Re: IPv6 Anycast has been killed by LINUX patch in 2016 - who cares?

[Tom Herbert <[e-mail]>]

On Sat, Aug 7, 2021 at 4:53 PM Töma Gavrichenkov <[e-mail]> wrote:
>
> Peace,
>
> On Sun, Aug 8, 2021 at 2:47 AM Robert Raszuk <[e-mail]> wrote:
> > if someone uses anycast to establish long lasting TCP sessions (say 12h) - Good Luck !
>
> Transport sessions cannot reliably last for 12 hours on the Internet.
> That's the thing.

Toma,

Transport connections follow the end-to-end model. Transport state is
maintained at the endpoints and not in the network, and hence as long
as there remains a reachable path between the running end points the
connection is viable. That is the fundamental architecture of the
Internet. Reliability diminishes when the network is invasive in the
transport layer and E2E model is not adhered to.

>
> An application should never rely on that.  12 minutes are already kind
> of long, but 12 hours sound ridiculous.

No application relies on connections being guaranteed to be viable for
any specified amount of time, but that is not the same thing as saying
that there aren't long lived connections that productively exist or
that the network should proactively disallow long lived connections.
If a connection is still useful after 12 hours, and everyone is happy,
why tear it down and start over?

> You can achieve that over multiple paths (MP-TCP was mentioned along
> the thread) with some higher level abstraction (also subject to
> eventually break) but that's it.
>
How does anycast work with Multipath TCP? Wouldn't each subflow
potentially be routed to a different anycast host?

Tom

> This is the limitation.  There's no way 'round it.
>
> --
> Töma