Should subscribeOn and observeOn only be invoked by the final subscriber?

Question

The Scheduling and Threading section of Intro to Rx says that

the use of SubscribeOn and ObserveOn should only be invoked by the final subscriber

It also says that in an UI application, the presentation layer, which is normally the final subscriber, should be the one to invoke these methods.

I am wondering if the advice is solid, since I see some situations in which this is not convenient:

1. For starters, I don't believe that the presentation layer should decide where an Observable coming from the data layer should be subscribed. In my opinion, the presentation layer should be unaware if the data is coming from a database, from a REST API, or from memory. For this reason, it's convenient for the data layer to call subscribeOn() before returning the Observable, passing the IO Scheduler or the immediate Scheduler as convenient.
2. If the presentation layer gets the Observable from some service or use case (which in turn gets it from the data layer) and this service decides that it needs to process the stream in some computation Scheduler, why should the presentation layer care about this?
3. What about a stream that is originally coming from the UI, so it needs to be subscribed in the UI thread. Then it will be sent to some service to do some work and finally come back to the presentation layer to be observed in the UI thread. That would require the UI stream to be subscribeOn() the UI Scheduler, then observeOn() some other Scheduler, and finally observeOn() the UI Scheduler. In this case, being able to invoke subscribeOn() and observeOn() only in the final subscriber would mean that the stream can only be processed in the UI thread.

Is there some good reason why I should sacrifice the architecture of my application and ignore Rx's ability to easily switch threads by invoking these two methods only by the final subscriber?

Answer 1

Great to see you have read the book and are taking the time to challenge some of the guidance there.

The reason I give this guidance is because

1. Concurrency is hard, and having a simple rule helps teams produce better code
2. Concurrency is hard, and having a single place to locate your concurrency concerns allows for a better mental model of your stack/layering and should simplify testing. The more layers that introduce concurrency in your application, the worse
3. Blocking the UI thread is not good news. Getting off the UI-Thread asap, and then delaying any processing of data back on the UI as late as possible is preferable. This pattern aims to serve that goal.

These are obviously my opinions but I have seen these simple guidelines help clean up code on dozens of projects, reduce code bases, improve test ability, improve predictability and in numerous case massively improve performance.

Sadly, it is difficult to put together case studies of these projects as most of them are protected by NDAs.

I would be keen to see how it works for you or how you apply an alternate pattern.