> I really miss Java’s exception safety in Kotlin.
Java is the only language that I'm aware of with checked exceptions (Swift has "throws" which I guess goes in that direction but is untyped) and it's in my experience not notably exception-safer than languages that use values for error handling (such as Rust, Haskell or Kotlin).
Swallowing exceptions because of bad ergonomics or misuse by library authors is unfortunately fairly common in a lot of Java code that I've seen. By contrast, most Kotlin code that I've seen uses exceptions only for unrecoverable errors and otherwise uses some sort of sum type (Result, Either or some sealed class/interface). These can be handled ergonomically and in a very type-safe manner - and importantly they don't break when used with higher-order functions. In the future, Kotlin will have Rich Errors which will have even better ergonomics (I'm personally fine with Either, but I know some people dislike a heavily functional style, so this will be good for them).
> Swallowing exceptions because of bad ergonomics or misuse by library authors is unfortunately fairly common in a lot of Java code that I've seen
This is because the language has not invested to make the ergonomics better. Checked exceptions/errors are still superior than runtime exceptions. "Misuse" can still occur with values. Bad developers can still bubble errors they can't handle and that you can't handle either. I've written elsewhere [0] the things that Java needs.
> By contrast, most Kotlin code that I've seen uses exceptions only for unrecoverable errors
Except for the whole Kotlin standard library. Like you said the errors as unions proposal is very far away. So either I need to adopt some library and then when unions come I now have 2-3 competing systems + whatever the libraries are using. I'd rather just use checked exceptions.
> - and importantly they don't break when used with higher-order functions.
This is not a property of values or exceptions. It's a property of the type system. Checked exceptions can work perfectly fine with higher order functions, Scala has done good research there [1] and I believe Swift's typed throws works with lambdas as well; Java just hasn't invested in the language.
Ultimately, there is no difference because Result<T, Error> and T func() throws Exception. One is not superior to the other. What it comes down to for me is whether a language has support for checked errors or not, and Kotlin does not.
Checked errors are not categorically better than runtime exceptions. There are lots of things that should never be checked exceptions: division by zero, illegal array access, OOM, and everything that can't be (at least not locally) recovered from.
I'm not sure what you mean. Most kotlin stdlib functions that can fail have some sort of "...orNull" variant to be used when you're not sure you're going to get something back, and there's "runCatching" which will wrap any exceptions in a Result type.
Result is part of the stdlib, so you don't have to use any library, although some people (me included) prefer Either. If/when they introduce Rich Errors, it's going to be semantically almost equivalent to Either, except you get somewhat improved ergonomics, so it's not really a paradigm shift.
> Ultimately, there is no difference because Result<T, Error> and T func() throws Exception. One is not superior to the other. What it comes down to for me is whether a language has support for checked errors or not, and Kotlin does not.
I don't understand how those two sentences don't contradict each other. Kotlin has support for Result<T> (stdlib) or Either<E, T> (from libraries like Arrow). Since you're agreeing that this is equivalent to checked exceptions I don't understand how you can claim that Kotlin doesn't support checked errors. It supports them just as well as Java does, it just has a different philosophy around it. (I guess the one point that I'll concede is that it can cause issues when interfacing with Java code and it would be nice if Kotlin had some better ergonomics around automatically inferring errors from Java library code the way it does for nullability checks; but in practice this hasn't been a big issue for me).
FWIW, I disagree that there isn't a big distinction. If the exception is part of the regular type, it is a value and it can be manipulated like any other value, making it very easy to write all sorts of code that manipulates exceptions. Of course, you could eventually write enough machinery so that you get the same power also with checked exceptions, but that just makes the language more complicated for questionable gain. And in any case, Java doesn't do that.
The Scala discussion you're linking is certainly interesting, but it's cutting-edge PL research (effect systems). I'm not sure how this matters in a discussion of tradeoffs between Java and Kotlin today.
Additionally, the above-linked JEP only proposes to make a small subset of types potentially null-safe. The goal there is performance, not correctness.
It will eventually be a goal. Java does things in steps. First will be introducing the syntax, then another JEP will take care of the standard library. Just like they introduced virtual threads and then introduced structured concurrency. Or like introducing records and then introducing pattern matching over records. Baby steps.
If Java ever ships Valhalla we might get null restricted types: https://openjdk.org/jeps/8316779