[Search] [txt|pdf|bibtex] [Tracker] [WG] [Email] [Diff1] [Diff2] [Nits]

Versions: 00 01 02 03 rfc2276                                           
Internet Draft                                          Karen R. Sollins
draft-ietf-urn-req-frame-01.txt                                  MIT/LCS
Expires September 28, 1997                                March 28, 1997

        Requirements and a Framework for URN Resolution Systems

Status of this draft
     This document is an Internet-Draft.  Internet-Drafts are working
     documents of the Internet Engineering Task Force (IETF), its
     areas, and its working groups.  Note that other groups may also
     distribute working documents as Internet-Drafts.

     Internet-Drafts are draft documents valid for a maximum of six
     months and may be updated, replaced, or obsoleted by other
     documents at any time.  It is inappropriate to use Internet-
     Drafts as reference material or to cite them other than as
     ``work in progress.''

     To learn the current status of any Internet-Draft, please check
     the ``1id-abstracts.txt'' listing contained in the Internet-
     Drafts Shadow Directories on ftp.is.co.za (Africa),
     nic.nordu.net (Europe), munnari.oz.au (Pacific Rim),
     ds.internic.net (US East Coast), or ftp.isi.edu (US West Coast).


This document addresses the issues of the discovery of local URN
resolver services that in turn will directly translate URNs into
URLs and URCs.  The document falls into three major parts, the
assumptions underlying the work, the requirements in order to be a
viable Resolver Discovery Service or RDS to help in finding URN
resolvers, and a framework for designing RDSs.  The requirements fall
into three major areas: evolvability, usability, and security and
privacy.  An RDS that is compliant with the framework will not
necessarily be compliant with the requirements.  Compliance with the
requirements will need to be validated separately.

1. Introduction

The purpose of this document is to lay out the engineering criteria
for what we will call here a Resolver Discovery Service (RDS), a
service to help in the learning about URN resolvers.  This is a
component of the realization of an information infrastructure.  In the
case of this work, that infrastructure is to be available, "in the
Internet" or globally, and hence the solutions to the problems we are
addressing must globally scalable.  In this work, we are focussing
specifically on naming of resources and resolution of those names to
the exclusion of other problems such as typing, resource access and
availability, security of the resources, etc.  Those are all important
problems, but not part of this effort.

The Uniform Resource Identifier Working Group defined a naming
architecture, as demonstrated in a series of three RFCs 1736[RFC1736],

                                 - 1 -

1737{RFC1737}, and 1738[RFC1738].  Although several further documents
are needed to complete the description of that architecture, it
incorporates three core functions often associated with "naming":
identification, location, and mnemonics or semantics.  By location, we
mean full-qualified Domain Names or IP addresses.  Names may provide
the ability to distinguish one resource from another, by
distinguishing their "names".  Names may help to provide access to a
resource by including "location" information.  Lastly, names may have
other semantic or mnemonic information that either helps human users
remember or figure out the names, or include other semantic
information about the resource being named.  The URI working group
concluded that there was need for persistent, globally unique
identifiers, distinct from location or other semantic information;
these "names" provide identity, in that if two of them are "the same"
(under some simple rule of canonicalization), they identify the same
resource.  Furthermore, the group decided that these "names" were
generally to be for machine, rather than human consumption.  One can
imagine a variety human-friendly naming (HFN) schemes supporting
different suites of applications and user communities.  These will
need to provide mappings to URNs in tighter or looser couplings,
depending on the namespace.  It is these that will be mnemonic,
content-full, and perhaps mutable, to track changes in use and
semantics.  They may provide nicknaming and other aliasing, relative
or short names, context sensitive names, descriptive names, etc.  The
URI naming architecture as described in the introductions to RFCs 1736
and 1737 lays out three sorts of components to the naming
architecture: identifiers called Uniform Resource Names (URNs),
locators called Uniform Resource Locators (URLs) and semantic
meta-information called Uniform Resource Characteristics (URCs).  This
document focusses on part of the problem of the translation from URN
to URL and/or URC.

Within the URI community there has been a concept used frequently that
for lack of a better term we will call a _hint_.  A hint is something
that helps in the resolution of a URN; we map URNs to hints as an
interim stage in accessing a resource.  A hint may also have
meta-information associated with it, such as an expiration_time or
certification of authenticity.  We expect that these will stay with a
hint rather than being managed elsewhere.  We will assume in all
further discussion of hints that they include any necessary
meta-information as well as the hint information itself.  Examples of
hints are: 1) the name of a resolver service that may further resolve
the URN, 2) the address of such a service, 3) a location at which the
resource was previously found.  The defining feature of hints is that
they are only hints; they may be out of date, temporarily invalid, or
only applicable within a specific locality.  They do not provide a
guarantee of access, but they probably will help in the resolution
process.  We must assume that most resolutions of URNs will be
provided by the use of locally stored hints, because maintaining a
database of globally available, completely up-to-date location
information is infeasible for performance reasons.  There are a number
of circumstances in which one can imagine that hints become invalid,
either because a resource has moved or because a different URN
resolver service has taken over the responsibility for resolution of
the URN.  Hints may be found in a variety of places.  It is generally
assumed that a well engineered system will maintain a set of hints for

                                 - 2 -

each URN at each location where that URN is found.  In addition, for
those situations in which those hints found locally fail, a
well-engineered system will provide a fall-back mechanism for
discovering further hints.  It is this fall-back mechanism, an RDS,
that is being addressed in this document.  As with all hints, there
can never be a guarantee that access to a resource will be available
to all clients, even if the resource is accessible to some.  However,
an RDS is expected to work with reasonably high reliability, and,
hence, may result in increased response time.

