Internet Architecture Board (IAB)                            J. Peterson
Request for Comments: 6950                                 NeuStar, Inc.
Category: Informational                                       O. Kolkman
ISSN: 2070-1721                                               NLnet Labs
                                                           H. Tschofenig
                                                  Nokia Siemens Networks
                                                                B. Aboba
                                                                   Skype
                                                            October 2013

Architectural Considerations on Application Features in the DNS

Abstract

   A number of Internet applications rely on the Domain Name System
   (DNS) to support their operations.  Many applications use the DNS to
   locate services for a domain; some, for example, transform
   identifiers other than domain names into formats that the DNS can
   process, and then fetch application data or service location data
   from the DNS.  Proposals incorporating sophisticated application
   behavior using DNS as a substrate have raised questions about the
   role of the DNS as an application platform.  This document explores
   the architectural consequences of using the DNS to implement certain
   application features, and it provides guidance to future application
   designers as to the limitations of the DNS as a substrate and the
   situations in which alternative designs should be considered.

Status of This Memo

   This document is not an Internet Standards Track specification; it is
   published for informational purposes.

   This document is a product of the Internet Architecture Board (IAB)
   and represents information that the IAB has deemed valuable to
   provide for permanent record.  It represents the consensus of the
   Internet Architecture Board (IAB).  Documents approved for
   publication by the IAB are not a candidate for any level of Internet
   Standard; see Section 2 of RFC 5741.

   Information about the current status of this document, any errata,
   and how to provide feedback on it may be obtained at
   http://www.rfc-editor.org/info/rfc6950.

Copyright Notice

   Copyright (c) 2013 IETF Trust and the persons identified as the
   document authors.  All rights reserved.

   This document is subject to BCP 78 and the IETF Trust's Legal
   Provisions Relating to IETF Documents
   (http://trustee.ietf.org/license-info) in effect on the date of
   publication of this document.  Please review these documents
   carefully, as they describe your rights and restrictions with respect
   to this document.

   1. Motivation ......................................................2
   2. Overview of DNS Application Usages ..............................4
      2.1. Locating Services in a Domain ..............................5
      2.2. NAPTR and DDDS .............................................6
      2.3. Arbitrary Data in the DNS ..................................8
   3. Challenges for the DNS .........................................10
      3.1. Compound Queries ..........................................10
           3.1.1. Responses Tailored to the Originator ...............12
      3.2. Using DNS as a Generic Database ...........................14
           3.2.1. Large Data in the DNS ..............................14
      3.3. Administrative Structures Misaligned with the DNS .........16
           3.3.1. Metadata about Tree Structure ......................18
      3.4. Domain Redirection ........................................20
   4. Private DNS and Split Horizon ..................................21
   5. Principles and Guidance ........................................23
   6. Security Considerations ........................................25
   7. IAB Members at the Time of Approval ............................26
   8. Acknowledgements ...............................................26
   9. Informative References .........................................27

1. Motivation

   The Domain Name System (DNS) has long provided a general means of
   translating domain names into Internet Protocol addresses, which
   makes the Internet easier to use by providing a valuable layer of
   indirection between names and lower-layer protocol elements.
   [RFC0974] documented a further use of the DNS: to locate an
   application service operating in a domain, via the Mail Exchange (MX)
   Resource Record; these records help email addressed to the domain to
   find a mail service for the domain sanctioned by the zone
   administrator.

   The seminal MX record served as a prototype for other DNS resource
   records that supported applications associated with a domain name.
   The SRV Resource Record [RFC2052] provided a more general mechanism
   for locating services in a domain, complete with a weighting system
   and selection among transports.  The Naming Authority Pointer (NAPTR)
   Resource Record (originally described in [RFC2168]), especially as it
   evolved into the more general Dynamic Delegation Discovery System
   (DDDS) [RFC3401] framework, added a generic mechanism for storing
   application data in the DNS.  Primarily, this involved a client-side
   algorithm for transforming a string into a domain name, which might
   then be resolved by the DNS to find NAPTR records.  This enabled the
   resolution of identifiers that do not have traditional host
   components through the DNS; the best-known examples of this are
   telephone numbers, as resolved by the DDDS application ENUM.  Recent
   work, such as DomainKeys Identified Mail (DKIM) [RFC6376], has
   enabled security features of applications to be advertised through
   the DNS, via the TXT Resource Record.

   The scope of application usage of the DNS has thus increased over
   time.  Applications in many environments require features such as
   confidentiality, and as the contexts in which applications rely on
   the DNS have increased, some application protocols have looked to
   extend the DNS to include these sorts of capabilities.  However, some
   proposed usages of, and extensions to, the DNS have become misaligned
   with both the DNS architecture and the DNS protocol.  If we take the
   example of confidentiality, we see that in the global public DNS, the
   resolution of domain names to IP addresses is an exchange of public
   information with no expectation of confidentiality.  Thus, the
   underlying query/response protocol has no encryption mechanism;
   typically, any security required by an application or service is
   invoked after the DNS query, when the resolved service has been
   contacted.  Only in private DNS environments (including split-horizon
   DNS) where the identity of the querier is assured through some
   external policy can the DNS maintain confidential records, by
   providing distinct answers to the private and public users of the
   DNS.  In support of load-balancing or other optimizations, a DNS
   server may return different addresses in response to queries from
   different sources, or even no response at all; see Section 3.1.1 for
   details.

   This document provides guidance to application designers and
   application protocol designers looking to use the DNS to support
   features in their applications.  It provides an overview of past
   application usage of the DNS as well as a review of proposed new
   usages.  It identifies concerns and trade-offs and provides guidance
   on the question, "Should I store this information in the DNS, or use
   some other means?" when that question arises during protocol
   development.  These guidelines remind application protocol designers

   of the strengths and weaknesses of the DNS in order to make it easier
   for designers to decide what features the DNS should provide for
   their application.

   The guidance in this document complements the guidance on extending
   the DNS given in [RFC5507].  Whereas [RFC5507] considers the
   preferred ways to add new information to the underlying syntax of the
   DNS (such as defining new resource records or adding prefixes or
   suffixes to labels), the current document considers broader
   implications of applications that rely on the DNS for the
   implementation of certain features, be it through extending the DNS
   or simply reusing existing protocol capabilities -- implications that
   may concern the invocation of the resolver by applications; the
   behavior of name servers, resolvers, or caches; extensions to the
   underlying DNS protocol; the operational responsibilities of zone
   administrators; security; or the overall architecture of names.  When
   existing DNS protocol fields are used in ways that their designers
   did not intend to handle new applications, those applications may
   demand further changes and extensions that are fundamentally at odds
   with the strengths of the DNS.

2. Overview of DNS Application Usages

   [RFC0882] identifies the original and fundamental connection between
   the DNS and applications.  It begins by describing how the
   interdomain scope of applications creates "formidable problems when
   we wish to create consistent methods for referencing particular
   resources that are similar but scattered throughout the environment".
   This motivated transitioning the "mapping between host names... and
   ARPA Internet addresses" from a global table (the original "hosts"
   file) to a "distributed database that performs the same function".
   [RFC0882] also envisioned some ways to find the resources associated
   with mailboxes in a domain: without these extensions, a user trying
   to send mail to a foreign domain lacked a discovery mechanism to
   locate the right host in the remote domain to which to connect.

   While a special-purpose service discovery mechanism could be built
   for each such application protocol that needed this functionality,
   the universal support for the DNS encourages installing these
   features into its public tree rather than inventing something new.
   Thus, over time, several other applications leveraged DNS resource
   records for locating services in a domain or for storing application
   data associated with a domain in the DNS.  This section gives
   examples of various types of DNS usage by applications to date.

2.1. Locating Services in a Domain

   The MX Resource Record provides the simplest example of an
   application advertising its domain-level resources in the Domain Name
   System.  The MX Resource Record contains the domain name of a server
   that receives mail on behalf of the administrative domain in
   question; that domain name must itself be resolved to one or more
   IP addresses through the DNS in order to reach the mail server.
   While naming conventions for applications might serve a similar
   purpose (a host might be named "mail.example.com", for example),
   approaching service location through the creation of a new resource
   record yields important benefits.  For example, one can put multiple
   MX records under the same name, in order to designate backup
   resources or to load-balance across several such servers (see
   [RFC1794]); these properties could not easily be captured by naming
   conventions (see [RFC4367], though more recently DNS-based Service
   Discovery (DNS-SD) [RFC6763] codifies service instance naming
   conventions for use across applications to locate services in a
   domain).

