Network Working Group P. Faltstrom
Internet-Draft R. Austein
Expires: April 11, 2005 IAB
October 11, 2004
Design Choices When Expanding DNS
draft-iab-dns-choices-00.txt
Status of this Memo
By submitting this Internet-Draft, I certify that any applicable
patent or other IPR claims of which I am aware have been disclosed,
and any of which I become aware will be disclosed, in accordance with
RFC 3668.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF), its areas, and its working groups. Note that
other groups may also distribute working documents as
Internet-Drafts.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
The list of current Internet-Drafts can be accessed at
http://www.ietf.org/ietf/1id-abstracts.txt.
The list of Internet-Draft Shadow Directories can be accessed at
http://www.ietf.org/shadow.html.
This Internet-Draft will expire on April 11, 2005.
Copyright Notice
Copyright (C) The Internet Society (2004). All Rights Reserved.
Abstract
This note discusses how to extend the DNS with new data for a new
application. DNS extension discussion too often circulate around
reuse of the TXT record. This document lists different mechanisms to
accomplish the extension to DNS, and comes to the conclusion use of a
new RR Type is the best solution.
Faltstrom & Austein Expires April 11, 2005 [Page 1]
Internet-Draft Design Choices When Expanding DNS October 2004
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 3
2. Background . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3. Extension mechanisms . . . . . . . . . . . . . . . . . . . . . 4
3.1 Place selectors inside the RDATA . . . . . . . . . . . . . 4
3.2 Add a prefix to the owner name . . . . . . . . . . . . . . 4
3.3 Add a suffix to the owner name . . . . . . . . . . . . . . 5
3.4 Add a new Class . . . . . . . . . . . . . . . . . . . . . 5
3.5 Add a new Resource Record Type . . . . . . . . . . . . . . 6
4. Conclusion (why adding a new RR type is the answer) . . . . . 7
5. Conclusion and Recommendation . . . . . . . . . . . . . . . . 9
6. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 10
7. Security Considerations . . . . . . . . . . . . . . . . . . . 10
8. References . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . 11
Intellectual Property and Copyright Statements . . . . . . . . 12
Faltstrom & Austein Expires April 11, 2005 [Page 2]
Internet-Draft Design Choices When Expanding DNS October 2004
1. Introduction
The DNS stores multiple categories of data. The two most commonly
used categories are infrastructure data for the DNS system itself (NS
and SOA records) and data which have to do with mappings between
domain names and IP addresses (A, AAAA and PTR). There are other
categories as well, some of which are tied to specific applications
like email (MX), while others are generic record types used to convey
information for multiple protocols (SRV, NAPTR).
When storing data in the DNS for a new application, the data are
usually tied to a "normal" domain name, so the application can query
for the data it wants, while minimizing the impact on existing
applications.
Historically, extension of DNS to store data for applications tied to
a domain name has been done in different ways at different times. MX
records were created as a new resource record type specifically
designed to support electronic mail. SRV records are a generic type
which use a prefixing scheme in combination with a base domain name.
Records associated with ENUM use a suffixing scheme. NAPTR records
add selection data inside the RDATA. It is clear the way of adding
new data to the DNS has been inconsistent, and the purpose of this
document is to attempt to clarify the implications of each of these
methods, both for the applications that use them and for the rest of
the DNS system.
2. Background
See RFC 2929 [RFC2929] for a brief summary of DNS query structure.
Readers interested in the full story should start with the base DNS
specification in RFC 1035 [RFC1035], and continue with the various
documents which update, clarify, and extend the base specification.
When composing a query into the DNS system, the parameters actually
used by the protocol are a triple: a DNS name, a DNS class, and a DNS
record type. Every resource record (RR) matching a particular name,
type and class is said to belong to the same resource record set
(RRset), and the whole RRset is always returned to the client which
queries for it. Splitting an RRset is a protocol violation, because
it results in coherency problems with the DNS caching mechanism.
Note however that this requirement has nothing to do with the MTU
size and choice of UDP or TCP as the transport protocol. The whole
RRset always has to be returned, and if it doesn't fit in the MTU in
UDP, TCP has to be used to not break this RRset atomic requirement.
Faltstrom & Austein Expires April 11, 2005 [Page 3]
Internet-Draft Design Choices When Expanding DNS October 2004
3. Extension mechanisms
The DNS protocol is intended to be extensible to support new kinds of
data. This section examines the various ways in which this sort of
extension can be accomplished.