The remainder of this document falls into three sections.  The first
identifies several sets of assumptions underlying this work.  The next
lays out the requirements for a Resolver Discovery Service.  This
section is probably the most critical of the document, because it is
this that provides the metric for whether or not a proposed scheme for
a RDS is adequate or not.  For the reader short on time, each of the
three major subsections of the requirements section begins with a
summary list of the requirements identified in that section.  The
final section of the document lays out a framework for such RDSs.  The
purpose of this last section is to bound the search space for RDS
schemes.  One must be careful not to assume that because an RDS scheme
fits within the framework that it necessarily meets the requirements.
As will be discussed further in this last section, designing within
the framework does not guarantee compliance, so compliance evaluation
must also be part of the process of evaluation of a scheme.

2. Assumptions

Based on previous internet drafts and discussion in both the URN BOFs
and on the URN WG mailing list, three major areas of assumptions are
apparent: longevity, delegation, and independence.  Each will be
discussed separately.

The URN requirements state that a URN is to be a "persistent
identifier".  It is probably the case that nothing will last forever,
but in the time frame of resources, users of those resources, and the
systems to support the resources, the identifier should be considered to
be persistent or have a longer lifetime than those other entities.
There are two assumptions that are implied by longevity of URNs:
mobility and evolution.  "Mobility" assumes that everything will move
over the life of a URN.  For example, resources will move from one
machine to another, because individual machines have a much shorter
lifetime than resources, generally measured in a number of years less
than a decade.  Owners of resources may move and wish their resources to
follow them.  The services themselves will move.  "Evolution" assumes
that the supporting infrastructure will evolve.  This may take the form
of entirely new transport protocols or new versions of existing
protocols.  Furthermore, services such as storage services may evolve;
it is even possible that within a human lifetime the Unix file system
model may no longer be in use!  Clearly there will be evolution of and
improvement in supporting authentication and security mechanisms.  These
are only examples.  In general, we must assume that almost any piece of
the supporting infrastructure of URN resolution will evolve.  In order
to deal with both the mobility and evolution assumptions that derive
from the assumption of longevity, we must assume that users and their

                                 - 3 -

applications can remain independent of these mutating details of the
supporting infrastructure.

The second and third assumptions are two forms of modularity: delegation
and isolation.  The delegation assumption is that an entity may
partition and pass off some of its authority or responsibility.  One of
those responsibilities is for assigning URNs; practically speaking,
there cannot be only a single authority for assigning URNs.  We expect
that there will be a multi-tiered naming authority delegation.
Furthermore, it is difficult to imagine a non-partitioned and delegated
global RDS, meaning that hint discovery and resolution will be
partitioned and delegated.  In some RDS schemes, the delegation of
naming authority will form a basis for delegating the management and
dispensing of location information.

The third assumption is independence or isolation of one authority from
another and, at least to some extent from its parent.  Underlying much
of the thinking and discussion in the URI and URN working groups has
been the assumption that when a component delegates authority to another
component, the delegatee can operate in that domain independently of its
peers and within bounds specified by the delegation, independently of
the delegator.  This isolation is critically important in order to allow
for independence of policy and mechanism.

There are a number of more specific assumptions that fall under this
rubric of isolation.  First, we assume that the publisher of a resource
can choose resolver services, independently of choices made by others.
At any given time, the owner of a namespace may choose a particular URN
resolver service for that delegated namespace.  Such a URN resolver
service may be outside the RDS service model, and just identified or
located by the RDS service.  Second, it must be possible to make a
choice among RDS services, perhaps based on different underlying
internal architectures.  The reason that this is an assumption is that
there must be an evolutionary path through a sequence of core RDS
services.  Although at any given time there is likely to be only one or
a small set of such services, the number is likely to increase during a
transition period from one architecture to another.  Thus, it must be
assumed that clients can make a choice among a probably very small set
of RDSs.  Third, there must be independence in the choice about levels
and models of security and authenticity required.  This choice may be
made by the owner of a naming subspace, in controlling who can modify
hints in that subspace.  A naming authority may delegate this choice to
the owners of the resources named by the names it has assigned.  There
may be limitations on this freedom of choice in order to allow other
participants to have the level of security and authenticity they
require, for example, in order to maintain the integrity of the RDS
infrastructure as a whole.  Fourth, there is an assumption of
independence of choice of the rule of canonicalization of URNs within a
namespace, limited by any restrictions or constraints that may have been
set by its parent namespace.  This is a choice held by naming
authorities over their own subnamespaces.  Rules for canonicalization
will be discussed further in the framework section below.  Thus, there
are assumptions of independence and isolation to allow for delegated,
independent authority in a variety of domains.

