|
1 | | -# Scala Adaptor for RxJava |
2 | | - |
3 | | -This adaptor allows to use RxJava in Scala with anonymous functions, e.g. |
| 1 | +Alternative Rx bindings for Scala |
| 2 | +================================= |
4 | 3 |
|
| 4 | +The current RxScala binding attempt to optimize for seamless interop between Scala and Java. |
| 5 | +The intended interop is illustrated by the following example where in Scala a class is defined that takes |
| 6 | +an `Observable[Movie]` that is transformed using RxScala operators: |
5 | 7 | ```scala |
6 | | -val o = Observable.interval(200 millis).take(5) |
7 | | -o.subscribe(n => println("n = " + n)) |
8 | | -Observable(1, 2, 3, 4).reduce(_ + _) |
| 8 | +class MovieLib(val moviesStream: Observable[Movie]) { |
| 9 | + val threshold = 1200 |
| 10 | + def shortMovies: Observable[Movie] = ??? |
| 11 | + def longMovies: Observable[Movie] = ??? |
| 12 | +} |
9 | 13 | ``` |
| 14 | +which is then called in Java, passing a Java `Observable<Movie>` to the constructor |
| 15 | +```java |
| 16 | +public void test() { |
| 17 | + MovieLib lib = new MovieLib(Observable.from(...)); |
10 | 18 |
|
11 | | -For-comprehensions are also supported: |
12 | | - |
13 | | -```scala |
14 | | -val first = Observable(10, 11, 12) |
15 | | -val second = Observable(10, 11, 12) |
16 | | -val booleans = for ((n1, n2) <- (first zip second)) yield (n1 == n2) |
| 19 | + lib.longMovies().subscribe(moviePrinter); |
| 20 | +} |
17 | 21 | ``` |
18 | | - |
19 | | -Further, this adaptor attempts to expose an API which is as Scala-idiomatic as possible. This means that certain methods have been renamed, their signature was changed, or static methods were changed to instance methods. Some examples: |
20 | | - |
| 22 | +The technique used to obtain this transparency is to use a value class with a private constructor that implements |
| 23 | +the Rx operators in an idiomatic Scala way, and a companion object that is used to construct instances in Scala |
21 | 24 | ```scala |
22 | | - // instead of concat: |
23 | | -def ++[U >: T](that: Observable[U]): Observable[U] |
24 | | - |
25 | | -// instance method instead of static: |
26 | | -def zip[U](that: Observable[U]): Observable[(T, U)] |
27 | | - |
28 | | -// the implicit evidence argument ensures that dematerialize can only be called on Observables of Notifications: |
29 | | -def dematerialize[U](implicit evidence: T <:< Notification[U]): Observable[U] |
30 | | - |
31 | | -// additional type parameter U with lower bound to get covariance right: |
32 | | -def onErrorResumeNext[U >: T](resumeFunction: Throwable => Observable[U]): Observable[U] |
| 25 | +object Observable { |
| 26 | + def apply[T](asJava: rx.Observable[_ <: T]): Observable[T] = { new Observable[T](asJava) } |
| 27 | +} |
33 | 28 |
|
34 | | -// curried in Scala collections, so curry fold also here: |
35 | | -def fold[R](initialValue: R)(accumulator: (R, T) => R): Observable[R] |
36 | | - |
37 | | -// using Duration instead of (long timepan, TimeUnit duration): |
38 | | -def sample(duration: Duration): Observable[T] |
39 | | - |
40 | | -// called skip in Java, but drop in Scala |
41 | | -def drop(n: Int): Observable[T] |
42 | | - |
43 | | -// there's only mapWithIndex in Java, because Java doesn't have tuples: |
44 | | -def zipWithIndex: Observable[(T, Int)] |
45 | | - |
46 | | -// corresponds to Java's toList: |
47 | | -def toSeq: Observable[Seq[T]] |
48 | | - |
49 | | -// the implicit evidence argument ensures that switch can only be called on Observables of Observables: |
50 | | -def switch[U](implicit evidence: Observable[T] <:< Observable[Observable[U]]): Observable[U] |
51 | | - |
52 | | -// Java's from becomes apply, and we use Scala Range |
53 | | -def apply(range: Range): Observable[Int] |
54 | | - |
55 | | -// use Bottom type: |
56 | | -def never: Observable[Nothing] |
| 29 | +class Observable[+T] private[scala] (val asJava: rx.Observable[_ <: T]) extends AnyVal { |
| 30 | + // Idiomatic Scala friendly definitions of Rx operators |
| 31 | +} |
57 | 32 | ``` |
58 | | - |
