WIP: ip per service

[thockin]

This is filled with cruft, but I wanted to float the idea. It actually works.

Overview: When a service object is created, allocate it an IP. Broadcast that IP to all kube-proxies. Each one will set up iptables rules to forward traffic for that IP:port to a random local port. Clients connect to a stable ip:port but no more port collisions!

[smarterclayton]

Very interesting. Scales pretty well logically as well. I assume then if you wanted your service to be exposed external you'd bind an external address as well and ensure it routes external to internal.

[thockin]

Exactly. It also allows for intra-pod ambassadors (proxies), rather than
machine-scoped proxies, eventually, which ties in with the config stuff.

On Tue, Aug 26, 2014 at 8:05 AM, Clayton Coleman notifications@github.com
wrote:

Very interesting. Scales pretty well logically as well. I assume then if
you wanted your service to be exposed external you'd bind an external
address as well and ensure it routes internal to external.

Reply to this email directly or view it on GitHub
#1038 (comment)
.

[lavalamp]

Ow, my eyes! ;) Can you make this work instead: http://golang.org/pkg/net/#IP

[smarterclayton]

Yeah - let's try to future proof ourselves for IPv6 here.

[lavalamp]

Oh yeah, that was my other comment which I forgot to actually say. IPv6!!

[kelseyhightower]

So what happens if the machine "holding" the service IP goes away? I thought the way it works today is that the service proxy would round-robin across N pods to ensure you always hit the service from anywhere in the cluster. Since the proxy runs on every host, all external clients can hit any member of the cluster and end up in the right place.

[lavalamp]

I think the model at the moment is that every minion just pretends (via IPTables magic) that X IP address gets routed to Y service, and nothing else changes. @thockin, is that right?

[bgrant0607]

It looks like I never filed an issue specifically for IP per service, but FYI this is a roadmap.md item and is covered in networking.md. It was perhaps first discussed in #146 and #188.

[thockin]

I'm going to write this up a bit before pushing further with this PR. I
have demonstrated that it works, now I want to offer some trade-offs in the
final design, and that is better served by a doc than a PR :)

On Thu, Aug 28, 2014 at 2:38 PM, bgrant0607 notifications@github.com
wrote:

It looks like I never filed an issue specifically for IP per service, but
FYI this is a roadmap.md
https://github.com/GoogleCloudPlatform/kubernetes/blob/master/docs/roadmap.md
item and is covered in networking.md
https://github.com/GoogleCloudPlatform/kubernetes/blob/master/docs/networking.md.
It was perhaps first discussed in #146
#146 and #188
#188.

Reply to this email directly or view it on GitHub
#1038 (comment)
.

[smarterclayton]

Such discipline!

[thockin]

It takes every ounce of my being not to write a 20 page doc with diagrams,
flowcharts, code, test cases, and cross-references. :P

#1107

On Thu, Aug 28, 2014 at 4:43 PM, Clayton Coleman notifications@github.com
wrote:

Such discipline!

Reply to this email directly or view it on GitHub
#1038 (comment)
.