                                 - 4 -

The modularity assumptions of delegation and isolation imply
independence of decision and implementation, leading to a
decentralization that provides a certain degree of safety from denial of
service.  Based on these these assumptions in conjunction with that of
longevity and those for URLs and URNs as detailed in RFCs 1736 and 1737,
we can now turn to the requirements for a Resolver Discovery Service.

3. Requirements

The requirements applying to a Resolver Discovery Service or RDS
center around three important design goals: evolvability, usability,
and security and privacy.  At its core the function of an RDS is to
provide hints for accessing a resource given a URN for it.  These
hints may range in applicability from local to global, and from
short-lived to long-lived.  They also may vary in their degree of
verifiable authenticity.  While it may be neither feasible nor
necessary that initial implementations support every requirement,
every implementation must support evolution to systems that do support
every requirement.

It is important to note that there are other requirements, not
applicable specifically to an RDS that must also be met.  A whole URN
system will consist of namespaces, the resolution information for them,
and the mapping from names in the namespaces to resolution information
(or hints).  URN schemes must meet the requirements of RFC 1737.
Resolution information, to the extent it is expressed as URLs must meet
the requirements of RFC 1736.  But this does not tell the whole story.
Although the URN working group will identify several acceptable
namespaces and the rules binding them, such as how delegation occurs,
how it is expressed in the names, how and to what extent binding to hint
information will be constrained by the namespace, in the long run a
document will be needed to guide the evaluation criteria for acceptance
of new namespaces.  These are not included in the list of requirements
below because they are not requirements for an RDS, but rather for naming
schemes themselves.

Each section below begins with a summary of the points made and discussed
in the following discussion.  It is worth noting here that there is
some degree of overlap in the areas of requirements, such as in
allowing for the evolution of security mechanisms, etc.  Issues may
appear in more than one requirement.  It is also important to
recognize that conformance with the requirements may often be
subjective.  Most of these requirements are not quantifiable and hence
conformance is a judgment call and a matter of degree.  Lastly, the
reader may find that some of the requirements are those of general
applicability to distributed systems and some are specific to URN
resolution.  Those of general applicability are included for
completeness and are not distinguished as such.

3.1 Evolution

The requirements in the area of evolvability are:

   [R1] To support evolution of mechanisms, specifically for
          {R1.1] a growing set of URN schemes;

                                 - 5 -

          [R1.2] new kinds local URN resolver services;
          [R1.3] new authentication schemes;
          [R1.4] alternative RDS schemes active simultaneously;
   [R2] To support the separation of global identification from
        location information.
   [R3] To allow for the support the development and deployment of
        administrative control mechanisms to manage human behavior
        with respect to limited resources.

One of the lessons of the Internet that we must incorporate into the
development of mechanisms for resolving URNs is that we must be prepared
for change.  Such changes may happen slowly enough to be considered
evolutionary modifications of existing services or dramatically enough
to be considered revolutionary.  They may permeate the Internet universe
bit by bit, living side by side with earlier services or they may take
the Internet by storm, causing an apparent complete transformation over
a short period of time.  There are several directions in which we can
predict the need for evolution, even at this time, prior to the
deployment of any such service.  At the very least, the community and
the mechanisms proposed should be prepared for these.

First, we expect there to be additions and changes to the mechanisms.
The community already understands that there must be a capacity for new
URN schemes.  A URN scheme will define a set of URNs that meet the URN
requirements[RFC1737], but may have further constraints on the internal
structure of the URN.  The requirements document would allow for an
overall plan in which URN schemes are free to specify parts of the URN
that are left opaque in the larger picture.  In fact, a URN scheme may
choose to make public the algorithms for any such "opaque" part of the
URN.  For example, although it may be unnecessary to know the structure
of an ISBN, the algorithm for understanding the structure of an ISBN has
been made public.  Other schemes may either choose not to make their
algorithms public, or choose a scheme in which knowledge of the scheme
does not provide any significant semantics to the user.  In any case, we
must be prepared for a growing number of URN schemes.

Often in conjunction with a new URN scheme, but possibly independently
of any particular URN scheme, new kinds of resolver services may
evolve.  For example, one can imagine a specialized resolver service
based on the particular structure of ISBNs that improves the
efficiency of finding documents given their ISBNs.  Alternatively, one
can also imagine a general purpose resolver service that trades
performance for generality; although it exhibits only average
performance resolving ISBNs, it makes up for this weakness by
understanding all existing URN schemes, so that its clients can use
the same service to resolve URNs regardless of naming scheme.  In this
context, there will always be room for improvement of services,
through improved performance, better adaptability to new URN schemes,
or lower cost.  In any case, new models for URN resolution will evolve
and we must be prepared to allow for their participation in the
overall resolution of URNs.

If we begin with one overall plan for URN resolution, into which the
enhancements described above may fit, we must also be prepared for an

                                 - 6 -

evolution in the authentication schemes that will be considered either
useful or necessary in the future.  There is no single globally accepted
authentication scheme, and there may never be one.  Even if one does
exist at some point in time, there will always be threats to it, and so
we must always be prepared to move on to newer and better schemes, as
the old ones become too easily spoofed or guessed.