3.1 Place selectors inside the RDATA
For a given query name, one might choose to have a single RRset (all
sharing the same name, type, and class) shared by multiple
applications, and have the different applications use selectors
within the RR data (RDATA) to determine which records are intended
for which applications. This sort of selector mechanism is usually
referred to "subtyping", because it is in effect creating an
additional type subsystem within a single DNS RR type.
Examples of subtyping include NAPTR RRs (see RFC 2916 [RFC2916]) and
the original DNSSEC KEY RR type (RFC 2535 [RFC2535]) (before it was
updated by RFC 3445 [RFC3445]).
All DNS subtyping schemes share a common weakness: With subtyping
schemes it is impossible for a client to query for just the data it
wants. Instead, the client must fetch the entire RRset, then select
the RRs in which it is interested. Furthermore, since DNSSEC
signatures operate on complete RRsets, the entire RRset must be
re-signed if any RR in it changes. As a result, each application
that uses a subtyped RR incurs higher overhead than any of the
applications would have incurred had they not been using a subtyping
scheme. The fact the RRset is always passed around as an indivisible
unit increases the risk the RRset will not fit in a UDP packet, which
in turn increases the risk that the client will have to retry the
query with TCP, which substantially increases the load on the name
server. More precisely: Having one query fail over to TCP is not a
big deal, but since the typical ratio of clients to servers in the
DNS system is very high, having a substantial number of DNS messages
fail over to TCP it will cause the relevant name servers to be
overloaded by TCP overhead.
The final result of using a subtyping scheme might be that the zone
administrator has to choose which of the services tied to one domain
name can actually be used, because not all of them can be announced
at the same time.
3.2 Add a prefix to the owner name
By adding an application-specific prefix to a domain name, we will
get different name/class/type triples, and therefore different
RRsets. The problem with adding prefixes has to do with wildcards,
Faltstrom & Austein Expires April 11, 2005 [Page 4]
Internet-Draft Design Choices When Expanding DNS October 2004
especially if one has records like
*.example.com. IN MX 1 mail.example.com.
and one wants records tied to those names. Suppose one creates the
prefix "_mail". One would then have to say something like
_mail.*.example.com. IN X-FOO A B C D
but DNS wildcards only work with the "*" as the leftmost token in the
domain name.
Even when a specific prefix is chosen, the data will still have to be
stored in some RR type. This RR type can either be a "kitchen-sink
record" or a new RR type. This implies that some other mechanism has
to be applied as well, with implications detailed in other parts of
this note (see also draft-ietf-dnsext-wcard-clarify [wcardclarify]
regarding use of wildcards and DNS).
3.3 Add a suffix to the owner name
Adding a suffix to a domain name changes the name/class/type triple,
and therefore the RRset. The query name can be set to exactly the
data one wants, and the size of the RRset is minimized. The problem
with adding a suffix is that it creates a parallel tree within the IN
class. There will be no technical mechanism to ensure that the
delegation for "example.com" and "example.com._bar" are made to the
same organization. Furthermore, data associated with a single entity
will now be stored in two different zones, such as "example.com" and
"example.com._bar", which, depending on who controls "_bar", can
create new synchronization and update authorization issues.
Even when using a different name, the data will still have to be
stored in some RR type. This RR type can either be a "kitchen-sink
record" or a new RR type. This implies that some other mechanism has
to be applied as well, with implications detailed in other parts of
this note.
In RFC 2163 [RFC2163] an infix token is inserted directly below the
TLD, but the result is the same as adding a suffix to the owner name
(and because of that creation of a new TLD).
3.4 Add a new Class
DNS zones are class-specific, in the sense that all the records in
that zone share the same class as the zone's SOA record, and the
existence of a zone in one class does not guarantee the existence of
the zone in any other class. In practice, only the IN class has ever
Faltstrom & Austein Expires April 11, 2005 [Page 5]
Internet-Draft Design Choices When Expanding DNS October 2004
seen widespread deployment, and the administrative overhead of
deploying an additional class would almost certainly be prohibitive.
Nevertheless, one could in theory use the DNS class mechanism to
distinguish between different kinds of data. However, since the DNS
delegation tree (represented by NS RRs) is itself tied to a specific
class, attempting to resolve a query by crossing a class boundary may
produce unexpected results, because there is no guarantee that the
name servers for the zone in the new class will be the same as the
name servers in the IN class. The MIT Hesiod system used a scheme
like this for storing data in the HS class, but only on a very small
scale (within a single institution), and with an administrative fiat
requiring that the delegation trees for the IN and HS trees be
identical.
