Abridged Compression for WebPKI Certificates

1.1. Motivation

When a server responds to a TLS Client Hello, the size of its initial flight of packets is limited by the underlying transport protocol. If the size limit is exceeded, the server must wait for the client to acknowledge receipt before concluding the flight, incurring the additional latency of a round trip before the handshake can complete. For TLS handshakes over TCP, the size limit is typically around 14,500 bytes. For TLS handshakes in QUIC, the limit is much lower at a maximum of 4500 bytes ([RFC9000], Section 8.1).¶

The existing compression schemes used in [TLSCertCompress] have been shown to deliver a substantial improvement in QUIC handshake latency [FastlyStudy], [QUICStudy] by reducing the size of server's certificate chain and so fitting the server's initial messages within a single flight. However, in a post-quantum setting, the signatures and public keys used in a TLS certificate chain will be typically 10 to 40 times their current size and cannot be compressed with existing TLS Certificate Compression schemes because most of the size of the certificate is in high entropy fields such as cryptographic keys and signatures.¶

Consequently studies [SCAStudy] [PQStudy] have shown that post-quantum certificate transmission becomes the dominant source of latency in PQ TLS with certificate chains alone expected to exceed even the TCP initial flight limit. This motivates alternative designs for reducing the wire size of post-quantum certificate chains.¶

1.2. Overview

This draft introduces a new TLS certificate compression scheme which is intended specifically for WebPKI applications and is negotiated using the existing certificate compression extension described in [TLSCertCompress]. It uses a predistributed dictionary consisting of all intermediate and root certificates contained in the root stores of major browsers which is sourced from the Common CA Database [CCADB]. As of May 2023, this dictionary would be 3 MB in size and consist of roughly 2000 intermediate certificates and 200 root certificates. The disk footprint can be reduced to near zero as many clients (such as Mozilla Firefox & Google Chrome) are already provisioned with their trusted intermediate and root certificates for compatibility and performance reasons.¶

Using a shared dictionary allows for this compression scheme to deliver dramatically more effective compression than previous schemes, reducing the size of certificate chains in use today by ~75%, significantly improving on the ~25% reduction achieved by existing schemes. A preliminary evaluation (Section 4) of this scheme suggests that 50% of certificate chains in use today would be compressed to under 1000 bytes and 95% to under 1500 bytes. Similarly to [SCA], this scheme effectively removes the CA certificates from the certificate chain on the wire but this draft achieves a much better compression ratio, since [SCA] removes the redundant information in chain that existing TLS Certificate Compression schemes exploit and is more fragile in the presence of out of sync clients or servers.¶

Note that as this is only a compression scheme, it does not impact any trust decisions in the TLS handshake. A client can offer this compression scheme whilst only trusting a subset of the certificates in the CCADB certificate listing, similarly a server can offer this compression scheme whilst using a certificate chain which does not chain back to a WebPKI root. Furthermore, new root certificates are typically included in the CCADB at the start of their application to a root store, a process which typically takes more than a year. Consequently, applicant root certificates can be added to new versions of this scheme ahead of any trust decisions, allowing new CAs to compete on equal terms with existing CAs as soon as they are approved for inclusion in a root program. As a result this scheme is equitable in so far as it provides equal benefits for all CAs in the WebPKI, doesn't privilege any particular end-entity certificate or website and allows WebPKI clients to make individual trust decisions without fear of breakage.¶

1.3. Relationship to other drafts

This draft defines a certificate compression mechanism suitable for use with TLS Certificate Compression [TLSCertCompress].¶

The intent of this draft is to provide an alternative to CA Certificate Suppression [SCA] as it provides a better compression ratio, can operate in a wider range of scenarios (including out of sync clients or servers) and doesn't require any additional error handling or retry mechanisms.¶

CBOR Encoded X.509 (C509) [I-D.ietf-cose-cbor-encoded-cert-05] defines a concise alternative format for X.509 certificates. If this format were to become widely used on the WebPKI, defining an alternative version of this draft specifically for C509 certificates would be beneficial.¶

Compact TLS, (cTLS) [I-D.ietf-tls-ctls-08] defines a version of TLS1.3 which allows a pre-configured client and server to establish a session with minimal overhead on the wire. In particular, it supports the use of a predefined list of certificates known to both parties which can be compressed. However, cTLS is still at an early stage and may be challenging to deploy in a WebPKI context due to the need for clients and servers to have prior knowledge of handshake profile in use.¶