Lastly, in terms of mechanism, although we may develop and deploy a
single RDS scheme initially, we must be prepared for that top level
model to evolve.  Thus, if the RDS model supports an apparently
centralized (from a policy standpoint) scheme for inserting and
modifying authoritative information, over time we must be prepared to
evolve to a different model, perhaps one that has a more distributed
model of authority and authenticity.  If the model has no core but
rather a cascaded partial discovery of information, we may find that
this becomes unmanageable with an increase in scaling.  Whatever the
core of the model, we must be prepared for it to evolve with changes in
scaling, performance, and policy constraints such as security and cost.

Second, in addition to the evolution of resolution mechanisms, we
expect that the community will follow an evolutionary path towards the
separation of location information from identification.  The URN
requirements document suggested this path as well, and there has been
general agreement in much of the community that such a separation is
desirable.  This is a problem that the public at large has generally
not understood.  Today we see the problem most clearly with the use of
URLs for identification.  When a web page moves, its URL becomes
invalid.  Suppose such a URL is embedded in some page, stored in long
term storage.  There are three possible outcomes to this scenario.
One possibility is that the client is left high and dry with some
message saying that the page cannot be found.  Alternatively, a
"forwarding pointer" may be left behind, in the form of an explicit
page requesting the client to click on a new URL.  Although this will
allow the client to find the intended page, the broken link cannot be
fixed because the URL is embedded in a file outside of the client's
control.  A third alternative is that the target server supplies an
HTTP redirect so that the new page is provided for the client
automatically.  In this case, the client may not even realize that the
URL is no longer correct.  The real problem with both of these latter
two situations is that they only work as long as the forwarding
pointer can be found at the old URL.  Location information, was
embedded in the identifier, and the resolution system was designed to
depend on that location information being correct.  There are few
cases in which we can expect such information to remain valid for a
long time, but in many cases references need to have long lifespans.
Most documents are only useful while their references still function.
To the extent that an RDS scheme supports the separation of global
identification from location information it will be encouraging the
longer utility of the identities.

A third evolutionary requirement is even more mechanical than the
others.  At any point in time, the community is likely to be
supporting a compromise position with respect to resolution.  We will
probably be operating in a situation balanced between feasibility and
the ideal, perhaps with policy controls used to help stabilize the

                                 - 7 -

service.  Ideally, the service would be providing exactly what the
customers wanted and they in turn would not request more support than
they need.  Since we will always be in a situation in which some
service provision resources will be in short supply, some form of
policy controls will always be necessary.  Some policy controls may be
realized as mechanisms within the servers or in the details of
protocols, while others may only be realized externally to the system.
For example, suppose hint entries are being submitted in such volume
that the hint servers are using up their excess capacity and need more
disk space.  An effective solution to this problem would be a
mechanism such as a pricing policy.  This pricing policy has the dual
effect of both encouraging conservative use of resources and
collecting revenue for the improvement and maintenance of the system.
We can also imagine administrative policy controls with the force of
laws or other social pressures behind them, but with no technical
mechanism enforcing or enabling them.  As technology changes and the
balance of which resources are in short supply changes, the mechanisms
and policies for controlling their use must evolve as well.

3.2 Usability and Feature Set Requirements

To summarize, the usability requirements fall into three areas based on
participation in hint management and discovery:

   * The publisher
      [R4] URN to hint resolution must be correct and efficient with
           very high probability;
      [R5] Publishers must be able to select and move among URN
           resolver services to locate their resources;
      [R6] Publishers should be able to arrange for multiple access
           points for their location information;
      [R7] Publishers must be able to provide for both long-lived and
           short-lived hints;
      [R8] It must be relatively easy for publishers to specify to the
           management and observer their hint information as well as
           any security constraints they need for their hints.
   * The client
      [R9] The interface to the RDS must be simple, effective, and
      [R10] The client and client applications must be able to understand
           the information stored in and provided by the RDS easily,
           in order to be able to make informed choices.
   * The management
      [R11] The management of hints must be as unobtrusive as possible,
           avoiding using too many network resources;
      [R12] The management of hints should allow for administrative
           controls that encourage certain sorts of behavior deemed
           necessary to meet other requirements;
      [R13] The configuration and verification of configuration of
           individual RDS servers must be simple enough not to
           discourage configuration and verification.

Usability can be evaluated from three distinct perspectives: those of a
publisher wishing to make a piece of information public, those of a

                                 - 8 -

client requesting URN resolution, and those of the provider or manager
of resolution information.  We will separately address the usability
requirements from each of these three perspectives.

It is worth noting that there are two additional sorts of participants
in the whole naming process, as discussed in the URN WG.  They are the
naming authorities which choose and assign names, and the authors who
include URNs in their resources.  These two are not relevant to the
design of an RDS and hence are not discussed further here.

3.2.1 The Publisher

The publisher must be able to make URNs known to potential customers.
From the perspective of a publisher, it is of primary importance that
URNs be correctly and efficiently resolvable by potential clients with
very high probability.  Publishers also stand to gain from long-lived
URNs, since they increase the chance that references continue to point
to their published resources.

