🚫 Errors

The greatest mistake is to imagine that we never err.

— Thomas Carlyle.

Errors in Rust are a bit ambiguous:

• the infamous Result<T, E> is a type that can either be Ok(T) or Err(E), where E is the error type in case something went wrong.
• the std::error::Error trait is a trait that represents errors that can be displayed and have a source (cause).

The ambiguity comes from the fact that the std::error::Error trait is not required to be implemented for the error type E in the Result<T, E> type. This means that one can have a Result<T, E> where E is not an error type. A common example of something else it can be is that it has the same type as the T type, which is not an error type. E.g. in case of a web service middleware a firewall could return a 403 Http response as the Err variant of the Result<T, Response>. Where T is most likely also a Response type. In which case you might as well have Result<Response, Infallible>.

Within Web Services we usually do not want an error type, as it does not make any sense. This is because the server has to respond something (unless you simply want to kill the connection), and so it makes much more sense to enforce the code type-wise to always return a response.

The most tricky scenario, if you can call it that, is what to do for middleware services. These situations are tricky because they can wrap any generic S type, where S is the service type. This means that the error type can be anything, and so it is not possible to create values of that type for scenarios where the error comes from the middleware itself.

There are several possibilities here and we'll go over them next. But before we do that, I do want to emphasise that while Rust's Result<T, E> does not enforce that E is an error type, it is still a good practice to use an error type for the E type. And that is also that as a rule of thumb we do in Rama.

Type Erasure

The BoxError type alias is a boxed Error trait object and can be used to represent any error that implements the std::error::Error trait and is used for cases where it is usually not that important what specific error type is returned, but rather that an error occurred. Boxed errors do allow to downcast to check for concrete error types, but this checks only the top-level error and not the cause chain.

Error Extension

The ErrorExt trait provides a set of methods to work with errors. These methods are implemented for all types that implement the std::error::Error trait. The methods are used to add context to an error, add a backtrace to an error, and to convert an error into an opaque error.

Opaque Error

The OpaqueError type is a type-erased error that can be used to represent any error that implements the std::error::Error trait. Using the OpaqueError::from_display you can even create errors from a displayable type.

The other advantage of OpaqueError over BoxError is that it is Sized and can be used in places where a Sized`` type is required, while [BoxError] is ?Sized` and can give you a hard time in certain scenarios.

Error Context

The ErrorContext allows you to add a context to Result and Option types:

• For Result types, the context is added to the error variant, turning Result<T, E> into Result<T, OpaqueError>;
• For Option types, the context is used as a DisplayError when the open is None, turning Option<T> into Result<T, OpaqueError>.

This is useful when you want to add custom context. And can also be combined with other ErrorExt methods, such as ErrorExt::backtrace to add even more info to the error case, if there is one.

It is also an easy way to turn an option value into the inner value, short-circuiting using ? with the new context (Display) error when the option was None.

Error Context Example

Option Example:

#![allow(unused)]
fn main() {
use rama::error::{ErrorContext, ErrorExt};

let value = Some(42);
let value = match value.context("value is None") {
   Ok(value) => assert_eq!(value, 42),
   Err(error) => panic!("unexpected error: {error}"),
};

let value: Option<usize> = None;
let result = value.context("value is None");
assert!(result.is_err());
}

Result Example:

#![allow(unused)]
fn main() {
use rama::error::{ErrorContext, ErrorExt, OpaqueError};

let value: Result<_, OpaqueError> = Ok(42);
let value = match value.context("get the answer") {
  Ok(value) => assert_eq!(value, 42),
  Err(error) => panic!("unexpected error: {error}"),
};

let value: Result<usize, _> = Err(OpaqueError::from_display("error"));
let result = value.context("get the answer");
assert!(result.is_err());
}

Error Composition

Sometimes it can be useful to compose errors with more expressive error types. In such cases OpaqueError is... too opaque.

In an early design of Rama we considered adding a compose_error function macro that would allow to create error types in a similar manner as the thiserror crate, but we decided against it as it would be an abstraction too much.

Rama was created to give developers the full power of the Rust language to develop proxies, and by extension also web services and http clients. In a similar line of thought it is also important that one has all tools available to create the error types for their purpose.

As such, if you want your own custom error types we recommend just creating them as you would any other type in Rust. The blog article https://sabrinajewson.org/blog/errors gives a good overview and background on this topic.

You can declare your own macro_rules in case there are common patterns for the services and middlewares that you are writing for your project. For inspiration you can see the http rejection macros we borrowed and modified from Axum's extract logic: https://github.com/plabayo/rama/blob/main/rama-http/src/utils/macros/http_error.rs

And of course... if you really want, against our advice in, you can use the thiserror crate, or even the anyhow crate. All is possible.