Roy Badami <roy(_at_)gnomon(_dot_)org(_dot_)uk> wrote:
EDNS0 (RFC2535) provides a mechanism for sending larger DNS
transactions over UDP. RFC3226 (which is also referenced by Sender
ID) says that EDNS0 implementations MUST allow at least 1220 bytes,
and SHOULD allow at least 4000 bytes.
this reference is questionable at best (read: DOC-BUG). RFC 3226 specifies
behaviour of RFC2874-conformant DNS implementations (A6 is kinda dead now)
and for DNSSEC-aware DNS implementations. None of both is relevant to the draft
in question and, more importantly, it is suboptimal to reference a
requirements document for DNS server implementations this way.
If at all, the requirement would have to be specified explicitly for the
'SPF' RR type, e.g.: all DNS servers and resolvers implementing 'SPF' are
required to support EDNS0 and ... size requirements ...
But then, please note that RFC 3226 is rather old and especially it predates
RFC 3597, so usually we do no longer assume that DNS implementations need
to explicitly support a certain RR type. And, you can't just 'require'
conformance for the intermediaries.
However, based on deployment surveys you may *assume* RFC 3597 conformance.
There has been work made available to this list (by Roy Arends, I guess),
showing which SW versions are conformant. An example survey for the distribution
of DNS server(!) versions can be found at
http://www.ripe.net/ripe/meetings/ripe-48/presentations/ripe48-dns-survey.pdf
Also, Jakob Schlyter has done an extensive interoperability test for
RFC 3597 implementations. Details can be found in the minutes of the DNSEXT
meeting of IETF60.
Bottom line: Size does matter. However, EDNS0 is there. Today on a moderate-
size server (~5000 zones, non-representative sample, of course) 2/3 of all
incoming queries (speaking of resolvers, not servers now) used EDNS0.
-Peter