Dave Crocker review of draft-iab-protocol-transition-05

[dthaler@…]

This ticket covers everything _except_ the IPv6 section which is a separate ticket.

Review of: draft-iab-protocol-transitions-05

Reviewer: D. Crocker
Date: 16 Jan 2017

Summary:

The draft provides an overview of the issues in achieving a transition from one capability to another. (It's worth considering whether this should include introduction of new capabilities -- that is, "transitioning" from no capability. Many adoption issues are the same.

The topic is of fundamental importance to IETF work and is often overlooked or viewed idealistically. So a document like this should be quite helpful (if folks will pay attention to it.)

The document is well-organized and well-written. There are some clarifications and expansions worth considering, as noted below, and some basic points cited here:

The document tends to merely mention essential issues, such as incentives, without giving much insight into either methods for adequately assessing incentives or deciding how to consider them in protocol design.

The document also seems to conflate "adoption" with "transition".

Much of the content of the document applies to initial adoption of a protocol, as well as to later transitions to revisions. While transitions carry significant additional burdens, beyond initial adoption, the IETF needs attention to initial adoption issues every bit as much as it needs attention to transition issues.

Although the document references open source implementations and the challenges of having a timeline, it should emphasize the role of the former more and the severe problems with the latter.

Might be worth adding some examples of highly successful transitions. MIME is, predictably, my favorite example. There had been multiple attempts to replace existing, text-based email with a new version that supported multi-media. MIME instead created an overlay that required no change to the infrastructure.

These above suggestions are in line with Eliot's call for more 'meat'. The document touches on essential issues. But for the IETF to deal with the issues well, there needs to be more detailed basis giving guidance for how to attend to them. This is particularly important for issues such as incentives analysis and aligning to incentives, since they are topics not normally within the purview of Internet engineers.

(If the feeling is that the meat should be added via later documents then there should at least, now, be development of some plan for those
documents.)

Stewart's call for considering the requirement of transition considerations -- I'd suggest 'adoption considerations' -- would press working groups to do far more due diligence about the barriers to adoption that is typically done now.

Detailed:

1. Introduction

A "transition" is "the process or period of changing from one state
or condition to another". There are several types of such

Use of quotation marks implies that the text comes from elsewhere. Where?

transitions, including both technical transitions (e.g., changing
protocols or deploying an extension) and organizational transitions
(e.g., changing what organization manages the IETF web site, or the
RFC production center). This document focuses solely on technical

It would be better for the examples to not be IETF-centric and especially not to require the reader to know about the internals of the IETF, such as about the RFC production center.

In this case, perhaps: changing what organization manages a web site that uses IETF specifications. Would authorizing a new network management team constitute a transition?

2. Transition vs. Co-existence

There is an important distinction between a strict "flag-day" style
transition where an old mechanism is immediately replaced with a new
mechanism, vs. a looser co-existence based approach where transition
proceeds in stages where a new mechanism is first added alongside an
existing one for some overlap period, and then the old mechanism is
removed at a later stage.

When a new mechanism is backwards compatible with an existing
mechanism, transition is easiest, and the difference between the two
types of transition is not particularly significant. However, when

I suspect you don't mean quite what is written. The differences still might be highly significant. More likely: transition is easiest because the timing of adoption by each party is not critical.

no backwards compatibility exists (such as in the IPv4 to IPv6
transition), a transition plan must choose either a "flag day" or a
period of co-existence. When a large number of entities are
involved, a flag day becomes impractical. Coexistence, on the other
hand, involves additional costs of maintaining two separate
mechanisms during the overlap period which could be quite long.
Furthermore, the longer the overlap period, the more the old
mechanism might get further deployment and thus increase the overall
pain of transition.

A phrase like "period of co-existence" encourages the reader to think that the period can be constrained. Besides making flag days impractical, large scale operation renders control over the length of a transition impractical. In fact it tends to ensure an extremely long adoption tail, measured in years and probably decades. This is not a small point, when designing for transitions. At base, 'transitions' for Internet scale are really long-term cohabitation.

Often the decision between a "flag day" and a sustained co-existence
period may be complicated when differing incentives are involved
(e.g., see the case studies in the Appendix).

For any IETF work, when has a flag day been specified and implemented successfully? While the idea of a flag day is appealing, it isn't ever practical both because of Internet scale and because multiple, independent administrations are (nearly) always involved.

3. Translation/Adaptation? Location

A translation or adaptation mechanism is often required if the old
and new mechanisms are not interoperable. Care must be taken when
determining whether one will work and where such a translator is best
placed.

