Getting unexpected eof after upgrade to 0.11.8

[mathstuf]

After upgrading our application (which only uses hyper directly for the header! macro in a dependent library and reqwest otherwise) to 0.11.8, we've started getting this error showing up:

pipeline error: unexpected end of file

which comes from tokio-proto. We've narrowed it down to hyper changing in commit 95e0164. Is there some way that we're using reqwest improperly that would cause this?

All of our calls using reqwest are contained in this file and this is a long-lived structure, though we're seeing errors right after startup. Does reqwest need changes related to its connection pool?

Cc: @bradking

[seanmonstar]

Does the connection send bytes it shouldn't? Such as, is there a body to a head request, or some extra bytes after the end of a response body?

[mathstuf]

I doubt it; we're talking to Gitlab which is using Grape for its REST API and I'd expect it to be well-behaved in that respect. I'll check the logs though to see if that's happening.

[mathstuf]

I'm only seeing 200 OK responses with full JSON data objects.

[seanmonstar]

Would it be possible to turn on logs from hyper and (after sanitizing) paste the ones that appear around the error log? I'm setting up a super simple program to just hit some public API end points on Gitlab and not seeing this error log, so I could use a little more on how to investigate this.

[mathstuf]

We see it when running the queries through Rayon, so it could be something to do with saturating reqwest's connection pool. Currently dumping logs out (there's lots of activity and it's taking a while to get a week's worth of logs).

[mathstuf]

Is there an email address that would be good to send it to? I anonymized it by removing all of the received data bits (they now just say Object() or Array()), but it is 220K which I'd rather not upload directly to github.

[seanmonstar]

You could send it to sean@seanmonstar.com.

[seanmonstar]

The most likely cause for this was that the connection was being closed immediately after the end of a response body was read. The internal implementation would not have adjusted its state that an EOF was OK at that point (only doing so in flush), and the tokio-proto dispatcher was calling poll again (since the previous one succeeded).

Merged to master is a fix to update its state that the message has completed and so an EOF is not unexpected. Would you be able to test using hyper master, to confirm?

[bradking]

Would you be able to test using hyper master, to confirm?

Yes, it works as of v0.11.9, thanks!