RFC 8324
DNS Privacy, Authorization, Special Uses, Encoding, Characters, Matching, and Root Structure: Time for Another Look?
Independent Submission J. Klensin Request for Comments: 8324 February 2018 Category: Informational ISSN: 2070-1721 DNS Privacy, Authorization, Special Uses, Encoding, Characters, Matching, and Root Structure: Time for Another Look?Abstract
The basic design of the Domain Name System was completed almost 30 years ago. The last half of that period has been characterized by significant changes in requirements and expectations, some of which either require changes to how the DNS is used or can be accommodated only poorly or not at all. This document asks the question of whether it is time to either redesign and replace the DNS to match contemporary requirements and expectations (rather than continuing to try to design and implement incremental patches that are not fully satisfactory) or draw some clear lines about functionality that is not really needed or that should be performed in some other way. Status of This Memo This document is not an Internet Standards Track specification; it is published for informational purposes. This is a contribution to the RFC Series, independently of any other RFC stream. The RFC Editor has chosen to publish this document at its discretion and makes no statement about its value for implementation or deployment. Documents approved for publication by the RFC Editor are not candidates for any level of Internet Standard; see Section 2 of RFC 7841. Information about the current status of this document, any errata, and how to provide feedback on it may be obtained at https://www.rfc-editor.org/info/rfc8324.
Copyright Notice Copyright (c) 2018 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 (https://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.
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 4 2. Background and Hypothesis . . . . . . . . . . . . . . . . . . 5 3. Warts and Tensions with the Current DNS . . . . . . . . . . . 6 3.1. Multi-type Queries . . . . . . . . . . . . . . . . . . . 6 3.2. Matching Part I: Case Sensitivity in Labels and Other Anomalies . . . . . . . . . . . . . . . . . . . . . . . . 7 3.3. Matching Part II: Non-ASCII ("Internationalized") Domain Name Labels . . . . . . . . . . . . . . . . . . . . . . . 7 3.4. Matching Part III: Label Synonyms, Equivalent Names, and Variants . . . . . . . . . . . . . . . . . . . . . . . . 8 3.5. Query Privacy . . . . . . . . . . . . . . . . . . . . . . 10 3.6. Alternate Namespaces for Public Use in the DNS Framework: The CLASS Problem . . . . . . . . . . . . . . . . . . . . 10 3.7. Loose Synchronization . . . . . . . . . . . . . . . . . . 10 3.8. Private Namespaces and Special Names . . . . . . . . . . 11 3.9. Alternate Query or Response Encodings . . . . . . . . . . 12 3.10. Distribution and Management of Root Servers . . . . . . . 12 3.11. Identifiers versus Brands and Other Convenience Names . . 13 3.12. A Single Hierarchy with a Centrally Controlled Root . . . 14 3.13. Newer Application Protocols, New Requirements, and DNS Evolution . . . . . . . . . . . . . . . . . . . . . . . . 14 3.13.1. The Extensions . . . . . . . . . . . . . . . . . . . 15 3.13.2. Extensions and Deployment Pressures -- The TXT RRTYPE . . . . . . . . . . . . . . . . . . . . . . . 15 3.13.3. Periods and Zone Cut Issues . . . . . . . . . . . . 16 3.14. Scaling of Reputation and Other Ancillary Information . . 17 3.15. Tensions among Transport, Scaling, and Content . . . . . 18 4. The Inverse Lookup Requirement . . . . . . . . . . . . . . . 19 5. Internet Scale, Function Support, and Incremental Deployment 20 6. Searching and the DNS -- An Historical Note . . . . . . . . . 20 7. Security Considerations . . . . . . . . . . . . . . . . . . . 21 8. References . . . . . . . . . . . . . . . . . . . . . . . . . 22 8.1. Normative References . . . . . . . . . . . . . . . . . . 22 8.2. Informative References . . . . . . . . . . . . . . . . . 22 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . 29 Author's Address . . . . . . . . . . . . . . . . . . . . . . . . 29
1. Introduction
This document explores contemporary expectations of the Internet's domain system (DNS) and compares them to the assumptions and properties of the DNS design, including both those documented in the RFC Series, an important early paper by the principal author of the original RFCs [Mockapetris-1988], and a certain amount of oral tradition. It is primarily intended to ask the question of whether the differences are causing enough stresses on the system, stresses that cannot be resolved satisfactorily by further patching, that the Internet community should be considering designing a new system, one that is better adapted to current needs and expectations, and developing a deployment and transition strategy for it. For those (perhaps the majority of us) for whom actually replacing the DNS is too radical to be realistic, the document may be useful in two other ways. It may provide a foundation for discussing what functions the DNS should not be expected to support and how those functions can be supported in other ways, perhaps via an intermediate system that then calls on the DNS or by using some other type of database technology for some set of functions while leaving the basic DNS functions intact. Or it may provide a basis for "better just get used to that and the way it works" discussions to replace fantasies about what the DNS might do in some alternate reality. There is a key design or philosophical question associated with the analysis in this document that the document does not address. It is whether changes to perceived requirements to DNS functionality as described here are, in most respects, evolutionary or whether many of them are instances of trying to utilize the DNS for new requirements