TLS Cached Information Extension [RFC7924] introduced a new extension allowing clients to signal they had cached certificate information from a previous connection and for servers to signal that the clients should use that cache instead of transmitting a redundant set of certificates. However this RFC has seen little adoption in the wild due to concerns over client privacy.¶

Handling long certificate chains in TLS-Based EAP Methods [RFC9191] discusses the challenges of long certificate chains outside the WebPKI ecosystem. Although the scheme proposed in this draft is targeted at WebPKI use, defining alternative shared dictionaries for other major ecosystems may be of interest.¶

1.4. Status

This draft is still at an early stage. Open questions are marked with the tag DISCUSS.¶

3.1. Pass 1: Intermediate and Root Compression

This pass relies on a shared listing of intermediate and root certificates known to both client and server. As many clients (e.g. Mozilla Firefox and Google Chrome) already ship with a list of trusted intermediate and root certificates, this pass allows their existing lists to be reused, rather than requiring them to have to be duplicated and stored in a separate format. The first subsection details how the certificates are enumerated in an ordered list. This ordered list is distributed to client and servers which use it to compress and decompress certificate chains as detailed in the subsequent subsection.¶

3.2. Pass 2: End-Entity Compression

This section describes a pass based on Zstandard [ZSTD] with application-specified dictionaries. The dictionary is constructed with reference to the list of intermediate and root certificates discussed earlier in Section 3.1.1, as well as several external sources of information.¶

[[DISCUSS: This draft is unopinionated about the underlying compression scheme is used as long as it supports application specified dictionaries. Is there an argument for using a different scheme?]]¶

5.1. Dictionary Versioning

The scheme defined in this draft is deployed with the static dictionaries constructed from the parameters listed in Section 3 fixed to a particular TLS Certificate Compression code point.¶

As new CA certificates are added to the CCADB and deployed on the web, new versions of this draft would need to be issued with their own code point and dictionary parameters. However, the process of adding new root certificates to a root store is already a two to three year process and this scheme includes untrusted root certificates still undergoing the application process in its dictionary. As a result, it would be reasonable to expect a new version of this scheme with updated dictionaries to be issued at most once a year and more likely once every two or three years.¶

A more detailed analysis and discussion of CA certificate lifetimes and root store operations is included in Appendix B, as well as an alternative design which would allow for dictionary negotiation rather than fixing one dictionary per code point.¶

[[DISCUSS: Are there concerns over this approach? Would using at most one code point per year be acceptable? Currently 3 of the 16384 'Specification Required' IANA managed code points are in use.]]¶

5.2. Version Migration

As new versions of this scheme are specified, clients and servers would benefit from migrating to the latest version. Whilst servers using CA certificates outside the scheme's listing can still offer this compression scheme and partially benefit from it, migrating to the latest version ensures that new CAs can compete on a level playing field with existing CAs. It is possible for a client or server to offer multiple versions of this scheme without having to pay twice the storage cost, since the majority of the stored data is in the pass 1 certificate listing and the majority of certificates will be in both versions and so need only be stored once.¶

Clients and servers SHOULD offer the latest version of this scheme and MAY offer one or more historical versions. Although clients and servers which fall out of date will no longer benefit from the scheme, they will not suffer any other penalties or incompatibilities. Future schemes will likely establish recommended lifetimes for sunsetting a previous version and adopting a new one.¶

As the majority of clients deploying this scheme are likely to be web browsers which typically use monthly release cycles (even long term support versions like Firefox ESR offer point releases on a monthly basis), this is unlikely to be a restriction in practice. The picture is more complex for servers as operators are often to reluctant to update TLS libraries, but as a new version only requires changes to static data without any new code and would happen infrequently, this is unlikely to be burdensome in practice.¶

5.3. Disk Space Requirements

Clients and servers implementing this scheme need to store a listing of root and intermediate certificates for pass 1, which currently occupies around ~3 MB and a smaller dictionary on the order of ~100 KB for pass 2. Clients and servers offering multiple versions of this scheme do not need to duplicate the pass 1 listing, as multiple versions can refer to same string.¶

