Re: Conclusions of Last Call for draft-ietf-spfbis-4408bis

[Douglas Otis <[e-mail]>]

On Sep 7, 2013, at 6:31 AM, Pete Resnick <[e-mail]> wrote:

> Below is the list of issues brought up during Last Call of draft-ietf-spfbis-4408bis. I have tried to collect together the common issues and tease out the ones that are slightly different. Below each issue, I've given what I take to be the answer to the issue (either the change that needs to be made, or the explanation of why no change is necessary). I've not put names to the objections or the answers, simply because multiple people gave different versions of the same objections or the same answers, and sorting that out seemed useless.
> 
> Issues:
> 
> 1. Overloading of the TXT RR for this use is bad.
>    - As far as I can tell, there is nobody that disagrees with this statement. However, it is also not in-and-of-itself an objection to the document: Nobody seems to have argued that this document should forbid use of the TXT RR completely in SPF. The only question is whether the document should provide a transition mechanism to the SPF RR or whether it is reasonable to go forward with this protocol using only the TXT RR. There is the "precedent"objection I will discuss in 2 below, but the specific technical objections seem mostly *not* to be about forbidding TXT RR use from the get-go. Some of the specific objections could be taken as arguments against the document going forward if it has no transition mechanism, but I believe all of those have either been addressed or can be addressed with clarifying language in the document:
>    - The only complete solution to many of the problems that fall under this category is if *all* misuse of TXT RR went away. There has been no convincing argument put forward that this is plausible in the foreseeable future.
> 
>    1a. TXT RR can cause large RDATA.
>        - In theory, that's certainly true. But I have not seen an argument that SPF is causing a major problem to date, nor that there is an expectation that it will in the future.
>        - There were some suggested text clarifications for section 3.4 to make it clear which size limitations relate to DNS response size, UDP payload size, or MTU size. They seem reasonable and were not objected to. I'll work with the editor and others to clarify that text.
> 
>    1b. Use of TXT RR can cause collisions with other applications.
>        - Again, there is no indication that this has caused a problem for SPF (or others) to date, and SPF requires rejection of TXT RR data that does not conform to the spec, so I see no evidence of the existing harm, nor of a solid reason to believe there will be future harm. (Again, I'm leaving aside the "precedent" arguments until 2.)
>        - There appears to be an effort underway to document (via an IANA registry) such uses to minimize the potential for these sorts of collision problems in the future.
> 
>    1c. Use of TXT RR for multiple purposes makes it impossible to do access control based on type of data (i.e., to allow delegation of the management for TXT RRs that are solely for SPF use).
>        - Many organizations have been managing these records already; no reasoning was given that this fine-grained management is necessary.
>        - Delegation is possible by pointing a TXT RR of (e.g.) example.com to _spf.example.com, delegating the latter.
> 
>    - All of the above was discussed extensively in the WG and taken into consideration. Given the charter limitations, a reasonable choice was made.
> 
> 2. Use of the TXT RR sets a bad precedent for future use.
>    - Several people responded that some additional text in 3.1 or elsewhere (either by way of some sort of applicability guidance or an overt IESG Statement) would address this issue. I think an IESG Statement is unnecessary since I have not heard significant objection to including some guidance, and I think something reasonable along these lines can be crafted. I'll work with the editor and others to get such text in the document.
>    - The impediments that caused SPF to use TXT RR in the first place are mostly gone. New protocols are unlikely to face the same challenges.
> 
> 3. Removing SPF RR support is a charter violation.
>    - Because the original spec has a non-interoperable mechanism for use, this constituted an "error" in the spec that was to be corrected.
> 
> 4. A new transition mechanism from TXT RR to SPF RR should be put into the spec.
>    - This was extensively considered by the WG.
>    - Backward compatibility would require support of TXT RR for the forseeable future anyway.
>    - The proposed transition mechanisms have technical issues: Doubling request traffic (e.g., doing queries in parallel), introducing delays (e.g., querying SPF RR first and running into firewalls, etc).
>    - This would be a new, unchartered requirement for the WG.
>    - The widely held conclusion was that such a transition plan would not be undertaken by implementers. (See also 5.)
> 
> 5. That there will be a lack of adoption of SPF RR was based on RFC 6686, which does not support the conclusion.
> 
>    5a. There is some current use of the SPF RR; it will increase if we put in a new transition mechanism.
>        - 6686 showed only minimal use, and reports of those in the industry shows the use dropping (i.e., the momentum is in the other direction).
>        - Bigger sites are conservative and will not transition for fear of breaking current usage.
>        - No solid case was made for why to believe that the transition would occur.
> 
>    5b. Use of SPF RR is on the increase now.
>        - Claims that use is increasing are only anecdotal, and disagree with the experience of those in the industry.
> 
>    5c. Things may have changed since 6686. We should do more data collection.
>        - There's no reason to believe that the small amounts of recently presented data are representative.
>        - Nobody presented any basis to doubt the folks working in the industry.
>        - There has been ample opportunity (and motivation) for folks outside of the WG to do more data collection; none has been presented.
>        - It is an unreasonable burden to place on the WG at this point.
> 
> There were a few smaller issues:
>    6. The term "SPF records" is confusing because it could refer to SPF RR.
>        - Because the document no longer uses SPF RR, this shouldn't be a problem in practice.
>    7. Clarifications are needed regarding the number of lookups to do in 4.6.4.
>        - This will be reviewed prior to publication.
>    8. There should be a limit on PTR lookups.
>        - This was already considered by the WG and rejected.
> 
> The only looming issue large issue is the architectural one:
> 
> 9. Using TXT RR for this purpose violates the architecture of the DNS.
> 
> I list this separately because this is not about the immediate technical implications (like objection 1) or a claim about precedence setting (like objection 2), but appears to be a claim that violating the architecture is in and of itself a reason to not put something on the standards track. The problem I have with this is, short of actual technical harm, I'm not sure how to judge this. Our processes judge whether standards should be adopted on the basis of interoperability, deployment, and solving a useful problem. We have many examples of protocols that violate architectural principles, but standardize them nonetheless. (We can all name our most hated.) We have found it more useful to document how to interoperate with protocols that may not be ideal but are widely deployed rather than reject them on principle. I can't see how to treat this protocol differently.
> 
> [There were a number of "straw men": There were responses to objections that never got brought up during Last Call, and a number of followups to responses to objections where the response was not one offered. Examples of these included:
>    - A response to the objection that we haven't let the Experiment run long enough. (Nobody ever argued that during Last Call.)
>    - A followup denying the contention that there is limited server support for SPF RRs. (Nobody ever said that the reason to use TXT RR now was because of limited server support.)
> I have ignored these threads.]
> 
> So, my conclusions in summary are:
> 
> - The document needs to make a statement in the document clarifying why the SPF RR is no longer used in the spec and making it clear that no precedent should be created by this protocol's continued use of TXT RR.
> 
> - A few clarifications are required in the text (size limitations, perhaps number of lookup limitations).
> 
> - The remainder of the objections were fully considered and understood by the WG, and were addressed to a reasonable extent, and therefore that there is rough consensus to go forward with this document.

