I mean, failure literally says that it is experimental, that's the first thing it says. And much of its explorations were then integrated into the standard Error trait.
anyhow makes it much clearer that it targets convenience for applications: it aims to be a better `Box<dyn Error>`, not to help with structured error handling but to propagate errors out of the way.
I've been using failure and would like to know if I should switch.
Also is it even possible to standardize error handling best practices across entirely different software components like an OS kernel, a web service and a CLI tool?
OS kernel should not panic? Service absolutely should panic and restart to maintain availability and consistency if something unexpected happens (see Erlang) and a CLI tool has a human in the loop, sometimes.
There is, and this is exactly what has been happening. Early crates showed some of the weaknesses in the Error trait, and then they were addressed. Failure was created, showed more holes, and they were addressed. Now anyhow/thiserr/other stuff have happened, and their improvements may or may not make it upstream.
> Also is it even possible to standardize error handling best practices across entirely different software components like an OS kernel, a web service and a CLI tool?
Some of the details differ, but it feels like we're at that point to me.
> OS kernel should not panic? Service absolutely should panic ...
Both of these are specific implementation choices, and some people will choose different options for each of them. It just really depends. But with Result (and to some extent, the Error trait), there are means to share the bits that make sense. The core of it is broadly applicable even if some of the details change.
I agree that "what to do about errors" differs a lot in those domains (and even more in some others, like embedded). That's not a reason not to standardize - it's something the standard needs to take into account.
Ideally, if I am writing a bit of code that could be in more than one of those domains, there is something standard I can do about errors that will provide sufficient flexibility to my users that I can act appropriately to the domain I actually find myself in. Without too much complexity.
To me, there is a large difference between them.
Result is for computations that can fail, such as opening a file or reading from a socket. Computations that fail should be handled by the caller (e.g. report a nice error which file could not be opened or attempting to open a new socket and retry the communication).
panic, in particular through assertions, are for catching programming errors, such as violation of invariants. These are cases where the program should just crash and burn, because it would be in an invalid state otherwise. E.g. splitting a Vec at an invalid index is a programming error. If a program attempts such a split, it is already in an invalid state, so it should crash and the bug should be fixed.
What I usually do in other languages is use exceptions as assertions, when it is just a simple one-off app, but then go back and catch the exceptions when possible to make it more user friendly and robust.
The Rust equivalent is letting Result::Err bubble up the call chain and have a
fn main() -> Result<(), Error>
or even nicer: fn main() -> anyhow::Result<()>It's possible to catch panics during unwinding[0], but that will not work in every case: `panic='abort'` is a completely valid configuration for a Rust compilation.
> What I usually do in other languages is use exceptions as assertions
Rust assertion macros will panic. `unwrap` methods are also, functionally, assertions. And they panic.
> when it is just a simple one-off app, but then go back and catch the exceptions when possible to make it more user friendly and robust. It seems that exceptions have been omitted on purpose, but is there another way to express this idiom in Rust?
That is very much not idiomatic rust[1] and I would recommend against it, but it is possible for applications as you'd hopefully be controlling the compilation configuration.
For libraries however it's a good way to shoot everyone in the foot: you simply can not assume that users of your library will `panic='unwind'`, so it is a very dangerous thing to do.
[0] https://doc.rust-lang.org/std/panic/fn.catch_unwind.html
[1] https://doc.rust-lang.org/book/ch09-03-to-panic-or-not-to-pa...
If you don’t want to handle Results, just ? all the way up and out of main. Then come back and deal with them as necessary when you’re ready.
It’s far nicer than dealing with exceptions.
Of course there are a few rough edges, but nothing major. I also don't mind having to wrap the successful result of a fn in Ok(...).
I hope "do a few minor things, add better documentation, but leave the language itself as it is" is also one of the possible outcomes...