The publisher must also be able to choose easily among a variety of
potential services that might translate URNs to location information.
In order to allow for this mobility among resolver services, the
architecture for resolver services specified within the IETF should
not result in a scenario in which changing from one resolver service
to another is an expensive operation.

The publisher should be able to arrange for multiple access points to a
published resource.  For this to be useful, resolver services should
be prepared to provide different resolution or hint information to
different clients, based on a variety of information including location
and the various access privileges the client might have.  For example,
companies might arrange for locally replicated copies of popular
resources, and would like to provide access to the local copies only for
their own employees.  This is distinct from access control on the
resource as a whole, and may be applied differently to different copies.

The publisher should be able to provide both long and short term
location information about accessing the resource.  Long term
information is likely to be such information as the long term or the
location or identity of a resolver service with which the publisher
has a long term relationship.  One can imagine that the arrangement
with such a long term "authoritative" resolver service might be a
guarantee of reliability, resiliency to failure, and atomic updates.
Shorter term information is useful for short term changes in services
or to avoid short lived congestion or failure problems.  For example,
if the actual repository of the resource is temporarily inaccessible,
the resource might be made available from another repository.  This
short term information can be viewed as temporary refinements of the
longer term information, and as such should be more easily and quickly
made available, but may be less reliable.

Lastly, the publishers will be the source of much hint information that
will be stored and served by the manager of the infrastructure.  Despite
the fact that many publishers will not understand the details of the RDS
mechanism, it must be easy and straightforward to install hint

                                 - 9 -

information.  The publisher must be able not only to express hints, but
also to verify that what is being served by the manager is correct.
Furthermore, to the extent that there are security constraints on hint
information, the publisher must be able to both express them and verify
compliance with them easily.

3.2.2 The Client

From the perspective of the client, simplicity and usability are
paramount.  Of critical importance to serving clients effectively is
that there be an efficient protocol through which the client can acquire
hint information.  Since resolving the name is only the first step on
the way to getting access to a resource, the amount of time spent on it
must be minimized.

Furthermore, it will be important to be able to build simple, standard
interfaces to the RDS so that both the client and applications on the
client's behalf can interpret hints and subsequently make informed
choices.  The client, perhaps with the assistance of the application,
must be able to specify preferences and priorities and then apply them.
If the ordering of hints is only partial, the client may become directly
involved in the choice and interpretation of them and hence they must be
understandable to that client.  On the other hand, in general it should
be possible to configure default preferences, with individual
preferences viewed as overriding any defaults.

From the client's perspective, although URNs will provide important
functionality, the client is most likely to interact directly only with
human friendly names (HFNs).  As in direct human interaction (not
computer mediated), the sharing of names will be on a small, private, or
domain specific scale.  HFNs will be the sorts of references and names
that are easy to remember, type, choose among, assign, etc.  There will
also need to be a number of mechanisms for mapping HFNs to URNs.  Such
services as "yellow pages" or "search tools" fall into this category.
Although we are mentioning HFNs here, it is important to recognize that
HFNs and the mappings from HFNs to URNs is and must remain a separate
functionality from an RDS.  Hence, although HFNs will be critical to
clients, they do not fall into the domain of this document.

3.2.3 The Management

Finally, we must address the usability concerns with respect to the
management of the hint infrastructure itself.  What we are terming
"management" is a service that is distinct from publishing; it is the
core of an RDS.  It involves the storage and provision of hints to the
clients, so that they can find published resources.  It also provides
security to the extent that there is a commitment for provision of such
security; this is addressed below.

The management of hints must be as unobtrusive as possible. First, its
infrastructure (hint storage servers and distribution protocols) should
have as little impact as possible on other network activities.  It must
be remembered that this is an auxiliary activity and must remain in the

                                 - 10 -

Second, in order to make hint management feasible, there may need to
be a system for administrative incentives and disincentives such as
pricing or legal restrictions.  Recovering the cost of running the
system is only one reason for levying charges.  The introduction of
payments often has a beneficial impact on social behavior.  It may be
necessary to discourage certain forms of behavior that when out of
control have serious negative impact on the whole community.  At the
same time, any administrative policies should encourage behavior that
benefits the community as a whole.  Thus, for example, a small
one-time charge for authoritatively storing a hint will encourage
conservative use of hints.  If we assume that there is a fixed cost
for managing a hint, then the broader its applicability across the URN
space, the more cost effective it is.  That is, when one hint can
serve for a whole collection of URNs, there will be an incentive to
submit one general hint over a large number of more specific hints.
Similar policies can be instituted to discourage the frequent changing
of hints.  In these ways and others, behavior benefitting the
community as a whole can be encouraged.

Lastly, symmetric to issues of usability for publishers, it must also
be simple for the management to configure the mapping of URNs to
hints.  It must be easy both to understand the configuration and to
verify that configuration is correct.  With respect to management,
this requirement may have an impact not only on the information itself
but also on how it is partitioned among network servers that
collaboratively provide the management service or RDS.  For example,
it should be straightforward to bring up a server and verify that the
data it is managing is correct.  Although this is not a requirement,
it is worth nothing that since we are discussing a global and probably
growing service, encouraging volunteer participants suggests that, as
with the DNS, such volunteers can feel confident about the service
they are providing and its benefit to both themselves and the rest of
the community.

