"duplication is cheaper than the wrong abstraction" is only true if the duplication: a) isn't also an abstraction (which, in go, it is in this case) and b) is "contained" (i.e. not overly used), which it often is not in non-trivial go code.
Go's errors package is riddled with `reflect` and other inefficient code constructs. Printing an error (e.g. to stdout or the log) is fine. But if you want to actually do anything with the error in the code (e.g. to discriminate/branch based on the error), you have to resort to bloated types that implement a poor interface or else rely on string inspection. In either case, you are using the underlying errors package. In small scale applications it's fine, but when you're doing anything at even modest scale and your application encounters errors, it's going to introduce measurable performance degradation.
>only true if the duplication: a) isn't also an abstraction (which, in go, it is in this case)
Can you give an example? I don't follow. I generally don't like abstractions for the sake of abstractions. I use Go's interfaces for a very specific reason: when I want dynamic dispatch, but with nice static typing guarantees. Interfaces in languages without structural typing force you to design a rigid, unflexible hierarchy/ontology well in advance, which only gets in your way when requirements change.
>is "contained" (i.e. not overly used)
Can you show why "containment" is necessary and why "overuse" is a bad thing?
We have a lot of Go services in production and errors have never been a performance issue. In the happy path, when there're no errors, errors are basically no-op. We don't use errors for control flow, though; only for exceptional situations, which aren't triggered often. If we want to branch based on an error, we use errors.Is. I don't remember ever having to inspect an error's string, that sounds like a hack. Usually, branching on an error's type is a rare scenario, even if it uses reflection, you usually just bubble up the error. In practice, at runtime, Go's error handling is just a bunch of TEST RAX, RAX instructions. Do you have benchmarks to show otherwise?
Go's errors package is riddled with `reflect` and other inefficient code constructs. Printing an error (e.g. to stdout or the log) is fine. But if you want to actually do anything with the error in the code (e.g. to discriminate/branch based on the error), you have to resort to bloated types that implement a poor interface or else rely on string inspection. In either case, you are using the underlying errors package. In small scale applications it's fine, but when you're doing anything at even modest scale and your application encounters errors, it's going to introduce measurable performance degradation.