GroupBy with flatMap and observeOn does not seem to support backpressure in 1.2.x

[sbarlabanov]

There are a plenty of issues fixing some sorts of backpressure problems in groupBy (e.g. #3428, #3425).
But nevertheless when I try the following code it requests all 10000 values from the source at once without waiting for processing of single items (example running with rxjava 1.2.x):

import rx.Observable;
import rx.schedulers.Schedulers;

import java.util.Random;

import static rx.Observable.range;

public class GroupByTest {
    public static void main(String[] args) throws Exception {
        Observable<Integer> source = range(1, 10000);
        source
                .doOnNext(i -> System.out.println("Requested " + i))
                .groupBy(v -> v % 5)
                .flatMap(g -> g.observeOn(Schedulers.io()).map(GroupByTest::calculation))
                .subscribe(i -> System.out.println("Got " + i));
        Thread.sleep(100000);
    }

    private static Integer calculation(Integer i) {
        sleep();
        System.out.println("Processing " + i);
        return i * 20;
    }

    private static void sleep() {
        try {
            Thread.sleep(new Random().nextInt(1000));
        } catch (InterruptedException e) {
            e.printStackTrace();
        }
    }
}

When running the code I see 10000 "Requested X" messages getting printed immediately and afterwards messages "Processing X" coming after some timeout.
With this behavior it does not seem to be possible to parallelize processing of single groups in rxjava in any useful way when working with large or unlimited sources.

[sbarlabanov]

It does work correctly with rxjava 2.x Flowable.

[akarnokd]

For historical reasons, flatMap in 1.x requests an unbounded amount of source elements for its 1 argument overload. There is another overload that takes a maxConcurrent parameter with which you can limit the number of active inner sources.

[sbarlabanov]

I know ;). But it does not work either :(. If I add maxConcurrent (1 or 2 or 3 etc.) to flatMap after groupBy the behavior stays the same - all 10000 items are requested at once.

    public static void main(String[] args) throws Exception {
        Observable<Integer> source = range(1, 10000);
        source
                .doOnNext(i -> System.out.println("Requested " + i))
                .groupBy(v -> v % 5)
                .flatMap(g -> g.observeOn(Schedulers.io()).map(GroupByTest::calculation), 4) // <---- maxConcurrent
                .subscribe(i -> System.out.println("Got " + i));
        Thread.sleep(100000);
    }

[sbarlabanov]

In RxJava 1.x there is OperatorGroupByTest which tests some backpressure scenarios.
If I adjust https://github.com/ReactiveX/RxJava/blob/1.x/src/test/java/rx/internal/operators/OperatorGroupByTest.java#L1027 by putting doOnNext with System.out.println I observe the same behavior as I described above - all items are requested at once.
So OperatorGroupByTest#testGroupByBackpressure does not test any backpressure ;).

[akarnokd]

OperatorGroupBy L247 keeps requesting from upstream and thus consumes it in an unbounded manner. Would you like to work out a fix?

[akarnokd]

This is a bug. Fix posted in #5030.

[sbarlabanov]

Looks good. Your fix seems to work :).
I tried my examples from above and also tested in a more complex scenario from a customer project (consuming from Kafka, grouping by partition and processing every partition in its own thread).
Is any action required from my side?