because it exists and is already deployed independent of whether the DNS is really appropriate or not. The latter might be an instance of a problem often described in the IETF as "when all you have is a hammer, everything looks like a nail". Other recent work, including a short article by Vint Cerf [Cerf2017], has discussed an overlapping set of considerations from a different perspective, reinforcing the view that it may be time to ask fundamental questions about the evolution and future of the DNS. While this document does not assume deep technical or operational knowledge of the DNS, it does assume some knowledge and at least general familiarity with the concepts of RFC 1034 [RFC1034] and RFC 1035 [RFC1035] and the terminology discussed in RFC 7719 [RFC7719] and elsewhere. Although some of the comments it contains might be taken as hints or examples of different ways to think about the design issues, it makes no attempt to explore, much less offer a tutorial on, alternate naming systems or database technologies.
It is perhaps worth noting that, while the perspective is different and more than a dozen years have passed, many of the issues discussed in this document were analyzed and described (most of them with more extensive explanations) in a 2005 US National Research Council report [NRC-Signposts]. Readers should note that several references are to obsolete documents. That was done because they are intended to show the documents and dates that introduced particular features or concepts. When current versions are intended, they are referenced.2. Background and Hypothesis
The Domain Name System (DNS) [RFC1034] was designed starting in the early 1980s [RFC0799] [RFC0881] [RFC0882] [RFC0883] with the main goal of replacing the flat, centrally administered, host table system [RFC0810] [RFC0952] [RFC0953] with a hierarchical, administratively distributed, system. The DNS design included some features that, after initial implementation and deployment, were judged to be unworkable and either replaced (e.g., the mail destination (MD) and mail forwarder (MF) approach [RFC0882] that were replaced by the MX approach [RFC0974]), abandoned (e.g., the mechanism for using email local parts as labels described in RFC 1034, Section 3.3), or deprecated (e.g., the WKS RR TYPE [RFC1123]). Newer ideas and requirements have identified a number of other features, some of which were less developed than others. Of course the original designers could not anticipate everything that has come to be expected of the DNS in the last 30 years. In recent years, demand for new and extended services and uses of the DNS have, in turn, led to proposals for DNS extensions or changes of various sorts. Some have been adopted, including a model for negotiating extended functionality [RFC2671] (commonly known as EDNS(0)) and to support IPv6 [RFC3596], others were found to be impracticable, and still others continue to be under consideration. Some examples of the latter two categories are discussed below. A few features of the original DNS specification, such as the CLASS property and label types, have also been suggested to be so badly specified that they should be deprecated [Sullivan-Class]. Unlike earlier changes such as the Internationalized Domain Names for Applications (IDNA) mechanisms for better incorporating non-ASCII labels without modifying the DNS structure itself [RFC3490] [RFC5890], some recent proposals require or strongly suggest changes to APIs, formats, or interfaces by programs that need to retrieve information from the DNS or interpret that information. Differences between the DNS architecture and the requirements that imply those proposals suggest that it may be time to stop patching the DNS or
trying to extend it in small increments. Instead, we should be considering moving some current or proposed functionality elsewhere or developing a new system that better meets today's needs and a transition strategy to it. The next section of this document discusses a number of issues with the current DNS design that could appropriately be addressed by a different and newer design model. In at least some cases, changing the model and protocols could bring significant benefits to the Internet and/or its administration. This document is not a proposal for a new protocol. It is intended to stimulate thought about how far we want to try to push the existing DNS, to examine whether expectations of it are already exceeding its plausible capabilities, and to start discussion of a redesign or alternatives to one if the time for that decision has come.3. Warts and Tensions with the Current DNS
As suggested above, there are many signs that the DNS is incapable of meeting contemporary expectations of how it should work and functionality it should support. Some of those expectations are unrealistic under any imaginable circumstances; others are impossible (or merely problematic) in the current DNS structure but could be accommodated in a redesign. These are examples, rather than a comprehensive list, and do not appear in any particular order.3.1. Multi-type Queries
The DNS does not gracefully support multi-type queries. The current case where this problem rears its head involves attempts at solutions that return both TYPE A (IPv4) and type AAA (IPv6) addresses collectively. The problem was originally seen with "QTYPE=MAILA" [RFC0882] for the original MA and MD RRTYPEs, an experience that strongly suggests that some very careful thinking about cache effects (and possibly additional DNS changes) would be needed. Other solutions might seem equally or more plausible. What they, including "two types of addresses", probably have in common is that they illustrate stresses on the system and that changing the DNS to deal with those stresses is not straightforward or likely to be problem- free.
