
buffer overloads with custom Collection supplier
cache(int) deprecated and dropped
collect RC3 returns Single<U>
collect(U, Action2<U, T>) disambiguated to collectInto and RC3 returns Single<U>
concatMap added overloads with prefetch
concatMapDelayError added overloads with prefetch, option to delay till the current ends or till the very end
concatMapEager added overloads with prefetch
concatMapEagerDelayError added overloads with prefetch, option to delay till the current ends or till the very end
count RC3 returns Single<Long> now
countLong dropped, use count
distinct overload with custom Collection supplier.
doOnCompleted renamed to doOnComplete, note the missing d!
doOnUnsubscribe renamed to Flowable.doOnCancel and doOnDispose for the others, additional info
N/A added doOnLifecylce to handle onSubscribe, request and cancel peeking
elementAt(int) RC3 no longer signals NoSuchElementException if the source is shorter than the index
elementAt(Func1, int) dropped, use filter(predicate).elementAt(int)
elementAtOrDefault(int, T) renamed to elementAt(int, T) and RC3 returns Single<T>
elementAtOrDefault(Func1, int, T) dropped, use filter(predicate).elementAt(int, T)
first() RC3 renamed to firstElement and returns Maybe<T>
first(Func1) dropped, use filter(predicate).first()
firstOrDefault(T) renamed to first(T) and RC3 returns Single<T>
firstOrDefault(Func1, T) dropped, use filter(predicate).first(T)
flatMap added overloads with prefetch
N/A added forEachWhile(Predicate<T>, [Consumer<Throwable>, [Action]]) for conditionally stopping consumption
groupBy added overload with bufferSize and delayError option, the custom internal map version didn't make it into RC1
ignoreElements RC3 returns Completable
isEmpty RC3 returns Single<Boolean>
last() RC3 renamed to lastElement and returns Maybe<T>
last(Func1) dropped, use filter(predicate).last()
lastOrDefault(T) renamed to last(T) and RC3 returns Single<T>
lastOrDefault(Func1, T) dropped, use filter(predicate).last(T)
nest dropped, use manual just
publish(Func1) added overload with prefetch
reduce(Func2) RC3 returns Maybe<T>
N/A added reduceWith(Callable, BiFunction) to reduce in a Subscriber-individual manner, returns Single<T>
N/A added repeatUntil(BooleanSupplier)
repeatWhen(Func1, Scheduler) dropped the overload, use subscribeOn(Scheduler).repeatWhen(Function) instead
retry added retry(Predicate), retry(int, Predicate)
N/A added retryUntil(BooleanSupplier)
retryWhen(Func1, Scheduler) dropped the overload, use subscribeOn(Scheduler).retryWhen(Function) instead
sample doesn't emit the very last item if the upstream completes within the period, added overloads with emitLast parameter
N/A added scanWith(Callable, BiFunction) to scan in a Subscriber-individual manner
single() RC3 renamed to singleElement and returns Maybe<T>
single(Func1) dropped, use filter(predicate).single()
singleOrDefault(T) renamed to single(T) and RC3 returns Single<T>
singleOrDefault(Func1, T) dropped, use filter(predicate).single(T)
skipLast added overloads with bufferSize and delayError options
startWith 2-9 argument version dropped, use startWithArray instead
N/A added startWithArray to disambiguate
subscribe No longer wraps all consumer types (i.e., Observer) with a safety wrapper, (just like the 1.x unsafeSubscribe no longer available). Use safeSubscribe to get an explicit safety wrapper around a consumer type.
N/A added subscribeWith that returns its input after subscription
switchMap added overload with prefetch argument
switchMapDelayError added overload with prefetch argument
takeLastBuffer dropped
N/A added test() (returns TestSubscriber subscribed to this) with overloads to fluently test
throttleLast doesn't emit the very last item if the upstream completes within the period, use sample with the emitLast parameter
timeout(Func0<Observable>, ...) signature changed to timeout(Publisher, ...) and dropped the function, use defer(Callable<Publisher>>) if necessary
toBlocking().y inlined as blockingY() operators, except toFuture
toCompletable RC3 dropped, use ignoreElements
toList RC3 returns Single<List<T>>
toMap RC3 returns Single<Map<K, V>>
toMultimap RC3 returns Single<Map<K, Collection<V>>>
N/A added toFuture
N/A added toObservable
toSingle RC3 dropped, use single(T)
toSortedList RC3 returns Single<List<T>>
unsafeSubscribe Removed as the Reactive Streams specification mandates the onXXX methods don't crash and therefore the default is to not have a safety net in subscribe. The new safeSubscribe method was introduced to explicitly add the safety wrapper around a consumer type.
withLatestFrom 5-9 source overloads dropped
zipWith added overloads with prefetch and delayErrors options

Different return types

Some operators that produced exactly one value or an error now return Single in 2.x (or Maybe if an empty source is allowed).

