It seems like SACK would make the problem less bad in theory, but wouldn’t 
eliminate the problem.   With DNS-over-DTLS, if a packet is lost, the next 
packet still gets an immediate answer, because DTLS is a datagram protocol, not 
a stream protocol.   The retry strategy for DNS-over-DTLS would be the same as 
DNS-over-UDP, once the Diffie-Hellman shared key has been derived.   So 
assuming you are using this to communicate to a resolver, there would be an 
initial handshake that’s extra compared to DNS-over-UDP, and then it would just 
be DNS-over-UDP until such time as re-keying is needed.

> On Mar 25, 2019, at 5:13 PM, Tony Finch <d...@dotat.at> wrote:
> 
> Ted Lemon <mel...@fugue.com> wrote:
>> On Mar 25, 2019, at 4:04 PM, Tony Finch <d...@dotat.at> wrote:
>>> If I understand it correctly, DTLS leaves MTU and fragmentation up to the
>>> application protocol. The DNS has horrible packet size problems, so it
>>> needs a lot more help than DTLS provides. QUIC is much better.
>> 
>> Indeed.  My point was simply that this avoids ordering problems, just as
>> QUIC does.  I suspect that it is worth having DNS-over-DTLS; QUIC is
>> great, but based on what I’ve been seeing, quite a lot, and probably not
>> appropriate for a lot of use cases where DNS-over-DTLS would do just
>> fine.
> 
> It isn't so much ordering that is the problem, but not delaying answer B
> when answer A suffers packet loss. I'm kind of curious about what the
> relative trade-offs might be between DoQ and DoT with a decent SACK
> implementation, when there is not much latency between the client and the
> resolver. DNS-over-DTLS will need its own application layer retry
> strategy. I kind of prefer the idea of DNS being able to re-use a good
> off-the-shelf transport protocol rather than doing its own thing.
> 
> Tony.
> -- 
> f.anthony.n.finch  <d...@dotat.at>  http://dotat.at/
> Irish Sea: West or northwest 4 or 5. Smooth or slight. Showers. Good.

_______________________________________________
DNSOP mailing list
DNSOP@ietf.org
https://www.ietf.org/mailman/listinfo/dnsop

Reply via email to