Hi Ben, On 17 Nov 2026, at 08:10, Ben Schwartz <[email protected]> wrote:
> This draft implies that a single resolver instance performs both ADoX > and non-ADoX queries to the same authorities. This creates a huge > number of complications. For example, some insecure records may be > present in cache that were fetched previously over UDP. Should they > be refetched over ADoX to prevent forgery, even though this reduces > privacy? I am not sure what you mean by "insecure records" here. Fetching over ADoX does not provide integrity protection; even if you authenticate the server you connect to (which might well be complicated) you need DNSSEC to validate the integrity of the response. The goal of this draft is to reduce the opportunity for on-path surveillance of queries, not to improve integrity protection. In this context, answering out of the cache does not reduce privacy. > What if they are only the glue records for a signed child > zone? I don't think the proposal is to change the way that the resolver functions with respect to using data cached after delivery in the additional section of an upstream response. If it's in the cache and you'd normally use it to satisfy a query from a client you can return it to a client who has requested privacy. > What is the privacy response status of a query whose recursive > stages used a mix of cached and uncached, secure and insecure, ADoX > and non-ADoX queries? Again, I don't think this is as complicated as you suggest. If a client requires privacy and you can answer using a mixture of cached records and upstream queries using secure transport, you can construct a response. If you can't construct a response without an insecure upstream query, then the client has told you they would rather the query fails. > What about QNAME-minimized queries that only > reveal a TLD? If the client requests secure transport, use secure transport or answer from the cache. If you can't construct a response without an upstream query using insecure transport, fail. > Should queries with EDNS Client Subnet be treated as > more sensitive in some way? I think this is a reasonable question that the draft could consider. It seems reasonable for a client that cares about privacy to care about whether a resolver sends ECS in an upstream query, regardless of transport. > It seems to me that the four proposed privacy-request levels are just > a handful of the zillions of plausible policies, and their semantics > are far from precise. Operational practice from Quad9, 1.1.1.1, and > NextDNS suggests a different way to expose these choices: by offering > different resolver identities with different behaviors. This gives > operators the flexibility to choose whatever combination of behaviors > they believe will be useful, and gives users a menu of possibilities > without requiring changes to applications and operating systems. I don't see zillions of options here. The options relate to a tightly-scoped set of upstream query behaviours for a resolver and the set seems complete to me. Privacy is complicated and a user might have a more extensive set of requirements for their application of which upstream query behaviour by the resolver might be just one component. Providing resolver service on different addresses is also possible, but seems clumsy. Four varieties of upstream query behaviour would take Cloudflare's current menu of resolvers (1.1.1.1, 1.1.1.2 and 1.1.1.3) and turn it into a choice of twelve (four upstream query behaviours for each) with no standard way to enumerate what each address does across different resolver operators. It seems much cleaner and less error-prone to provide a standard client signal that can be used with multiple resolvers. > I would love to see public resolver operators offer a "strict ADoX" > resolver option, as a bridge to a world where that is the expected > default, but I don't think we should define a signal for this in the > query. Why? The existence of this draft suggests we can. > On the response side, I do agree that we may need an ADoX equivalent > of the "AD" bit, but I don't foresee a need for more than one bit of > signal (meaning "this response is fully authenticated by some > combination of DNSSEC and ADoX"). I don't think I understand what you mean by this. Are you saying that resolvers should no longer validate but should instead trust authority servers to assert that validation is not necessary? Earlier you suggested that four was too small a number to enumerate the privacy options in Leon's draft, but here you seem to say that one bit is sufficient to cover all the combinations of DNSSEC validation, secure transport and certificate validation requirements? (I don't think you probably mean any of those things; I am just illustrating my lack of understanding :-) Joe _______________________________________________ DNSOP mailing list -- [email protected] To unsubscribe send an email to [email protected]
