Hi Geoff,

On 1 Nov 2026, at 20:09, Geoff Huston <[email protected]> wrote:

> Over in the world of paying for infrastructure "cheap" rules. I have listened 
> to a number of presentations over the years that note that a DNS server's 
> throughput drops by some 60% when the same query load is presented over TCP 
> rather than UDP. My experience with DNS servers supports this quite radical 
> drop in capacity (in my case throughput dropped by 90%!). So where does the 
> money come from to upgrade DNS infrastructure to support a transition to TCP 
> everywhere in the DNS?

I agree that a wholesale shift to a new transport protocol is a disruption, but 
I am not so convinced that the economics are as simple as you suggest above.

DNS servers are commonly provisioned with 80% headroom to be able to tolerate 
unwanted traffic without affecting service quality, and the nature and cost of 
unwanted traffic changes when you have source addresses you can believe. The 
cost for the victims of reflection attacks is reduced if a large number of your 
amplifiers are able to suppress suspicious stateless traffic more aggressively. 
The cost of stateful transport on individual endpoints can be expected to 
reduce as we adopt more efficient practices in pipeline and bundle management. 
No doubt there are many other effects.

The web used to be provisioned around small objects being delivered over 
dedicated sessions. Pipelining allowed session reuse. Successive versions of 
the HTTP allowed objects to be interleaved in the same session, reducing 
serialisation delay. HTTP used to be overwhelmingly plain-text. Today, less 
than 10% of the web uses HTTP/1.x and the proportion not using TLS is a 
rounding error. I am not suggesting that HTTP and DNS are the same or will have 
the same challenges, but I think it's clear that similar transitions have 
happened.

Since HTTPS is a defined transport protocol for DNS, it's perhaps even an 
option for some to stand precisely upon those HTTP scaling adaptations when 
provisioning authority servers in the future, if a useful population of 
resolver traffic can use it.

We are not out of options. We can imagine ways of making this happen, even if 
there is still a lot to learn.

> Maybe things would be different if we had started with DNS over TCP at the 
> outset, but now we are in the UDP room there is a price to be paid to move to 
> TCP and no clear view as to who and how we pay that price.

I do not disagree, but I think the answer lies in finding clarity and not 
avoiding the problem.


Joe
_______________________________________________
DNSOP mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to