- From: Mark Nottingham <mnot@mnot.net>
- Date: Thu, 10 Sep 2026 09:02:33 +1000
- To: Tim Bruijnzeels <tbruijnzeels@ripe.net>
- Cc: ietf-http-wg@w3.org, draft-ietf-sidrops-publication-server-bcp.all@ietf.org, last-call@ietf.org, sidrops@ietf.org
Hi Tim,
Responses below.
> On 9 Sep 2026, at 11:57 pm, Tim Bruijnzeels <tbruijnzeels@ripe.net> wrote:
>
> Hi Mark,
>
> Thank you for the review and comments.
>
>> On 9 Sep 2026, at 07:28, Mark Nottingham via Datatracker <noreply@ietf.org> wrote:
>>
>> Document: draft-ietf-sidrops-publication-server-bcp
>> Title: Best Practices for Operating Resource Public Key Infrastructure (RPKI)
>> Publication Services Reviewer: Mark Nottingham Review result: Ready with Issues
>>
>> [Drafted with AI assistance from a structured review over the document and the
>> WG's record, then reviewed and edited by me.]
>>
>> The draft is mostly careful with HTTP -- no fixed paths, no new methods or
>> status codes, no redefinition of the caching model, and RFC 9110 is the
>> reference used throughout. Two things are worth fixing before publication.
>>
>> ## Issues
>>
>> 1. Section 3.4 says RFC 8181 has no way to signal backoff, but HTTP does
>>
>> "Unfortunately, the publication protocol specified in [RFC8181] has no adequate
>> support for rate limiting or signaling requests to CAs to backoff. Therefore,
>> publishers SHOULD NOT perform this resynchronisation more frequently than once
>> every 10 minutes..."
>>
>> RFC 8181 runs over HTTP POST. 429 or 503 with Retry-After (Section 10.2.3 of
>> RFC 9110) lets a publication engine enforce and communicate exactly this limit,
>> rather than asking publishers to self-limit on trust.
>>
>> Reword to say the PDUs themselves carry no backoff signal, but a compliant
>> server can still respond with 429/503 plus Retry-After -- and recommend doing
>> so.
>
> I agree that this would be the best way to do this. (and similarly for RFC6492
> by the way).
>
> How about:
>
> OLD:
> Unfortunately, the publication protocol specified in [RFC8181] has no
> adequate support for rate limiting or signaling requests to CAs to backoff.
> Therefore, publishers SHOULD NOT perform this resynchronisation more
> frequently than once every 10 minutes, unless otherwise agreed with the
> publication engine operator.
>
> NEW:
> Unfortunately, the publication PDUs specified in [RFC8181] have no
> adequate support for rate limiting or signaling requests to CAs to backoff.
> It is RECOMMENDED that overloaded publication servers respond with an
> 429 (Too Many Requests) response (see section 4 of [RFC6585]), or with
> a 503 (Service Unavailable) response with Retry-After (see section 10.2.3
> of [RFC9110]).
>
> However, since there is no guarantee that publication servers implement
> this, and that publishers support these responses, publishers SHOULD NOT
> perform this resynchronisation more frequently than once every 10 minutes,
> unless otherwise agreed with the publication engine operator.
>
> What I am trying to capture here is that - while I agree with you on the
> way to do this - this may not (yet) be supported. So having an overall limit
> still makes sense.
Seems reasonable to me, thanks.
>> 2. Compression with no mention of Vary
>>
>> Section 5.3: "RRDP endpoints SHOULD support compression. At a minimum, gzip
>> content coding (see Section 8.4.1.3 of [RFC9110]) SHOULD be supported..."
>> Section 5.4 separately recommends serving RRDP content through a CDN.
>>
>> Without Vary: Accept-Encoding, a shared cache has no basis to key on the
>> client's Accept-Encoding -- whichever variant it stores first is what every RP
>> behind that cache gets, regardless of what it asked for. An RP that never
>> implemented decompression, because it never sends Accept-Encoding: gzip, could
>> still be handed a gzip response.