3.2. Matching Part I: Case Sensitivity in Labels and Other Anomalies
The DNS specifications assume that labels are octet strings and octets with the high bit zero have seven-bit ASCII codes in the remaining bits. They require that, when a domain name used in a query is matched to one stored in the database, those ASCII characters be interpreted in a case-independent way, i.e., upper- and lower-case letters are treated as equivalent (digits and symbols are not affected) [RFC4343]. For non-ASCII octets, i.e., octets in labels with the first bit turned on, there are no assumptions about the character coding used, much less any rules about character case equivalence -- strings must be compared by matching bits in sequence. Even though the current model for handling non-ASCII (i.e., "internationalized") domain name labels (IDNs) [RFC5890] (see Section 3.3 below) encodes information so the DNS is not directly affected, the notion that some characters in labels are handled in a case-insensitive way and that others are case sensitive (or that upper case must be prohibited entirely as IDNA does) has caused a good deal of confusion and resentment. Those concerns and complaints about inconsistent behavior and mishandling (or suboptimal handling) of case relationships for some languages have not been mitigated by repeated explanations that the relationships between "decorated" lower-case characters and their upper-case equivalents are often sensitive to language and locality and therefore not deterministic with information available to DNS servers.3.3. Matching Part II: Non-ASCII ("Internationalized") Domain Name Labels
Quite independent of the case-sensitivity problem, one of the fundamental properties of Unicode [Unicode] is that some abstract characters can be represented in multiple ways, such as by a single, precomposed, code point or by a base code point followed by one or more code points that specify combining characters. While Unicode Normalization can be used to eliminate many (but not all) of those distinctions for comparison (matching) purposes, it is best applied during matching rather than by changing one string into another. The first version of IDNA ("IDNA2003") made the choice to change strings during processing for either storage or retrieval [RFC3490] [RFC3491]; the second ("IDNA2008") required that all strings be normalized and that upper-case characters are not allowed at all [RFC5891]. Neither is optimal, if only because, independent of where they are changed if they are changed at all, transforming the strings themselves implies that the input string in an application may not be the same as the string used in processing and perhaps later display.