Dear Pete,

You missed two concerns raised in the last call:

10. As DNSSEC becomes more broadly used with email, SPF is likely to introduce problems neither foreseen nor properly considered by the WG.   "Not a problem for SPF" does not mean SPF will not impose problems for DNS which may include DNSSEC.  It should be obvious which protocol SPF or DNS should receive greater consideration  The SPF protocol also ignores the number of PTR Resource Records returned by a query made by a recipient on behalf of unknown senders against various third-party domains.  These domains can be constructed and modulated by message elements via macro expansion greatly increasing the reflected amplification this protocol enables.  The same issue could occur with a series of otherwise innocent TXT resource records that the SPF protocol also permits. This "feature" could prove highly problematic and extremely difficult if not impossible to defend against.  Even limits on NX domains can be easily sidestepped by malefactors leveraging the practice of using synthetic domains to track users without the use of cookies.

11.  The continued specification of SPF macros inhibits interchange.  Because it is common for SPF records not to produce a pass, this issue is not likely to have been given adequate attention.  When SPF macros are not implemented by receivers for any number of very valid reasons, such as ensuring effective caching of DNS, their required use can and will lead to interchange issues.  SPF may impose complex macro handling over multiple DNS responses determined by a sequence of queries that can not be directly handled by DNS itself.  Even the operation of SPF macros represents security concerns threatening the integrity of the associated SMTP and DNS servers.   Since the publication of SPF macros is well below the level used to justify removal of the SPF RR record type, the same consideration should have been given SPF macros.  Use of SPF macros also interferes with forensic efforts at handling interchange problems.  As such, it is not surprising to find extremely few domains publish SPF records using macros and large providers not processing SPF macros.

The lack of consideration given DNS by the SPF protocol offers overwhelming justification not to consider this protocol suitable for endorsing as a standard.  Going from experimental to informational should not represent any hardship, but would serve as a warning protocols should pay attention to their impact on underlying infrastructure.  There are limits on making it easy to send messages since the Internet is not suffering from a message scarcity.

Regards,
Douglas Otis