RxJava重试时出现奇怪的行为

我正在玩 RxJava重试时间 http://reactivex.io/documentation/operators/retry.html操作员。网上关于它的资料很少，唯一值得一提的是this http://blog.danlew.net/2016/01/25/rxjavas-repeatwhen-and-retrywhen-explained。这也不足以探索我想了解的各种用例。我还加入了异步执行和回退重试，以使其更加现实。

我的设置很简单：我有一堂课ChuckNorrisJokesRepository从 JSON 文件返回随机数量的 Chuck Norris 笑话。我的测试类是ChuckNorrisJokesService如下所示。我感兴趣的用例如下：

1. 第一次尝试成功（无需重试）
2. 重试 1 次后失败
3. 尝试重试 3 次，但第二次成功，因此不会重试第三次
4. 第三次重试成功

Note: 该项目可以在我的GitHub https://github.com/asarkar/java/blob/master/concurrency-learning/src/main/java/org/abhijitsarkar/rx/ChuckNorrisJokesService.java.

ChuckNorrisJokesService.java:

@Slf4j
@Builder
public class ChuckNorrisJokesService {
    @Getter
    private final AtomicReference<Jokes> jokes = new AtomicReference<>(new Jokes());

    private final Scheduler scheduler;
    private final ChuckNorrisJokesRepository jokesRepository;
    private final CountDownLatch latch;
    private final int numRetries;
    private final Map<String, List<String>> threads;

    public static class ChuckNorrisJokesServiceBuilder {
        public ChuckNorrisJokesService build() {
            if (scheduler == null) {
                scheduler = Schedulers.io();
            }

            if (jokesRepository == null) {
                jokesRepository = new ChuckNorrisJokesRepository();
            }

            if (threads == null) {
                threads = new ConcurrentHashMap<>();
            }

            requireNonNull(latch, "CountDownLatch must not be null.");

            return new ChuckNorrisJokesService(scheduler, jokesRepository, latch, numRetries, threads);
        }
    }

    public void setRandomJokes(int numJokes) {
        mergeThreadNames("getRandomJokes");

        Observable.fromCallable(() -> {
            log.debug("fromCallable - before call. Latch: {}.", latch.getCount());
            mergeThreadNames("fromCallable");
            latch.countDown();

            List<Joke> randomJokes = jokesRepository.getRandomJokes(numJokes);
            log.debug("fromCallable - after call. Latch: {}.", latch.getCount());

            return randomJokes;
        }).retryWhen(errors ->
                errors.zipWith(Observable.range(1, numRetries), (n, i) -> i).flatMap(retryCount -> {
                    log.debug("retryWhen. retryCount: {}.", retryCount);
                    mergeThreadNames("retryWhen");

                    return Observable.timer(retryCount, TimeUnit.SECONDS);
                }))
                .subscribeOn(scheduler)
                .subscribe(j -> {
                            log.debug("onNext. Latch: {}.", latch.getCount());
                            mergeThreadNames("onNext");

                            jokes.set(new Jokes("success", j));
                            latch.countDown();
                        },
                        ex -> {
                            log.error("onError. Latch: {}.", latch.getCount(), ex);
                            mergeThreadNames("onError");
                        },
                        () -> {
                            log.debug("onCompleted. Latch: {}.", latch.getCount());
                            mergeThreadNames("onCompleted");

                            latch.countDown();
                        }
                );
    }

    private void mergeThreadNames(String methodName) {
        threads.merge(methodName,
                new ArrayList<>(Arrays.asList(Thread.currentThread().getName())),
                (value, newValue) -> {
                    value.addAll(newValue);

                    return value;
                });
    }
}

为简洁起见，我将仅显示第一个用例的 Spock 测试用例。看我的GitHub https://github.com/asarkar/java/blob/master/concurrency-learning/src/test/groovy/org/abhijitsarkar/rx/ChuckNorrisJokesServiceSpec.groovy对于其他测试用例。

def "succeeds on 1st attempt"() {
    setup:
    CountDownLatch latch = new CountDownLatch(2)
    Map<String, List<String>> threads = Mock(Map)
    ChuckNorrisJokesService service = ChuckNorrisJokesService.builder()
            .latch(latch)
            .threads(threads)
            .build()

    when:
    service.setRandomJokes(3)
    latch.await(2, TimeUnit.SECONDS)

    Jokes jokes = service.jokes.get()

    then:
    jokes.status == 'success'
    jokes.count() == 3

    1 * threads.merge('getRandomJokes', *_)
    1 * threads.merge('fromCallable', *_)
    0 * threads.merge('retryWhen', *_)
    1 * threads.merge('onNext', *_)
    0 * threads.merge('onError', *_)
    1 * threads.merge('onCompleted', *_)
}

这失败了：

Too few invocations for:

1 * threads.merge('fromCallable', *_)   (0 invocations)
1 * threads.merge('onNext', *_)   (0 invocations)

我期待的是fromCallable被调用一次，就成功了，onNext被调用一次，然后是onCompleted。我缺少什么？

P.S.：完全披露 - 我也将这个问题发布在RxJava GitHub https://github.com/ReactiveX/RxJava/issues/4207.

经过几个小时的故障排除并在 ReactiveX 成员 David Karnok 的帮助下，我解决了这个问题。

retryWhen是一个复杂的，甚至可能是有缺陷的操作符。官方文档和至少一个答案在这里使用range运算符，如果没有重试，则立即完成。看我的讨论 https://github.com/ReactiveX/RxJava/issues/4207与大卫·卡诺克。

该代码可以在我的GitHub https://github.com/asarkar/java/blob/master/concurrency-learning/src/main/java/org/abhijitsarkar/rx/ChuckNorrisJokesService.java完成以下测试用例：

1. 第一次尝试成功（无需重试）
2. 重试 1 次后失败
3. 尝试重试 3 次，但第二次成功，因此不会重试第三次
4. 第三次重试成功