Even when using a different class, the data will still have to be
stored in some RR type or another. This RR type can either be a
"kitchen-sink record" or a new RR type. This implies that some other
mechanism has to be applied as well, with implications detailed in
other parts of this note.
3.5 Add a new Resource Record Type
When adding a new Resource Record type to the system, entities in
four different roles have to be able to handle the new type:
1. There must be a way to insert the new resource records into the
zone of the Primary Master name server. For some server
implementations, the user interface only accepts record types
which it understands (perhaps so that the implementation can
attempt to validate the data). Other implementations allow the
zone administrator to enter an integer for the resource record
type code and the RDATA in Base64 or hexadecimal encoding (or
even as raw data). RFC 3597 [RFC3597] specifies a standard
generic encoding for this purpose.
2. A slave authoritative name server must be able to do a zone
transfer, receive the data from some other authoritative name
server, and serve data from the zone even though the zone
includes records of unknown types. Historically, some
implementations have had problems parsing stored copies of the
zone file after restarting, but those problems have not been seen
for a few years.
3. A full service resolver will cache the records which are
responses to queries. As mentioned in RFC 3597 [RFC3597],there
are various pitfalls where a recursive name server might end up
having problems.
4. The application must be able to get the record with a new
resource record type. The application itself may understand the
Faltstrom & Austein Expires April 11, 2005 [Page 6]
Internet-Draft Design Choices When Expanding DNS October 2004
RDATA, but the resolver library might not. Support for a generic
interface for retrieving arbitrary DNS RR types has been a
requirement since 1989 (see RFC 1123 [RFC1123] Section 6.1.4.2).
Some stub resolver library implementations neglect to provide
this functionality and cannot handle unknown RR types, but
implementation of a new stub resolver library is not particularly
difficult, and open source libraries that already provide this
functionality are available.
4. Conclusion (why adding a new RR type is the answer)
By now, the astute reader will be wondering about how to use the
issues presented so far. We will now attempt to clear up the
reader's confusion by following the thought processes of a typical
application designer who wishes to store stuff in the DNS, showing
how such a designer almost inevitably hits upon the idea of just
using TXT RR, and why this is a bad thing, and instead why a new RR
type is to be allocated.
A typical application designer is not interested in the DNS for its
own sake, but rather as a distributed database in which application
data can be stored. As a result, the designer of a new application
is usually looking for the easiest way to add whatever new data the
application needs to the DNS in a way that naturally associates the
data with a DNS name.
As explained in Section 3.4, using the DNS class system as an
extension mechanism is not really an option, and in fact most users
of the system don't even realize that the mechanism exists. As a
practical matter, therefore any extension is likely to be within the
IN class.
Adding a new RR type is the technically correct answer from the DNS
protocol standpoint (more on this below), doing so requires some DNS
expertise, due to the issues listed in Section 3.5. As a result,
this option is usually considered too hard.
The application designer is thus left with the prospect of reusing
some existing DNS type within the IN class, but when the designer
looks at the existing types, almost all of them have well-defined
semantics, none of which quite match the needs of the new
application. This has not completely prevented proposals to reuse
existing RR types in ways incompatible with their defined semantics,
but it does tend to steer application designers away from this
approach.
RR Type 40 was registered for the SINK record type. This RR Type was
discussed in the DNSIND working group of the IETF, and it was decided
Faltstrom & Austein Expires April 11, 2005 [Page 7]
Internet-Draft Design Choices When Expanding DNS October 2004
at the 46th IETF to not move the I-D forward to become an RFC.
Reason was the risk for large RR sets and the ability for application
creators to use the SINK RR Type instead of registering a new RR
Type.
Eliminating all of the above leaves the TXT RR type in the IN class.
The TXT RDATA format is free form text, and there are no existing
semantics to get in the way. Furthermore, the TXT RR can obviously
just be used as a bucket in which to carry around data to be used by
some higher level parser, perhaps in some human readable programming
or markup language. Thus, for many applications, TXT RRs are the
"obvious" choice. Unfortunately, this conclusion, while
understandable, is also wrong, for several reasons.
The first reason why TXT RRs are not well suited to such use is
precisely the lack of defined semantics that make them so attractive.
Arguably, the TXT RR is misnamed, and should have been called the
Local Container record, because the lack of defined semantics means
that a TXT RR means precisely what the data producer says it means.
