Benefit of marking classes as final?

[gaborcsardi]

I am wondering about the benefits of marking some classes final, e.g. Span in the SDK.

I am embedding opentelemetry-cpp into an R package, to bring OpenTelemetry to R. It is working great in general. Some things would be easier if I could subclass some of the SDK classes, though. E.g. I could work around #2667 easily. Especially if some private members could move to protected.

So I am wondering why these classes are marked as final, and if that restricts creating opentelemetry-cpp wrappers in other languages. If that's a goal in opentelemetry-cpp at all.

[lalitb]

Marking classes as final and keeping things private helps keep the SDK’s public API small and manageable. If we made things protected, we’d end up exposing a lot more of the internals, which means we’d have to support and maintain those details forever—making future changes harder.

The idea is that if you want to extend OpenTelemetry C++, you do it by implementing the public API interfaces, not by subclassing the SDK itself. This way, it’s clear how you can plug in your own logic, and we can keep improving the SDK without worrying about breaking user subclasses.

[gaborcsardi]

Thanks, this makes sense.

For the case where opentelemetry-cpp is embedded (=vendored), breaking changes are not that important, because they only materialize when the vendored opentelemetry-cpp is updated.

But I guess it is also true that I can patch opentelemetry-cpp in this case. This would be some form of "implementing the public API interfaces".

[ThomsonTan]

Could you please share the concrete example, like why need extends the SpanData class in SDK. Just in case if there is alternative for solving the problem.

[gaborcsardi]

For Span I would like to have an API in R that automatically sets the status code of a span. This is possible because in R you can attach a hook to the termination of an R function, and then this hook would set the status depending on whether the R function is terminating successfully or not. But I also want to support setting the status code manually, because some spans are errors even if there is no exception thrown, e.g. a span for a HTTP 404 in the client. So I want to disable setting the automatic status if the user has set the status manually. But there is no way to query a Span object if its status has been set or not, and I cannot extend the class to add this functionality.

Maybe there is another way to do this, IDK.

Another example is making the minimum severity work in the logger. This is broken in opentelemetry-cpp AFAICT, and I cannot fix it because the classes are final.

But as I said, I understand your reasons, and I am fine with patching the vendored library.