>>
>> Suggestion: "RRDP endpoints SHOULD support compression through proactive
>> content negotiation (see Section 12.1 of [RFC9110]). At a minimum, gzip content
>> coding (see Section 8.4.1.3 of [RFC9110]) SHOULD be supported. Additionally,
>> servers are RECOMMENDED to support other widely used compression algorithms
>> where feasible. Note that this implies that the Vary header field be sent on
>> responses; see Section 12.5.5 of [RFC9110]."
>
> Sounds good to me. Although setting the Vary header may be out of control of
> the operator when using a CDN. It looks like it may need special attention when
> one uses a local proxy server - so I think it's good to call this out.
If the CDN doesn't set Vary, it's broken. Your text looks fine to me.
>
> See commit:
> https://github.com/timbru/draft-sidrops-publication-server-bcp/commit/61c23c17c90da2ef23de985fe1e2ba4c606c4de4
>
>
>
>>
>> ## Comments
>>
>> Section 5.4's "unless the backend RRDP server is unavailable, in which case it
>> is RECOMMENDED that stale files are served" needs a mechanism: RFC 9111 Section
>> 4.2.4 only lets a connected cache serve stale content when the origin or client
>> explicitly permits it (e.g. stale-if-error, RFC 5861, or an out-of-band CDN
>> contract).
>
> Ok, how about:
>
> OLD:
> unless the backend RRDP server is unavailable, in which case it
> is RECOMMENDED that stale files are served.
>
> NEW:
> unless the backend RRDP server is unavailable, in which case it
> is RECOMMENDED that operators permit that stale files are served
> (see section 4.2.4 of [RFC9111]).
Yes.
>> ARTART (Fältström) already asked that Section 5.1's "same host" be pinned down
>> precisely. This document already normatively cites RFC 9110, whose Section
>> 4.3.1 gives the same canonicalisation ARTART is asking for (lowercase scheme
>> and host, drop leading zeros from the port, substitute the default port when
>> elided) without pulling in RFC 6454.
>
> This was addressed in another thread by changing 'host' to 'origin'.
>
>> Section 8 says no new security issues arise beyond RFC 8181/8182/9589/9674, but
>> Section 5.4's own CDN recommendation exposes RP fetch patterns to a third party
>> -- none of those documents' security considerations account for that. SECDIR's
>> review is still outstanding; may already be on their list.
>
> Technically this concern applies to RFC 8182 (RRDP). Do you have a suggestion
> for this?
>
> Operators should be aware that, in case they use a third party CDN for
> their RRDP repository, that third party will have the ability to modify
> data presented to RPs as well as track RP fetch patterns.
Nothing specific, maybe let's see what the SEC review says.
>> ## Nits
>>
>> * Section 5.2: "HTTP GET parameters" isn't HTTP terminology -- query
>> parameters, or the query component of the request-target?
>
> Okay, how about:
>
> OLD:
> ..if the repository operator knows HTTP query parameters are not
> used to provide service, then the operator can safely block any
> requests containing query parameters.
>
> NEW:
> ..if the repository operator knows HTTP query parameters are not
> used to provide service, then the operator can safely block any
> requests containing query parameters.
>
> (== s/GET/query)
Being super-nitpicky, at the HTTP/URI layer it's "query component" - see eg
https://httpwg.org/specs/rfc9110.html#rfc.section.4.2.1
The a=b&c=d convention is introduced by HTML, not HTTP.
So I think the right approach might be:
... if the repository operator knows that the query component of the request's URI is not used to provide service, they can safely block requests containing a query component.
("request-target" is a bit obscure)
>> * Section 5.9.2:
>> "HTTP keepalive" -- RFC 9112 Section 9.3 calls this persistent connections. *
>
> How about:
>
> OLD:
> Due to HTTP keepalive,..
>
> NEW:
> Due to HTTP persistent connections (see section 9.3 of [RFC9112]),..
>
> and:
>
> OLD:
> ...limit HTTP keepalive...
> NEW:
> ...limit HTTP persistent connections...
Yes.
>> Section 5.4: "HTTP 404 responses" -- editorial convention for status codes is
>> using the bare code plus reason phrase in parenthesis: "404 (Not Found)".
>
> Ok. So, just s/HTTP 404 responses/HTTP 404 (Not Found) responses/ right?
Yes.
Cheers,
--
Mark Nottingham https://mnot.net/
Received on Wednesday, 9 September 2026 23:02:45 UTC