It would almost certainly be preferable, and more consistent with Unicode recommendations, to use normalization (and perhaps other techniques if they are appropriate) at matching time rather than altering the strings at all, even if there were still only a single matching algorithm, i.e., normalization were added to the existing ASCII-only case folding. However, even Unicode's discussion of normalization [Unicode-UAX15] indicates that there are special, language-dependent, cases (the most commonly cited example is the dotless "i" (U+0131)). Not only does the DNS lack any information about languages that could be used in a mapping algorithm, but, as long as there is a requirement that there be only one mapping algorithm for the entire system, that information could not be used even if it were available. One could imagine a successor system that would use information stored at nodes in the hierarchy to specify different matching rules for subsidiary nodes (or equivalent arrangements for non-hierarchical systems). It is not clear whether that would be a good idea, but it certainly is not possible with the DNS as we know it.3.4. Matching Part III: Label Synonyms, Equivalent Names, and Variants
As the initial phases of work on IDNs started to conclude, it became obvious that the nature and evolution of human language and writing systems required treating some names as "the same as" others. The first important example of this involved the relatively recent effort to simplify the Chinese writing system, thereby creating a distinction between "Simplified" and "Traditional" Chinese even though the meaning of the characters remained the same in almost all cases (in so-called ideographic character sets, characters have meaning rather than exclusively representing sounds). A joint effort among the relevant Country Code Top-Level Domain (ccTLD) registries and some other interested parties produced a set of recommendations for dealing with the issues with that script [RFC3743] and introduced the concept of "variant" characters and domain names. However, when names are seen as having meanings, rather than merely being mnemonics, especially when they represent brands or the equivalent, or when spelling for a particular written language is not completely standardized, demands to treat different strings as exact equivalents are obvious and inevitable. As a trivial English- language example, it is widely understood that "colour" and "color" represent the same word, so does that imply that, if they are used as DNS labels in domain names all of whose other labels are identical, the two domain names should be treated as identical? Examples for other languages or writing systems, especially ones in which some or all markings that distinguish characters or words by sound or tone or that change the pronunciation of words are optional, are often more numerous and more problematic than national spelling differences in
English, but they are harder to explain to those unfamiliar with those other languages or writing systems (and hard to illustrate in ASCII-only Internet-Drafts and RFCs). Although approximations are possible, the DNS cannot handle that requirement: not only do its aliasing mechanisms (CNAME, DNAME, and various proposals for newer and different types of aliasing [DNS-Aliases] [DNS-BNAME]) not provide a strong enough binding, but the ability to use those aliases from a subtree controlled by one administrative entity to that of another one implies that there is little or no possibility of the owner (in either the DNS sense or the registrar-registrant one) of a particular name to control the synonyms for it. Some of that issue can be dealt with at the application level, e.g., by redirects in web protocols, but taking that approach, which is the essential characteristic of "if both names belong to the same owner, everything is OK" approaches, results in names being handled in inconsistent ways in different protocols. A different way of looking at part of this issue (and, to some degree, of the one discussed above in Section 3.3) is that these perceived equivalences and desired transformations are context- dependent, but the DNS resolution process is not [RFC6912]. Similar problems arise as people notice that some characters are easily mistaken for others and that might be an opportunity for user confusion and attacks. Commonly cited examples include the Latin and Cyrillic script "a" characters, which are identical [CACM-Homograph], the characters in many scripts that look like open circles or vertical or horizontal lines, and even the Latin script letter "l" and the European digit "1", but examples abound in other scripts and combinations of scripts as well. The most common proposed solution within the DNS context has been to treat these cases, as well as those involving orthographic variations, as "variants" (but variants different from the system for Chinese characters mentioned above) and either ban all but one (or a few) of the possible labels from the DNS (possibly on a first come, first served basis) or ensure that any collection of such strings that are delegated as assigned to the same ownership (see above). Neither solution is completely satisfactory: if all but one string is excluded, users who guess at a different form, perhaps in trying to transcribe characters from written or printed form, don't find what they are looking for and, as pointed out above, "same ownership" is sufficient only with carefully designed and administered applications protocol support, and sometimes not then. Some of these issues are discussed at more length in an ICANN report [ICANN-VIP].