3.3 Security and Privacy Requirements

SUMMARY: Security and privacy requirements can be identified as some
degree of protection from threats.  These requirements are all stated
in terms of possibilities or options for users of the service to
require and utilize.  Hence they are requirements for the availability
of functionality, but not for the use of it.  We recognize that all
security is a matter of degree and compromise.  These may not satisfy
all potential customers, and there is no intention here to prevent
them from building more secure servers with more secure protocols to
suit their needs.  These are intended to satisfy the needs of the
general public.

   [R14] It must be possible to create authoritative versions of a hint
         with access-to-modification privileges controlled;
   [R15] It must be possible to determine the identity of servers or avoid
         contact with unauthenticated servers;
   [R16] It must be possible to reduce the threat of denial of service
         by broad distribution of information across servers.

                                 - 11 -

   [R17] It must be possible within the bounds of organization policy
         criteria to provide at least some degree of privacy for
   [R18] It must be possible for publishers to keep private certain
         information such as an overall picture of the resources they are
         publishing and the identity of their clients;
   [R19] It must be possible for publishers to be able to restrict
         access to the resolution of the URNs for the resources they
         publish, if they wish.

When one discusses security, one of the primary issues is an
enumeration of the threats being considered for mitigation.  The
tradeoffs often include cost in money and computational and
communications resources, ease of use, likelihood of use, and
effectiveness of the mechanisms proposed.  With this in mind, let us
consider a set of threats.

A good place to begin is with the early work of Voydock and Kent
[VK83].  They identify unauthorized release of information as a
passive attack, and all three of unauthorized modification of
information, denial of service, and spurious association initiation as
active attacks.  An intruder at any protocol layer can attack at any
of the links or computational elements (hosts, routers, etc.) at that
layer.  Attacks at one layer can be achieved by subverting or
attacking the lower layers.  An unauthorized release of information is
a violation of privacy or confidentiality.  This may be achieved by a
release of the information itself.  Additional passive threats are
from secondary information through traffic analysis or other
violations of transmission security, such as noticing lengths and/or
sources and destinations of traffic.  Moving to the active threats,
unauthorized modification of information can be partitioned into
problems with authenticity, integrity and ordering.  Denial of service
may take the form of discarding information before it reaches its
destination or some degree of delay in delivering information.
Finally, spurious association may occur when a previous legitimate
association initiation is played back or an initiation is made under
false identity.  Security measures may take the form of either
detection or prevention of each of these threats.  Within the scope of
this work, we must identify those threats that are both of concern and
that we expect to be able to mediate.  Of these threats the prevention
of passive attacks is known to be a particularly difficult problem to
address in the general case.

Of these threats, the passive threats to privacy or confidentiality
and the active threats of authenticity and integrity are probably the
most important to consider here.  To the extent that spurious
association causes threats to the privacy, authenticity, or integrity
with respect to information within servers managing data, it is also
important.  Because updates to hint information are idempotent, at
least with short periods of time, we will set aside the problems of
ordering for this analysis.  Denial of service is probably the most
difficult of these areas of threats both to detect and to prevent, and
we will therefore set it aside for the present as well, although it
will be seen that solutions to other problems will also mitigate some
of the problems of denial of service.  Furthermore, because this is

                                 - 12 -

intended to be provide a global service to meet the needs of a variety
of communities, the engineering tradeoffs will be different for
different clients.  Hence the requirements are stated in terms of,
"It must be possible..."  It is important to note that the
information of concern here is hint information, which by nature is
not guaranteed to be correct or up-to-date; therefore, it is unlikely
to be worth putting too much expense into the correctness of hints,
because there is no guarantee that they are still correct anyway.  But
the exact choice of degree of privacy, authenticity, and integrity
must be determined by the needs of the client and the availability of
services from the server.

There is one further issue to address at this point, the distinction
between mechanism and policy.  In general, a policy is realized by
means of a set of mechanisms.  In the case of an RDS there may be
policies internal to the RDS that it needs to have supported in order
to do its business as it sees fit.  Since, in general it is in the
business of storing and distributing information, most of its security
policies may have to do with maintaining its own integrity, and are
rather limited.  Beyond that, to the degree possible, it should impose
no policy on its customers, the publishers and users.  It is they that
may have policies that they would like supported by the RDS.  To that
end, an RDS should provide a spectrum of "tools" or mechanisms that
the customers can cause to be deployed on their behalf to realize
policies.  An RDS may not provide all that is needed by a customer.  A
customer may have different requirements within his or her
administrative bounds than outside.  Thus, "it must be possible..."
captures the idea that the RDS must generally provide the tools to
implement policies as needed by the customers.