A translation mechanism may not work for every use case. For
example, if a translation from one protocol (or protocol version) to
another produces indeterminate results, translation will not work
reliably. In addition, if translation always produces a downgraded
protocol result, the incentive considerations in Section 4.2 will be
relevant.

Requiring a translator in the middle of the path can hamper end-to-
end security and reliability. For example, see the discussion of
network-based filtering in [RFC7754].

On the other hand, requiring a translation layer within an endpoint
can be a resource issue in some cases, such as if the endpoint could
be a constrained node [RFC7228].

In addition, when a translator is within an endpoint, it can can
attempt to hide the difference between an older protocol and a newer
protocol, either by exposing one of the two sets of behavior to
applications and internally mapping it to the other set of behavior,
or by exposing a higher level of abstraction which is then
alternatively mapped to either one depending on detecting which is
needed. In contrast, when a translator is in the middle of the path,
typically only the first approach can be done since the middle of the
path is typically unable to provide a higher level of abstraction.

Any transition strategy for a non-backward-compatible mechanism
should include a discussion of where it is placed and a rationale.
The transition plan should also consider the transition away from the
use of translation and adaptation technologies.

This discussion should also consider the complexity of translation required. It is sometimes possible to make the new design easier to translate to/from the old, or to make it more difficult. Enthusiasm for new features often causes this point to be ignored.

The original Deering IPv6 design was pretty easy to translate. In fact, if IPv6 addressing had been made a superset of existing IPv4, translation would have been trivial.

The major challenge in translation is for semantic differences. Often, syntactic differences can be translated seamlessly. Semantic ones almost never.

Hence, attention to transition, when there is any interest in translation, should include documenting the syntactic and semantic differences;

4. Transition Plans

A review of the case studies described in Appendix A suggests that a
good transition plan includes at least the following components: an
understanding of what is already deployed and in use, an explanation
of incentives for each entity involved, a description of the phases
of the transition along with a proposed timeline, a method for

Overall, quite a good list.

However for IETF efforts -- that is, for anything to be deployed at Internet scale -- any concept of a transition timeline is misleading, at best. There is no history of succeeding with an attempt at timely transition, nevermind attempting to predict it.

Hence, trying to set a schedule distracts from the well-established track showing that transitions essentially take forever. Hence the most practical approach is to talk about adoption milestones and, in particular, considering what constitutes 'critical mass'. That is, when is it reasonable to consider adoption sufficient to ensure the continued use and further adoption of the capability?

Also, there is almost always need for an entity facilitating the transition. The issue here isn't one of authority but of advocacy and focus. Otherwise -- even with a good understanding of incentives -- the effort is left to happenstance. This is an entity independent of the IETF.

Thaler Expires July 8, 2017 [Page 5]

Internet-Draft Planning for Transition January 2017

measuring the transition's success, a contingency plan for failure of
the transition, and an effective method for communicating the plan to
the entities involved and incorporating their feedback thereon. We
recommend that such criteria be considered when evaluating proposals
to transition to new or updated protocols. Each of these components
is discussed in the subsections below.

4.1. Understanding of Existing Deployment

Often an existing mechanism has variations in implementations and
operational deployments. For example, a specification might include
optional behaviors that may or may not be implemented or deployed.
In addition, there may also be implementations or deployments that
deviate from, or include vendor-specific extensions to, various
aspects of a specification. It is important when considering a
transition to understand what variations one is intending to
transition from or co-exist with, since the technical and non-
technical issues may vary greatly as a result.

4.2. Explanation of Incentives

A transition plan should explain the incentives to each involved
entity to support the transition. Note here that many entities other
than the endpoint applications and their users may be affected, and
the barriers to transition may be nontechnical as well as technical.
When considering these incentives, also consider network operations
tools, practices, and processes, personnel training, accounting and
billing dependencies, and legal and regulatory incentives.

It's worth noting that an analysis of incentives is too easily led astray by wishful thinking and by a failure to adequately consider the realities of the entities being described.

An obvious (and frequent) example of misjudging incentives is ever thinking that any commercial operation adopts something out of a sense of civic obligation or long-term benefit. Although there are, of course, exceptions, the pattern is never encouraging.

Consequently, analysis of incentives should carefully justify the /basis/ for claiming the incentives. This is aided by an honest consideration of the barriers to adoption for each entity. What could cause them to fail to adopt or take longer?

If there is opposition to a particular new protocol (e.g., from
another standards organization, or a government, or some other
affected entity), various non-technical issues arise that should be
part of what is planned and dealt with. Similarly, if there are
significant costs or other disincentives, the plan needs to consider
how to overcome them.

