RFC 0201: Error chaining

libs | lang (traits | macros | error-handling)

Summary

This RFC improves interoperation between APIs with different error types. It proposes to:

The proposed changes are all library changes; no language changes are needed -- except that this proposal depends on multidispatch happening.

Motivation

Typically, a module (or crate) will define a custom error type encompassing the possible error outcomes for the operations it provides, along with a custom Result instance baking in this type. For example, we have io::IoError and io::IoResult<T> = Result<T, io::IoError>, and similarly for other libraries. Together with the try! macro, the story for interacting with errors for a single library is reasonably good.

However, we lack infrastructure when consuming or building on errors from multiple APIs, or abstracting over errors.

Consuming multiple error types

Our current infrastructure for error handling does not cope well with mixed notions of error.

Abstractly, as described by this issue, we cannot do the following:

fn func() -> Result<T, Error> {
    try!(may_return_error_type_A());
    try!(may_return_error_type_B());
}

Concretely, imagine a CLI application that interacts both with files and HTTP servers, using std::io and an imaginary http crate:

fn download() -> Result<(), CLIError> {
    let contents = try!(http::get(some_url));
    let file = try!(File::create(some_path));
    try!(file.write_str(contents));
    Ok(())
}

The download function can encounter both io and http errors, and wants to report them both under the common notion of CLIError. But the try! macro only works for a single error type at a time.

There are roughly two scenarios where multiple library error types need to be coalesced into a common type, each with different needs: application error reporting, and library error reporting

Application error reporting: presenting errors to a user

An application is generally the "last stop" for error handling: it's the point at which remaining errors are presented to the user in some form, when they cannot be handled programmatically.

As such, the data needed for application-level errors is usually related to human interaction. For a CLI application, a short text description and longer verbose description are usually all that's needed. For GUI applications, richer data is sometimes required, but usually not a full enum describing the full range of errors.

Concretely, then, for something like the download function above, for a CLI application, one might want CLIError to roughly be:

struct CLIError<'a> {
    description: &'a str,
    detail: Option<String>,
    ... // possibly more fields here; see detailed design
}

Ideally, one could use the try! macro as in the download example to coalesce a variety of error types into this single, simple struct.

Library error reporting: abstraction boundaries

When one library builds on others, it needs to translate from their error types to its own. For example, a web server framework may build on a library for accessing a SQL database, and needs some way to "lift" SQL errors to its own notion of error.

In general, a library may not want to reveal the upstream libraries it relies on -- these are implementation details which may change over time. Thus, it is critical that the error type of upstream libraries not leak, and "lifting" an error from one library to another is a way of imposing an abstraction boundaries.

In some cases, the right way to lift a given error will depend on the operation and context. In other cases, though, there will be a general way to embed one kind of error in another (usually via a "cause chain"). Both scenarios should be supported by Rust's error handling infrastructure.

Abstracting over errors

Finally, libraries sometimes need to work with errors in a generic way. For example, the serialize::Encoder type takes is generic over an arbitrary error type E. At the moment, such types are completely arbitrary: there is no Error trait giving common functionality expected of all errors. Consequently, error-generic code cannot meaningfully interact with errors.

(See this issue for a concrete case where a bound would be useful; note, however, that the design below does not cover this use-case, as explained in Alternatives.)

Languages that provide exceptions often have standard exception classes or interfaces that guarantee some basic functionality, including short and detailed descriptions and "causes". We should begin developing similar functionality in libstd to ensure that we have an agreed-upon baseline error API.

Detailed design

We can address all of the problems laid out in the Motivation section by adding some simple library code to libstd, so this RFC will actually give a full implementation.

Note, however, that this implementation relies on the multidispatch proposal currently under consideration.

The proposal consists of two pieces: a standardized Error trait and extensions to the try! macro.

The Error trait

The standard Error trait follows very the widespread pattern found in Exception base classes in many languages:

pub trait Error: Send + Any {
    fn description(&self) -> &str;

    fn detail(&self) -> Option<&str> { None }
    fn cause(&self) -> Option<&Error> { None }
}

Every concrete error type should provide at least a description. By making this a slice-returning method, it is possible to define lightweight enum error types and then implement this method as returning static string slices depending on the variant.

The cause method allows for cause-chaining when an error crosses abstraction boundaries. The cause is recorded as a trait object implementing Error, which makes it possible to read off a kind of abstract backtrace (often more immediately helpful than a full backtrace).

