Why I Prefer Exceptions to Error Values

I think I disagree with everything here.

Exceptions Are Much Easier to Work With

Well, they're "easier" in the same way that dynamic typing is easier. It's obviously less work initially just to say "screw it; any error gets caught in main()". But that's short term easiness. In the long term its much more painful because:

1. You don't know which functions might produce errors, and therefore you don't know where you should be even trying to handle errors. (Checked exceptions are the answer here but they are relatively rarely used in practice.)
2. Actually handling errors properly often involves responding to errors from individual function calls - at least adding human readable context to them. That is stupidly tedious with exceptions. Every line turns into 5. Sometime it makes the code extremely awkward:

try {
   int& foo = bar();
} catch (...) {
   std::cout << "bar failed, try ...\n";
   return nullopt;
}
foo = 5;

(It actually gets worse than that but I can't think of a good example.)

Over 100× less code! [fewer exception catching in their C++ database than error handling in a Go database]

Well... I'm guessing your codebase is a lot smaller than the other one for a start, and you're comparing with Go which is kind of worst case... But anyway this kind of proves my point! You only actually have proper error handling in 140 places - apparently mostly in tests. In other words you just throw all exceptions to main().

System errors [he's mainly talking about OOM, stack overflow & arithmetic errors like integer overflow]

Kind of a fair point I guess. I dunno how you can reasonably stack overflows without exceptions. But guess what - Rust does have panic!() for that, and you can catch panics. I'd say that's one of the few reasonable cases to use catch_unwind.

Exceptions Lead to Better Error Messages

Hahahahahaha. I dunno if a bare stack trace with NullPointerException counts as a "better error message". Ridiculous.

Exceptions Are More Performant

Sure maybe in error handling microbenchmarks, or particularly extreme examples. In real world code it clearly makes little difference. Certainly not enough to choose an inferior error handling system.

I would say one real reason to prefer exceptions over Result<>s is they are a fair bit easier to debug because you can just break on throw. That's tricky with Result<> because creating a Err is not necessarily an error. At least I have not found a way to "break on Err". You can break on unwrap() but that is usually after the stack has been unwound quite a bit and you lose all context.

The guy keeps on picking on Go, which is infamous for having terrible error handling, and then he has the nerve to even pick on the UNIX process return convention, which was designed in the 70s.
The few times he mentions Rust, for whatever reason he keeps on assuming that .unwrap() is the only choice, which's use is decidedly discouraged in production code.

I do think there is room for debate here. But error handling is a hellishly complex topic, with different needs between among others:

• short- vs. long-running processes
• API vs. user-facing
• small vs. big codebase
• library vs. application code
• prototyping vs. production phase

And even if you pick out a specific field, the two concepts are not clearly separated.
Error values in Rust usually have backtraces these days, for example (unless you're doing embedded where this isn't possible).
Or Java makes you list exceptions in your function signature (except for unchecked exceptions), so you actually can't just start throwing new exceptions in your little corner without the rest of the codebase knowing.
I find it quite difficult to properly define the differences between the two.

The more I read about these kind of article the more I am amazed that our digital future is at hand in utterly incompetent people.

This person clearly have no understanding of monadic error (AKA Maybe/option monad or slightly more advanced Either monad), which is the first monad we teach at a class targeting second year undergrad.

The performance comparison is just plain factual error. The functional error code will continue to compute n2 when computation of n1 failed; the same do not happen in the exception version. If you compare codes with completely different traces, of course they will have different performance...

A properly implemented monadic error will return as soon as compute for n1 failed, and never execute the rest of the code. This is the default and idiomatic behavior in Haskell, OCaml, F#, and rust. This performance problem doesn't even happen in LINQ-style handling like in C# and Kotlin (maybe also Typescript?).

The point of monadic error is that its control flow is local, whereas exception is non-local. Specifically, the exception can be handled and occur anywhere in the code base, with no indication on the type level. So programmers will be constantly worrying about whether the exception in a function call is properly handled.

Even worse, when you try to catch a certain error, there is always the risk to accidentally catch similar exceptions in a library call or someone else's code. Writing good code with try-catch requires a lot of principle and style guides. But unlike monads, these principle and rules cannot be enforced by the type system, adding extra burden to programmers.

In fact, we have known for a long time that non-local control flows (goto, break, contiune, exception, long jump) are the breeding ground for spaghetti code. As an evidence, many non-local control flows (goto, long jump) are baned in most languages.

That being said, there are certainly cases, with proper documentation, that exception style is easy to write and understand. But I think they are very specific scenarios, which have to be justified on a case-by-case basis.

