Fix the Error trait

text/0000-fix-error.md

@@ -0,0 +1,246 @@
+- Feature Name: fix_error
+- Start Date: 2018-07-18
+- RFC PR: (leave this empty)
+- Rust Issue: (leave this empty)
+
+# Summary
+[summary]: #summary
+
+Change the `std::error::Error` trait to improve its usability. Introduce a
+backtrace module to the standard library to provide a standard interface for
+backtraces.
+
+# Motivation
+[motivation]: #motivation
+
+The `Error` trait has long been known to be flawed in several respects. In
+particular:
+
+1. The required `description` method is limited, usually, to returning static
+strings, and has little utility that isn't adequately served by the required
+`Display` impl for the error type.
+2. The signature of the `cause` method does not allow the user to downcast the
+cause type, limiting the utility of that method unnecessarily.
+3. It provides no standard API for errors that contain backtraces (as some
+users' errors do) to expose their backtraces to end users.
+
+We propose to fix this by deprecating the existing methods of `Error` and adding
+two new, provided methods. As a result, the undeprecated portion of the `Error`
+trait would look like this:
+
+```rust
+trait Error: Display + Debug {
+    fn backtrace(&self) -> Option<&Backtrace> {
+        None
+    }
+
+    fn source(&self) -> Option<&dyn Error + 'static> {
+        None
+    }
+}
+```
+
+# Guide-level explanation
+[guide-level-explanation]: #guide-level-explanation
+
+## The new API of the Error trait
+
+The new API provides three main components:
+
+1. The Display and Debug impls, for printing error messages. Ideally, the
+   Display API would be focused on *end user* messages, whereas the Debug impl
+   would contain information relevant to the programmer.
+2. The backtrace method. If the Error contains a backtrace, it should be exposed
+   through this method. Errors are not required to contain a backtrace, and are
+   not expected to.
+3. The `source` method. This returns another Error type, which is the underlying
+   source of this error. If this error has no underlying source (that is, it is
+   the "root source" of the error), this method should return none.
+
+## The backtrace API
+
+This RFC adds a new `backtrace` module to std, with one type, with this API:
+
+```rust
+pub struct Backtrace {
+    // ...
+}
+
+impl Backtrace {
+    pub fn new() -> Backtrace {
+        // ...
+    }
+}
+
+impl Display for Backtrace {
+    // ...
+}
+
+impl Debug for Backtrace {
+    // ...
+}
+```
+
+This minimal initial API is just intended for printing backtraces for end users.
+In time, this may grow the ability to visit individual frames of the backtrace.
+
+### Backtrace controlling environment variables
+
+Today, the `RUST_BACKTRACE` controls backtraces generated by panics. After this
+RFC, it also controls backtraces generated in the standard library: no backtrace
+will be generated when calling `Backtrace::new` unless this variable is set.
+
+Two additional variables will be added: `RUST_PANIC_BACKTRACE` and
+`RUST_STD_BACKTRACE`: these will independently override the behavior of
+`RUST_BACKTRACE` for backtraces generated for panics and from the std API.
+
+## The transition plan
+
+Deprecating both `cause` and `description` is a backward compatible change, and
+adding provided methods `backtrace` and `source` is also backward compatible.
+We can make these changes unceremoniously, and the `Error` trait will be much
+more functional.
+
+We also change the default definition of `cause`, even though it is deprecated:
+
+```rust
+fn cause(&self) -> Option<&dyn Error> {
+    self.source()
+}
+```
+
+This way, if an error defines `source`, someone calling the deprecated `cause`
+API will still get the correct cause type, even though they can't downcast it.
+
+# Reference-level explanation
+[reference-level-explanation]: #reference-level-explanation
+
+## Why `cause` -> `source`
+
+The problem with the existing `cause` API is that the error it returns is not
+`'static`. This means it is not possible to downcast the error trait object,
+because downcasting can only be done on `'static` trait objects (for soundness
+reasons).