59 | | -Also, the Scala Observable is fully covariant in its type parameter, whereas the Java Observable only achieves partial covariance due to limitations of Java's type system (or if you can fix this, your suggestions are very welcome). |
60 | | - |
61 | | -For more examples, see [RxScalaDemo.scala](https://github.com/Netflix/RxJava/blob/master/language-adaptors/rxjava-scala/src/examples/scala/rx/lang/scala/examples/RxScalaDemo.scala). |
62 | | - |
63 | | -Scala code using Rx should only import members from `rx.lang.scala` and below. |
64 | | - |
65 | | - |
66 | | -## Documentation |
67 | | - |
68 | | -The API documentation can be found [here](http://rxscala.github.io/scaladoc/index.html#rx.lang.scala.Observable). |
69 | | - |
70 | | -You can build the API documentation yourself by running `./gradlew scaladoc` in the RxJava root directory. |
71 | | - |
72 | | -Then navigate to `RxJava/language-adaptors/rxjava-scala/build/docs/scaladoc/index.html` to display it. |
73 | | - |
74 | | - |
75 | | -## Binaries |
76 | | - |
77 | | -Binaries and dependency information for Maven, Ivy, Gradle and others can be found at [http://search.maven.org](http://search.maven.org/#search%7Cga%7C1%7Ca%3A%22rxjava-scala%22). |
78 | | - |
79 | | -Example for Maven: |
80 | | - |
81 | | -```xml |
82 | | -<dependency> |
83 | | - <groupId>com.netflix.rxjava</groupId> |
84 | | - <artifactId>rxjava-scala</artifactId> |
85 | | - <version>x.y.z</version> |
86 | | -</dependency> |
| 33 | +Since `rx.lang.scala.Observable[T] extends AnyVal`, the underlying representation of `rx.lang.scala.Observable[T]` |
| 34 | +is the same as `rx.Observable<T>`. Because `rx.lang.scala.Observable[T]` is an opaque type in Scala, |
| 35 | +the Scala programmer only sees the Scala-friendly operators. |
| 36 | + |
| 37 | +However, in the current the illusion of interop is quickly lost when going beyond this simple example. |
| 38 | +For example but type `Notification[T]` and `Scheduler[T]` are defined using wrappers, |
| 39 | +and hence they are not compatible with `Notification<T>` respectively `Scheduler<T>`. |
| 40 | +For instance, when materializing an `Observable[T]` in Scala to an `Observable[Notification[T]]`, |
| 41 | +we lost the seamless interop with `Observable<Notification<T>>` on the Java side. |
| 42 | + |
| 43 | +However, the real problems with seamless interop show up when we try to creating bindings for other Rx types. |
| 44 | +In particular types that have inheritance or more structure. |
| 45 | + |
| 46 | +For example, RxScala currently defines a type synonym `type Observer[-T] = rx.Observer[_ >: T]`, |
| 47 | +but no further bindings for observers. |
| 48 | +Similarly, for subjects RxScala defines `type Subject[-T, +R] = rx.subjects.Subject[_ >: T, _ <: R]`. |
| 49 | +The problem with these definitions is that on the Java side, subjects are defined as: |
| 50 | +```scala |
| 51 | +public abstract class Subject<T, R> extends Observable<R> implements Observer<T> { …} |
87 | 52 | ``` |
| 53 | +without binding any of the Rx subjects. |
| 54 | + |
| 55 | +The consequence is that `Subject[S,T]` in Scala is unrelated to `rx.lang.scala.Observable[T]` in Scala, |
| 56 | +but shows up as a `rx.Observable[T]`. The problem however is that if we want to expose subjects in Scala |
| 57 | +such that they derive from both `Observable[S]` and `Observer[T]` we cannot use the `extend AnyVal` trick |
| 58 | +we used for `Observable[T]` and immediately lose transparent interop with Java. |
| 59 | + |
| 60 | +The problem is even worse because `AsyncSubject<T>`, `BehaviorSubject<T>`, … all derive from `Subject<T,T>`, |
| 61 | +so if we want them to derive from a common base `Subject[T,T]` type in Scala we lose transparency for those as well. |
| 62 | +And again, if we expose the various subjects by extending `AnyVal`, they are useless in Scala because they do not inherit |
| 63 | +from a common base type. To avoid implementing all methods of observable and observer on each specific subject |
| 64 | +we might add implicit conversions to `Observable[T]` and `Observer[T]` but that still does not give Scala users |
| 65 | +a native `Subject[S,T]` type. |
| 66 | +```scala |
| 67 | +object AsyncSubject { |