3.5. Query Privacy
There has been growing concern in recent years that DNS queries occur in cleartext on the public Internet and that, if those queries can be intercepted, they can expose a good deal of information about interests and contacts that could compromise individual privacy. While a number of proposals, including query name minimization [RFC7816] and running DNS over an encrypted tunnel [RFC7858], have been made to mitigate that problem, they all appear to share the common properties of security patches rather than designed-in security or privacy mechanisms. While experience may prove otherwise once (and if) they are widely deployed, it does not appear that any of them are as satisfactory as a system with query privacy designed in might be. More general tutorials on this issue have appeared recently [Huston2017a].3.6. Alternate Namespaces for Public Use in the DNS Framework: The CLASS Problem
The DNS standards include specification of a CLASS value, which "identifies a protocol family or instance of a protocol" (RFC 1034, Section 3.6, and elsewhere). While CLASS was used effectively in the early days of the DNS to manage different protocol families within the same administrative environment, recent attempts to use it to either partition the DNS namespace in other ways such as for non-ASCII names (partially to address the issues in Sections 3.2 and 3.3) or use DNS mechanisms for entirely different namespaces have exposed fundamental problems with the mechanism [Sullivan-Class]. Perhaps the most fundamental of those problems is disagreement about whether multiple CLASSes were intended to exist within a given zone (with records within RRSETs) or whether different CLASSes implied different zones. Different implementations make different assumptions [Faltstrom-2004] [Vixie-20170704]. These problems have led to recommendations that it be dropped entirely [Sullivan-Class], but discussions on the IETF list and in WGs in mid-2017 made it clear that there is no clear consensus on that matter.3.7. Loose Synchronization
The DNS model of master and slave servers, with the latter initiating updates based on expiration interval values, and local caches with updates based on TTL values, depends heavily on an approach that has come to be called "loose synchronization", i.e., that there can be no expectation that all of the servers that might reasonably answer a query will have exactly the same data unless those data have been unchanged for a rather long period. Put differently, if some or all of the records associated with a particular node in the DNS
(informally, a fully qualified domain name (FQDN)) change, one cannot expect those changes to be propagated immediately. That model has worked rather well since the DNS was first deployed, protecting the system from requirements for mechanisms that are typical where a simultaneous update of multiple systems is needed. Such mechanisms include elaborate locking, complex update procedures and handshaking, or journaling. As has often been pointed out with the Internet, implementation and operational complexity are often the enemy of stability, security, and robustness. Loose synchronization has helped keep the DNS as simple and robust as possible. A number of recent ideas about using the DNS to store data for which important changes occur very rapidly are, however, largely incompatible with loose synchronization. Efforts to use very short (or zero) refresh times (in SOA records for slave updates from masters) and TTLs (for caches) to simulate nearly simultaneous updating may work up to a point but appear to impose very heavy loads on servers and distribution mechanisms that were not designed to accommodate that style of working. Similar observations can be made about attempts to use the NOTIFY extension [RFC1996] or dynamic, "server-push", updating rather than the traditional DNS mechanisms. While the NOTIFY and push mechanisms normally provide refresh times and update mechanisms faster than those specified in RFCs 1034 and 1035, they imply that a "master" server must know the identities of (and have good connectivity to all of) its slaves. That defeats at least some of the advantages associated with stealth slaves, particularly those associated with reduction of query traffic across the Internet. Those mechanisms do nothing for cache updates: unless servers keep track of the source of every query for names associated with a specific zone and then somehow notify the query source systems, the only alternative to having information that might be obsolete stored in caches is to use very short or zero TTLs so the cached data time out almost immediately after being stored (or are not stored at all), requiring a new query to an authoritative server each time a resolver attempts to look up a name.3.8. Private Namespaces and Special Names
Almost since the DNS was first deployed, there have been situations in which it is desirable to use DNS-like names, and often DNS resolution mechanisms or modifications of them, with namespaces for which globally available and consistent resolution using the public DNS is either unfeasible or undesirable (and for which the use of CLASS is not an appropriate mechanism). The need to isolate names and addresses on LANs from the public Internet, typically via "split horizon" approaches, is one example of this requirement although often not recognized as such. Another example that has generated a
good deal of controversy involves "special names" -- labels or pseudo-labels, often in TLD positions, that signal that the full name should not be subject to normal DNS resolution or other processing [RFC6761] [RFC8244]. Independent of troublesome policy questions about who should allocate such names and the procedures to be used, they almost inherently require either a syntax convention to identify them (there actually was such a convention, but it was abandoned many years ago and there is no plausible way to reinstitute it) or tables of such names that are known to, and kept updated on, every resolver on the Internet, at least if spurious queries to the root servers are to be avoided. If the DNS were to be redesigned and replaced, we could recognize this requirement as part of the design and handle it much better than it is possible to handle it today.3.9. Alternate Query or Response Encodings
The DNS specifies formats for queries and data responses, based on the state of the art and best practices at the time it was designed. Recent work has suggested that there would be significant advantages to supporting at least a description of the DNS messages in one or more alternate formats, such as JSON [Hoffman-DNS-JSON] [Hoffman-SimpleDNS-JSON]. While that work has been carefully done to avoid requiring changes to the DNS, much of the argument for having such a JSON-based description format could easily be turned into an argument that, if the DNS were being revised, that format might be preferable as a more direct alternative to having DNS queries and responses in the original form.3.10. Distribution and Management of Root Servers
The DNS model requires a collection of root servers that hold, at minimum, information about top-level domains. Over the years, that requirement has evolved from a technically fairly minor function, normally carried out as a service to the broader Internet community and its users and systems, to a subject that is intensely controversial with regard to control of those servers, including how they should be distributed and where they should be located. While a number of mechanisms, most recently including making the information more local [RFC7706], have been proposed and one (anycast [RFC7094]) is in very active use to mitigate some of the real and perceived problems, it seems obvious that a DNS successor, designed for today's global Internet and perceived requirements, could handle these problems in a technically more appropriate and less controversial way. Some additional discussion of the issues involved appears in a recent paper [Huston2017b].
