Two different versions of a crate interacting leads to unhelpful error messages

[huonw]

https://github.com/huonw/bad-error-messages-a contains a crate a. https://github.com/huonw/bad-error-messages-bc contains two crates, b and c.

• a contains a trait Foo and a function test that requires that trait.
• b depends on a specific revision (8b532f5, not the HEAD) of a and contains a type Bar that implements a::Foo.
• c depends on a's HEAD (happens to be 84cfe230) and b, and tries to use b::Bar with a::test.

c fails to compile despite Bar "clearly" implementing Foo... it's even displayed in the docs!

Cloning https://github.com/huonw/bad-error-messages-bc and running cargo build in the root (which is c) gives

   Compiling a v0.0.1 (https://github.com/huonw/bad-error-messages-a#84cfe230)
   Compiling a v0.0.1 (https://github.com/huonw/bad-error-messages-a?rev=8b532f5#8b532f51)
   Compiling b v0.0.1 (file:///home/huon/projects/test-rust/error-messages/c)
   Compiling c v0.0.1 (file:///home/huon/projects/test-rust/error-messages/c)
src/lib.rs:5:5: 5:12 error: the trait `a::Foo` is not implemented for the type `b::Bar` [E0277]
src/lib.rs:5     a::test(b::Bar);
                 ^~~~~~~
error: aborting due to previous error

There's no indication of the fundamental problem from either rustdoc or rustc: that there's two different versions of a being used, meaning a#84cfe230::Foo and a#8b532f5::Foo are different. Bar only implements the latter, but in c, the name a refers to a#84cfe230 so a::test(...) requires the former. (Using name#rev to represent the crate with that version.)

There is an indication of the difference in the example above, since a is compiled twice, but that comes from cargo, not rustc. This issue is focusing on the fact that rustc itself does not give helpful errors and/or allow tools like cargo to control those errors.

It would be nice if rustc:

• disambiguated ambiguous names, probably by filepath/name by default
• allowed cargo to specify semantically relevant info for the disambiguation e.g. pass in the version/revision/..., so that rustc can print foo#0.2.3 instead of the filepath/name
• detected errors that may be caused by ambiguous names and noted it explicitly, e.g. when there is a error involving an ambiguous crate print (once) some extra lines like "note: there's multiple versions of crate ... in use": along with a listing of the possibilities.

(NB. I've used cargo to construct this example because it is the easiest way I can get two different versions of a crate. This appears in rust-lang/rust's makefiles too, where the std test runner is called "std" internally, but links against another crate called "std", i.e. the conventional standard library.)

[Hoverbear]

This bit me today with uuid! It took me a few minutes before I even understood that it wasn't "my" code that was giving this off.

[Kimundi]

Related, its also annoying if only crate local paths get emitted, like in backtraces.

We need a way for crates to identify themself uniquely, and not just by an identifier. Maybe a scheme like <crate foo #ef12f212>::bar::baz?

[golddranks]

I was bitten by this too.

Kimundi: Agreed for the need to unique identification. But then again, in some cases the local path tells more directly where the problem is. I wonder if there are cases where path is better and cases where a hash is better. Or maybe some kind of hybrid: If version information from cargo is available, use that. If not, if version control information is available, use commit SSH, and if no other information exists, use a path...?

[huonw]

Nominating for high priority. I think this is one of the main pain points people have when using cargo, and one of the major reasons for people complaining about cargo allowing multiple versions of a crate. It's usually easy to identify once you've seen it a few times, but it is a horribly confusing and annoying experience.

(Tagged with both compiler and tools since cargo presumably plays into this, but this is probably most relevant to the compiler subteam.)

@golddranks: the question of connecting semantically relevant info to crates is discussed a little in the issue. I think the best plan would external tools to pass in arbitrary strings connected to each crate, defaulting to the path, meaning rustc doesn't have to actually understand anything, just print strings (the compiler already has an --extern NAME=PATH argument that tools like cargo use, it could be extended to also allow --extern NAME#"ID"=PATH or something).

[steveklabnik]

I've also seen an increasing number of IRC questions about this.

On Sep 7, 2015, at 20:58, Huon Wilson notifications@github.com wrote:

Nominating for high priority. I think this is one of the main pain points people have when using cargo, and one of the major reasons for people complaining about cargo allowing multiple versions of a crate. It's usually easy to identify once you've seen it a few times, but it is a horribly confusing and annoying experience.

(Tagged with both compiler and tools since cargo presumably plays into this, but this is probably most relevant to the compiler subteam.)

@golddranks: the question of connecting semantically relevant info to crates is discussed a little in the issue. I think the best plan would external tools to pass in arbitrary strings connected to each crate, defaulting to the path, meaning rustc doesn't have to actually understand anything, just print strings (the compiler already has an --extern NAME=PATH argument that tools like cargo use, it could be extended to also allow --extern NAME#"ID"=PATH or something).

―
Reply to this email directly or view it on GitHub.

[Manishearth]

I'll take a crack at this

[Manishearth]

I think at the very least, for mismatched type and unimplemented trait errors, if we detect a crate mismatch, we should mention it. Any other such errors?

