How to communicate between agent context and application context

[wallezhang]

The situation is: I create a new context in my application by Context.current(), and set this new context into a Map for cross thread communication. While read the context from Map by the agent instrumentation in another thread, a ClassCastException is thrown: class io.opentelemetry.javaagent.instrumentation.opentelemetryapi.context.AgentContextWrapper cannot be cast to class io.opentelemetry.javaagent.shaded.io.opentelemetry.context.Context.

I wonder if I should not put Context directly into the map, but replace it with the text converted by TextMapPropagator. Or is there any other good way?

[mateuszrzeszutek]

Hey @wallezhang ,

The javaagent should propagate context automatically over threads in most cases (e.g. Context.current() called in a Runnable should return the context captured at the time when the runnable was submitted to an executor). Does that not work for you?

I create a new context in my application by Context.current(), and set this new context into a Map for cross thread communication. While read the context from Map by the agent instrumentation in another thread, a ClassCastException is thrown: class io.opentelemetry.javaagent.instrumentation.opentelemetryapi.context.AgentContextWrapper cannot be cast to class io.opentelemetry.javaagent.shaded.io.opentelemetry.context.Context.

The application and the agent are using two different copies of the OpenTelemetry API. The agent one is shaded (thus the io.opentelemetry.javaagent.shaded package prefix). The application OTel API is being bridged by the opentelemetry-api instrumentation; it effectively implements the application's interfaces and forwards all the calls to the (shaded) agent API.
The AgentContextWrapper class that you're seeing is a part of the opentelemetry-api instrumentation; it implements the application Context interface and bridges all the calls over to the agent Context (the shaded one).

When you put the value of Context.current() into some kind of shared map (on the application side), you actually put an instance of AgentContextWrapper there. So, when you retrieve it from the javaaagent code space, the object not implement the agent's Context interface (and thus the class cast error). You could use AgentContextStorage.getAgentContext() to extract the actual agent context from an AgentContextWrapper. In general, I'd recommend avoiding this kind of app-agent communication if possible (just try to put all the parts either on the application side or the agent side, not both of them at once), as coordinating it gets rather bothersome.

[mateuszrzeszutek]

Ah, good point. We publish unshaded instrumentations to make referencing/using them easier - if we shaded the application... package, we'd have to shade the agent copy of the API too, otherwise they'd conflict. And in general we don't want our users to worry about shading - in most cases, custom extension developers don't really have to think about it, they can reference unshaded APIs and the agent will handle the shading/translation in the runtime.

[wallezhang]

Oh, thanks, I see why published dependency was not shaded. I'll just have to go and make a duplicate of shaded OTel API module then, or put all the logic into the agent extension.

[wallezhang]

😢 There would be another problem here, I make a copy of the opentelemetry-api-shaded module, I need to shade application.io.opentelemetry to io.opentelemetry and then shade the io.opentelemetry used in the agent extension to io.opentelemetry.shaded when agent extension is compiled. But then, when the extension is loaded via -Dotel.javaagent.extension, it will replace io.opentelemetry in the class file with io.opentelemetry.shaded again by the io.opentelemetry.javaagent.tooling. RemappingUrlConnection class.

If I don't shade application.io.opentelemetry, extension will refer to application.io.opentelemetry... at runtime, this couldn't work. Maybe also shade application.io.opentelemetry in RemappingUrlConnection class?

[mateuszrzeszutek]

If I don't shade application.io.opentelemetry, extension will refer to application.io.opentelemetry... at runtime, this couldn't work. Maybe also shade application.io.opentelemetry in RemappingUrlConnection class?

Hm, that's a fair point. Can you create a bug for this?

[wallezhang]

Ok, I'll create a bug and glad to fix it.