net: enable happy eyeballs by default

[bobrik]

What version of Go are you using (go version)?

go version go1.9.1 linux/amd64

Does this issue reproduce with the latest release?

Yes.

What operating system and processor architecture are you using (go env)?

I suspect any OS, but here's what I'm using:

GOARCH="amd64"
GOBIN=""
GOEXE=""
GOHOSTARCH="amd64"
GOHOSTOS="linux"
GOOS="linux"
GOPATH="/home/ivan/go"
GORACE=""
GOROOT="/usr/local/go"
GOTOOLDIR="/usr/local/go/pkg/tool/linux_amd64"
GCCGO="gccgo"
CC="gcc"
GOGCCFLAGS="-fPIC -m64 -pthread -fmessage-length=0"
CXX="g++"
CGO_ENABLED="1"
CGO_CFLAGS="-g -O2"
CGO_CPPFLAGS=""
CGO_CXXFLAGS="-g -O2"
CGO_FFLAGS="-g -O2"
CGO_LDFLAGS="-g -O2"
PKG_CONFIG="pkg-config"

What did you do?

Run the following program:

package main

import (
	"log"
	"net"
	"time"
)

func main() {
	d := net.Dialer{
		Timeout: 8 * time.Second,
	}

	s := time.Now()

	// IPv4 is example.com, IPv6 is a blackhole
	//
	// $ host broken-ipv6-working-ipv4.bobrik.name
	// broken-ipv6-working-ipv4.bobrik.name has address 93.184.216.34
	// broken-ipv6-working-ipv4.bobrik.name has IPv6 address 100::1
	c, err := d.Dial("tcp", "broken-ipv6-working-ipv4.bobrik.name:80")
	if err != nil {
		log.Fatal(err)
	}

	log.Printf("Dial to %v completed in %v", c.RemoteAddr(), time.Since(s))
}

What did you expect to see?

Connection succeeding to IPv4 address with no human perceptible delay, working IPv6 or not.

What did you see instead?

2017/10/12 02:39:42 Dial to 93.184.216.34:80 completed in 4.016407843s

This is the half of net.Dialer.Timeout.

What happened here is that my custom dialer did not set net.Dialer.DualStack to true explicitly.

Current docs: https://github.com/golang/go/blob/d2826d3e06/src/net/dial.go#L46-L51

// DualStack enables RFC 6555-compliant "Happy Eyeballs"
// dialing when the network is "tcp" and the host in the
// address parameter resolves to both IPv4 and IPv6 addresses.
// This allows a client to tolerate networks where one address
// family is silently broken.
DualStack bool

This setting is off by default in zero-value net.Dialer, net.Dial() and friends. However, it was turned on in http.DefaultTransport: #15324.

We fixed an issue in Grafana that caused 15s stalls for us:

• Enable dualstack in every net.Dialer, fixes #9364 grafana/grafana#9367

Wikipedia says:

Implementations of Happy Eyeballs stacks exist in Google's Chrome web browser, Opera 12.10, Firefox version 13, OS X, and cURL.

• https://en.wikipedia.org/wiki/Happy_Eyeballs

I think Go should also have this on by default. This protects users from indefinite (half of default infinite connect timeout is still infinity) stalls if resolver returns both A and AAAA (or even AAAAA), but the network drops fancy IPv6 packets on the floor.

Current name DualStack is also slightly confusing, because Go dials both stacks in either case. Setting only controls Happy Eyeballs part, so having NoHappyEyeballs instead would allow zero-valued dialers to work seamlessly with broken IPv6.

If we end up not enabling Happy Eyeballs by default, this issue will at least hold the reason why.

cc @bradfitz

[rsc]

If we change the default we want to make the zero Dialer pick up that default, because various code including package net does things like

 d := Dialer{Timeout: xxx}

to override one part of the default and expect the rest of the default behavior to carry forward. That would mean a second DisableDualStack bool and ignoring the current DualStack bool or something like that. Let's leave this API decision for Go 1.11, when @bradfitz and @mikioh can have more time to discuss it (Brad is away right now and will be until at least the freeze).

[bobrik]

@rsc see the 2nd paragraph from the end in the original ticket description. I think we should go with DisableHappyEyeballs instead of DisableDualStack.

[bobrik]

@bradfitz, @mikioh is this a better time when 1.11 window is open?

[mikioh]

I'm fine with having appropriate default values for stateful/connection-oriented communication setup using net.Dial and its variants.

Right now I have no concrete idea which variables should have default values but it seems like Timeout and DualStack might be mandatory. The most extreme case I've ever heard is some experiment using MPTCP over IP over MLPPP over some satellite communication protocol, but we can exclude such possibility.

Well, for naming thing, in early draft CLs the existing DualStack of Dialer was HappyEyeballs (or EnableHappyEyeballs) and @rsc proposed renaming; see What's next in networking? Joyful spleens? Ebullient elbows? Raving kneecaps perhaps.

[bobrik]

I don't think that renaming RFC6555 is a good idea: https://tools.ietf.org/html/rfc6555

Timeouts seem like a whole big separate story, I'd rather focus this issue on Happy Eyeballs. Can we start by having DisableHappyEyeballs addition to the struct and deprecation of DualStack?

Suggested algorithm:

useHappyEyeballs := d.DualStack || !d.DisableHappyEyeballs

DisableHappyEyeballs	DualStack	Happy Eyeballs enabled
false	false	true (new default)
true	false	false (explicitly disabled)
false	true	true (explicitly enabled via DualStack)
true	true	false (explicitly weird)

The change here is that if user explicitly disables DualStack, they'll have to explicitly set DisableHappyEyeballs to true now to retain the same behavior.

The behavior is "connect to IPv6 and stall indefinitely if it doesn't respond". It's a fairly safe choice to assume this to be very exotic (if existing in practice at all) use case.

[mikioh]

IIUC, Eyeballs or Eyeball network is slang or a password in telco sector or some inner circle. I'm still not sure the reason why we need to stick with it, even if it's a standardized technology name, though I still don't have a good replacement for it.

If we need a new control knob that defaults to being enabled, it's better to name FeatureNameDisabled instead of DisableFeatureName.

Timeouts seem like a whole big separate story

I don't think so because in my understanding the purpose of Happy Eyeballs is just to establish a TCP connection in a reasonable period and to report an error as quickly as possible without having unreasonable long failover delays on funky circumstances. Timeouts in the go runtime (Dialer.Timeout, Dialer.FallbackDelay) and the kernel (max # of syn-rexmit * rexmit-delay) are part of the delays.

However, what you want here is just to run a connection setup race by default, that's probably fine.

[mikioh]

that's probably fine.

I mean, it's fine to send a CL containing:

• "Updates net: enable happy eyeballs by default #22225", not "Fixes net: enable happy eyeballs by default #22225" so that we are able to navigate issues complaining "Happy Eyeballs is broken" or "Happy Eyeballs doesn't work on my circumstance" to this issue,
• a new variable in net.Dialer with a nice name: not so pedantic, not so private to some inner circle.

[pavlo-v-chernykh]

Happy Eyeballs aka Fast Fallback. FastFallbackDisabled could be a good name for a var.