Improving support for transports other than UDP such as TCP and TLS
would be a step in the right direction.

It would not be ideal to consider DNS over UDP as "legacy". DNS is used
at the head of almost all internet activity, and there are all sorts of
small devices that connect over myriad IP implementations to IP
networks. The UDP single request single response datagram mechanism
should be preserved. It may not be possible to access DNSSEC records or
perform validation over it in the future, but a device could rely on a
local resolver to perform the validation even though the AD signaling at
the last hop/mile may not be secure.

There are two aspects to connected transports: the network latency due
to additional roundtrips, and additional server/client state. DNS over
other transports can perform better than it currently does, and some of
this is due to gaps in implementation. I'm not sure higher spec'd
hardware would get as far as improving the software and better use of
protocol would. There are some protocol facilities that we don't take
significant advantage of today such as TCP fastopen even though there's
some limited support in implementations. (IIRC, TCP fastopen is not even
enabled by default in many current Linux distributions.) For TLS there's
session resumption (and even 0-RTT in 1.3 if one doesn't mind the
possible PFS gap for the early data). These only help after the first
connection, but still they would help.

Stubs can set AD=1 without setting DO=1 or even sending an EDNS query
and get AD=0/1 in the response from a validating resolver indicating
whether the answer was validated without receiving any DNSSEC
records. So a client need not be validating, but still can benefit from
authenticated data. We want clients to benefit from DNSSEC, and this
should be the behavior even if clients don't need DNSSEC RRsets. They
should request authenticated data. E.g., glibc has the "trust-ad" option
in /etc/resolv.conf for controlling it.

I have pondered about the effect of having MLDSA44 RRSIGs in resolver
caches. The RRSIG blows up a 4-octet A RDATA record+RRSIG by about
500x. This showed up in some stark contrast when I recently signed a 1
million RRs zone of address records with MLDSA44 to show off GPU signing
and the signed zone was several GB in size, so much so it wouldn't fit
on an NS VM. Very large zone operators can use some methods like online
signing with a cache to avoid having a large signed zone in memory, but
large validating resolver operators would have to provision far more
memory if MLDSA44 becomes common. Some other PQ signatures are even
larger.

Some others talked about improving TCP support in the recent PQ meeting
and on this list and +1 to that which should be what we target
most. Establish a group to understand what works well, make best
practices and document it. Point out deficiences. IIRC some DNS over TCP
testing was done about a decade ago by John and Sara Dickinson at
Sinodun. Although implementations have improved since, this is still a
problem. There is still time. We have all these innovative protocol
proposals, but why not establish a group to review how support for DNS
over TCP and other secure transports can be improved in deployment which
will go a long way in supporting the future PQ landscape keeping the
protocols simple as they are?

                Mukund


On Fri, Oct 30, 2026 at 05:33:37PM -0500, Warren Kumari wrote:
> Draft: draft-wagope-dnsop-do-bit - "Clearing the DO bit for DNSSEC"
> <https://datatracker.ietf.org/doc/draft-wagope-dnsop-do-bit/>
> 
> Some of us were chatting at the recent ICANN meeting about impediments to
> adoption of PQC algorithms in DNSSEC.
> 
> One of the big concerns is that PQC signatures will almost certainly
> cause a fallback to TCP, and this may cause significant load issues for
> both authoritative and recursive servers.
> 
> While only something like 40% of resolvers worldwide perform DNSSEC
> validation, because of the way the DO bit works, all recursives request
> this data, whether they need it or not.
> 
> This draft updates[0] RFC4034 - "Resource Records for the DNS Security
> Extensions" <https://datatracker.ietf.org/doc/rfc4034/> to recommend that
> recursive resolvers only set the DO bit if they, or their clients, are
> validating.  This does mean that the resolver will need to track how the
> data entered it's cache, and will have to additional queries if an incoming
> query with DO arrives for an cache entry which was populated without DNSSEC
> info. This trades complexity and additional queries for less TCP fallback.
> 
> The authors cannot figure out if this is the best idea ever, or the worst
> idea ever, but we wanted to share it and get your feedback.
> 
> Warren, Peter and Geoff.
> —
> [0]: More correctly, this draft will be updated to update RFC4034 if the WG
> decides that the idea is worth pursuing. I had some initial OLD / NEW text,
> but I pulled it out to keep the initial version short and readable (and
> because the draft-cutoff arrived faster than expected, and I didn't have
> time to tidy it up more :-) )

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

Attachment: signature.asc
Description: PGP signature

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

Reply via email to