The pragmatics of the incentives analysis is facilitated by looking at whatever advocacy group has formed to promoted the adoption. Who are the folk promoting the transition and what are their capabilities for making it likely to succeed. Here, too, the challenge is to avoid wishful thinking...

4.3. Description of Phases and Proposed Timeline

Transition phases might include pilot/experimental deployment,
coexistence, deprecation, and removal phases for a transition from
one technology to another incompatible one.

Hmmm. Rather than attempting a timeline, it probably helps more to consider specifying criteria that need to be satisfied, to go from one phase to the next. So a term like "phases" and "milestones" makes more sense.

Timelines are notoriously difficult to predict and impossible to
impose on uncoordinated transitions at the scale of the Internet, but
rough estimates can help all involved entities to understand the
intended duration of each phase.

So, yes, good that this is in the document, but I'll suggest it show up earlier and, if anything, even stronger.

Thaler Expires July 8, 2017 [Page 6]

Internet-Draft Planning for Transition January 2017

4.4. Measurement of Success

The degree of deployment of a given protocol or feature at a given
phase in its transition can be measured differently, depending on its
design. For example, server-side protocols and options which
identify themselves through a versioning or negotiation mechanism can
be discovered through active Internet measurement studies.

4.5. Contingency Planning

A contingency plan can be as simple as providing for indefinite
coexistence between an old and new protocol.

This seems an unusual enough topic to warrant more detail.

What types of contingency have been done and proved useful? What other
sorts might be considered?

4.6. Communicating the Plan

Many of the entities involved in a protocol transition may not be
aware of the IETF or the RFC series, so dissemination through other
channels is key for sufficiently broad communication of the
transition plan. While flag days are impractical at Internet scale,
coordinated "events" such as World IPv6 Launch may improve general
awareness of an ongoing transition.

Yes, but... Is there any basis for believing that that event was
actually useful in gaining wider adoption? If so, it's worth citing the
documentation. How do the IPv6 statistics support this?

My point is that events should be considered with a skeptical eye
towards pragmatics. It is far too easy to conduct an event that feels
encouraging to those putting it on but which has little practical
benefit. The downside of this is that, at the least, it drains energy
from the community promoting adoption.

Appendix A. Case Studies

Appendix A of [RFC5218] describes a number of case studies that are
relevant to this document and highlight various transition problems
and strategies (see for instance the Inter-Domain Multicast case
study in Section A.4 of [RFC5218]). We now include several
additional case studies that focus on transition problems and
strategies. Many other equally good case studies could have been
included, but, in the interests of brevity, only a sampling is
included here that is sufficient to justify the conclusions in the
body of this document.

A.1. Explicit Congestion Notification

This one sounds more like "adoption" than "transition". It's a new
mechanism and the adjustments were to find a way to get /any/ stable use.

A.2. Internationalized Domain Names

The deployment of Internationalized Domain Names (IDN) has a long and
complicated history. This should not be surprising, since
internationalization deals with language and cultural issues
regarding differing expectations of users around the world, thus
making it inherently difficult to agree on common rules.
Furthermore, because human languages evolve and change over time,
even if common rules can be established, there is likely to be a need
to review and update them regularly.

There have been multiple technical transitions related to IDNs,

There have been multiple specifications. From what I've seen, the
specification process has paid little attention to transition.

(There's an ICANN initiative to get better /adoption/, but that's not
strictly the same as is meant here for /transition/.)

This section highlights the challenge of distinguishing between the fact
of specification evolution, versus the process of transitioning between
versions. This section seems to cite the specifications rather than
transition details.

including the introduction of non-ASCII in DNS, the transition to
each new version of Unicode, and the transition from IDNA 2003 to
IDNA 2008. A brief history of the introduction of non-ASCII in DNS

Thaler Expires July 8, 2017 [Page 12]

Internet-Draft Planning for Transition January 2017

and the various complications that arose therein, can be found in
section 3 of [RFC6055]. While IDNA 2003 was limited to Unicode
version 3.2 only, one of the IDNA 2008 changes was to decouple its
rules from any particular version of Unicode (see [RFC5894],
especially section 1.4, for more discussion of this point, and see
[RFC4690] for a list of other issues with IDNA 2003 that motivated
IDNA 2008). However, the transition from IDNA 2003 to IDNA 2008
itself presented a problem since IDNA 2008 did not preserve backwards
compatibility with IDNA 2003 for a couple of codepoints.
Investigations and discussions with affected parties led to the IETF
ultimately choosing IDNA 2008 because the overall gain by moving to
IDNA 2008 to fix the problems with IDNA 2003 was seen to be much
greater than the problems due to the few incompatibilities at the
time of the change, as not many IDNs were in use, and even fewer that
might see incompatibilities.