My anecdotal experience is that Rust code, for example, has more calls to unwrap than I’d like. The problem here is that simply unwrapping results will crash the program on errors that could have been a user-visible error message with exceptions.

unwrap() is explicitly not handling the error in a Result type. If you must do this, then at least use except(), to unwrap the code but with an error message if program crashes. Its the equivalent of having Exceptions and then not handling that exception. Therefore your critique is not valid here.

One problem with Exceptions is, you never know what code your function or library calls that can produce an exception. It's not encoded in the type system or signature of the function. So you need to pray and try catch all possible exceptions (I look at this from Pythons perspective), if you don't want a Catch..All, which then you wouldn't know what error this actually is. And you still don't know where this error came from or happened in the code, how deep in the function call chain? Instead Errors as Values means its encoded in type system and you can directly see what errors the function can cause and (in Rusts case) you must handle the error, otherwise program won't compile. You don't need to handle anything else in this context. Compiler ensures that all possible errors are handled (again within context of our discussion). Vast improvement!

In C++, ignoring anything that any other language provides...

I mean, yeah, if your language does not support error values, do not use them.

One problem with exceptions is composability.

You have to rely on good and up-to-date documentation or you have to dig into the source code to figure out what exceptions are possible. For a lot of third party dependencies (which constitute a huge part of modern software), both can be missing.

Error type is a mitigation, but you are free to e.g. panic in Rust if you think the error is unrecoverable.

A third option is to have effect types like Koka, so that all possible exceptions (or effects) can be checked at type level. A similar approach can be observed in practical (read: non-academic) languages like Zig. It remains to be seen whether this style can be adopted by the mainstream.

Memory isn’t infinite, CPUs can’t process all integers, and Santa isn’t real

Wait, what? Need a spoiler tag.

It would be nice to include Zig's approach in the comparison. I've only just begun learning it, but the syntax seems pretty elegant from what I've seen so far.

Upvoting not because I share author's preference, but because I'm interested in reading other people's perspectives on the topic.

The article made a few good points, but a good amount of it was conjecture. I liked the part about comparing the two functions and showing that exceptions are faster but I think a big thing he's not getting is readability. Even in the functions he showed, you can directly see that the one using std::expected has the happy path and error path directly in the function signature, whereas the exception one doesn't.

As for the "error kind" trap he was talking about, that definitely exists, but ignores the fact that you can also get this same kind of error from exceptions. I've definitely gotten exceptions that I didn't understand from Python or Java libraries, but it's not a problem with exceptions but a problem with how they're shown. If there's nothing to tell me that I should have thought of that error, it shouldn't be an expectation for a dev to have thought of it.

Look at gleam and elixir. Both are functional. Both use exceptions, but both also use error values as well. There is no reason why we can't have both. These are incredibly fault tolerant systems.

They both have their place. I just recently discovered a bug in lemmy bot I wrote where the lemmy API module will raise an Exception if login fails (response status code != 200), which feels extremely out of place, as the error/status code do matter in that case.

Other times exceptions make more sense as Phillip pointed out. It's easier faster to ask for forgiveness than permission after all.

I didn’t read everything, but I mostly agree with the author, especially on this point:

While you can definitely abuse exceptions, functional-style error values are not a one-size-fits-all solution.

There are time and place for both. I think exceptions are good for bigger errors. Like database connection errors. Things that shouldn’t happen without any easy backup plan. Those errors might need to be escalated as high as possible where proper action can be made (like resetting the database connection and everything relying on it).

Functional style is great for smaller stuff. Like key not found in hash maps. In many cases there might be good defaults that can be used instead.

They make valid points, and maybe it makes sense to always prefer them in their context.

I don't think exceptions always lead to better error handling and messages though. It depends on what you're handling.

A huge bin of exception is detailed and has a lot of info, but often lacks context and concise, obvious error messages. When you catch in outer code, and then have a "inaccessible resource" exception, it tells you nothing. You have to go through the stack trace and analyze which cases could be covered.

If explicit errors don't lead to good handling I don't think you can expect good exception throwing either. Both solutions need adequate design and implementation to be good.

Having a top-level (in their server context for one request or connection) that handles and discards one context while the program continues to run for others is certainly simple. Not having to propagate errors simplifies the code. But it also hides error states and possibilities across the entire stack between outer catch and deep possible throw.

In my (C#) projects I typically make conscious decisions between error states and results and exceptional exceptions where basic assumptions or programming errors exist.

Does the performance cost of error checking/result types they discovered in C++ apply to languages that have native result and option types like Rust?

I would hope they were able to find efficient, performant implementations, and that branch prediction picks the expected non-error branch in most cases.

Oof, some of these comments. Sorry on behalf of the edge lords, OP.

I'm just going to comment on the face value of the title itself, and make assumptions otherwise.

• Exceptions are control flow mechanism. I.e. that can be used for code execution flow, in the same application.

• Error codes are useful across some API boundary.

Does this adequately cover whatever it is they figured out was a good tradeoff?