(Remark: this is "experimental" in RC2 and RC3 to see how it feels to program with such mixed-type sequences and whether or not there has to be too much toObservable/toFlowable back-conversion.)

Operator Old return type New return type Remark
all(Predicate) Observable<Boolean> Single<Boolean> Emits true if all elements match the predicate
any(Predicate) Observable<Boolean> Single<Boolean> Emits true if any elements match the predicate
count() Observable<Long> Single<Long> Counts the number of elements in the sequence
elementAt(int) Observable<T> Maybe<T> Emits the element at the given index or completes
elementAt(int, T) Observable<T> Single<T> Emits the element at the given index or the default
elementAtOrError(int) Observable<T> Single<T> Emits the indexth element or a NoSuchElementException
first(T) Observable<T> Single<T> Emits the very first element or NoSuchElementException
firstElement() Observable<T> Maybe<T> Emits the very first element or completes
firstOrError() Observable<T> Single<T> Emits the first element or a NoSuchElementException if the source is empty
ignoreElements() Observable<T> Completable Ignore all but the terminal events
isEmpty() Observable<Boolean> Single<Boolean> Emits true if the source is empty
last(T) Observable<T> Single<T> Emits the very last element or the default item
lastElement() Observable<T> Maybe<T> Emits the very last element or completes
lastOrError() Observable<T> Single<T> Emits the lastelement or a NoSuchElementException if the source is empty
reduce(BiFunction) Observable<T> Maybe<T> Emits the reduced value or completes
reduce(Callable, BiFunction) Observable<U> Single<U> Emits the reduced value (or the initial value)
reduceWith(U, BiFunction) Observable<U> Single<U> Emits the reduced value (or the initial value)
single(T) Observable<T> Single<T> Emits the only element or the default item
singleElement() Observable<T> Maybe<T> Emits the only element or completes
singleOrError() Observable<T> Single<T> Emits the one and only element, IndexOutOfBoundsException if the source is longer than 1 item or a NoSuchElementException if the source is empty
toList() Observable<List<T>> Single<List<T>> collects all elements into a List
toMap() Observable<Map<K, V>> Single<Map<K, V>> collects all elements into a Map
toMultimap() Observable<Map<K, Collection<V>>> Single<Map<K, Collection<V>>> collects all elements into a Map with collection
toSortedList() Observable<List<T>> Single<List<T>> collects all elements into a List and sorts it

Removals

To make sure the final API of 2.0 is clean as possible, we remove methods and other components between release candidates without deprecating them.

Removed in version Component Remark
RC3 Flowable.toCompletable() use Flowable.ignoreElements()
RC3 Flowable.toSingle() use Flowable.single(T)
RC3 Flowable.toMaybe() use Flowable.singleElement()
RC3 Observable.toCompletable() use Observable.ignoreElements()
RC3 Observable.toSingle() use Observable.single(T)
RC3 Observable.toMaybe() use Observable.singleElement()

Miscellaneous changes

doOnCancel/doOnDispose/unsubscribeOn

In 1.x, the doOnUnsubscribe was always executed on a terminal event because 1.x' SafeSubscriber called unsubscribe on itself. This was practically unnecessary and the Reactive-Streams specification states that when a terminal event arrives at a Subscriber, the upstream Subscription should be considered cancelled and thus calling cancel() is a no-op.

For the same reason, unsubscribeOn is not called on the regular termination path but only when there is an actual cancel (or dispose) call on the chain.

Therefore, the following sequence won't call doOnCancel:

Flowable.just(1, 2, 3)
.doOnCancel(() -> System.out.println("Cancelled!"))
.subscribe(System.out::println);

However, the following will call since the take operator cancels after the set amount of onNext events have been delivered:

Flowable.just(1, 2, 3)
.doOnCancel(() -> System.out.println("Cancelled!"))
.take(2)
.subscribe(System.out::println);

If you need to perform cleanup on both regular termination or cancellation, consider the operator using instead.

Alternatively, the doFinally operator (introduced in 2.0.1 and standardized in 2.1) calls a developer specified Action that gets executed after a source completed, failed with an error or got cancelled/disposed:

Flowable.just(1, 2, 3)
.doFinally(() -> System.out.println("Finally"))
.subscribe(System.out::println);

Flowable.just(1, 2, 3)
.doFinally(() -> System.out.println("Finally"))
.take(2) // cancels the above after 2 elements
.subscribe(System.out::println);
Clone this wiki locally