On Sun, Oct 25, 2026 at 02:13:03AM +0200, Steffen Nurpmeso wrote:

>  |Its DNSKEY RRset
>  |(including an ML-DSA-44 KSK, ZSK and RRSIG) arrives over TCP:
>  |
>  |    ;; Query time: 1395 msec
>  |    ;; SERVER: 127.0.0.1#53(127.0.0.1) (TCP)
>  |    ;; WHEN: Sat Oct 24 03:46:45 UTC 2026
>  |    ;; MSG SIZE  rcvd: 6546
> 
> That is in my opinion just as strange as trying to use this for
> DKIM, .. but likely the TTL is even lower.

It's TTL is 1 hour, and the deployment is I expect a deliberate
pilot exercise.

> I want to believe this is an operator error of someone who is
> either misinformed, or even hysteric.

There's no reason to think so.  The algorithm is being tested at
Cloudflare (with domains that are thus far either private or in
any case off the DANE-survey's eTLD+1 radar).

> Yes TCP retries forever, but the musl C library for example only
> ever implemented it in the not too distant past.  (Four years now
> that i look.)

TCP support in the stub resolver is not what's important here, it
is the system's validating recursive resolver that needs TCP, even
prior to the registration of algorithm 18:

    https://www.iana.org/assignments/dns-sec-alg-numbers
    https://datatracker.ietf.org/doc/html/draft-westerbaan-dnssec-mldsa-03/

> And then, it is "very expensive", at least in my world; there it
> is a fallback if the UDP did not work out.
> In a modern system this thus likely sends maybe even two UDP
> packets (IPv4 and IPv6), and then (hopefully only one) a TCP
> fallback retry.

That's not how DNS works.  The 1st UDP query elicits a truncated
response (TC=1) (one RTT) and then the client switches to TCP,
(3-way handshake, query, response, which is 2-RTT in the absence
of TFO and/or a cached TCP connection for a domain that is queried
often).  When I suppress TCP with "+ignore", dig queries to my
local cache report:

    ;; global options: +cmd
    ;; Got answer:
    ;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 36254
    ;; flags: qr tc rd ra ad; QUERY: 1, ANSWER: 0, AUTHORITY: 0, ADDITIONAL: 1

    ...

    ;; Query time: 0 msec
    ;; SERVER: 127.0.0.1#53(127.0.0.1) (UDP)
    ;; WHEN: Sun Oct 25 01:40:31 UTC 2026
    ;; MSG SIZE  rcvd: 48

> No no no no.  DNSSEC, i do not dig that myself.
> But yes, likely there the new algorithms will improve the
> situation a lot.

For larger zone's like .COM there's work in progress to implement Merkle
Ladder DNSSEC, with just the tree roots signed with ML-DSA-44.

> (In times where certificate rotation intervals get shrunken all
> the time, coming over with PQ for DNSSEC in 2026 seems very
> strange, sorry.)

Not everything one does not yet understand is "strange" or a "mistake".

> But likely before 2030 i would at least think, if they get chosen
> in 2027.

You need to include multi-year latencies for LTS releases of the
libraries to appear in LTS releases of an OS, which are then adopted
by enterprise users and operated for up to a decade.

> Some/most are already implemented in an OpenSSL compatible
> variant, in that PQ playground library where also the others came
> from (or where first).  I have read the communication on the FAEST
> algorithm (there).

Implementation in liboqs has little relevance in real-world operations.

> But one thing is plain: i will not add a DNS/UDP blowing algorithm
> in 2026 when 2027 brings algorithms which easily fit there,
> including ("normal") DNSSEC plus ("normal") DTLS, at least with
> EDNS.

Nobody is suggesting that *you* or any other "typical user" switch
to MLDSA44(18).  It is in a pilot phase to explore the deployment
challenges.

-- 
    Viktor.  🇺🇦 Слава Україні!
_______________________________________________
Ietf-dkim mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to