3.11. Identifiers versus Brands and Other Convenience Names
A key design element of the original network object naming systems for the ARPANET, largely inherited by the DNS, was that the names, while expected to be mnemonic, were identifiers and their being highly distinguishable and not prone to ambiguity was important. That led to restrictive rules about what could appear in a name. Those restrictions originated with the host table and even earlier [RFC0236] [RFC0247] and came to the DNS (largely via SMTP) as the "preferred syntax" (RFC 1034, Section 3.5) or what we now often call the letter-digit-hyphen (LDH) rule. Similar rules to make identifiers easier to use, less prone to ambiguity, or less likely to interfere with syntax occur frequently in more formal languages. For example, almost every programming language has restrictions on what can appear in an identifier, and Unicode provides general recommendations about identifier composition [Unicode-USA31]. Both are quite restrictive as compared to the number of characters and total number of strings that can be written using that character coding system. That model, which originally prohibited labels starting with digits in order to avoid any possible confusion with IP addresses, began to break down in 1987 or 1988 when a company named 3Com wanted to use its corporate name as a label within the COM TLD, and the rule was relaxed [RFC1123]. In the last decade or two, the perspective that company names should be supported if possible has expanded and done so largely without its limits, if any, being explicitly understood or acknowledged. In the current form, the DNS is really (and primarily) a system for expressing thoughts and concepts. Those include free expression of ideas in as close to natural language as possible as well as representation of product names and brands. That view requires letter-like characters that might not be reasonable in identifiers along with a variety of symbols and punctuation. It may also require indicators of preferred type styles to provide information in a form that exactly matches personal or legal preferences. At least if carried to an extreme, that perspective would argue for standardizing word and sentence separators, removing the limit of 63 octets per label and probably the limit of 255 octets on the total length of a domain name, and perhaps even eliminating the hierarchy or allowing separators for labels in presentation form (now fixed at "." for the DNS) to be different according to context. It suggests that, at least, the original design was defective in not prioritizing those uses over the more restrictive approach associated with prioritizing unique and unambiguous identifiers.
So we have two or, depending on how one counts, three very different use cases. The historical one is support for unique identifiers. The other is expression of ideas and, if one considers them separate, presentation of brand and product names. Because they inherently involve different constraints, priorities, and success criteria, these perspectives are, at best, only loosely compatible. We cannot simultaneously optimize both the identifier perspective and either or both of the others in the same system. At best, there are some complex trade-offs involved. Even then, it is not clear that the same DNS (or other system) can accommodate all of them. Until we come to terms with that, the differences manifest themselves with friction among communities, most often with tension between "we want to do (or use or sell) these types of labels" and "not good for the operational Internet or the DNS".3.12. A Single Hierarchy with a Centrally Controlled Root
A good many Internet policy discussions in the last two decades have revolved around such questions of how many top-level domains there should be, what they should be, who should control them and how, how (or if) their individual operations and policy decisions should be accountable to others, and what processes should be used (and by what entities or organizational structures) to make those decisions. Several people have pointed out that, if we were designing a next- generation DNS using today's technology, it should be possible to remove the technical requirement for a central authority over the root (some people have suggested that blockchain approaches would be helpful for this purpose; others believe they just would not scale adequately, at least at acceptable cost, but that other options are possible). Whether elimination of a single, centrally controlled, root would be desirable or not is fairly obviously a question of perspective and priorities.3.13. Newer Application Protocols, New Requirements, and DNS Evolution
New work done in other areas has led to demands for new DNS features, many of them involving data values that require recursively referencing the DNS. Early record types that did that were accompanied by restrictions that reduced the risk of looping references or other difficulties. For example, while the MX RRTYPE has a fully qualified domain name as its data, SMTP imposes "primary name" restrictions that prevent the name used from being, e.g., a CNAME. While loops with CNAMEs are possible, Section 3.6 of RFC 1034 includes a discussion about ways to avoid problems and how they should be handled. Some newer protocols and conventions can cause more stress. There are separate issues with additions and with how the DNS has been extended to try to deal with them.
