rxJava 2 instrumentation

[mcculls]

This instrumentation takes roughly the same approach as the latest reactor instrumentation:

• active parent spans are captured when observables are created
• captured parent spans are activated when the observable sends events to observers
• no spans are created for the observables themselves

The general idea is to propagate context so that spans resulting from reactive work created under an active span use that span as their parent, even if the reactive work is triggered later on in the application.

In the future it would be nice to store some kind of continuation structure in the context store rather than the parent span. I did look into using State/ConcurrentState but ran into activation issues when an observable sends multiple events or has multiple observers - so I'd prefer to get this initial approach merged first (with tests) and then iterate on it.

[william-tran]

Speak of the devil, I was just looking at issues around not seeing traces propagate in Java Specialagent after using an RxJava2 based MQTT client. Your implementation is quite different from https://github.com/opentracing-contrib/java-specialagent/tree/master/rule/rxjava-2. Is that because Specialagent is missing a lot of cases?

[mcculls]

@william-tran I haven't looked at that instrumentation, but afaict it only handles observables and not all of the other rxJava constructions. Also the goal of this particular instrumentation is not so much to do with adding named spans for each step in the stream, but propagating the relevant context around.

[bantonsson]

I get a slightly uneasy feeling from Spans being moved around and reactivated, and potentially finished early by other code. Looking forward to the follow-up with State.

[richardstartin]

@bantonsson For the record, that's my position too.

[mcculls]

FWIW I agree the eventual solution will involve something that looks very much like State but fits better with the reactive model - at which point we can also apply it to the reactor instrumentation.