The first approach to URN resolution is to discover local hints.  In
order for hints to be discovered locally, they will be as widely
distributed to what is considered to be local for every locale.  The
drawback of such wide distribution is the wide distribution of
updates, causing network traffic problems or delays in delivering
updates.  An alternative model would concentrate hint information in
servers, thus requiring that update information only be distributed to
these servers.  In such a model the vulnerable points are the sources
of the information and the distribution network among them.  Attackers
on the integrity of the information stored in a server may come in the
form of other a fake owner of the information or a fake server to the
extent that servers exchange updates with each other.  Wide
replication of information among servers increases the difficult of
masquerading at all the locations of the information as well as
reducing the threat of denial service.  These lead us to three
identifiable goals for our security model:

* ACCESS CONTROL ON HINTS: It must be possible to create an
  authoritative version of each hint with change control limited only
  to those principals with the right to modify it.  The choice of who
  those principals are or whether they are unlimited must be made by
  the publisher of a hint.

                                 - 13 -

* SERVER AUTHENTICITY: Servers and clients must be able to learn the
  identity of the servers with which they communicate.  This will be a
  matter of degree and it is possible that there will be more
  trustworthy, but less accessible servers, supported by a larger
  cluster of less authenticatable servers that are more widely
  available.  In the worst case, if the client receives what appears to
  be unvalidated information, the client should assume that the hint
  may be inaccurate and confirmation of the data might be sought from
  more reliable but less accessible data.

* SERVER DISTRIBUTION: Broad availability will provide resistance to
  denial of service.  It is only to the extent that the services are
  available that they provide any degree of trustworthiness.  In
  addition, the distribution of services will reduce vulnerability
  of the whole community, by reducing the trust put in any single
  server.  This must be mitigated by the fact that to the extent trust
  is based on a linked set of servers, if any one fails, the whole
  chain of trust fails; the more elements there are in such a chain,
  the more vulnerable it may become.

Privacy is a more difficult problem to address.  It may be a
double-edged sword; for example, an organization may consider it
critically important that its competitors not be able to read its
traffic, while it may also consider it important to be able to monitor
exactly what its employees are transmitting to and from whom, for a
variety of reasons such as reducing the probability that its employees
are giving or selling the company's secrets to verifying that
employees are not using company resources for private endeavor.  Thus,
although there are likely to be needs for privacy and confidentiality,
what they are, who controls them and how, and by what mechanisms vary
widely enough that it is difficult to say anything concrete about them

The privacy of publishers is much easier to safeguard.  Since they are
trying to publish something, in general privacy is probably not
desired.  However, publishers do have information that they might like
to keep private: information about who their clients are, and
information about what names exist in their namespace.  The
information about who their clients are may be difficult to collect
depending on the implementation of the resolution system.  For
example, if the resolution information relating to a given publisher
is widely replicated, the hits to _each_ replicated copy will need to
be recorded.  Of course, determining if a specific client is
requesting a given name can be approached from the other direction, by
watching the client as we saw above.

The other privacy issue for publishers has to do with access control
over URN resolution.  This issue is dependent on the implementation of
the publisher's authoritative URN resolver server.  URN resolver
servers can be designed to require proof of identity in order to be
issued resolution information; if the client does not have permission
to access the URN requested, the service denies that such a URN
exists.  An encrypted protocol can also be used so that both the
request and the response are obscured.  Encryption is possible in this

                                 - 14 -

case because the identity of the final recipient is known (i.e. the
URN server).

4. The Framework

With these assumptions and requirements in mind, one can conclude with a
general framework within which RDS designs will fall.  As stated
earlier, although this framework is put forth as a suggested guide for
RDS designers, compliance with it will in no way guarantee compliance
with the requirements.  Such an evaluation must be performed separately.
It is also understood that there may be RDS services that do not meet
the requirements in clearly identified ways.  This may be true
especially with early plans and experiments.  For example, although a
careful threat analysis may have been done to understand security
requirements, not all those security requirements may be addressed, in
order to use existing facilities to allow for early deployment for
experimentation purposes.  All such lack of compliance should be clearly

The design of the framework is based on a simple assumption about the
syntax of a URN a documented in RFC-XXX[RFCXXX}.  This assumed syntax


where URN: is a prefix on all URNs, NID is the namespace identifier,
and NSS is the namespace specific string.  The prefix identifies each
URN as such.  The NID determines the general syntax for all URNs
within its namespace.  The NSS is probably partitioned into a set of
delegated and subdelegated namespaces, and this is probably reflected
in further syntax specifications.  In more complex environments, each
delegated namespace will be permitted to choose the syntax of the
variable part of the namespace that has been delegated to it.  In
simpler namespaces, the syntax will be restricted completely by the
parent namespace.  For example, although the DNS does not meet all the
requirements for URNs, it has a completely restricted syntax, such
that any further structuring must be done only by adding further
refinements to the left, maintaining the high order to low order,
right to left structure.  A delegated syntax might be one in which a
host is named by the DNS, but to the right of that and separated by an
"@" is a string whose internal ordering is defined by the file system
on the host, which may be defined high order to low order, left to
right.  Of course, much more complex and nested syntaxes should be
possible, especially given the need to grandfather namespaces.  In
order to resolve URNs, rules will be needed for two reasons.  One is
simply to canonicalize those namespaces that do not fall into a
straightforward (probably right to left or left to right) ordering of
the components of a URN, as determined by the delegated naming
authorities involved.  It is also possible that rules will be needed
in order to derive from URNs the names of RDS servers to be used in

