Box error alias

text/0000-box-error-alias.md

@@ -0,0 +1,170 @@
+- Feature Name: `box_error_alias`
+- Start Date: 2019-11-20
+- RFC PR: [rust-lang/rfcs#0000](https://github.com/rust-lang/rfcs/pull/0000)
+- Rust Issue: [rust-lang/rust#0000](https://github.com/rust-lang/rust/issues/0000)
+
+# Summary
+[summary]: #summary
+
+Add a type alias to std::error of the form
+
+```rust
+pub type BoxError = Box<dyn Error + Send + Sync + 'static>;
+```
+
+# Motivation
+[motivation]: #motivation
+
+Rust error handling has a long and complicated history, and it is still evolving right now. This
+reflects the fact that errors perform a multitude of functions, and these different functions
+require different implementations: for example, if errors are regularly encountered in a real-time
+thread, then one would not want their creation to involve an allocation or the indirection of a
+trait object, whereas for a cli app, most errors may not occur on the happy path, and so
+their performance is inconsequential. Likewise, a low-level library may want fine-grained control
+of error handling and recovery, whereas a cli app may only want to handle a small subset of
+errors, simply printing a message to screen for the rest.
+
+The error trait in rust's standard library provides a helpful way to perform two tasks: firstly it
+requires a Display implementation - so all errors can be logged/written/etc, and secondly it
+provides the error with the option of exposing another inner error, thereby allowing chains of
+errors to be created. [Research into the best way to represent errors](https://internals.rust-lang.org/t/thoughts-on-error-context-in-error-handling-libraries/10349/4)
+is ongoing in rust, and this RFC does not attempt to resolve the discussion around how error
+handling in rust should evolve. Instead, it proposes a simple type alias to serve the following
+motivations
+
+ 1. A unified name for boxed errors, to make them easier to recognise,
+ 2. A short descriptive name for error trait objects, to reduce code noise and cognitive load,
+ 3. A place to show the value of making errors `Send + Sync + 'static`, and
+ 3. A place in the standard library to document the pattern of using a trait object to return any
+    error, aiding discoverability.
+
+Currently what happens in practice is rust programmers either create an alias for
+`Box<dyn Error + Send + Sync + 'static>` in some utility module, where each user will use a
+sligtly different name, or just write out the full type that `BoxFuture` aliases.
+
+This idea evolved on a [thread in irlo](https://internals.rust-lang.org/t/proposal-add-std-boxerror/10953)
+before this RFC was written, and there names such as `AnyError` were suggested, as well as using an
+opaque object with a raw pointer as is the case in the [`failure` crate][`failure`].
+However, this RFC only satisfies the narrow goal of providing a trait alias for `Box<Error + ..>`,
+allowing more experimentation to take place regarding the specific design of an opaque and more
+full-featured error type in the crate ecosystem.
+
+# Guide-level explanation
+[guide-level-explanation]: #guide-level-explanation
+
+Once you start doing anything more advanced than arithmetic with a computer, you will encounter
+routines of the computer that may fail. The source of these failures may be hardware-related (e.g.
+a resistor has shorted in your temperature sensor peripheral), a bug in some code you (or someone
+else) wrote, the misuse of an interface, or a network failure, amongst many others. Ideally we
+design programming languages and libraries so that programming mistakes are caught when the program
+is compiled rather than at run-time, but obviously this is not always possible. It is therefore
+necessary to have run-time error handling.
+
+Not all errors are the same: sometimes an error can be corrected on the fly, sometimes an operation
+must be re-tried (e.g. a tcp packet failed to arrive), and sometimes the error represents something
+catastrophic and the best response is to terminate the program with a useful error message. These
+different errors require different types of data: an error that will be instantly corrected should
+be able to choose whether it is stored on the heap or the stack, whereas for an error that will
+terminate the program the performance impact of a trait object/allocation will be negligable. The
+functionality of errors may be further constrained by the requirements of the programming
+environment: an embedded chip with low memory performing tasks in realtime may not have run-time
+memory allocation as an option.