The Any bound is needed to allow downcasting of errors. This RFC stipulates that it must be possible to downcast errors in the style of the Any trait, but leaves unspecified the exact implementation strategy. (If trait object upcasting was available, one could simply upcast to Any; otherwise, we will likely need to duplicate the downcast APIs as blanket impls on Error objects.)

It's worth comparing the Error trait to the most widespread error type in libstd, IoError:

pub struct IoError {
    pub kind: IoErrorKind,
    pub desc: &'static str,
    pub detail: Option<String>,
}

Code that returns or asks for an IoError explicitly will be able to access the kind field and thus react differently to different kinds of errors. But code that works with a generic Error (e.g., application code) sees only the human-consumable parts of the error. In particular, application code will often employ Box<Error> as the error type when reporting errors to the user. The try! macro support, explained below, makes doing so ergonomic.

An extended try! macro

The other piece to the proposal is a way for try! to automatically convert between different types of errors.

The idea is to introduce a trait FromError<E> that says how to convert from some lower-level error type E to Self. The try! macro then passes the error it is given through this conversion before returning:

// E here is an "input" for dispatch, so conversions from multiple error
// types can be provided
pub trait FromError<E> {
    fn from_err(err: E) -> Self;
}

impl<E> FromError<E> for E {
    fn from_err(err: E) -> E {
        err
    }
}

impl<E: Error> FromError<E> for Box<Error> {
    fn from_err(err: E) -> Box<Error> {
        box err as Box<Error>
    }
}

macro_rules! try (
    ($expr:expr) => ({
        use error;
        match $expr {
            Ok(val) => val,
            Err(err) => return Err(error::FromError::from_err(err))
        }
    })
)

This code depends on multidispatch, because the conversion depends on both the source and target error types. (In today's Rust, the two implementations of FromError given above would be considered overlapping.)

Given the blanket impl of FromError<E> for E, all existing uses of try! would continue to work as-is.

With this infrastructure in place, application code can generally use Box<Error> as its error type, and try! will take care of the rest:

fn download() -> Result<(), Box<Error>> {
    let contents = try!(http::get(some_url));
    let file = try!(File::create(some_path));
    try!(file.write_str(contents));
    Ok(())
}

Library code that defines its own error type can define custom FromError implementations for lifting lower-level errors (where the lifting should also perform cause chaining) -- at least when the lifting is uniform across the library. The effect is that the mapping from one error type into another only has to be written one, rather than at every use of try!:

impl FromError<ErrorA> MyError { ... }
impl FromError<ErrorB> MyError { ... }

fn my_lib_func() -> Result<T, MyError> {
    try!(may_return_error_type_A());
    try!(may_return_error_type_B());
}

Drawbacks

The main drawback is that the try! macro is a bit more complicated.

Unresolved questions

Conventions

This RFC does not define any particular conventions around cause chaining or concrete error types. It will likely take some time and experience using the proposed infrastructure before we can settle these conventions.

Extensions

The functionality in the Error trait is quite minimal, and should probably grow over time. Some additional functionality might include:

Features on the Error trait

The main reason not to include the above two features is so that Error can be used with extremely minimal data structures, e.g. simple enums. For such data structures, it's possible to produce fixed descriptions, but not mutate descriptions or other error properties. Allowing generic creation of any Error-bounded type would also require these enums to include something like a GenericError variant, which is unfortunate. So for now, the design sticks to the least common denominator.

Concrete error types

On the other hand, for code that doesn't care about the footprint of its error types, it may be useful to provide something like the following generic error type:

pub struct WrappedError<E> {
    pub kind: E,
    pub description: String,
    pub detail: Option<String>,
    pub cause: Option<Box<Error>>
}

impl<E: Show> WrappedError<E> {
    pub fn new(err: E) {
        WrappedErr {
            kind: err,
            description: err.to_string(),
            detail: None,
            cause: None
        }
    }
}

impl<E> Error for WrappedError<E> {
    fn description(&self) -> &str {
        self.description.as_slice()
    }
    fn detail(&self) -> Option<&str> {
        self.detail.as_ref().map(|s| s.as_slice())
    }
    fn cause(&self) -> Option<&Error> {
        self.cause.as_ref().map(|c| &**c)
    }
}

This type can easily be added later, so again this RFC sticks to the minimal functionality for now.