   While the MX record represents a substantial improvement over naming
   conventions as a means of service location, it remains specific to a
   single application.  Thus, the general approach of the MX record was
   adapted to fit a broader class of applications through the Service
   (SRV) Resource Record (originally described in [RFC2052]).  The SRV
   record allows DNS resolvers to query for particular services and
   underlying transports (for example, HTTP running over Transport Layer
   Security (TLS) [RFC2818]) and to learn a host name and port where
   that service resides in a given domain.  It also provides a weighting
   mechanism to allow load-balancing across several instances of a
   service.

   The reliance of applications on the existence of MX and SRV records
   has important implications for the way that applications manage
   identifiers and the way that applications pass domain names to
   resolvers.  Email identifiers of the form "user@domain" rely on MX
   records to provide the convenience of simply specifying a "domain"
   component rather than requiring an application to guess which
   particular host handles mail on behalf of the domain.  While naming
   conventions continue to abound ("www.example.com") for applications
   like web browsing, SRV records allow applications to query for an
   application-specific protocol and transport in the domain.  For the
   Lightweight Directory Access Protocol (LDAP), the SRV service name
   corresponds to the URL scheme of the identifier invoked by the
   application (e.g., when "ldap://example.com" is the identifier, the
   SRV query passed to the resolver is for "_ldap._tcp.example.com");
   for other applications, the SRV service name that the application
   passes to the resolver may be implicit in the identifier rather than

   explicit.  In either case, the application delivers the service name
   to the DNS to find the location of the host of that service for the
   domain, the port where the service resides on that host, additional
   locations or ports for load-balancing and fault tolerance, and
   related application features.

   Locating specific services for a domain was the first major function
   for which applications started using the DNS beyond simple name
   resolution.  SRV broadened and generalized the precedent of MX to
   make service location available to any application, rather than just
   to mail.  As applications that acquire MX (or SRV) records might need
   to perform further queries or transformations in order to arrive at
   an eventual domain name that will resolve to the IP addresses for the
   service, [RFC1034] allowed that the Additional (data) section of DNS
   responses may contain the corresponding address records for the names
   of services designated by the MX record; this optimization, which
   requires support in the authoritative server and the resolver, is an
   initial example of how support for application features requires
   changes to DNS operation.  At the same time, this is an example of an
   extension of the DNS that cannot be universally relied on: many DNS
   resolver implementations will ignore the addresses in the additional
   section of the DNS answers because of the trustworthiness issues
   described in [RFC2181].

2.2. NAPTR and DDDS

   The NAPTR Resource Record evolved to fulfill a need in the transition
   from Uniform Resource Locators (URLs) to the more mature Uniform
   Resource Identifier (URI) [RFC3986] framework, which incorporated
   Uniform Resource Names (URNs).  Unlike URLs, URNs typically do not
   convey enough semantics internally to resolve them through the DNS,
   and consequently a separate URI-transformation mechanism is required
   to convert these types of URIs into domain names.  This allows
   identifiers with no recognizable domain component to be treated as
   domain names for the purpose of name resolution.  Once these
   transformations result in a domain name, applications can retrieve
   NAPTR records under that name in the DNS.  NAPTR records contain a
   far more rich and complex structure than MX or SRV Resource Records.
   A NAPTR record contains two different weighting mechanisms ("order"
   and "preference"), a "service" field to designate the application
   that the NAPTR record describes, and then two fields that can contain
   translations: a "replacement" field or a "regexp" (regular
   expression) field, only one of which appears in a given NAPTR record
   (see [RFC2168]).  A "replacement", like NAPTR's ancestor the PTR
   record, simply designates another domain name where one would look
   for records associated with this service in the domain.  The

   "regexp", on the other hand, allows regular expression
   transformations on the original URI intended to turn it into an
   identifier that the DNS can resolve.

   As the abstract of [RFC2915] says, "This allows the DNS to be used to
   lookup services for a wide variety of resource names (including URIs)
   which are not in domain name syntax".  Any sort of hierarchical
   identifier can potentially be encoded as a domain name, and thus
   historically the DNS has often been used to resolve identifiers that
   were never devised as a name for an Internet host.  A prominent early
   example is found in the in-addr domain [RFC0883], in which IPv4
   addresses are encoded as domain names by applying a string
   preparation algorithm that required reversing the octets and treating
   each individual octet as a label in a domain name -- thus, for
   example, the address 192.0.2.1 became 1.2.0.192.in-addr.arpa.  This
   allowed resolvers to query the DNS to learn name(s) associated with
   an IPv4 address.  The same mechanism has been applied to IPv6
   addresses [RFC3596] and other sorts of identifiers that lack a domain
   component.  Eventually, this idea connected with activities to create
   a system for resolving telephone numbers on the Internet, which
   became known as ENUM (originally described in [RFC2916]).  ENUM
   borrowed from an earlier proposal, the "tpc.int" domain [RFC1530],
   which provided a means for encoding telephone numbers as domain names
   by applying a string preparation algorithm that required reversing
   the digits and treating each individual digit as a label in a domain
   name -- thus, for example, the number +15714345400 became
   0.0.4.5.4.3.4.1.7.5.1.tpc.int.  In the ENUM system, in place of
   "tpc.int" the special domain "e164.arpa" was reserved for use.

   In the more mature form of the NAPTR standard, in the Dynamic
   Delegation Discovery System (DDDS) [RFC3401] framework, the initial
   transformation of an identifier (such as a telephone number) to a
   domain name was called the "First Well Known Rule".  The address-
   reversing mechanism, whereby a query name is formed by reversing an
   IPv4 address and prepending it to the in-addr.arpa domain, is
   generalized for the use of NAPTR: each application defines a "First
   Well Known Rule" that translates a specific resource into a query
   name.  Its flexibility has inspired a number of proposals beyond ENUM
   to encode and resolve unorthodox identifiers in the DNS.  Provided
   that the identifiers transformed by the "First Well Known Rule" have
   some meaningful structure and are not overly lengthy, virtually
   anything can serve as an input for the DDDS structure: for example,
   civic addresses.  Though [RFC3402] stipulates regarding the
   identifier that "The lexical structure of this string must imply a
   unique delegation path", there is no requirement that the identifier
   be hierarchical nor that the points of delegation in the domain name

   created by the "First Well Known Rule" correspond to any points of
   administrative delegation inherent in the structure of the
   identifier.

   While this ability to look up names "which are not in domain name
   syntax" does not change the underlying DNS protocol -- the names
   generated by the DDDS algorithm are still just domain names -- it
   does change the context in which applications pass name to resolvers
   and can potentially require very different operational practices of
   zone administrators (see Section 3.3).  In terms of the results of a
   DNS query, the presence of the "regexp" field of NAPTR records
   enabled unprecedented flexibility in the types of identifiers that
   applications could resolve with the DNS.  Since the output of the
   regular expression frequently took the form of a URI (in ENUM
   resolution, for example, a telephone number might be converted into a
   SIP URI [RFC3261]), anything that could be encoded as a URI might be
   the result of resolving a NAPTR record -- which, as the next section
   explores, essentially means arbitrary data.

2.3. Arbitrary Data in the DNS

   URI encoding has ways of encapsulating basically arbitrary data: the
   most extreme example is a data URL [RFC2397].  Thus, the returned
   NAPTR record might be interpreted to produce output other than a
   domain name that would subsequently be resolved to IP addresses and
   contacted for a particular application -- it could give a literal
   result that would be consumed by the application.  Originally, as
   discussed in [RFC2168], the intended applicability of the regular
   expression field in NAPTR was narrower: the "regexp" field contained
   a "substitution expression that is applied to the original URI in
   order to construct the next domain name to lookup", in order to
   "change the host that is contacted to resolve a URI" or as a way of
   "changing the path or host once the URL has been assigned".  The
   regular expression tools available to NAPTR record authors, however,
   grant much broader powers to alter the input string, and thus
   applications began to rely on NAPTR to perform more radical
   transformations that did not serve any of those aforementioned needs.
   According to [RFC3402], the output of DDDS is wholly application-
   specific: "the Application must define what the expected output of
   the Terminal Rule should be", and the example given in the document
   is one of identifying automobile parts by inputting a part number and
   receiving at the end of the process information about the
   manufacturer.