[steveklabnik]

Mismatched type is the big one.

Being able to understand this would also help Rustdoc a lot, I think there are bugs open. (That way, we could know if something is a re-export or not)

[Manishearth]

No no no github, I said "partial fix"

[Manishearth]

Anyway, #28300 fixes the error when there's a type mismatch. "Can't find trait" will be a tad trickier

[hexsel]

I had a corner-case posted on Reddit that was even more confusing:

src/main.rs:9:45: 9:47 error: mismatched types:
 expected `&cookie::jar::CookieJar<'_>`,
 found `&cookie::jar::CookieJar<'_>`

I was using Hyper with the latest cookie crate. I did it because even though 'hyper' depends on 'cookie=0.1.21' or somesuch, I couldn't use 'cookie' on my program. I'll suggest a change on cargo if it doesn't yet exist.

[Manishearth]

Yeah, this should be fixed but the just-landed PR.

[nikomatsakis]

triage: P-medium -- we think that the most important cases here have been covered. If this issue continues to arise frequently on IRC, we could bump up the priority to high.

[crumblingstatue]

Sometimes this can suggest private paths, which makes this even more confusing.

This snippet is taken from an IRC conversation with a (rightly) confused user:

expected struct `gfx_graphics::glyph::GlyphCache`, found struct `gfx_graphics::GlyphCache`
                 ^ --- This path is private                      ^ --- This is the publicly available path

Ref: #21934

[crumblingstatue]

I suggest we at least temporarily mitigate this by applying @huonw's third suggestion from the list, which shouldn't require interfacing with anything external.

detected errors that may be caused by ambiguous names and noted it explicitly, e.g. when there is a error involving an ambiguous crate print (once) some extra lines like "note: there's multiple versions of crate...in use": along with a listing of the possibilities.

[jmesmon]

Some of the issues here (around misleading module paths being reported) may be related to #34769

[shepmaster]

This has started popping up more frequently again on Stack Overflow due to a slow shift towards Serde 1.0 as some crates depend on Serde 0.8, 0.9, or 1.0, all of which are incompatible.

[Mark-Simulacrum]

@Manishearth I'm unassigning you since I don't think you are actively working on this and probably don't have it in a todo list or anything like that?

[frankmcsherry]

Experienced this again ("doesn't implement trait", due to disagreement about which type we were talking about). Asked about status on IRC and got "Lints are not really a priority", which I think would be too bad.

[Manishearth]

This isn't a lint.

[frankmcsherry]

I'm not sure what we are talking about, but what I meant by "this" is the missed opportunity to explain an error to the user. I agree that this is in no sense a lint, and in the future I'll explain this to people on IRC who try and speak for the project?

[zmanian]

What @Manishearth is trying to communicate is that there is a policy level of pushing most lints to Clippy or other tools rather than integrating them into the compiler.

This seems more like an area where the type errors could be improved.

[steveklabnik]

New lints need an RFC, but this isn't a lint; it's an error message that should be improved. They're different things.

[frankmcsherry]

Let me try again. I experienced a form of the issue described, though the version where a trait is implemented for one version of a type and not another (which may not be what was partially fixed). When asking about this issue on IRC, it was stated (by a regular, possibly not in a formal capacity) that "Lints are not really a priority".

This is only meant to be another "ping" about how Rust's usually quite good error messages still have some limitations, and wasn't meant to be about who thinks it was a lint, or whether its standing as a lint makes it less interesting, or anything like that. Obviously, feel free to hash that out amongst yourselves, but not on my account. :)

[Manishearth]

I'm not sure what we are talking about,

I'm meant that this is not a lint, this is diagnostics, so "lints are not a priority" doesn't apply here and whoever gave you that status was wrong about it. Though it's possible that diagnostics are no longer a priority here either (I'd be surprised, given the ergonomics initiative and prior focus on diagnostics by the docs team).

[JustAPerson]

So it appears there should be a note for this, but I don't see it in the case when you try to call a trait method and the trait is implemented for the wrong version of the struct.

error[E0599]: no method named `build_vk_surface` found for type `winit::WindowBuilder` in the current scope
  --> src/main.rs:27:46
   |
27 |     let window = winit::WindowBuilder::new().build_vk_surface(&events_loop, Arc::clone(&instance));
   |                                              ^^^^^^^^^^^^^^^^

Here my crate was accidentally depending on winit 0.11 whereas the trait was defined on version 0.7.
No chance at better diagnostics here?

[Kixunil]

My friend just ran into this and one thing I'd like to see in error messages is explanation where those crates came from.

E.g. the crate depends on a and b, a depends on c 1.0 and b depends on c 1.1.

I'd like to see error message like this:

Note: two versions of the crate `c` are present
Note: crate `c` of version 1.0 required by crate `a`
Note: crate `a` required by this crate
Note: crate `c` of version 1.1 required by crate `a`
Note: crate `a` required by this crate

I'd like to see both dependency chains in full to be able to investigate.