This is fine, so long as TXT RRs are being used by human beings or by
private agreement between data producer and data consumer. However,
it becomes a problem once one starts using them for standardized
protocols in which there is no prior relationship between data
producer and data consumer. Reason for this is that there is nothing
to prevent collisions with some other incompatible use of TXT RRs.
This is even worse than the general subtyping problem described in
Section 3.1, because TXT RRs don't even have a standardized selector
field in which to store the subtype. RFC1464 [RFC1464] tried, but it
was not a success. At best a definition of a subtype is reduced to
hoping that whatever scheme one has come up with will not accidently
conflict with somebody else's subtyping scheme, and that it will not
be possible to mis-parse one application's use of TXT RRs as data
intended for a different application. Any attempt to come up with a
standardized format within the TXT RR format would be at least
fifteen years too late even if it were put into effect immediately.
Using one of the naming modifications discussed in Section 3.2 and
Section 3.3 would address the subtyping problem, but each of these
approaches brings in new problems of its own. The prefix approach
(such as SRV RRs use) does not work well with wildcards, which is a
particular problem for mail-related applications, since MX RRs are
probably the most common use of DNS wildcards. The suffix approach
doesn't have wildcard issues, but, as noted previously, it does have
synchronization and update authorization issues, since it works by
creating a second subtree in a different part of the global DNS name
space.
The next reason why TXT RRs are not well suited to protocol use has
Faltstrom & Austein Expires April 11, 2005 [Page 8]
Internet-Draft Design Choices When Expanding DNS October 2004
to do with the limited data space available in a DNS message. As
alluded to briefly in Section 3.1, typical DNS query traffic patterns
involve a very large number of DNS clients sending queries to a
relatively small number of DNS servers. Normal path MTU discovery
schemes do little good here, because, from the server's perspective,
there isn't enough repeat traffic from any one client for it to be
worth retaining state. UDP-based DNS is an idempotent query, whereas
TCP-based DNS requires the server to keep state (in the form of TCP
connection state, usually in the server's kernel) and roughly triples
the traffic load. Thus, there's a strong incentive to keep DNS
messages short enough to fit in a UDP datagram, preferably a UDP
datagram short enough not to require IP fragmentation.
Subtyping schemes are therefore again problematic, because they
produce larger RRsets than necessary, but verbose text encodings of
data are also wasteful, since the data they hold can usually be
represented more compactly in a resource record designed specifically
to support the application's particular data needs. If the data that
need to be carried are so large that there is no way to make them fit
comfortably into the DNS regardless of encoding, it is probably
better to move the data somewhere else, and just use the DNS as a
pointer to the data, as with NAPTR.
5. Conclusion and Recommendation
Given the problems detailed in Section 4, it is worth reexamining the
oft-jumped-to conclusion that specifying a new RR type is hard.
Historically, this was indeed the case, but recent surveys suggest
that support for unknown RR types [RFC3597] is now widespread, and
that lack of support for unknown types is mostly an issue for
relatively old software that would probably need to be upgraded in
any case as part of supporting a new application. One should also
remember that deployed DNS software today should support DNSSEC, and
software recent enough to do so will have higher chance of being able
to also support RFC3597.
Of all the issues detailed in Section 3.5, provisioning the data is
in some respects the most difficult. The problem here is less the
authoritative name servers themselves than the front-end systems used
to enter (and perhaps validate) the data. Hand editing does not work
well for maintenance of large zones, so some sort of tool is
necessary, and the tool may not be tightly coupled to the name server
implementation itself. Note, however, that this provisioning problem
exists to some degree with any new form of data to be stored in the
DNS, regardless of data format, RR type, or naming scheme. Adapting
front-end systems to support a new RR type may be a bit more
difficult than reusing an existing type, but this appears to be a
minor difference in degree rather than a difference in kind.
Faltstrom & Austein Expires April 11, 2005 [Page 9]
Internet-Draft Design Choices When Expanding DNS October 2004
Given the various issues described in this note, we believe that:
o there is no magic solution which allows a completely painless
addition of new data to the DNS, but
o on the whole, the best solution is still to use the DNS type
mechanism designed for precisely this purpose, and
o of all the alternate solutions, the "obvious" approach of using
TXT RRs is almost certainly the worst.
6. IANA Considerations
This document does not require any IANA actions.
7. Security Considerations
DNS RRsets can be signed using DNSSEC. DNSSEC is almost certainly
necessary for any application mechanism that stores authorization
data in the DNS itself. DNSSEC signatures significantly increase the
size of the messages transported, and because of this, the DNS
message size issues discussed in Section 3.1 and Section 4 are more
serious than they might at first appear.