A couple browser vendors in particular were concerned about the

couple of

differences between IDNA 2003 and IDNA 2008, and the fact that if a
browser stopped being able to get to some site, or unknowingly sent a
user to a different (e.g., phishing) site instead, the browser would
be blamed. As such, any user-perceivable change from IDNA 2003
behavior would be painful to the vendor to deal with, and hence they
could not depend on solutions that would need action by other
entities.

Thus, to deal with issues like such incompatibilities, applications
and client-side frameworks often want to map one string into another
(namely, a string that would give the same result as when IDNA 2003
was used) before invoking DNS.

"want"? this sounds more like prescription than about a case study of
what actually was done.

To provide such mapping (and some other functioanlity), the Unicode
Consortium published [TR46] that continued down the path of IDNA 2003
with a code point by code point selection mechanism. This was
implemented by some, but never adopted by the IETF.

Meanwhile, the IETF did not publish any mapping mechanism, but
[RFC5895] was published on the Independent Submission stream. In
discussions around mapping, one of the key topics was about how long
the transition should last. At one end of the duration spectrum is a
flag day where some entities would be broken initially but the change
would happen before IDN usage became even more ubiquitous. At the
other end of the spectrum is the need to maintain mappings
indefinitely. Local incentives at each entity who needed to change,
however, meant that a short timeframe was impractical.

I don't understand the above.

Again, it appears to be a discussion of possibilities rather than the
details of something that was part of a case study.

There are many affected types of entities with very different
incentives. For example, the incentives affecting browser vendors,
registries, domain name marketers and applicants, app developers, and

Thaler Expires July 8, 2017 [Page 13]

Internet-Draft Planning for Transition January 2017

protocol designers are each quite different, and the various
solutions require changes by multiple types of entities, where the

The substance of listing these entities is in talking about actual
incentives, not merely saying they will be different. Readers need to
see enough detail to learn something about applying the concern they
should have.

benefits do not always align with the costs. If there is some group
(or even an individual) that is opposed to a change/transition and
able to put significant resources behind their opposition,
transitions get a lot harder.

Finally, it is worth pointing out that there are multiple naming
contexts, and the protocol behavior within each naming context can be

Huh? How is this statement relevant to IDN?

different. Hence applications and frameworks often encounter a
variety of behaviors and may or may not be designed to deal with
them. See sections 2 and 3 of [RFC6055] for more discussion.

In summary, all this diversity can cause problems for each affected
entity, especially if a competitor does not have such a problem,
e.g., for browser vendors if competing browsers do not have the same
problems, or for an email server provider if competing server
providers do not have the same problems.

A.3. IPv6

...
[covered by separate ticket]

A.4. HTTP/2

HTTP/2 [RFC7540] is a new version of the popular HTTP protocol
[RFC7230]. The original versions of HTTP (0.9 [HTTP0.9], 1.0
[RFC1945], and 1.1 [RFC2616]) have only small differences; each
iteration made small improvements over the previous version without
making major changes.

The changes in HTTP/2 are largely aimed at improving performance.
The primary improvement is request multiplexing, which is supported

is to

by request prioritization and flow control. HTTP/2 includes
efficiency improvements with header compression [RFC7541] and binary
framing.

A.4.1. Bundling of Features with New Versions

The bundling of additional constraints on a new version of a protocol
could affect adoption by making the transition more costly. However,
the transition to a new version also represents an opportunity to
improve multiple aspects of a protocol at the same time.

The HTTP working group decided that a new version of the protocol
represented an opportunity to improve security posture. HTTP/2 is
much stricter about its use of TLS. In particular, a long list of

"to improve security posture"? is a word missing?

Thaler Expires July 8, 2017 [Page 16]

Internet-Draft Planning for Transition January 2017

TLS cipher suites are prohibited, constraints are placed on the key
exchange method, and renegotiation is prohibited. These changes did
cause deployment problems. Though most were minor and transitory,
disabling renegotiation caused problems for deployments that relied
on the feature to authenticate clients and prompted new work to
replace the feature.

