Dariusz Jędrzejczyk

Related issue: #3633; PR ignoring this test for the time being: #3690

`shouldNotReportSchedulerScheduledFutureTask` is also flaky.

Interestingly, the combination of `replay` with `refCount` might be problematic as pointed out in [one summary by Simon](https://github.com/reactor/reactor-core/issues/3017#issuecomment-1108809550). This reflection came a few years after the test was introduced to...

For debugging locally, I added new RuntimeException("onError called!").printStackTrace(); to `FluxRefCountGrace$RefCountInner.onError` and got this during a failing run: ``` java.lang.RuntimeException: onError called! at reactor.core.publisher.FluxRefCountGrace$RefCountInner.onError(FluxRefCountGrace.java:270) at reactor.core.publisher.FluxReplay$SizeBoundReplayBuffer.replayNormal(FluxReplay.java:869) at reactor.core.publisher.FluxReplay$SizeBoundReplayBuffer.replay(FluxReplay.java:970) at reactor.core.publisher.FluxReplay$ReplaySubscriber.dispose(FluxReplay.java:1405) at...

Thanks @MikkelHJuul – please have a look at the linked PR (#3707), I suppose it incorporates the same cancellation validation, although using `Flux#next()` which implicitly cancels the source.

#3707 resolves this.

Hey @Hailong it's been a while, are you still willing to push the contribution forward?

@Hailong I incorporated my suggestion and merged. Thank you for the contribution.

@OlegDokuka the commit [0334959](https://github.com/reactor/reactor-core/pull/3634/commits/033495922f1a92396def5ea6a240afd2035ba756) adds significant improvements to the test suite that helps catch the racy situations. In the next commit I will add temporary fixes for the identified issues,...

For the latest changes, I added a JMH benchmark. The idea is to simply test a non-contended, single threaded case of pulling 100 items and packaging them in 1, 10,...