+
+## Note on backtraces
+
+The behavior of backtraces is somewhat platform specific, and on certain
+platforms backtraces may contain strange and inaccurate information. The
+backtraces provided by the standard library are designed for user display
+purposes only, and not guaranteed to provide a perfect representation of the
+program state, depending on the capabilities of the platform.
+
+## How this impacts failure
+
+The failure crate defines a `Fail` trait with an API similar to (but not
+exactly like) the API proposed here. In a breaking change to failure, we would
+change that trait to be an extension trait to `Error`:
+
+```rust
+// Maybe rename to ErrorExt?
+trait Fail: Error + Send + Sync + 'static {
+    // various provided helper methods
+}
+
+impl<E: Error + Send + Sync + 'static> Fail for E {
+
+}
+```
+
+Instead of providing a derive for Fail, failure would provide a derive for the
+std library Error trait, e.g.:
+
+```rust
+#[derive(Debug, Display, Error)]
+#[display = "My display message."]
+struct MyError {
+    #[error(source)]
+    underlying: io::Error,
+    backtrace: Backtrace,
+}
+```
+
+The exact nature of the new failure API would be determined by the maintainers
+of failure, it would not be proscribed by this RFC. This section is just to
+demonstrate that failure could still work using the std Error trait.
+
+# Drawbacks
+[drawbacks]: #drawbacks
+
+This causes some churn, as users who are already using one of the deprecated
+methods will be encouraged (by warnings) to change their code, and library
+authors will need to revisit whether they should override one of the new
+methods.
+
+# Rationale and alternatives
+[rationale-and-alternatives]: #rationale-and-alternatives
+
+## Provide a new error trait
+
+The most obvious alternative to this RFC would be to provide an entirely new
+error trait. This could make deeper changes to the API of the trait than we are
+making here. For example, we could take an approach like `failure` has, and
+impose stricter bounds on all implementations of the trait:
+
+```rust
+trait Fail: Display + Debug + Send + Sync + 'static {
+    fn cause(&self) -> Option<&dyn Fail> {
+        None
+    }
+
+    fn backtrace(&self) -> Option<&Backtrace> {
+        None
+    }
+}
+```
+
+Doing this would allow us to assemble a more perfect error trait, rather than
+limiting us to the changes we can make backwards compatibly to the existing
+trait.
+
+However, it would be much more disruptive to the ecosystem than changing the
+existing trait. We've already seen some friction with failure and other APIs
+(like serde's) that expect to receive something that implements `Error`. Right
+now, we reason that the churn is not worth slight improvements that wouldn't be
+compatible with the Error trait as it exists.
+
+In the future, if these changes are not enough to resolve the warts with the
+Error trait, we could follow this alternative: we would deprecate the Error
+trait and introduce a new trait then. That is, accepting this RFC now does not
+foreclose on this alternative later.
+
+## Bikeshedding the name of `source`
+
+The replacement for `cause` could have another name. The only one the RFC author
+came up with is `origin`.
+
+# Prior art
+[prior-art]: #prior-art
+
+This proposal is largely informed by our experience with the existing Error
+trait API, and from the experience of the `failure` experiment. The main
+conclusions we drew:
+
+1. The current Error trait has serious flaws.
+2. The `Fail` trait in failure has a better API.
+3. Having multiple error traits in the ecosystem (e.g. both `Error` and `Fail`)
+can be very disruptive.
+
+This RFC threads the needle between the problem with the Error trait and the
+problems caused by defining a new trait.
+
+# Unresolved questions
+[unresolved-questions]: #unresolved-questions
+
+This RFC intentionally proposes a most minimal API. There are a number of API
+extensions we could consider in the future. Prominent examples:
+
+1. Extending the backtrace API to allow programmatic iteration of backtrace
+frames and so on.
+2. Providing derives for traits like `Display` and `Error` in the standard
+libray.
+3. Providing helper methods on `Error` that have been experimented with in
+failure, such as the causes iterator.
+
+None of these are proposed as a part of *this* RFC, and would have a future RFC
+discussion.