   Historically speaking, NAPTR did not pioneer the storage of arbitrary
   data in the DNS.  At the start, [RFC0882] observed that "it is
   unlikely that all users of domain names will be able to agree on the
   set of resources or resource information that names will be used to

   retrieve", and consequently places little restriction on the
   information that DNS records might carry: it might be "host
   addresses, mailbox data, and other as yet undetermined information".
   [RFC1035] defined the TXT record, a means to store arbitrary strings
   in the DNS; [RFC1035] also specifically stipulates that a TXT
   contains "descriptive text" and that "the semantics of the text
   depends on the domain where it is found".  The existence of TXT
   records has long provided new applications with a rapid way of
   storing data associated with a domain name in the DNS, as adding data
   in this fashion requires no registration process.  [RFC1464]
   experimented with a means of incorporating name/value pairs to the
   TXT record structure, which allowed applications to distinguish
   different chunks of data stored in a TXT record -- surely not just
   "descriptive text" as the TXT originally specified.  In this fashion,
   an application that wants to store additional data in the DNS can do
   so without registering a new resource record type, though [RFC5507]
   points out that it is "difficult to reliably distinguish one
   application's record from others, and for its parser to avoid
   problems when it encounters other TXT records".

   While open policies surrounding the use of the TXT record have
   resulted in a checkered past for standardizing application usage of
   TXT, TXT has been used as a technical solution for many applications.
   Recently, DKIM [RFC6376] sidestepped the problem of TXT ambiguity by
   storing keys under a specialized DNS naming structure that includes
   the component "_domainkeys", which serves to restrict the scope of
   that TXT solely to DKIM use.  Storing keys in the DNS became the
   preferred solution for DKIM for several reasons: notably, because
   email applications already queried the DNS in their ordinary
   operations, because the public keys associated with email required
   wide public distribution, and because email identifiers contain a
   domain component that applications can easily use to consult the DNS.
   If the application had to negotiate support for the DKIM mechanism
   with mail servers, it would give rise to bid-down attacks (where
   attackers misrepresent that DKIM is unsupported on the originating
   side) that are not possible if the DNS delivers the keys (provided
   that DNSSEC [RFC4033] guarantees authenticity of the data).  However,
   there are potential issues with storing large data in the DNS, as
   discussed in Section 3.2.1, as well as with the DKIM namespace
   conventions that complicate the use of DNS wildcards (as discussed in
   Section 6.1.2 of [RFC6376] and in more general terms in [RFC5507]).
   If prefixes are used to identify TXT records used by an application,
   potentially the use of wildcards may furthermore cause leakages that
   other applications will need to detect.

3. Challenges for the DNS

   The methods discussed in the previous section for transforming
   arbitrary identifiers into domain names and returning arbitrary data
   in response to DNS queries both represent significant departures from
   the basic function of translating host names to IP addresses, yet
   neither fundamentally alters the underlying semantics of the DNS.
   When we consider, however, that the URIs returned by DDDS might be
   base-64-encoded binary data in a data URL, the DNS could effectively
   implement the entire application feature set of any simple query-
   response protocol.  Effectively, the DDDS framework considers the DNS
   a generic database -- indeed, the DDDS framework was designed to work
   with any sort of underlying database; as [RFC3403] says, the DNS is
   only one potential database for DDDS to use.  Whether the DNS as an
   underlying database can support the features that some applications
   of DDDS require, however, is a more complicated question.

   As the following subsections will show, the potential for
   applications to rely on the DNS as a generic database gives rise to
   additional requirements that one might expect to find in a database
   access protocol: authentication of the source of queries for
   comparison to access control lists, formulating complex relational
   queries, and asking questions about the structure of the database
   itself.  The global public DNS was not designed to provide these
   sorts of properties, and extending the DNS protocols to encompass
   them could result in a fundamental alteration to its model.
   Ultimately, this document concludes that efforts to retrofit these
   capabilities into the DNS would be better invested in selecting, or
   if necessary inventing, other Internet services with broader powers
   than the DNS.  If an application protocol designer wants these
   properties from a database, in general this is a good indication that
   the DNS cannot, or can only partly, meet the needs of the application
   in question.

   Since many of these new requirements have emerged from the ENUM
   space, the following sections use ENUM as an illustrative example;
   however, any application using the DNS as a feature-rich database
   could easily end up with similar requirements.

3.1. Compound Queries

   Traditionally, DNS RRsets are uniquely identified by domain name,
   resource record type, and class.  DNS queries are based on this
   3-tuple, and the replies are resource record sets that are to be
   treated as atomic data elements (see [RFC2181]); to applications, the
   behavior of the DNS has traditionally been that of an exact-match
   query-response lookup mechanism.  Outside of the DNS space, however,
   there are plenty of query-response applications that require a

   compound or relational search, one taking into account more than one
   factor in formulating a response or one that uses no single factor as
   a key to the database.  For example, in the telephony space,
   telephone call routing often takes into account numerous factors
   aside from the dialed number, including originating trunk groups,
   interexchange carrier selection, number portability data, time of
   day, and so on.  All are considered simultaneously in generating a
   route.  While in its original conception ENUM hoped to circumvent the
   traditional Public Switched Telephone Network (PSTN) and route
   directly to Internet-enabled devices, the infrastructure ENUM effort
   to support the migration of traditional carrier routing functions to
   the Internet aspires to achieve feature parity with traditional
   number routing.  However, [RFC3402] explicitly states that "it is an
   assumption of the DDDS that the lexical element used to make a
   delegation decision is simple enough to be contained within the
   Application Unique String itself.  The DDDS does not solve the case
   where a delegation decision is made using knowledge contained outside
   the AUS and the Rule (time of day, financial transactions, rights
   management, etc.)".  Consequently, some consideration has been given
   to ways to append additional data to ENUM queries to give the DNS
   server sufficient information to return a suitable URI (see
   Section 3.1.1).

   From a sheer syntactical perspective, however, domain names do not
   admit of this sort of rich structure.  Several workarounds have
   attempted to instantiate these sorts of features in DNS queries.  For
   example, the domain name itself could be compounded with the
   additional parameters: one could take a name like
   0.0.4.5.4.3.4.1.7.5.1.e164.arpa and append a trunk group identifier
   to it, for example, of the form
   tg011.0.0.4.5.4.3.4.1.7.5.1.e164.arpa.  While in this particular case
   a DNS server can adhere to its traditional behavior in locating
   resource records, the syntactical viability of encoding additional
   parameters in this fashion is dubious, especially if more than one
   additional parameter is required and the presence of parameters is
   optional so that the application needs multiple queries to assess the
   completeness of the information it needs to perform its function.

   As an alternative, it has been proposed that we piggyback additional
   query parameters as Extension Mechanisms for DNS (EDNS(0)) extensions
   (see [RFC6891]).  This might be problematic for three reasons.
   First, supporting EDNS(0) extensions requires significant changes to
   name server behavior; these changes need to be supported by the
   authoritative and recursive name servers on which the application
   relies and might be very hard to realize on a global scale.  In
   addition, the original stated applicability of the EDSN(0) mechanism,
   as [RFC2671] states, was to "a particular transport level message and
   not to any actual DNS data", and consequently the OPT Resource

   Records it specifies are never to be forwarded.  The use of EDNS(0)
   for compound queries, however, clearly is intended to discriminate
   actual DNS data rather than to facilitate transport-layer handling.
   Finally, [RFC6891] also specifies that "OPT RRs MUST NOT be cached,
   forwarded, or stored" (see the next paragraph).  For these reasons,
   this memo recommends against crafting compound DNS queries by using
   EDNS(0).