3.13.1. The Extensions
Some examples of DNS extensions for new protocol demands that illustrate, or have led to, increased stress include: NAPTR: Requires far more complex data in the DNS for ENUM (e.g., Voice over IP (VoIP), specifically SIP) support, including URI information and hence recursive or repeated lookups, than any of the RRTYPEs originally supported. The RRSET associated with these records can become quite large because the separator between the various records is part of the RDATA, and not the {owner, class, type} triple (a problem slightly related to the problem with overloading of TXT RRTYPE discussed in Section 3.13.2). This problem, and similar ones for some of the cases below. may suggest that any future design is in need of a different TYPE model such as systematic arrangements for subtypes or some explicit hierarchy in the TYPEs. URI: Has a URI as its data, typically also requiring recursive or repeated lookups. Service location (SRV) and credential information (including Sender Policy Framework (SPF) and DomainKeys Identified Mail (DKIM)): Require structured data and, especially for the latter two, significantly more data than most original RRTYPEs. URI/URL: The early design decision for the World Wide Web that its mechanism for identifying digital web content (now known as Uniform Resource Identifiers [RFC3986]) did so by using domain names and hence the network location of the information or other material. That, in turn, has required systems intended to improve web performance by locating and retrieving a "nearest copy" (rather than the single copy designated by the URL) to intercept DNS queries and respond with values that are not precisely those stored for the designated domain name in the DNS or to otherwise access information in a way not supported by the DNS itself.3.13.2. Extensions and Deployment Pressures -- The TXT RRTYPE
Unfortunately (but unsurprisingly), and despite IETF efforts to make things easier [RFC6895], DNS support libraries have often been slow to add full support for new RRTYPEs. This has impeded deployment of applications that depend on those types and that must ask (query) explicitly for them. Both to get faster deployment and, at least until recently, to avoid burdensome IETF approval procedures, many application designers have chosen to push protocol-critical
information into records with TXT RRTYPE, a record type that was originally intended to include only information equivalent to comments. This causes two problems. First, TXT records used this way tend to get long and complex, which is a problem in itself if one is trying to minimize TCP connections. Second, applications that are attempting to obtain data cannot merely ask for the relevant QTYPE; they must obtain all of the records with QTYPE TXT and parse them to determine which ones are of interest. That would be easier if there was some standard for how to do that parsing, but, at least in part because the clear preference in the DNS design is for distinct RRTYPEs for different kinds of information, there is no such standard. (There was a proposal in 1993 to structure the TXT DATA in a way that would have addressed the issue [RFC1464], but it apparently never went anywhere.) On the other hand, this issue is somewhat different from most of the others described in this document because (as the IETF has recommended several times) the problem is easily solved within the current DNS design by allocating and supporting new RRTYPEs when needed rather than using TXT as a workaround (that does not mean that other solutions are impossible, either with the current DNS or with some other design). The problem then lies in the implementations and/or mechanisms that deter or impede rapid deployment of support for new RRTYPEs.3.13.3. Periods and Zone Cut Issues
One of the DNS characteristics that is poorly understood by non-experts is that the period (".", U+002E) character can be used in four different ways: o As a label separator in the presentation form that also designates a "zone break" (delegation boundary). For example, foo.bar.example.com indicates the owner, "foo", of records in the "bar.example.com" zone. o As a label separator in the presentation form that does not designate a zone break. For example, foo.bar.example.com indicates the owner, "foo.bar", of records in the "example.com" zone. o As a character within a label, including as a substitute for an at-sign ("@") when an email address appears in an SOA record or in a label that denotes such an address (see Section 2 above). The ability to embed periods in labels in this way has also led to attacks in which, e.g., a domain name consisting of the labels
"example" followed by "com" is deliberately confused with the
single label "example.com" with an embedded period.
o At the end of a fully qualified domain name to designate the root
zone, e.g., "example.com." (RFC 1034, Section 3.1).