The NID defines a top level syntax.  This syntax will determine
whether the NID alone or in conjunction with some extraction from the
NSS (for the top level naming authority name) is to be used to

                                 - 15 -

identify the first level server to be contacted.  At each stage of the
lookup either a new rule for generating the strings used in yet
another lookup (the strings being the identity of another RDS server
and possibly a string to be resolved if it is different than the
original URN) or a reference outside the RDS to a private URN
resolver service, sidestepping any further use of the RDS scheme.
Figure 1 depicts this process.

                       |Global NID registry|
              (return rule or URN resolver service reference)
                                 |                                  |
                       +->(apply rule to determine RDS server)      |
                       |         |                                  |
                       |         |                                  |
                       |         |                                  |
                       |    +----------+                            |
                       |    |RDS server|          +-----------------+
                       |    +----------+          |
                       |      |   |               v
                       |      |   |   (set of choices)
                       |      |   +----+----------(...)--------+
                       |   (rule)      |                       |
                       |      |        |                       |
                       |      |        |                       |
                       +------+        |                       |
                                       v                       v
                                  +----------+            +----------+
                                  |private   |            |private   |
                                  |URN       |            |URN       |
                                  |resolver  |            |resolver  |
                                  |service   |            |service   |
                                  +----------+            +----------+

        Figure 1: An RDS framework

There are several points worth noting about the RDS framework.  First,
it leaves open the determination of the protocols, data organization,
distribution and replication needed to support a particular RDS

                                 - 16 -

scheme.  Second, it leaves open the location of the computations
engendered by the rules.  Third, it leaves open the possibility that
partitioning (distribution) of the RDS database need not be on the
same boundaries as the name delegation.  This may seem radical to
some, but if the information is stored in balanced B-trees for
example, the partitioning may not be along those naming authority
delegation boundaries.  Lastly, it leaves open access to the Global
NID Registry.  Is this distributed to every client, or managed in
widely distributed servers?

One concept that has not been addressed in Figure 1 is that there may
be more than one RDS available at any given time, in order to allow
for evolution to new schemes.  Thus, the picture should probably look
more like Figure 2.

                   |                               |
                   |                               |
                   |                               |
                   v                               v
         +---------------------+        +---------------------+
         |Global NID registry 1|        |Global NID registry N|
         +---------------------+        +---------------------+
                   .                               .
                   .                               .
                   .                               .

        Figure 2: More than one co-existing RDS scheme

If we are to support more than one co-existing RDS scheme, there will
need to be coordination between them with respect to storage and
propagation of information and modifications.  The issue is that
generally it should be assumed that all information should be available
through any operational RDS scheme.  One cannot expect potential
publishers to submit updates to N RDS schemes.  Hence there will need to
be a straightforward mapping of information from one to the other of
these schemes.  It is possible that that transformation will only go in
one direction, because a newer RDS service is replacing an older one,
which is not kept up to date, in order to encourage transfer to the
newer one.  Thus, at some point, updates may be made only to the newer
one and not be made available to the older one.  Such a situation should
probably be avoided, if possible.

This framework is presented in order to suggest to RDS scheme
designers a direction in which to start designing.  It should be
obvious to the reader that adherence to this framework will in no way
guarantee compliance with the requirements or even assumption
described in Sections 2 and 3.  These must be reviewed independently
as part of the design process.  There is no single correct design that

                                 - 17 -

will meet these requirements.  Furthermore, it is assumed that
preliminary proposals may not meet all the requirements, but should be
expected to itemized and justify any lack of compliance.

5. Acknowledgments

Foremost acknowledgment for this document goes to Lewis Girod, as my
co-author on a previous URN requirements document and for his insightful
comments on this version of the document.  In addition, I recognize the
contributors to a previous URN framework document, the "Knoxville"
group.  There are too many of you to acknowledge here individually, but
thank you.  Finally, I must thank the contributors to the URN working
group mailing list (urn-ietf@bunyip.com), for their animated discussions
on these and related topics.

6. References

[RFC1736] Kunze, J., "Functional Recommendations for Internet Resource
Locators", RFC 1736, February, 1995.

[RFC1737] Sollins, K. and Masinter, L., "Functional Requirements for
Uniform Resource Names", RFC 1738, December, 1994.

[RFC1738] Berners-Lee, T., Masinter, L., McCahill, M., "Uniform
Resource Locators (URL)", RFC 1738, December, 1994.

[RFCXXX] Moats, Ryan, "URN Syntax", currently available as
draft-ietf-urn-syntax-04.txt, March, 1997.

[VK83] Voydock, V. L., and Kent, S. T., "Security Mechanisms in
High-Level Protocols", ACM Computing Surveys, v. 15, No. 2, June,
1983, pp. 135-171.

7. Contact information:

Karen Sollins
MIT Laboratory for Computer Science
545 Technology Sq.
Cambridge, MA 02139

Tel: +1 617 253 6006
Email: sollins@lcs.mit.edu

This Internet Draft expires on September 28, 1997.

                                 - 18 -