   The implications of these sorts of compound queries for recursion and
   caching are potentially serious.  The logic used by the authoritative
   server to respond to a compound query may not be understood by any
   recursive servers or caches; intermediaries that naively assume that
   the response was selected based on the domain name, type, and class
   alone might serve responses to queries in a different way than the
   authoritative server intends.  Therefore, were EDNS(0) to be employed
   this way, its attributes would not be transitive, and if this were
   not considered where intermediaries are employed, as is normally the
   case in the global DNS, brokenness might occur.

3.1.1. Responses Tailored to the Originator

   DNS responses tailored to the identity of their originator, where
   some sort of administrative identity of the originator must be
   conveyed to the DNS, constitute the most important subcase of these
   compound queries.  We must first distinguish this from cases where
   the originating IP address or a similar indication is used to serve a
   location-specific name.  For those sorts of applications, which
   generally lack security implications, relying on factors like the
   source IP address introduces little harm; for example, when providing
   a web portal customized to the region of the client, it would not be
   a security breach if the client saw the localized portal of the wrong
   country.  Because recursive resolvers may obscure the origination
   network of the DNS client, a recent proposal suggested introducing a
   new DNS query parameter to be populated by DNS recursive resolvers in
   order to preserve the originating IP address (see [EDNS-CLIENT-IP]).
   However, aside from purely cosmetic uses, these approaches have known
   limitations due to the prevalence of private IP addresses, VPNs, and
   so on, which obscure the source IP address and instead supply the IP
   address of an intermediary that may be very distant from the
   originating endpoint.  Implementing technology such as the one
   described by [EDNS-CLIENT-IP] would require significant changes in
   the operation of recursive resolvers and the authoritative servers
   that would rely on the original source IP address to select resource
   records, and moreover a fundamental change to caching behavior as
   well.  As a result, such technology cannot be rolled out in an
   incremental, unilateral fashion but could only be successful when
   implemented bilaterally (by authoritative server and recursive
   resolver); this is a significant bar to deployment.

   In other deployments in use today, including those based on the BIND
   "views" feature, the source IP address is used to grant access to a
   selected, and potentially sensitive, set of resource records.  The
   security implications of trusting the source IP address of a DNS
   query have prevented most solutions along these lines from being
   standardized (see [RFC6269]), though the practice remains widespread
   in "split horizon" private DNS deployments (see Section 4), which
   typically rely on an underlying security layer, such as a physical
   network, a clear perimeter demarcation at a network perimeter point
   (with network-layer anti-spoofing countermeasures), or an IPsec VPN,
   to prevent spoofing of the source IP address.  These deployments do
   have a confidentiality requirement to prevent information intended
   for a constrained audience (internal to an enterprise, for example)
   from leaking to the public Internet -- while these internal network
   resources may use private IP addresses that should not be useful on
   the public Internet anyway, in some cases this leakage would reveal
   topology or other information that the name server administrator
   hopes to keep private.  More recently, TSIG [RFC2845] has been
   employed as a way of selecting among "views" in BIND; this provides a
   stronger level of security than merely relying on the source IP
   address, but typically many users share the same secret to access a
   given view, and moreover TSIG does not provide confidentiality
   properties to DNS messages -- without network-layer separation
   between users of different views, eavesdroppers might capture the DNS
   queries and responses.

   The use of source IP addresses as a discriminator to select DNS
   resource records, regardless of its lack of acceptance by the
   standards community, has widespread acceptance in the field.  Some
   applications, however, go even further and propose extending the DNS
   to add an application-layer identifier of the originator; for
   example, [EDNS-OPT-CODE] provides a SIP URI in an EDNS(0) parameter.
   Effectively, this conveyance of application-layer information about
   the administrative identity of the originator through the DNS is a
   weak authentication mechanism, on the basis of which the DNS server
   makes an authorization decision before sharing resource records.
   This can approximate a confidentiality mechanism per resource record,
   where only a specific set of originators is permitted to see resource
   records, or a case where a query for the same name by different
   entities results in completely different resource record sets.
   However, without any underlying cryptographic security, this
   mechanism must rely on external layers for security (such as VPNs)
   rather than any direct assurance.  Again, caching, forwarding, and
   recursion introduce significant challenges for applications that
   attempt to offload this responsibility to the DNS.  Achieving feature
   parity with even the simplest authentication mechanisms available at
   the application layer would likely require significant rearchitecture
   of the DNS.

3.2. Using DNS as a Generic Database

   As previously noted, applications can use a method like the "First
   Well Known Rule" of DDDS to transform an arbitrary string into a
   domain name and then receive from the DNS arbitrary data stored in
   TXT RRs, in the "regexp" of NAPTRs, or even in custom records.  Some
   query-response applications, however, require queries and responses
   that simply fall outside the syntactic capabilities of the DNS.  For
   example, domain names themselves must consist of labels that do not
   exceed 63 octets, while the total length of the encoded name may not
   exceed 255 octets, and applications that use label characters outside
   the traditional ASCII set may run into problems (however, see the
   discussion in [RFC6055], Section 3 for definitive guidance on the use
   of non-ASCII in the DNS).  The DNS therefore cannot be a completely
   generic database.  Similar concerns apply to the size of DNS
   responses.

3.2.1. Large Data in the DNS

   While the "data" URL specification [RFC2397] notes that it is "only
   useful for short values", unfortunately it gives no particular
   guidance about what "short" might mean.  Some applications today use
   quite large data URLs (containing a megabyte or more of data) as
   workarounds in environments where only URIs can syntactically appear
   (for example, in Apple iOS, to pass objects between applications).
   The meaning of "short" in an application context is probably very
   different from how we should understand it in a DNS message.
   Referring to a typical public DNS deployment, [RFC5507] observes that
   "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".  And while EDNS(0) allows for mechanisms
   to negotiate DNS message sizes larger than the traditional 512
   octets, there is a high risk that a long payload will cause UDP
   fragmentation, in particular when the DNS message already carries
   DNSSEC information.  If EDNS(0) is not available, or the negotiated
   EDNS(0) packet size is too small to fit the data, or UDP fragments
   are dropped, the DNS may (eventually) resort to using TCP.  While TCP
   allows DNS responses to be quite long, this requires stateful
   operation of servers, which can be costly in deployments where
   servers have only fleeting connections to many clients.  Ultimately,
   there are forms of data that an application might store in the DNS
   that exceed reasonable limits: in the ENUM context, for example,
   something like storing base-64-encoded mp3 files of custom ringtones.

   Designs relying on storage of large amounts of data within DNS RRs
   furthermore need to minimize the potential damage achievable in a
   reflection attack (see [RFC4732], Section 3), in which the attacker
   sends UDP-only DNS queries with a forged source address, and the

   victim receives the response.  The attacker relies on amplification,
   where a small query generates a large response directed at the
   victim.  Where the responder supports EDNS(0), an attacker may set
   the requester maximum payload size to a larger value while querying
   for a large resource record, such as a certificate [RFC4398].  Thus,
   the combination of large data stored in DNS RRs and responders
   supporting large payload sizes has the potential to increase the
   potential damage achievable in a reflection attack.

   Since a reflection attack can be launched from any network that does
   not implement source address validation, these attacks are difficult
   to eliminate absent the ubiquitous deployment of source address
   validation or "heavier" transport protocols such as TCP.  The
   bandwidth that can be mustered in a reflective amplification attack
   directed by a botnet reflecting off a recursive name server on a
   high-bandwidth network is sobering.  For example, if the responding
   resolver could be directed to generate a 10KB response in reply to a
   50-octet query, then magnification of 200:1 would be attainable.
   This would enable a botnet controlling 10000 hosts with 1 Mbps of
   bandwidth to focus 200 Gbps of traffic on the victim, more than
   sufficient to congest any site on today's Internet.

   DNS reflection attacks typically utilize UDP queries; it is
   prohibitively difficult to complete a TCP three-way handshake begun
   from a forged source address for DNS reflection attacks.  Unless the
   attacker uses EDNS(0) [RFC6891] to enlarge the requester's maximum
   payload size, a response can only reach 576 octets before the
   truncate bit is set in the response.  This limits the maximum
   magnification achievable from a DNS query that does not utilize
   EDNS(0).  As the large disparity between the size of a query and size
   of the response creates this amplification, techniques for mitigating
   this disparity should be further studied, though this is beyond the
   scope of this memo (for an analysis of the effects of limiting
   EDNS(0) responses while still accommodating DNSSEC, see [Lindsay]).
   For example, some implementations could limit EDNS(0) responses to a
   specific ratio compared to the request size, where the precise ratio
   can be configured on a per-deployment basis (taking into account
   DNSSEC response sizes).  Without some means of mitigating the
   potential for amplification, EDNS(0) could cause significant harm.