In general, these cases cannot be distinguished by looking at them.
The third is problematic for non-DNS reasons, e.g.,
"john.doe.example.net" can be interpreted as either a simple FQDN or
as a notation for john@doe.example.net, john.doe@example.net, or even
(at least in principle) john.doe.example@net.
The distinction between the FQDN interpretation and the first email-
like one was probably not important as the DNS was originally
intended to be used. However, as soon as RRTYPEs (other than NS
records that define the zone cut) are used that are sensitive to the
boundaries between zones, the distinctions become important to people
other than the relevant zone administrators. DNSSEC [RFC4033]
involves one such set of relationships. It increases the importance
of questions about what should go in a parent zone and what should go
in child zones and how much difference it makes if NS records in a
parent zone for a child zone are consistent with the records and data
in the child zone. This also causes application issues and may raise
questions about relationships between registrars and one or more
registries or, if they are separate, DNS operators.
3.14. Scaling of Reputation and Other Ancillary Information
The original design for DNS administration, reflected in RFC 1591
[RFC1591] and elsewhere, assumed that all domains would exhibit a
very high level of responsibility toward and for the community and
that level of responsibility would be enforced if necessary.
More recent decisions, many of them associated with commercialization
of the DNS, have eroded those very strong assumptions of registry
responsibility and accountability to the point that many consider
decisions about delegation of names, identification of registrants,
and relationships among names to be matters of "registrant beware"
and even "user and applications beware". While some recent protocols
and proposals at least partially reflect that original model of a
high level of responsibility (see, e.g., IDNA [RFC5890] and a more
recent discussion [Klensin-5891bis]), other decisions and actions
tend to shift responsibility to the registrant or try to avoid
accountability entirely. One possible approach to the problems,
especially security problems, that are enabled by those new trends
and the associated environment is to establish reputation systems
associated with clearly defined administrative boundaries and with
warnings to users, even if those reputation systems are managed by parties not directly associated with the DNS. The IETF DBOUND WG [IETF-DBOUND] addressed ways to establish and document boundaries more precise than simple dependencies on TLDs, but it was not successful in producing a standard. A TLD reputation-based approach was adopted by some web browsers after IDNs and a growing number of Generic Top-Level Domains (gTLDs) were introduced; that approach was based on a simple list and does not scale to the current size of the DNS or even the DNS root.3.15. Tensions among Transport, Scaling, and Content
The original design for the DNS envisaged a simple query and response protocol where both the command and the response could be readily mapped into a single IP packet. The host requirements specification [RFC1123] required all DNS applications to accept a UDP query or response over UDP with up to 512 octets of DNS payload. TCP was seen as a fallback when the response was greater than this 512-octet limit, and this fallback to use TCP as the transport protocol was considered to be the exception rather than the rule. Over the intervening years, we have seen the rise of a common assumption of an Internet-wide Maximum Transmission Unit (MTU) size of 1,500 octets, accompanied with an assumption that UDP fragmentation is generally viable. This underpins the adoption of the Extension Mechanisms for DNS (EDNS(0)) [RFC6891] to, among other things, specify a UDP buffer size larger than 512 octets and a suggestion within that specification to use 4,096 as a suitable compromise for the UDP payload size. This has proved to be fortuitous for the DNSSEC security extensions where the addition of DNSSEC security credentials in DNS responses [RFC4034] can lead to the use of large DNS responses. However, this exposes some tensions over the handling of fragmentation in IP, where UDP fragments have been observed to be filtered by various firewalls. Additionally for IPv6, there are the factors of filtering the ICMPv6 Packet Too Big diagnostic messages and discarding the IPv6 packets that contain extension headers [RFC7872]. More generally, fragmented UDP packets appear to have a lower level of reliability than unfragmented TCP packets. Behind this observation about relative reliability of delivery is the tension between the lightweight load of UDP and the downside of elevated probability of discarding of packet fragments as compared to TCP, which offers increased levels of assurance of content delivery, but with the associated imposition of TCP session state and the downside of reduced DNS scalability and increased operational cost.
(next page on part 2)
