Skip to main content

1.x to 2.x migration: Observable vs. Observable: RxJava FAQ

View from Basilica of Notre-Dame de Fourvière
The title is not a mistake. rx.Observable from RxJava 1.x is a completely different beast than io.reactivex.Observable from 2.x. Blindly upgrading rx dependency and renaming all imports in your project will compile (with minor changes) but does not guarantee the same behavior. In the very early days of the project Observable in 1.x had no notion of backpressure but later on backpressure was included. What does it actually mean? Let's imagine we have a stream that produces one event every 1 millisecond but it takes 1 second to process one such item. You see it can't possibly work this way in the long run:

import rx.Observable;  //RxJava 1.x
import rx.schedulers.Schedulers;

Observable
        .interval(1, MILLISECONDS)
        .observeOn(Schedulers.computation())
        .subscribe(
                x -> sleep(Duration.ofSeconds(1)));
MissingBackpressureException creeps in within few hundred milliseconds. But what does this exception mean? Well, basically it's a safety net (or sanity check if you will) that prevents you from hurting your application. RxJava automatically discovers that producer is overflowing the consumer and proactively terminates the stream to avoid further damage. So what if we simply search and replace few imports here and there?

import io.reactivex.Observable;     //RxJava 2.x
import io.reactivex.schedulers.Schedulers;

Observable
        .interval(1, MILLISECONDS)
        .observeOn(Schedulers.computation())
        .subscribe(
                x -> sleep(Duration.ofSeconds(1)));
The exception is gone! So is our throughput... The application stalls after a while, staying in an endless GC loop. You see, Observable in RxJava 1.x has assertions (bounded queues, checks, etc.) all over the place, making sure you are not overflowing anywhere. For example observeOn() operator in 1.x has a queue limited to 128 elements by default. When backpressure is properly implemented across the whole stack, observeOn() operator asks upstream to deliver not more than 128 elements to fill in its internal buffer. Then separate threads (workers) from this scheduler are picking up events from this queue. When queue becomes almost empty, observeOn() operator asks (request() method) for more. This mechanism breaks apart when producer does not respect backpressure requests and sends more data than it was allowed, effectively overflowing the consumer. The internal queue inside observeOn() operator is full, yet interval() operator keeps emitting new events - because that's what interval() is suppose to do.

Observable in 1.x discovers such overflow and fails fast with MissingBackpressureException. It literally means: I tried so hard to keep the system in healthy state, but my upstream is not respecting backpressure - backpressure implementation is missing. However Observable in 2.x has no such safety mechanism. It's a vanilla stream that hopes you will be a good citizen and either have slow producers or fast consumers. When system is healthy, both Observables behave the same way. However under load 1.x fails fast, 2.x fails slowly and painfully.

Does it mean RxJava 2.x is a step back? Quite the contrary! In 2.x an important distinction was made:

  • Observable doesn't care about backpressure, which greatly simplifies its design and implementation. It should be used to model streams that can't support backpressure by definition, e.g. user interface events
  • Flowable does support backpressure and has all the safety measures in place. In other words all steps in computation pipeline make sure you are not overflowing the consumer.
2.x makes an important distinction between streams that can support backpressure ("can slow down if needed" in simple words) and those that don't. From the type system perspective it becomes clear what kind of source are we dealing with and what are its guarantees. So how should we migrate our interval() example to RxJava 2.x? Easier than you think:

Flowable
        .interval(1, MILLISECONDS)
        .observeOn(Schedulers.computation())
        .subscribe(
                x -> sleep(Duration.ofSeconds(1)));
That simple. You may ask yourself a question, how come Flowable can have interval() operator that, by definition, can't support backpressure? After all interval() is suppose to deliver events at constant rate, it can't slow down! Well, if you look at the declaration of interval() you'll notice:

@BackpressureSupport(BackpressureKind.ERROR)
Simply put this tells us that whenever backpressure can no longer be guaranteed, RxJava will take care of it and throw MissingBackpressureException. That's precisely what happens when we run Flowable.interval() program - it fails fast, as opposed to destabilizing whole application.

So, to wrap up, whenever you see an Observable from 1.x, what you probably want is Flowable from 2.x. At least unless your stream by definition does not support backpressure. Despite same name, Observables in these two major releases are quite different. But once you do a search and replace from Observable to Flowable you'll notice that migration isn't that straightforward. It's not about API changes, the differences are more profound.

There is no simple Flowable.create() directly equivalent to Observable.create() in 2.x. I made a mistake myself to overuse Observable.create() factory method in the past. create() allows you to emit events at an arbitrary rate, entirely ignoring backpressure. 2.x has some friendly facilities to deal with backpressure requests, but they require careful design of your streams. This will be covered in the next FAQ.

Comments

  1. FYI, there is the Flowables.intervalBackpressure operator in the extensions project that doesn't signal MissingBackpressureException and emits values based on time and demand:

    https://github.com/akarnokd/RxJava2Extensions#flowablesintervalbackpressure

    The thing to note is that the emission pattern may no longer be completely periodic and the signals may arrive in quick succession if there was a temporal lack of requests.

    ReplyDelete
  2. I once asked you a related question on Twitter, but still unsure - does your RxJava book cover 2.x flavor? If not - do you consider it still relevant or you plan for 2.x edition?

    ReplyDelete
    Replies
    1. Dear Igor, our RxJava book unfortunately covers 1.x. I have no plans so far to publish a 2nd edition, but we'll see :-).

      Delete

Post a Comment