   In summary, there are two operational forces that tend to drive the
   practically available EDNS(0) sizes down: possible UDP fragmentation
   and minimizing amplification in case of reflection attacks.  DNSSEC
   data will use a significant fraction of the available space in a DNS
   packet.  Therefore -- appreciating that given the current DNSSEC and

   EDNS(0) deployment experience, precise numbers are impossible to give
   -- the generic payload available to other DNS data, given the premise
   that TCP fallback is to be minimized, is likely to be closer to
   several hundred octets than a few thousand octets.

3.3. Administrative Structures Misaligned with the DNS

   While the DDDS framework enables any sort of alphanumeric data to
   serve as a domain name through the application of the "First Well
   Known Rule", the delegative structure of the resulting domain name
   may not reflect any administrative division of responsibilities
   inherent in the original data.  While [RFC3402] requires only that
   the "Application Unique String has some kind of regular, lexical
   structure that the rules can be applied to", DDDS is first and
   foremost a delegation system: its abstract stipulates that
   "Well-formed transformation rules will reflect the delegation of
   management of information associated with the string".  Telephone
   numbers in the United States, for example, are assigned and delegated
   in a relatively complex manner.  Historically, the first six digits
   of a nationally specific number (called the "NPA/NXX") reflected a
   point of administrative delegation from the number assignment agency
   to a carrier; from these blocks of ten thousand numbers, the carrier
   would in turn delegate individual assignments of the last four digits
   (the "XXXX" portion) to particular customers.  However, after the
   rise of North American telephone number portability in the 1990s, the
   first point of delegation went away: the delegation is effectively
   from the number authority to the carrier for each complete ten-digit
   number (NPA/NXX-XXXX).  While technical implementation details differ
   from nation to nation, number portability systems with similar
   administrative delegations now exist worldwide.

   To render these identifiers as domain names in accordance with the
   DDDS Rule for ENUM yields a large flat administrative domain with no
   points of administrative delegation from the country-code
   administrator, e.g., 1.e164.arpa, down to the ultimate assignee of a
   number.  Under the assumption that one administrative domain is
   maintained within one DNS zone containing potentially over five
   billion names, scalability difficulties manifest in a number of
   areas: the scalability that results from caching depends on these
   points of delegation, so that cached results for intermediate servers
   take the load off higher-tier servers.  If there are no such
   delegations, then as in the telephone number example the zone apex
   server must bear the entire load for queries.  Worse still, number
   portability also introduces far more dynamism in number assignment,
   where in some regions updated assignees for ported numbers must
   propagate within fifteen minutes of a change in administration.
   Jointly, these two problems make caching the zone extremely
   problematic.  Moreover, traditional tools for DNS replication, such

   as the zone transfer protocols AXFR [RFC1034] and IXFR [RFC1995],
   might not scale to accommodate zones with these dimensions and
   properties.

   In practice, the maximum sizes of telephone number administrative
   domains are closer to 300M (the current amount of allocated telephone
   numbers in the United States today -- still more than three times the
   number of .com domain names), and one can alleviate some of the
   scalability issues mentioned above by artificially dividing the
   administrative domain into a hierarchy of DNS zones.  Still, the fact
   that traditional DNS management tools no longer apply to the
   structures that an application tries to provision in the DNS is a
   clue that the DNS might not be the right place for an application to
   store its data.

   While DDDS is the most obvious example of these concerns, the point
   is more generic: for example, were address portability to be
   implemented for IP addresses and their administration thus to become
   non-hierarchical, the same concerns would apply to in-addr.arpa
   names.  The difficulty of mapping the DNS to administrative
   structures can even occur with traditional domain names, where
   applications expect clients to infer or locate zone cuts.  In the web
   context, for example, it can be difficult for applications to
   determine whether two domains represent the same "site" when
   comparing security credentials with URLs (see Section 3.4 below for
   more on this).  This has also caused known problems in determining
   the scope of web cookies, in contexts where applications must infer
   where administrative domains end in order to grant cookies that are
   as narrowly scoped as possible.

   In summary, the "First Well Known Rule" of DDDS provides a capability
   that transforms arbitrary strings into domain names, but those names
   play well with the DNS only when the input strings have an
   administrative structure that maps to DNS delegations.  In the first
   place, delegation implies some sort of hierarchical structure.  Any
   mechanism to map a hierarchical identifier into a domain name should
   be constructed such that the resulting domain name does match the
   natural hierarchy of the original identifier.  Although telephone
   numbers, even in North America, have some hierarchical qualities
   (like the geographical areas corresponding to their first three
   digits), after the implementation of number portability these points
   no longer mapped onto an administrative delegation.  If the input
   string to the DDDS does not have a hierarchical structure
   representing administrative delegations that can map onto the DNS
   distribution system, then that string probably is not a good
   candidate for translating into a domain name.

3.3.1. Metadata about Tree Structure

   There are also other ways in which the delegative structure of an
   identifier may not map well onto the DNS.  Traditionally, DNS
   resolvers assume that when they receive a domain name from an
   application the name is complete -- that it is not a fragment of a
   domain name that a user is still in the middle of typing.  For some
   communications systems, however, this assumption does not apply.
   ENUM use cases have surfaced a couple of optimization requirements to
   reduce unnecessary calls and queries; proposed solutions include
   metadata in the DNS that describes the contents and structure of ENUM
   DNS trees to help applications handle incomplete queries or queries
   for domains not in use.

   In particular, the "send-n" proposal [ENUM-Send-N] hoped to reduce
   the number of DNS queries sent in regions with variable-length
   numbering plans.  When a dialed number potentially has a variable
   length, a telephone switch ordinarily cannot anticipate when a dialed
   number is complete, as only the numbering plan administrator (who may
   be a national regulator, or even in open number plans a private
   branch exchange) knows how long a telephone number needs to be.
   Consequently, a switch trying to resolve such a number through a
   system like ENUM might send a query for a telephone number that has
   only partially been dialed in order to test its completeness.  The
   send-n proposal installs in the DNS a hint informing the telephone
   switch of the minimum number of digits that must be collected by
   placing in zones corresponding to incomplete telephone numbers some
   resource records that state how many more digits are required --
   effectively how many steps down the DNS tree one must take before
   querying the DNS again.  Unsurprisingly, those boundaries reflect
   points of administrative delegation, where the parent in a number
   plan yields authority to a child.  With this information, the
   application is not required to query the DNS every time a new digit
   is dialed but can wait to collect sufficient digits to receive a
   response.  As an optimization, this practice thus saves the resources
   of the DNS server, though the call cannot complete until all digits
   are collected, so this mechanism simply reduces the time the system
   will wait before sending an ENUM query after collecting the final
   dialed digit.  A tangentially related proposal, [ENUM-UNUSED],
   similarly places resource records in the DNS that tell the
   application that it need not attempt to reach a number on the
   telephone network, as the number is unassigned -- a comparable
   general DNS mechanism would identify, for a domain name with no
   records available in the DNS, whether or not the domain had been
   allocated by a registry to a registrant (which is a different
   condition than a name merely being unresolvable, per the semantics of
   NXDOMAIN).

   Both proposals optimize application behavior by placing metadata in
   the DNS that predicts the success of future queries or application
   invocation by identifying points of administrative delegation or
   assignment in the tree.  In some respects, marking a point in the
   tree where a zone begins or ends has some features in common with the
   traditional parent zone use of the NS record type, except that
   instead of pointing to a child zone these metadata proposals point to
   distant grandchildren.  While this does not change resolver behavior
   as such (instead, it changes the way that applications invoke the
   resolver), it does have implications for the practices for zone
   administrators.  Metadata in one administrative domain would need to
   remain synchronized with the state of the resources it predicts in
   another administrative domain in the DNS namespace, in a case like
   overlap dialing where the carrier delegates to a zone controlled by
   an enterprise.  When dealing with external resources associated with
   other namespaces, like number assignments in the PSTN or the
   databases of a registry operator, other synchronization requirements
   arise; maintaining that synchronization requires that the DNS have
   "semi-real time" updates that may conflict with scale and caching
   mechanisms of the DNS.

