A way to determine if a Service closed correctly or errored about tower

[seanmonstar]

Reconnect is an interesting case:

• If the service is established, and svc.poll_ready() is an error, Reconnect swallows it, tosses the svc, and calls new_service().
• If the Future from new_service itself errors, then Reconnect determines it cannot ever get back up, and returns the connect (NewService::InitError) error.

I don't think the proposed change hurts Reconnect. If anything, it gives it a little more context. It could be adjusted to always reconnect when Status::Closed, but perhaps keep a count of Errs and return after a limit, or something else entirely.

[seanmonstar]

As a comparison, futures mpsc channel has poll_ready as well, and currently encodes the closed state as a Disconnected SendError.

[olix0r]

As a part of this change, I think we should explicitly document that poll_ready errors should be considered fatal. If that can't be the case today (i.e. in reconnect), then we may want to figure out if Status needs to handle the case of benign failures.

[carllerche]

Before moving forward with this, I would like to discover a case that is made possible with this change.

A service closing normally can be represented with an Error type today, however there is no standard way to do this. Middleware implementations can always define a trait and require that the inner service's error type implement that trait. In order to justify a change like this, it should provide benefit to the majority of cases.

The original example was a connection pool, where the connection pool needs to know if an inner service failed unexpectedly or closed gracefully. In this example, it is not obvious to me that the connection pool needs to know this. The pool should shield its caller from the details of each individual connection and individual connections may fail without any implication to other connections.

So, I would say this should be a "won't fix" until there is a solid use case driving it.

[carllerche]

I'm going to close this as it seems that we are leaning towards "won't fix"