Rename 4-tuple routing?
martinduke opened this issue · comments
The '11' codepoint might be better defined. It would be good, for instance, if a stateful load balancer actually stored the observed CID so that the connection can survive a NAT rebinding.
Actually this is a bit harder than we'd like.
If the load balancer is stateless, then the short-header processing would have to be based on the destination connection ID to be robust to NAT rebinding. But then, as the connection ID will change mid-handshake, long-header processing needs a different approach. I don't see how, logically, a stateless LB can maintain consistency between these two results unless we force the unconfigured server to retain most of the client-generated CID.
Having reflected on this a bit more, this is what I think we should do:
Keep 0b11 as 4-tuple routing. However Load Balancers MAY observe destination connection IDs and keep state on them, to later survive a NAT rebinding.
I see two use cases for 0b11:
(1) a general purpose server that always uses QUIC-LB, whether or not it's behind an LB, for simplicity. So when it's not behind a Layer 4 LB it just generates random CIDs except for the first byte. As an added bonus, if there is an unknown L4 LB in front of it the behavior will not be pathological, at least until there's a migration. But in general, the rebinding performance is not all that important.
(2) As a panic mode when things are misconfigured or the server has run out of nonces. Again, I'm not sure rebinding performance is super-critical.
In other words, if a server is behind an L4 load balancer is and is using 0b11, something has gone very wrong.