| 68 | + def apply[T](): AsyncSubject[T] = |
| 69 | + new AsyncSubject[T](rx.subjects.AsyncSubject.create()) |
| 70 | +} |
88 | 71 |
|
89 | | -and for Ivy: |
| 72 | +class AsyncSubject[T] private [scala] (val inner: rx.subjects.AsyncSubject[T]) |
| 73 | + extends AnyVal |
| 74 | +{ … } |
90 | 75 |
|
91 | | -```xml |
92 | | -<dependency org="com.netflix.rxjava" name="rxjava-scala" rev="x.y.z" /> |
| 76 | +implicit final def asObservable[T](subject: AsyncSubject[T]): Observable[T] = |
| 77 | + Observable(subject.inner) |
| 78 | + |
| 79 | +implicit final def asObserver[T](subject: AsyncSubject[T]): Observer[T] = |
| 80 | + subject.inner |
93 | 81 | ``` |
| 82 | +The inheritance problem is not just limited to subjects, but also surfaces for subscriptions. |
| 83 | +Rx scala currently defines `type Subscription = rx.Subscription` using a type synonym as well, |
| 84 | +and we run into exactly the same problems as with subjects when we try to bind the |
| 85 | +various Rx subscriptions `BooleanSubscription`, `SerialSubscription`, etc. |
94 | 86 |
|
95 | | -and for sbt: |
| 87 | +Since we cannot wrap Rx types in Scala such that they are both (a) transparently interoperable with Java, |
| 88 | +and (b) feel native and idiomatic to Scala, we should decide in favor of optimizing RxScala for Scala |
| 89 | +and consumption of Rx values from Java but not for Scala as a producer. |
96 | 90 |
|
| 91 | +If we take that approach, we can make bindings that feels like a completely native Scala library, |
| 92 | +without needing any complications of the Scala side. |
97 | 93 | ```scala |
98 | | -libraryDependencies ++= Seq( |
99 | | - "com.netflix.rxjava" % "rxjava-scala" % "x.y.z" |
100 | | -) |
| 94 | +object Observer { …} |
| 95 | +trait Observable[+T] { |
| 96 | + def asJavaObservable: rx.Observable[_ <: T] |
| 97 | +} |
| 98 | + |
| 99 | +object Observer {…} |
| 100 | +trait Observer[-T] { |
| 101 | + def asJavaObserver: rx.Observer[_ >: T] |
| 102 | +} |
| 103 | + |
| 104 | +object Subject {…} |
| 105 | +trait Subject[-T, +R] extends Observable[R] with Observer[T] { |
| 106 | + val asJavaSubject: rx.subjects.Subject[_ >: T, _<: R] |
| 107 | +} |
| 108 | + |
| 109 | +object Scheduler {…} |
| 110 | +trait Scheduler { |
| 111 | + def asJavaScheduler: rx.Scheduler; |
| 112 | +} |
| 113 | + |
| 114 | +object Notification {…} |
| 115 | +trait Notification[+T] { |
| 116 | + def asJavaNotification: rx.Notification[_ <: T] |
| 117 | +} |
| 118 | + |
| 119 | +object Subscription {…} |
| 120 | +trait Subscription { |
| 121 | + def asJavaSubscription: rx.Subscription |
| 122 | +} |
| 123 | +``` |
| 124 | +You pay the price when crossing the Scala/Java interop boundary, which is where it should be. |
| 125 | +The proper way is to put the burden of interop on the Scala side, in case you want to create |
| 126 | +a reusable Rx-based library in Scala, or wrap and unwrap on the Java side. |
| 127 | +```java |
| 128 | +public static void main(String[] args) { |
| 129 | + |
| 130 | + Observable<Movie> movies = Observable.from(new Movie(3000), new Movie(1000), new Movie(2000)); |
| 131 | + MovieLib lib = new MovieLib(toScalaObservable(movies)); |
| 132 | + lib.longMovies().asJavaObservable().subscribe(m -> |
| 133 | + System.out.println("A movie of length " + m.lengthInSeconds() + "s") |
| 134 | + ); |
| 135 | +} |
101 | 136 | ``` |
| 137 | +Delegation versus Inheritance |
| 138 | +----------------------------- |
| 139 | +The obvious thought is that using delegation instead of inheritance (http://c2.com/cgi/wiki?DelegationIsInheritance) |
| 140 | +will lead to excessive wrapping, since all Scala types wrap and delegate to an underlying RxJava implementation. |
| 141 | +Note however, that the wrapping happens at query generation time and incurs no overhead when messages are flowing |
| 142 | +through the pipeline. Say we have a query `xs.map(f).filter(p).subscribe(o)`. Even though the Scala types are wrappers, |
| 143 | +the callback that is registered with xs is something like `x => { val y = f(x); if(p(y)){ o.asJavaObserver.onNext(y) }}` |
| 144 | +and hence there is no additional runtime penalty. |
0 commit comments