A number of other features or characteristics of HTTP were identified
as potentially undesirable. Several such features were considered
for removal during the design process. This included trailers, the
1xx series of responses, certain modes of request forms, and the
unsecured (​http://) variant of the protocol. For each of these, the
risk to the successful deployment of the new version was considered
to be too great to justify removing the feature. However, deployment
of the unsecured variant of HTTP/2 remains extremely limited.

I'm not understanding the basis for having the 'However' here.

How does that sentence connect with the preceding text? For that
matter, what is the 'unsecured variant'?

A.4.2. Planning for Replacement

HTTP/1.1 provides a mechanism, Upgrade, to transition to an entirely
different protocol. That same facility was little used other than to
enable the use of WebSockets? [RFC6455]. However, with performance
being a primary motivation for HTTP/2, a new mechanism was needed to
avoid spending an additional round trip on this negotiation. A new
mechanism was added to TLS to permit the negotiation of the new
version of HTTP: Application Layer Protocol Negotiation (ALPN)
[RFC7301]. Upgrade was used only for the unsecured variant of the
protocol.

This highlights a problem with a mechanism that is put into a protocol
'for future use' and without having adequate sense of how it will be
used. It tends not to work very well (or at all.) This also happened
with SNMP's original 'security' field.

ALPN was identified as the way in which future protocol versions
would be negotiated. The mechanism was well-tested during
development of the specification, which proved that new versions
could be deployed safely and easily using ALPN. Several draft
versions of the protocol were successfully deployed during protocol
development, and version negotiation was never shown to be an issue.

Confidence that new versions would be easy to deploy if necessary
lead to a particular design stance that might be considered unusual
in light of the advice in RFC 5218 [RFC5218], though is completely
consistent with RFC 6709 [RFC6709]: many of the ways in which the
protocol might be extended were removed unless an immediate need was
understood. This decision was made on the basis that it would be
easier to revise the entire protocol than it would be to ensure that
an extension point was correctly specified and implemented such that
it would be available when needed.

This is far to important an observation to have it buried at the end of
the appendix.

[dthaler@…]

I have accepted the editorial nits, and made other changes as noted below.

The document is well-organized and well-written. There are some clarifications and expansions worth considering, as noted below, and some basic points cited here:

The document tends to merely mention essential issues, such as incentives, without giving much insight into either methods for adequately assessing incentives or deciding how to consider them in protocol design.

Unfortunately we may not have a lot of experience with such methods. However, I have made several notable changes in response to this and others’ comments.
An Extensibility section now discusses extensibility in more detail and refers to the IAB’s recent RFC 6709 regarding protocol design.
A Conclusions section now encourages the community to share such insights and future experiences, on the architecture-discuss list.

The document also seems to conflate "adoption" with "transition".

Much of the content of the document applies to initial adoption of a protocol, as well as to later transitions to revisions. While transitions carry significant additional burdens, beyond initial adoption, the IETF needs attention to initial adoption issues every bit as much as it needs attention to transition issues.

Added terminology point to the Introduction saying that transition is generic, with four types of transition (adoption of a new protocol with no precedent, transitioning from an old to a new protocol, transitioning from an old to a new version of a protocol, and decommissioning of an obsolete protocol). Of course some points in the doc apply more strongly to certain types of transition.

Although the document references open source implementations and the challenges of having a timeline, it should emphasize the role of the former more and the severe problems with the latter.

I rewrote the section about timelines and criteria for phrases based on your feedback.

Might be worth adding some examples of highly successful transitions. MIME is, predictably, my favorite example. There had been multiple attempts to replace existing, text-based email with a new version that supported multi-media. MIME instead created an overlay that required no change to the infrastructure.

When the discussed it, the consensus was that additional case studies in the doc were not necessary. Unless of course there’s a key point that would be called out that isn’t captured at all in the other case studies.

These above suggestions are in line with Eliot's call for more 'meat'. The document touches on essential issues. But for the IETF to deal with the issues well, there needs to be more detailed basis giving guidance for how to attend to them. This is particularly important for issues such as incentives analysis and aligning to incentives, since they are topics not normally within the purview of Internet engineers.

(If the feeling is that the meat should be added via later documents then there should at least, now, be development of some plan for those
documents.)

Stewart's call for considering the requirement of transition considerations -- I'd suggest 'adoption considerations' -- would press working groups to do far more due diligence about the barriers to adoption that is typically done now.

Perhaps, but that’s feedback for the IESG to consider, rather than the IAB’s purview per se.

Detailed:

I believe I’ve incorporated your various detailed points into the new draft, soon to be posted…

BTW, the case studies were written by others (I’m the editor) and so there may still be
additional changes after that to address some of your points in the IPv6 case study.