Backpressure sensitivity of cache() in Hystrix

[akarnokd]

I've recently implemented backpressure in RxJava's cache() operator which unfortunately failed in Netfix' canary environment: threads got accumulated over time probably due to hangs in the downstream. I've also run Hystrix' unit test with 1.0.13RC and everything passed.

I've looked at the AbstractCommand but it looks complicated and I don't really see what is subscribing to the cached Observable that could hang. The addition of backpressure causes the operator to not send any values unless requested; if there is a downstream operator that forgets to request, it will hang (previously, it just got the next value). The second change is that cache replays its buffer live instead of accumulating a synchronous source first, causing changes in timing of the value delivery.

[mattrjacobs]

Thanks for the report, @akarnokd. Sorry for the delay in responding - was out of the country on vacation and just got back. I will investigate and pick up the relevant context from @benjchristensen and @abersnaze.

[mattrjacobs]

To add some more context to this thread, the initial change to RxJava was ReactiveX/RxJava#2969. It was canaried in a release candidate in: ReactiveX/RxJava#3076. The revert was done here: ReactiveX/RxJava#3081

[mattrjacobs]

This was not an issue with Hystrix. I don't know the precise issue that manifested in the Netflix internal canary, but it was resolved without a Hystrix change