   Placing metadata in the DNS may also raise questions about the
   authority and delegation model.  Who gets to supply records for
   unassigned names?  While in the original but little-used e164.arpa
   root of ENUM this would almost certainly be a numbering plan
   administrator, it is far less clear who that would be in the more
   common and successful "infrastructure" ENUM models (see Section 4).
   Ultimately, these metadata proposals share some qualities with DNS
   redirection services offered by ISPs (for example, [DNS-REDIRECT])
   that "help" web users who try to browse to sites that do not exist.
   Similarly, metadata proposals like [ENUM-UNUSED] create DNS records
   for unallocated zones that redirect to a service provider's web page.
   However, in the [DNS-REDIRECT] cases, at least the existence or
   non-existence of a domain name is a fact about the Internet
   namespace, rather than about an external namespace like the telephony
   E.164 namespace (which must be synchronized with the DNS tree in the
   metadata proposals).  In send-n, different leaf zones that administer
   telephone numbers of different lengths can only provision their hints
   at their own apex, which provides an imperfect optimization: they
   cannot install it themselves in a parent, both because they lack the
   authority and because different zones would want to provision
   contradictory data.  The later the hint appears in the tree, however,
   the less optimization will result.  An application protocol designer
   managing identifiers whose administrative model does not map well
   onto the DNS namespace and delegations structure would be better
   served to implement such features outside the DNS.

3.4. Domain Redirection

   Most Internet application services provide a redirection feature --
   when one attempts to contact a service, the service may refer the
   person to a different service instance, potentially in another
   domain, that is for whatever reason better suited to service a
   request.  In HTTP and SIP, for example, this feature is implemented
   by the 300 class responses containing one or more URIs, which may
   indicate that a resource has moved temporarily or permanently to
   another service.  Several tools in the DNS, including the SRV record,
   can provide a similar feature at a DNS level, and consequently some
   applications as an optimization offload the responsibility for
   redirection to the DNS; NAPTR can also provide this capability on a
   per-application basis, and numerous DNS resource records can provide
   redirection on a per-domain basis.  This can prevent the unnecessary
   expenditure of application resources on a function that could be
   performed as a component of a DNS lookup that is already a
   prerequisite for contacting the service.  Consequently, in some
   deployment architectures this DNS-layer redirection is used for
   virtual hosting services.

   Implementing domain redirection in the DNS, however, has important
   consequences for application security.  In the absence of universal
   DNSSEC, applications must blindly trust that their request has not
   been hijacked at the DNS layer and redirected to a potentially
   malicious domain, unless some subsequent application mechanism can
   provide the necessary assurance.  By way of contrast, application-
   layer protocols supporting redirection, such as HTTP and SIP, have
   available security mechanisms, including TLS, that can use
   certificates to attest that a 300 response came from the domain that
   the originator initially hoped to contact.

   A number of applications have attempted to provide an after-the-fact
   security mechanism that verifies the authority of a DNS delegation in
   the absence of DNSSEC.  The specification for dereferencing SIP URIs
   ([RFC3263], reaffirmed in [RFC5922]), requires that during TLS
   establishment the site eventually reached by a SIP request present a
   certificate corresponding to the original URI expected by the user;
   this requires a virtual hosting service to possess a certificate
   corresponding to the hosted domain.  (In other words, if example.com
   redirects to example.net in the DNS, this mechanism expects that
   example.net will supply a certificate for example.com in TLS, per the
   HTTP precedent in [RFC2818]).  This restriction rules out many styles
   of hosting deployments common in the web world today, however.
   [HARD-PROBLEM] explores this problem space.  [RFC6125] proposes a
   solution for all applications that use TLS, which relies on new
   application-specific identifiers in certificates, as does [RFC4985]);
   note, however, that support for such certificates would require

   changes to existing certificate authority practices as well as
   application behavior.  With DNSSEC in place, DNS-based Authentication
   of Named Entities (DANE) [RFC6394] offers another way to bind
   certificates to particular applications and services.

   All of these application-layer measures attempt to mirror the
   delegation of administrative authority in the DNS, when the DNS
   itself serves as the ultimate authority on how domains are delegated.
   (Note: changing the technical delegation structure by changing NS
   records in the DNS is not the same as administrative delegation,
   e.g., when a domain changes ownership.)  Synchronizing a static
   instrument like a certificate with a delegation in the DNS, however,
   is problematic because delegations are not static: revoking and
   reissuing a certificate every time an administrative delegation
   changes is cumbersome operationally.  In environments where DNSSEC is
   not available, the problems with securing DNS-layer redirections
   would be avoided by performing redirections in the application layer.
   This inevitably gives rise to various design trade-offs involving
   performance, load, and related factors, but in these application
   environments, the security properties typically take priority.

4. Private DNS and Split Horizon

   The classic view of the uniqueness of domain names in the DNS is
   given in [RFC2826]:

      DNS names are designed to be globally unique, that is, for any one
      DNS name at any one time there must be a single set of DNS records
      uniquely describing protocol addresses, network resources and
      services associated with that DNS name.  All of the applications
      deployed on the Internet which use the DNS assume this, and
      Internet users expect such behavior from DNS names.

   [RFC2826] "does not preclude private networks from operating their
   own private name spaces" but notes that if private networks "wish to
   make use of names uniquely defined for the global Internet, they have
   to fetch that information from the global DNS naming hierarchy".
   There are various DNS deployments outside of the global public DNS,
   including "split horizon" deployments and DNS usages on private (or
   virtual private) networks.  In a split horizon, an authoritative
   server gives different responses to queries from the public Internet
   than they do to internal resolvers; while some deployments
   differentiate internal queries from public queries by the source IP
   address, the concerns in Section 3.1.1 relating to trusting source IP
   addresses apply to such deployments.  When the internal address space
   range is private [RFC1918], this makes it both easier for the server
   to discriminate public from private and harder for public entities to
   impersonate nodes in the private network.  Networks that are made

   private physically, or logically by cryptographic tunnels, make these
   private deployments more secure.  The most complex deployments along
   these lines use multiple virtual private networks to serve different
   answers for the same name to many distinct networks.

   The use cases that motivate split-horizon DNS typically involve
   restricting access to some network services -- intranet resources
   such as internal web sites, development servers, or directories, for
   example -- while preserving the ease of use offered by domain names
   for internal users.  While for many of these resources the split
   horizon would not return answers to public resolvers for those
   internal resources (those records are kept confidential from the
   public), in some cases the same name (e.g., "mail.example.com") might
   resolve to one host internally but another externally.  The
   requirements for multiple-VPN private deployments, however, are
   different: in this case the authoritative server gives different, and
   confidential, answers to a set of resolvers querying for the same
   name.  While these sorts of use cases rarely arise for traditional
   domain names, where, as [RFC2826] says, users and applications expect
   a unique resolution for a name, they can easily arise when other
   sorts of identifiers have been translated by a mechanism such as the
   "First Well Known Rule" of DDDS into "domain name syntax".  Telephone
   calls, for example, are traditionally routed through highly mediated
   networks, in which an attempt to find a route for a call often
   requires finding an appropriate intermediary based on the source
   network and location rather than finding an endpoint (see the
   distinction between the Look-Up Function and Location Routing
   Function in [RFC5486]).  Moreover, the need for responses tailored to
   the originator, and for confidentiality, is easily motivated when the
   data returned by the DNS is no longer "describing protocol addresses,
   network resources and services" [RFC2826] but instead is arbitrary
   data.  Although, for example, ENUM was originally intended for
   deployment in the global public root of the DNS (under e164.arpa),
   the requirements of maintaining telephone identifiers in the DNS
   quickly steered operators into private deployments.