As popular web browsers already ship a complete list of trusted intermediate and root certificates, their additional storage requirements are minimal. Servers offering this scheme are intended to be 'full-fat' web servers and so typically have plentiful storage already. This draft is not intended for use in storage-constrained IoT devices, but alternative versions with stripped down listings may be suitable.¶

[[DISCUSS: The current draft priorities an equitable treatment for every recognized and applicant CA over minimizing storage requirements. The required disk space could be significantly reduced by only including CAs which meet a particular popularity threshold via CT log sampling.]]¶

5.4. Implementation Complexity

Although much of this draft is dedicated to the construction of the certificate list and dictionary used in the scheme, implementations are indifferent to these details. Pass 1 can be implemented as a simple string substitution and pass 2 with a single call to permissively licensed and widely distributed Zstandard implementations such as [ZstdImpl]. Future versions of this draft which vary the dictionary construction then only require changes to the static data shipped with these implementations and the use of a new code point.¶

There are several options for handling the distribution of the associated static data. One option is to distribute it directly with the TLS library and update it as part of that library's regular release cycle. Whilst this is easy for statically linked libraries written in languages which offer first-class package management and compile time feature selection (e.g. Go, Rust), it is trickier for dynamically linked libraries who are unlikely to want to incur the increased distribution size. In these ecosystems it may make sense to distribute the dictionaries are part of an independent package managed by the OS which can be discovered by the library at run-time. Another promising alternative would be to have existing automated certificate tooling provision the library with both the full certificate chain and multiple precompressed chains during the certificate issuance / renewal process.¶

B.1. CCADB Churn

Typically around 10 or so new root certificates are introduced to the WebPKI each year. The various root programs restrict the lifetimes of these certificates, Microsoft to between 8 and 25 years (3.A.3), Mozilla to between 0 and 14 years (Summary). Chrome has proposed a maximum lifetime of 7 years in the future (Blog Post). Some major CAs have objected to this proposed policy as the root inclusion process currently takes around 3 years from start to finish (Digicert Blog). Similarly, Mozilla requires CAs to apply to renew their roots with at least 2 years notice (Summary).¶

Typically around 100 to 200 new WebPKI intermediate certificates are issued each year. No WebPKI root program currently limits the lifetime of intermediate certificates, but they are in practice capped by the lifetime of their parent root certificate. The vast majority of these certificates are issued with 10 year lifespans. A small but notable fraction (<10%) are issued with 2 or 3 year lifetimes. Chrome's Root Program has proposed that Intermediate Certificates be limited to 3 years in the future (Update). However, the motivation for this requirement is unclear. Unlike root certificates, intermediate certificates are only required to be disclosed with a month's notice to the CCADB (Mozilla Root Program Section 5.3.2, Chrome Policy).¶

B.2. Dictionary Negotiation

This draft is currently written with a view to being adopted as a particular TLS Certificate Compression Scheme. However, this means that each dictionary used in the wild must have an assigned code point. A new dictionary would likely need to be issued no more than yearly. However, negotiating the dictionary used would avoid the overhead of minting a new draft and code point. A sketch for how dictionary negotiation might work is below.¶

This draft would instead define a new extension, which would define TLS Certificate Compression with ZStandard Dictionaries. Dictionaries would be identified by an IANA-assigned identifier of two bytes, with a further two bytes for the major version and two more for the minor version. Adding new certificates to a dictionary listing would require a minor version bump. Removing certificates or changing the pass 2 dictionary would require a major version bump.¶

struct {
  uint16 identifier;
  uint16 major_version;
  uint16 minor_version;
} DictionaryId

struct {
  DictionaryId supported_dictionaries<6..2^16-1>
} ZStdSharedDictXtn

The client lists their known dictionaries in an extension in the ClientHello. The client need only retain and advertise the highest known minor version for any major version of a dictionary they are willing to offer. The server may select any dictionary it has a copy of with matching identifier, matching major version number and a minor version number not greater than the client's minor version number.¶

The expectation would be that new minor versions would be issued monthly or quarterly, with new major versions only every year or multiple years. This reflects the relative rates of when certificates are added or removed to the CCADB listing. This means in practice clients would likely offer a single dictionary containing their latest known version. Servers would only need to update their dictionaries yearly when a new major version is produced.¶