Afnic

7 avenue du 8 mai 1945 Guyancourt 78280 FR bortzmeyer+ietf@nic.fr https://www.afnic.fr/

NLnet Labs

Science Park 400 Amsterdam 1098 XH NL willem@nlnetlabs.nl https://nlnetlabs.nl/

Quad9

Werdstrasse 2 Zürich 8004 CH babak@farrokhi.net https://quad9.net/

The FreeBSD Foundation

3980 Broadway St Boulder CO 80304 US bofh@freebsd.org https://freebsdfoundation.org/

Internet Systems Consortium

Czech Republic ondrej@isc.org

PowerDNS

NL otto.moerbeek@powerdns.com

General Internet Engineering Task Force DNS DNSSEC optimization caching Networks of cooperating and mutually trusting DNS resolvers could benefit from cache sharing, where one resolver would distribute the result of a resolution to other resolvers. This document standardizes a protocol to do so.

Introduction When an organisation operates a big network of DNS resolvers , for instance for an important public resolver (), it may be a performance improvement to distribute the result of the resolution process between the resolvers. The same applies to resolvers operated by distinct organisations which have agreed to cooperate. This document standardizes how to do so, using unicast messages to a set of pre-configured peers.

Requirements Language The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

Terminology

Network of resolvers

A set of mutually trusting resolvers working together under a common policy, often but not necessarily under the same administration

Peer (or peer resolver)

One of the other resolvers in the network

Originating resolver

A resolver sending data to its peers in the network

Receiving resolver (or receiving peer)

A resolver receiving data from one of its peers in the network

Resolver

As used in

The protocol When completing a successful DNS resolution, the resolver transmits a DNS message (with the Q/R bit set, since it is a response) to the pre-configured peers. No acknowledgment is sent or expected. To save work, the resolver MAY send the data only if the TTL is higher than some predefined value. TSIG is mandatory to implement but optional to use, following . Implementations MUST support it, so that peers always have one authentication mechanism in common. Whether a given network of resolvers enables it is outside of the scope of this document. The resolver MUST send only data that it is sure of (for instance by DNSSEC validation or because it came with the AA bit from the queried server). All the resolvers of the network MUST agree on the same policy for this assessment. When they are not under the same administration, this policy MUST be agreed explicitly before deployment, since it cannot be inferred from a shared administrative practice. Negative answers (, section 3) MUST NOT be transmitted to peers. A message of this protocol is a response signed like a query. Since there is no request MAC (), the MAC MUST be computed over the digest components listed for a request in . This distinguishes it from an ordinary DNS response. The TSIG key SHOULD nevertheless be specific to this protocol. Peers that do not use TSIG need another discriminator, for instance a dedicated port. This message MUST be a new message (not the message received by the resolver from the authoritative name servers) with data composed from data already obtained and validated by the originating resolver. It MUST contain an Answer section and MAY contain other sections. (So, the Question section is not mandatory.) The EDNS section MUST be a new one, created to fit the needs of successful transmission to the peer. Each peer then MAY store the data in its cache. The peer is not supposed to do DNSSEC validation (there is not always all the necessary data in the message). After all, the goal is to save work for the peers, so does not apply here. (Remember all peers trust each other, and have a consistent policy. The data is as trustworthy as if you validated it yourself.) A resolver which cannot rely on its peers applying the agreed policy MUST NOT accept their data. The receiver MAY cache only what is in the Answer section.

IANA Considerations None. [RFC-Editor: you may delete this section]

Security Considerations The integrity and authenticity of the cached data is of course critical. DNSSEC would help but it is not yet universally deployed and, moreover, the peer resolvers should not have to redo the validation. So, trust between the peer resolvers is expected because it is the only way for the receiver to be sure of the data. Having all of the peers under the same administration is the simplest way to obtain this trust, but it is not the only one. The channel between peers also needs protection, preferably with cryptography. ACL and other network techniques are of course useful. Sharing one TSIG key across the whole network lets any peer impersonate any other (). Using a distinct key per pair of peers is more important when the peers are not under the same administration.

Privacy Considerations The records transmitted are public, but the messages contain the names that the resolvers have resolved, and therefore reveal to an eavesdropper what their clients looked up . Removing the question section does not hide this, since the answer section contains the owner name. Using an encrypted transport, for instance DoT or DoQ , or otherwise protecting the network path between peers, defeats this leakage and is RECOMMENDED when the messages travel over the public Internet. If the originating resolver sends the original question section in its messages to receiving peers, it can have privacy consequences , for instance in the case of negative answers. When all the peers are under the same administration, these consequences are limited, and the originator SHOULD remove this section or replace it with dummy data. When the peers are under distinct administrations, the question section would disclose the queries of one party's clients to another, and the originator MUST remove it or replace it with dummy data.

Operational Considerations It is reminded that all resolvers in the network need to trust each other. This specification is not meant to be deployed between unrelated resolvers. The network of peer resolvers has to be configured out-of-band beforehand. The way to do it is out-of-scope for this specification.

Related and future work

Related work describes a mechanism to fill DNS caches with data. The format is, like in this document, standard DNS as seen on the wire.

Future work

Actual measurement Measuring the efficiency of caching optimizations is hard because the performance depends a lot on the actual queries sent and their timeline. It would be interesting to test on real-world data to measure the improvement brought by this technology.

Negative answers In the future, we may allow negative answers to be sent, which would permit the receivers to use and/or to synthesize negative answers. Be careful of the risk of overloading receiving peers for instance when there is a dictionary attack.

ECS Today, we don't transmit the EDNS record since EDNS is not end-to-end. But ECS may be an important information for the participating resolvers.

Authentication Standardizing how peers select an alternative authentication mechanism, such as SIG(0) or DoT , is future work.

Dispatching of messages Today, messages following this specification are sent to the proper handler based on the TSIG key used. It could be an interesting alternative to use a separate transport port instead.

Transport of messages Messages could be transmitted in long-lived TCP sessions, too. If there are 1,000 servers, sending 1,000 messages, or having a full mesh of 1,000 TCP connections may be too much. It may be interesting to replace the unicast messages by multicast (the issues of multicast on the public Internet do not apply here since we envision work under only one administration. A network spanning several administrations would have to face them). Other protocols may be considered such as MQTT which is well suited for publish-by-one/consume-by-many, raw protocol buffers or dnstap over them. Also, a different organisation of the network of resolvers could be done in a hub-and-spoke way, in the future.

Packing of messages It could be interesting to optimize by packing the data in a C-DNS flow, sent with TCP (with TLS) or QUIC. (Of course, other formats/protocols are possible.)

Different responses When the authoritative servers send different replies depending on the client, the various peers may send different (and under-optimized) responses to a receiving peer.

References Normative References Informative References OASIS Google Developers

Acknowledgements Original idea at the DNS hackathon (RIPE-NCC / Netnod / DNS-OARC) in march 2025 at the Netnod office in Stockholm.