When calling `to_string()` on an identifier, the returned string will
have an `r#` prefix if the original identifer was a raw identifier. This
prefix needs to be removed if the identifier is interpolated into
another identifier (e.g. `__some_prefix__r#other_ident`) in order for
the new identifier to be valid.
This issue was previously masked due to the fact that the relevant test
uses `r#i32`. It's possible to use `i32` as a normal (non-raw) ident -
due to a bug in rustc, this means that the user-supplied `r#` prefix
will be lost when calling `to_string()`. This bug will eventually be
fixed, causing `to_string()` to start returning `r#i32` instead of
`i32`.
This commit strips the `r#` prefix (if present) from an identifier
before using it to construct a new identifier. The relevant test is
updated to use `r#match`, which actually requires the `r#` prefix to be
a valid identifier. This causes the test to fail without this patch on
current versions of Rust.
This completely overhauls the error system used in async-graphql.
- `Error` has been renamed to `ServerError` and `FieldError` has been
renamed to just `Error`. This is because `FieldError` is by far the most
common error that users will have to use so it makes sense to use the
most obvious error name. Also, the current name didn't make sense as it
was used for things other than field errors, such as the data callback
for websockets.
- `ServerError` has been made completely opaque. Before it was an enum
of all the possible errors, but now it just contains an error message,
the locations, the path and extensions. It is a shame that we lose
information, it makes more sense as _conceptually_ GraphQL does not
provide that information. It also frees us to change the internals of
async-graphql a lot more.
- The path of errors is no longer an opaque JSON value but a regular
type, `Vec<PathSegment>`. The type duplication of `PathSegment` and
`QueryPathSegment` is unfortunate, I plan to work on this in the future.
- Now that `ServerError` is opaque, `RuleError` has been removed from
the public API, making it simpler.
- Additionally `QueryError` has been completely removed. Instead the
error messages are constructed ad-hoc; I took care to never repeat an
error message.
- Instead of constructing field-not-found errors inside the
implementations of field resolvers they now return `Option`s, where a
`None` value is representative of the field not being found.
- As an unfortunate consequence of the last change, self-referential
types based on the output of a subscription resolver can no longer be
created. This does not mean anything for users, but causes lifetime
issues in the implementation of merged objects. I fixed it with a bit of
a hack, but this'll have to be looked into further.
- `InputValueError` now has a generic parameter - it's kind of weird but
it's necessary for ergonomics. It also improves error messages.
- The `ErrorExtensions` trait has been removed. I didn't think the
`extend` method was necessary since `From` impls exist. But the
ergonomics are still there with a new trait `ExtendError`, which
is implemented for both errors and results.
- `Response` now supports serializing multiple errors. This allows for
nice things like having multiple validation errors not be awkwardly
shoved into a single error.
- When an error occurs in execution, data is sent as `null`. This is
slightly more compliant with the spec but the algorithm described in
<https://spec.graphql.org/June2018/#sec-Errors-and-Non-Nullability> has
yet to be implemented.