That seems super naive... However, most of the methods on hyper::Error are private and I could not figure out a way to determine that this error is Error(Body, Error { kind: Io(Kind(BrokenPipe)) })

I think we can actually use hyper::Error::is_closed for this:

https://github.com/hyperium/hyper/blob/9a8413d91081ad5a949276f05337e984c455e251/src/error.rs#L114-L129

Though, we'd have to check a few other methods besides that one. Perhaps we can use Error::into_cause?

Not sure we can. In this particular case we get a kind of Io, so right here https://github.com/hyperium/hyper/blob/9a8413d91081ad5a949276f05337e984c455e251/src/error.rs#L317. If we to into_cause, this does not give us much apart from being able to look at the textual form of this error. Not sure we want to be comparing strings.

(As we discussed offline...) It looks like into_cause should be able to return the underlying io::Error.

After looking more closely, I believe we can probalby modify set_grpc_status to consume the error... We'll have to change some of those downcast_ref to downcast.

Will give it a shot

@olix0r The main head scratcher for me is the fact that we are handling a ServiceError here, that holds an Arc to its inner error. So we can only take a shared ref of it and cannot consume it, thereby preventing us from calling the method recursively. Can we at that point of the code assume that .try_unwrap on the Arc holding the inner error will succeed ?

No, that won't work, because the error is shared with the lock/buffer--both of these layers are used to share an inner service, and when its inner service fails, we have to share that error with all outer instances... So try_unwrap won't have exclusive ownership of the error.

As I said, I'm fine with just matching on hyper::Error in this case to get this to master. I'd rather get the fix out and refine it on our own time. I'll take a quick look at the error handler and see if there's a way to simplify the problem, but I'm happy to submit that as a followup. If I can figure it out, that is.

Honestly, the best option might be to submit a PR to hyper to add an Error::cause like

    /// Returns a reference to the cause
    pub fn cause(&self) -> Option<&Box<dyn StdError + Send + Sync>> {
        self.inner.cause.as_ref()
    }

We probably won't pick up this change until tokio 0.2 merges, but it would allow us to clean this up later...