   In private environments, it is also easier to deploy any necessary
   extensions than it is in the public DNS: in the first place,
   proprietary non-standard extensions and parameters can more easily be
   integrated into DNS queries or responses, as the implementations of
   resolvers and servers can likely be controlled; secondly,
   confidentiality and custom responses can be provided by deploying,
   respectively, underlying physical or virtual private networks to
   shield the private tree from public queries, and effectively
   different virtual DNS trees for each administrative entity that might
   launch a query; thirdly, in these constrained environments, caching
   and recursive resolvers can be managed or eliminated in order to
   prevent any unexpected intermediary behavior.  While these private

   deployments serve an important role in the marketplace, there are
   risks in using techniques intended only for deployment in private and
   constrained environments as the basis of a standard solution.  When
   proprietary parameters or extensions are deployed in private
   environments, experience teaches us that these implementations will
   begin to interact with the public DNS and that the private practices
   will leak.

   While such leakage is not a problem when using the mechanisms
   described in Sections 3.1, 3.2, and 3.5 (with private RR types) of
   [RFC5507], other extension mechanisms might cause confusion or harm
   if leaked.  The use of a dedicated suffix (Section 3.3 of [RFC5507])
   in a private environment might cause confusion if leaked to the
   public Internet, for example.

   That this kind of leakage of protocol elements from private
   deployments to public deployments does happen has been demonstrated,
   for example, with SIP: SIP implemented a category of supposedly
   private extensions ( the "P-" headers) that saw widespread success
   and use outside of the constrained environments for which they were
   specifically designed.  There is no reason to think that
   implementations with similar "private" extensions to the DNS
   protocols would not similarly end up in use in public environments.

5. Principles and Guidance

   The success of the global public DNS relies on the fact that it is a
   distributed database, one that has the property that it is loosely
   coherent and offers lookup instead of search functionality.  Loose
   coherency means that answers to queries are coherent within the
   bounds of data replication between authoritative servers (as
   controlled by the administrator of the zone) and caching behavior by
   recursive name servers.  Today, this term increasingly must also
   include load-balancing or related features that rely on the source IP
   address of the resolver.  It is critical that application designers
   who intend to use the DNS to support their applications consider the
   implications that their uses have for the behavior of resolvers;
   intermediaries, including caches and recursive resolvers; and
   authoritative servers.

   It is likely that the DNS provides a good match whenever the needs of
   applications are aligned with the following properties:

   o  Data stored in the DNS can be propagated and cached using
      conventional DNS mechanisms, without intermediaries needing to
      understand exceptional logic (considerations beyond the name,
      type, and class of the query) used by the authoritative server to
      formulate responses

   o  Data stored in the DNS is indexed by keys that do not violate the
      syntax or semantics of domain names

   o  Applications invoke the DNS to resolve complete names, not
      fragments

   o  Answers do not depend on an application-layer identity of the
      entity doing the query

   o  Ultimately, applications invoke the DNS to assist in communicating
      with a service whose name is resolved through the DNS

   Whenever one of the five properties above does not apply to one's
   data, one should seriously consider whether the DNS is the best place
   to store actual data.  On the other hand, if one has to worry about
   the following items, then these items are good indicators that the
   DNS is not the appropriate tool for solving problems:

   o  Populating metadata about domain boundaries within the tree -- the
      points of administrative delegation in the DNS are something that
      applications are not in general aware of

   o  Domain names derived from identifiers that do not share a semantic
      or administrative model compatible with the DNS

   o  Selective disclosure of data stored in and provided by the DNS

   o  DNS responses not fitting into UDP packets, unless EDNS(0) is
      available, and only then with the caveats discussed in
      Section 3.2.1

   In cases where applications require these sorts of features, they are
   likely better instantiated by independent application-layer protocols
   than the DNS.  For example, the objects that HTTP can carry in both
   queries and responses can easily contain the necessary structure to
   manage compound queries.  Many applications already use HTTP because
   of widespread support for it in middleboxes.  Similarly, HTTP has
   numerous ways to provide the necessary authentication, authorization,
   and confidentiality properties that some features require, as well as

   the redirection properties discussed in Section 3.4.  These
   differences highlight the fact that the DNS and HTTP offer very
   different services and have different applicabilities; while both are
   query-response protocols, HTTP should not be doing the job of DNS,
   and DNS should not be doing the job of HTTP.  Similarly, DNS should
   not be doing the job of Diameter, LDAP, or other application-layer
   protocols.  The overhead of using any application-layer protocol may
   not be appropriate for all environments, of course, but even in
   environments where a more lightweight protocol is appropriate, DNS is
   usually not the only alternative.

   Where the administrative delegations of the DNS form a necessary
   component in the instantiation of an application feature, there are
   various ways that the DNS can bootstrap access to an independent
   application-layer protocol better suited to field the queries in
   question.  For example, since NAPTR or URI [URI-RR] Resource Records
   can contain URIs, those URIs can in turn point to an external query-
   response service such as an HTTP service where more syntactically and
   semantically rich queries and answers might be exchanged.  Any
   protocol designer considering where to implement features must
   perform their own gap analysis and determine whether or not
   implementing some features is worth the potential cost in increased
   network state, latency, and so on, but implementing some features
   simply requires heavier structures than others.

6. Security Considerations

   Many of the concerns about how applications use the DNS discussed in
   this document involve security.  The perceived need to authenticate
   the source of DNS queries (see Section 3.1.1) and authorize access to
   particular resource records also illustrates the fundamental security
   principles that arise from offloading certain application features to
   the DNS.  As Section 3.2.1 observes, large data in the DNS can
   provide a means of generating reflection attacks, and without the
   remedies discussed in that section (regarding the use of EDNS(0) and
   TCP) the presence of large sets of records (e.g., ANY queries) is not
   recommended.  Section 3.4 discusses a security problem concerning
   redirection that has surfaced in a number of protocols (see
   [HARD-PROBLEM]).

   While DNSSEC, were it deployed universally, can play an important
   part in securing application redirection in the DNS, DNSSEC does not
   provide a means for a resolver to authenticate itself to a server,
   nor a framework for servers to return selective answers based on the
   authenticated identity of resolvers, nor a confidential mechanism.
   Some implementations may support authenticating users through TSIG,
   provided that the security association with a compliant server has
   been pre-established, though authentication is typically not needed

   for queries in the global public DNS.  The existing feature set of
   DNSSEC is, however, sufficient for providing security for most of the
   ways that applications traditionally have used the DNS.  The
   deployment of DNSSEC ([RFC4033] and related specifications) is
   heartily encouraged.  Nothing in this document is intended to
   discourage the implementation, deployment, or use of Secure DNS
   Dynamic Updates [RFC3007], though this document does recommend that
   large data in the DNS be treated in accordance with the guidance in
   Section 3.2.1.

7. IAB Members at the Time of Approval

   Internet Architecture Board Members at the time this document was
   approved were:

   Bernard Aboba
   Jari Arkko
   Marc Blanchet
   Ross Callon
   Alissa Cooper
   Spencer Dawkins
   Joel Halpern
   Russ Housley
   David Kessens
   Danny McPherson
   Jon Peterson
   Dave Thaler
   Hannes Tschofenig

8. Acknowledgements

   The IAB appreciates the comments and often spirited disagreements of
   Eric Osterweil, John Levine, Stephane Bortzmeyer, Ed Lewis, Dave
   Crocker, Ray Bellis, Lawrence Conroy, Ran Atkinson, Patrik Faltstrom,
   and Eliot Lear.

9. Informative References

   [DNS-REDIRECT]
              Creighton, T., Griffiths, C., Livingood, J., and R. Weber,
              "DNS Redirect Use by Service Providers", Work in Progress,
              October 2010.

   [EDNS-CLIENT-IP]
              Contavalli, C., van der Gaast, W., Leach, S., and D.
              Rodden, "Client IP information in DNS requests", Work in
              Progress, May 2010.

   [EDNS-OPT-CODE]
              Kaplan, H., Walter, R., Gorman, P., and M. Maharishi,
              "EDNS Option Code for SIP and PSTN Source Reference Info",
              Work in Progress, October 2011.

   [ENUM-Send-N]
              Bellis, R., "IANA Registrations for the 'Send-N'
              Enumservice", Work in Progress, June 2008.

