On Mon, Aug 9, 2021 at 3:40 PM Christian Huitema <huitema@xxxxxxxxxxx> wrote:
On 8/9/2021 11:54 AM, Warren Kumari wrote:
> There are dire proclamations that IPv6 TCP anycast cannot work.... And yet
> there are a bunch of existing implementations where it is clearly working,
> and people have built their business models around it, showing that it is
> working fine.
Or something in between. Whether TCP anycast works or not depends on how
long one needs the TCP connection to last. If the connection lasts long
enough, route flaps end up happening, and then the TCP connection
breaks. How long that lasts also depends on how long the route is. If
the anycast destination is just one AS hop away from the TCP source,
then you can expect BGP paths to be fairly stable. If the anycast target
is multiple ASes away, the frequency of route flaps is going to increase
markedly.
For big CDN deployments, or for the edge servers of big tech companies,
there are many servers and the AS paths are quite short. TCP sessions
will mostly work.
Yup. In many cases the traffic to/from CDN is also HTTP/similar, and the application layer either expects to have short lived connections, and / or handles retries gracefully.
If the path changes while my browser is fetching https://www.example.com/favicon.ico, no-harm, no-foul -- my browser gets the RST and tries again.
On the other hand, if you start a deployment with just
a couple servers, say one per continent, the AS paths will be long and
the rate of TCP failure will be high.
Yup, or if I were, for example, doing SSH over TCP over IPv6 over Anycast IP (for some weird convoluted reason :-)). But, that would be silly, and I wouldn't try and do it in v4 either...
W
-- Christian Huitema
The computing scientist’s main challenge is not to get confused by the
complexities of his own making.
-- E. W. Dijkstra
complexities of his own making.
-- E. W. Dijkstra