+
+In general, rust programmers will create different types to handle these different situations. The type
+for a recoverable error may just be some fixed-size block of data, whereas a more serious error may
+contain references to descriptions of the error or other errors that cause it. Rust provides the
+`Error` trait, which includes converting the error into a string and referencing any underlying
+error, but rust programmers can still implement this trait even though they don't want to be allocating
+strings or inner errors, since it is the caller of the methods who causes the allocations/etc. In
+this way, rust allows all errors to be able to describe themselves, which is very useful during
+development and prototyping, as well as helping to support different use cases for library code.
+
+Whist the `Error` trait helps error authors to provide self-describing errors, it does not help
+code using these errors to easily combine them with other errors of different types, and displaying
+the contents of an error that might be one of multiple types. The standard way to solve this
+problem is to create a trait object behind a thick pointer, hence the signature `Box<dyn Error..>`.
+The trait object obscures the actual type, instead presenting the Error interface, and is easily
+created from any error using the `impl From<'a, E: Error + Send + Sync + 'a> for Box<dyn Error +
+Send + Sync + 'a>`.
+
+`BoxError` is simply a synonym for `Box<dyn Error + Send + Sync + 'static>`. Here are some examples
+showing its usefulness.
+
+## Example: Returning from `main`
+
+Here we are getting a database version from a file and returning it. We are in the prototyping
+stage and don't want to devote too much time to complex error handling.
+
+```rust
+fn db_version() -> Result<usize, BoxError> {
+    let version = fs::read_to_string("/path/to/version/file")?
+        .parse()?;
+    Ok(version)
+}
+```
+
+# Reference-level explanation
+[reference-level-explanation]: #reference-level-explanation
+
+The implementation is very straight forward, simply the code in the [summary].
+
+# Drawbacks
+[drawbacks]: #drawbacks
+
+ 1. Increases std api surface area, although semantically there is no change since it is a type
+    alias.
+ 2. Could potentially encourage people to use it instead of a specialized type like
+    [`anyhow::Error`](https://docs.rs/anyhow/latest/anyhow/struct.Error.html).
+
+   - I would challenge this drawback with the argument that even after the addition of a
+     hypothetical `AnyError`, there will still be times when a `BoxError` is appropriate for its
+     simplicity. It may turn out that this satisfies 99% of use cases for type-erased errors, and
+     any more specialized solutions can live in crates.
+
+# Rationale and alternatives
+[rationale-and-alternatives]: #rationale-and-alternatives
+
+This RFC proposes a very small addition that provides a type alias that many people define
+manually. The only real alternative is to do nothing and maybe land a more specialized type-erased
+error type in the future.
+
+The rationale for not including a specialized error type is that it is not clear what the best
+design for it would be, especially since the story around backtraces is not yet finalised. It's
+also not clear if a specialized type is necessary. The disadvantages of `BoxError` over a
+hypothetical `AnyError` are
+
+  1. An increased memory footprint (increasing the size of `Result` in the case that the `Ok`
+     variant is small), but with fewer memory indirections. If this is an issue, then a concrete
+     error type is more appropriate.
+  2. The `Debug` implementation defers to the inner error type. The hypothetical `AnyError` could
+     use the inner error's `Display` type, leading to better messages in the
+     `fn main() -> Result<(), AnyError>` case.
+
+
+# Prior art
+[prior-art]: #prior-art
+
+There are many error models in the rust ecosystem, including many type-erased error types. Crates
+include [`anyhow`] and [`failure`] among others. Prior art for a type synonym is the many crates
+that define it internally.
+
+An example of prior art of name choice is the [`BoxFuture`] type in the [`futures`] crate.
+
+# Unresolved questions
+[unresolved-questions]: #unresolved-questions
+
+ - Should the error be `Send + Sync`. It seems like a good default, and nothing is stopping rust
+   programmers simply not using the alias if they want a `Box<dyn Error + 'static>`.
+
+# Future possibilities
+[future-possibilities]: #future-possibilities
+
+There are many ways that the Error trait could be enhanced, for example by providing a dedicated
+`AnyError` type, or providing a method like `Error::wrap` that would take an error, and make it the
+source of an `AnyError`. This RFC attempts to be very focused on a small change, and so these ideas
+are better discussed elsewhere.
+
+[`anyhow`]: https://github.com/dtolnay/anyhow
+[`failure`]: https://github.com/rust-lang-nursery/failure
+[`BoxFuture`]: https://docs.rs/futures/0.3.1/futures/future/type.BoxFuture.html
+[`futures`]: https://github.com/rust-lang-nursery/futures-rs