Adding new RR types (as discussed in Section 3.5) might conceivably
trigger bugs and other bad behavior in software which is not
compliant with RFC 3597 [RFC3597], but most such software is old
enough and insecure enough that it should be updated for other
reasons in any case. Basic API support for retrieving arbitrary RR
types has been a requirement since 1989 (see RFC 1123 [RFC1123]).
Any new protocol that proposes to use the DNS to store data used to
make authorization decisions would be well advised not only to use
DNSSEC but also to encourage upgrades to DNS server software recent
enough not to be riddled with well-known exploitable bugs. Because
of this, support for new RR Types will not be as hard as people might
think at first.
8 References
[RFC1035] Mockapetris, P., "Domain names - implementation and
specification", STD 13, RFC 1035, November 1987.
[RFC1123] Braden, R., "Requirements for Internet Hosts - Application
and Support", STD 3, RFC 1123, October 1989.
[RFC1464] Rosenbaum, R., "Using the Domain Name System To Store
Arbitrary String Attributes", RFC 1464, May 1993.
[RFC2163] Allocchio, C., "Using the Internet DNS to Distribute MIXER
Conformant Global Address Mapping (MCGAM)", RFC 2163,
Faltstrom & Austein Expires April 11, 2005 [Page 10]
Internet-Draft Design Choices When Expanding DNS October 2004
January 1998.
[RFC2535] Eastlake, D., "Domain Name System Security Extensions",
RFC 2535, March 1999.
[RFC2606] Eastlake, D. and A. Panitz, "Reserved Top Level DNS
Names", BCP 32, RFC 2606, June 1999.
[RFC2671] Vixie, P., "Extension Mechanisms for DNS (EDNS0)", RFC
2671, August 1999.
[RFC2916] Faltstrom, P., "E.164 number and DNS", RFC 2916, September
2000.
[RFC2929] Eastlake, D., Brunner-Williams, E. and B. Manning, "Domain
Name System (DNS) IANA Considerations", BCP 42, RFC 2929,
September 2000.
[RFC3445] Massey, D. and S. Rose, "Limiting the Scope of the KEY
Resource Record (RR)", RFC 3445, December 2002.
[RFC3597] Gustafsson, A., "Handling of Unknown DNS Resource Record
(RR) Types", RFC 3597, September 2003.
[wcardclarify]
Halley, B. and E. Lewis,
"draft-ietf-dnsext-wcard-clarify-02.txt, Clarifying the
Role of Wild Card Domains in the Domain Name System, work
in progress", September 2003.
Authors' Addresses
Patrik Faltstrom
IAB
EMail: paf@cisco.com
Rob Austein
IAB
EMail: sra@isc.org
Faltstrom & Austein Expires April 11, 2005 [Page 11]
Internet-Draft Design Choices When Expanding DNS October 2004
Intellectual Property Statement
The IETF takes no position regarding the validity or scope of any
Intellectual Property Rights or other rights that might be claimed to
pertain to the implementation or use of the technology described in
this document or the extent to which any license under such rights
might or might not be available; nor does it represent that it has
made any independent effort to identify any such rights. Information
on the procedures with respect to rights in RFC documents can be
found in BCP 78 and BCP 79.
Copies of IPR disclosures made to the IETF Secretariat and any
assurances of licenses to be made available, or the result of an
attempt made to obtain a general license or permission for the use of
such proprietary rights by implementers or users of this
specification can be obtained from the IETF on-line IPR repository at
http://www.ietf.org/ipr.
The IETF invites any interested party to bring to its attention any
copyrights, patents or patent applications, or other proprietary
rights that may cover technology that may be required to implement
this standard. Please address the information to the IETF at
ietf-ipr@ietf.org.
Disclaimer of Validity
This document and the information contained herein are provided on an
"AS IS" basis and THE CONTRIBUTOR, THE ORGANIZATION HE/SHE REPRESENTS
OR IS SPONSORED BY (IF ANY), THE INTERNET SOCIETY AND THE INTERNET
ENGINEERING TASK FORCE DISCLAIM ALL WARRANTIES, EXPRESS OR IMPLIED,
INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE
INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED
WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
Copyright Statement
Copyright (C) The Internet Society (2004). This document is subject
to the rights, licenses and restrictions contained in BCP 78, and
except as set forth therein, the authors retain all their rights.
Acknowledgment
Funding for the RFC Editor function is currently provided by the
Internet Society.
Faltstrom & Austein Expires April 11, 2005 [Page 12]