   [ENUM-UNUSED]
              Stastny, R., Conroy, L., and J. Reid, "IANA Registration
              for Enumservice UNUSED", Work in Progress, March 2008.

   [HARD-PROBLEM]
              Barnes, R. and P. Saint-Andre, "High Assurance
              Re-Direction (HARD) Problem Statement", Work in Progress,
              July 2010.

   [Lindsay]  Lindsay, G., "DNSSEC and DNS Amplification Attacks",
              April 2012.

   [RFC0882]  Mockapetris, P., "Domain names: Concepts and facilities",
              RFC 882, November 1983.

   [RFC0883]  Mockapetris, P., "Domain names: Implementation
              specification", RFC 883, November 1983.

   [RFC0974]  Partridge, C., "Mail routing and the domain system",
              STD 10, RFC 974, January 1986.

   [RFC1034]  Mockapetris, P., "Domain names - concepts and facilities",
              STD 13, RFC 1034, November 1987.

   [RFC1035]  Mockapetris, P., "Domain names - implementation and
              specification", STD 13, RFC 1035, November 1987.

   [RFC1464]  Rosenbaum, R., "Using the Domain Name System To Store
              Arbitrary String Attributes", RFC 1464, May 1993.

   [RFC1530]  Malamud, C. and M. Rose, "Principles of Operation for the
              TPC.INT Subdomain: General Principles and Policy",
              RFC 1530, October 1993.

   [RFC1794]  Brisco, T., "DNS Support for Load Balancing", RFC 1794,
              April 1995.

   [RFC1918]  Rekhter, Y., Moskowitz, B., Karrenberg, D., de Groot, G.,
              and E. Lear, "Address Allocation for Private Internets",
              BCP 5, RFC 1918, February 1996.

   [RFC1995]  Ohta, M., "Incremental Zone Transfer in DNS", RFC 1995,
              August 1996.

   [RFC2052]  Gulbrandsen, A. and P. Vixie, "A DNS RR for specifying the
              location of services (DNS SRV)", RFC 2052, October 1996.

   [RFC2168]  Daniel, R. and M. Mealling, "Resolution of Uniform
              Resource Identifiers using the Domain Name System",
              RFC 2168, June 1997.

   [RFC2181]  Elz, R. and R. Bush, "Clarifications to the DNS
              Specification", RFC 2181, July 1997.

   [RFC2397]  Masinter, L., "The "data" URL scheme", RFC 2397,
              August 1998.

   [RFC2671]  Vixie, P., "Extension Mechanisms for DNS (EDNS0)",
              RFC 2671, August 1999.

   [RFC2818]  Rescorla, E., "HTTP Over TLS", RFC 2818, May 2000.

   [RFC2826]  Internet Architecture Board, "IAB Technical Comment on the
              Unique DNS Root", RFC 2826, May 2000.

   [RFC2845]  Vixie, P., Gudmundsson, O., Eastlake 3rd, D., and B.
              Wellington, "Secret Key Transaction Authentication for DNS
              (TSIG)", RFC 2845, May 2000.

   [RFC2915]  Mealling, M. and R. Daniel, "The Naming Authority Pointer
              (NAPTR) DNS Resource Record", RFC 2915, September 2000.

   [RFC2916]  Faltstrom, P., "E.164 number and DNS", RFC 2916,
              September 2000.

   [RFC3007]  Wellington, B., "Secure Domain Name System (DNS) Dynamic
              Update", RFC 3007, November 2000.

   [RFC3261]  Rosenberg, J., Schulzrinne, H., Camarillo, G., Johnston,
              A., Peterson, J., Sparks, R., Handley, M., and E.
              Schooler, "SIP: Session Initiation Protocol", RFC 3261,
              June 2002.

   [RFC3263]  Rosenberg, J. and H. Schulzrinne, "Session Initiation
              Protocol (SIP): Locating SIP Servers", RFC 3263,
              June 2002.

   [RFC3401]  Mealling, M., "Dynamic Delegation Discovery System (DDDS)
              Part One: The Comprehensive DDDS", RFC 3401, October 2002.

   [RFC3402]  Mealling, M., "Dynamic Delegation Discovery System (DDDS)
              Part Two: The Algorithm", RFC 3402, October 2002.

   [RFC3403]  Mealling, M., "Dynamic Delegation Discovery System (DDDS)
              Part Three: The Domain Name System (DNS) Database",
              RFC 3403, October 2002.

   [RFC3596]  Thomson, S., Huitema, C., Ksinant, V., and M. Souissi,
              "DNS Extensions to Support IP Version 6", RFC 3596,
              October 2003.

   [RFC3986]  Berners-Lee, T., Fielding, R., and L. Masinter, "Uniform
              Resource Identifier (URI): Generic Syntax", STD 66,
              RFC 3986, January 2005.

   [RFC4033]  Arends, R., Austein, R., Larson, M., Massey, D., and S.
              Rose, "DNS Security Introduction and Requirements",
              RFC 4033, March 2005.

   [RFC4367]  Rosenberg, J., Ed., and IAB, "What's in a Name: False
              Assumptions about DNS Names", RFC 4367, February 2006.

   [RFC4398]  Josefsson, S., "Storing Certificates in the Domain Name
              System (DNS)", RFC 4398, March 2006.

   [RFC4732]  Handley, M., Ed., Rescorla, E., Ed., and IAB, "Internet
              Denial-of-Service Considerations", RFC 4732,
              December 2006.

   [RFC4985]  Santesson, S., "Internet X.509 Public Key Infrastructure
              Subject Alternative Name for Expression of Service Name",
              RFC 4985, August 2007.

   [RFC5486]  Malas, D., Ed., and D. Meyer, Ed., "Session Peering for
              Multimedia Interconnect (SPEERMINT) Terminology",
              RFC 5486, March 2009.

   [RFC5507]  IAB, Faltstrom, P., Ed., Austein, R., Ed., and P. Koch,
              Ed., "Design Choices When Expanding the DNS", RFC 5507,
              April 2009.

   [RFC5922]  Gurbani, V., Lawrence, S., and A. Jeffrey, "Domain
              Certificates in the Session Initiation Protocol (SIP)",
              RFC 5922, June 2010.

   [RFC6055]  Thaler, D., Klensin, J., and S. Cheshire, "IAB Thoughts on
              Encodings for Internationalized Domain Names", RFC 6055,
              February 2011.

   [RFC6125]  Saint-Andre, P. and J. Hodges, "Representation and
              Verification of Domain-Based Application Service Identity
              within Internet Public Key Infrastructure Using X.509
              (PKIX) Certificates in the Context of Transport Layer
              Security (TLS)", RFC 6125, March 2011.

   [RFC6269]  Ford, M., Ed., Boucadair, M., Durand, A., Levis, P., and
              P. Roberts, "Issues with IP Address Sharing", RFC 6269,
              June 2011.

   [RFC6376]  Crocker, D., Ed., Hansen, T., Ed., and M. Kucherawy, Ed.,
              "DomainKeys Identified Mail (DKIM) Signatures", RFC 6376,
              September 2011.

   [RFC6394]  Barnes, R., "Use Cases and Requirements for DNS-Based
              Authentication of Named Entities (DANE)", RFC 6394,
              October 2011.

   [RFC6763]  Cheshire, S. and M. Krochmal, "DNS-Based Service
              Discovery", RFC 6763, February 2013.

   [RFC6891]  Damas, J., Graff, M., and P. Vixie, "Extension Mechanisms
              for DNS (EDNS(0))", STD 75, RFC 6891, April 2013.

   [URI-RR]   Faltstrom, P. and O. Kolkman, "The Uniform Resource
              Identifier (URI) DNS Resource Record", Work in Progress,
              July 2013.

Authors' Addresses

   Jon Peterson
   NeuStar, Inc.

   EMail: jon.peterson@neustar.biz

   Olaf Kolkman
   NLnet Labs

   EMail: olaf@nlnetlabs.nl

   Hannes Tschofenig
   Nokia Siemens Networks

   EMail: Hannes.Tschofenig@gmx.net

   Bernard Aboba
   Skype

   EMail: Bernard_aboba@hotmail.com