feat(client): implement rfc 6555 (happy eyeballs)

[hban]

First PR, criticism welcome!

Update client connector to attempt a parallel connection using alternative address family, if connection using preferred address family takes too long.

Some questions:

• What should be a default value for timeout before we attempt other connections? Currently is 300ms.
• How to write a test for this? For manual testing I used edited /etc/hosts and iptables to simulate non-functioning IPv6. Maybe same could be done for Travis? Unsure about Appveyor.

Closes: #1316

[seanmonstar]

Awesome work, thanks! With regards to testing, I peeked at what Golang does. They make use of some special reserved IP addresses to test their happy eyeballs implementation: https://github.com/golang/go/blob/master/src/net/dial_test.go#L136

Would that help inspire some tests here?

[scrogson]

It appears that RFC 6555 has been obsoleted by https://tools.ietf.org/html/rfc8305

[hban]

I've added a test, but I'm not really happy how it turned out. I needed to provide custom name resolver, which Hyper doesn't support, so I hacked one in, hidden behind non-default feature flag. It's crude and ugly solution.

Instead, I can write a test directly for ConnectingTcp. It would remove need for this test-only custom resolver, but would mix this integration test with other unit tests.

Anyway, I'm not really sure which options is better, or if there is a third one I missed.

Regarding RFC 8305 @scrogson mentioned, I see it as a extension of work done here (correct me if I'm wrong). DNS part would require support for querying A records separately from AAAA records. ToSocketAddrs doesn't support specifying address family AFAIK, so we would need some other resolver. This would also impact implementing custom resolvers since they would either need to receive address family as parameter or return pair-of-futures/stream. Staggering connections could be done now, but I'd prefer for them to be a separate PR (if they are desired feature).

[hban]

Apparently Travis doesn't enable IPv6 by default (travis-ci/travis-ci#8361), so I copied workaround from one of the comments to make it work. I hope that's OK.

[seanmonstar]

Instead, I can write a test directly for ConnectingTcp. It would remove need for this test-only custom resolver, but would mix this integration test with other unit tests.

I think this way sounds much better! I have no problem with inner tests when testing very specific behavior that is difficult to simulate in the integration tests.

[hban]

Done (and commits squashed).

[seanmonstar]

Phenomenal! The code is clear, the comments are useful for "why", and test is thorough! ❤️

[mehcode]

@seanmonstar @hban RFC 8305 (which obsoletes RFC 6555) seems to recommend 250ms for the timeout (section 5).

I'm not super clear on what big changes are between 6555 and 8305 but it might be worth looking into.

https://tools.ietf.org/html/rfc8305

---

See golang/go#23841 for some more concrete details

---

Really awesome work by the way 💯 I can remove a bunch of hacks I currently have.