Errors not being delivered in Android Scheduler

[vinc3m1]

I'm running into intermittent cases when using AndroidSchedulers.mainThread() (which is just a wrapper around AndroidSchedulers.handerThread()) where the onError event is not being delivered to the main thread.

I have a gist showing my logs: https://gist.github.com/vinc3m1/dc9232653a22f8f2d900

It seems the onError event is propagated all the way up until .observeOn(AndroidSchedulers.mainThread(). After which in some (seemingly random) cases it's not delivered. And the cause appears to be that innerSubscription is reporting itself to be unsubscribed.

So my main questions are:

1. Should we even be doing the isUnsubscribed check on innerSubscription if we already have handler.removeCallbacks(runnable) to handle unsubscribe events?
2. Should innerSubscription be reporting isUnsubscribed == true when nothing else is unsubscribing from it?

Also, I'm using rather complex combination of map, flatmap, and combineLatest upstream, is it possible that some combinations of these are causing innerSubscription to be unsubscribed?

[vinc3m1]

Ok, dug into it a bit further: looks like unsubscribe() is being called on the Worker immediately after the onError is posted to the Handler thread. This means if the Handler doesn't execute the Runnable fast enough, innerSubscription.isUnsubscribed() will be true and no message delivered.

I can see this being useful for delayed subscriptions where no further messages should be delivered after unsubscribing, but technically all messages are delayed if the main thread gets blocked long enough.

I'm not sure how the non-Andorid schedulers handle this?

[zsxwing]

looks like unsubscribe() is being called on the Worker immediately after the onError is posted to the Handler thread.

I feel there is some operator may handle unsubscribe incorrectly. Could you provide the operators you are using so that we can review them?

Just found you already provided them: map, flatmap, and combineLatest...

[vinc3m1]

also nest, doOnNext, subscribeOn, but I think #1242 is the only place where unsubscribe is handled incorrectly, or do you think there's more to it than that?

[vinc3m1]

Fixed in 0.19