]> The Varnish Cache Project

phk@varnish-cache.org

An abstract data model for HTTP headers, “Common Structure”, and a HTTP/1 serialization of it, generalized from current HTTP headers.

The HTTP protocol does not impose any structure or datamodel on the information in HTTP headers, the HTTP/1 serialization is the datamodel: An ASCII string without control characters. HTTP header definitions specify how the string must be formatted and while families of similar headers exist, it still requires an uncomfortable large number of bespoke parser and validation routines to process HTTP traffic correctly. In order to improve performance HTTP/2 and HPACK uses naive text-compression, which incidentally decoupled the on-the-wire serialization from the data model. During the development of HPACK it became evident that significantly bigger gains were available if semantic compression could be used, most notably with timestamps. However, the lack of a common data structure for HTTP headers would make semantic compression one long list of special cases. Parallel to this, various proposals for how to fulfill data-transportation needs, and to a lesser degree to impose some kind of order on HTTP headers, at least going forward were floated. All of these proposals, JSON, CBOR etc. run into the same basic problem: Their serialization is incompatible with ’s ABNF definition of ‘field-value’. For binary formats, such as CBOR, a wholesale base64/85 reserialization would be needed, with negative results for both debugability and bandwidth. For textual formats, such as JSON, the format must first be neutered to not violate field-value’s ABNF, and then workarounds added to reintroduce the features just lost, for instance UNICODE strings, and suddenly it is no longer JSON anymore. This proposal starts from the other end, and builds and generalizes a data structure definition from existing HTTP headers, which means that HTTP/1 serialization and ‘field-value’ compatibility is built in. If all new HTTP headers are defined to fit into this Common Structure we have at least halted the proliferation of bespoke parsers and started to pave the road for semantic compression serializations of HTTP traffic.

In this document, the key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “MAY”, and “OPTIONAL” are to be interpreted as described in BCP 14, RFC 2119 .

The data model of Common Structure is an ordered sequence of named dictionaries. Please see for how this model was derived. The definition of the data model is on purpose abstract, uncoupled from any protocol serialization or programming environment representation, meant as the foundation on which all such manifestations of the model can be built. Common Structure in ABNF:

In ABNF:

" h1_common_structure "<" ) # Self-identifying HTTP headers ( field-name ":" OWS h1_common_structure ) / # legacy HTTP headers on white-list, see {{iana}} h1_common_structure = h1_element * ("," h1_element) h1_element = identifier * (";" identifier ["=" h1_value]) h1_value = identifier / number / h1_ascii_string / h1_unicode_string / h1_blob / h1_timestamp / h1_common-structure h1_ascii_string = DQUOTE *( ( "\" DQUOTE ) / ( "\" "\" ) / 0x20-21 / 0x23-5B / 0x5D-7E ) DQUOTE # This is a proper subset of h1_unicode_string # NB only allowed backslash escapes are \" and \\ h1_unicode_string = DQUOTE *( ( "\" DQUOTE ) ( "\" "\" ) / ( "\" "u" 4*HEXDIG ) / 0x20-21 / 0x23-5B / 0x5D-7E / 0x80-F7 ) DQUOTE # XXX: how to say/import "UTF-8 encoding" ? # HTTP1 unfriendly codepoints (00-1f, 7f) must be # encoded with \uXXXX escapes h1_blob = "'" base64 "'" # XXX: where to import base64 from ? h1_timestamp = number # UNIX/POSIX time_t semantics. # fractional seconds allowed. h1_common_structure = ">" h1_common_structure "<" ]]>

XXX: Allow OWS in parsers, but not in generators ? In programming environments which do not define a native representation or serialization of Common Structure, the HTTP/1 serialization should be used.

All future standardized and all private HTTP headers using Common Structure should self identify as such. In the HTTP/1 serialization by making the first character “>” and the last “<”. (These two characters are deliberately “the wrong way” to not clash with exsisting usages.) Legacy HTTP headers which fit into Common Structure, are marked as such in the IANA Message Header Registry (see {iana}), and a snapshot of the registry can be used to trigger parsing according to Common Structure of these headers.

All new HTTP headers SHOULD use the Common Structure if at all possible.

Should we allow splitting common structure data over multiple headers ? Pro: Avoids size restrictions, easier on-the-fly editing Contra: Cannot act on any such header until all headers have been received. We must define where headers can be split (between identifier and dictionary ?, in the middle of dictionaries ?) Most on-the-fly editing is hackish at best.

The HTTP/1 serializations self-identification mechanism makes it possible to extend the definition of existing headers into Common Structure. For instance one could imagine:

1475061449.201< ]]>

Which would be faster to parse and validate than the current definition of the Date header and more precise too. Some kind of signal/negotiation mechanism would be required to make this work in practice.

A machine-readable specification of the legal contents of HTTP headers would go a long way to improve efficiency and security in HTTP implementations.

The IANA Message Header Registry will be extended with an additional field named “Common Structure” which can have the values “True”, “False” or “Unknown”. The RFC723x headers listed in will get the value “True” in the new field. The RFC723x headers listed in will get the value “False” in the new field. All other existing entries in the registry will be set to “Unknown” until and if the owner of the entry requests otherwise.

&RFC2119; &RFC7230; &RFC5234;

Several proposals have been floated in recent years to use some preexisting structured data serialization or other for HTTP headers, to impose some sanity. None of these proposals have gained traction and no obvious candidate data serializations have been left unexamined. This effort tries to tackle the question from the other side, by asking if there is a common structure in existing HTTP headers we can generalize for this purpose.

The RFC723x family of HTTP/1 standards control 49 entries in the IANA Message Header Registry, and they share two common motifs. The majority of RFC723x HTTP headers are lists. A few of them are ordered, (‘Content-Encoding’), some are unordered (‘Connection’) and some are ordered by ‘q=%f’ weight parameters (‘Accept’) In most cases, the list elements are some kind of identifier, usually derived from ABNF ‘token’ as defined by . A subgroup of headers, mostly related to MIME, uses what one could call a ‘qualified token’::

The second motif is parameterized list elements. The best known is the “q=0.5” weight parameter, but other parameters exist as well. Generalizing from these motifs, our candidate “Common Structure” data model becomes an ordered list of named dictionaries. In pidgin ABNF, ignoring white-space for the sake of clarity, the HTTP/1.1 serialization of Common Structure is is something like:

Nineteen out of the RFC723x’s 48 headers, almost 40%, can already be parsed using this definition, and none the rest have requirements which could not be met by this data model. See and for the full survey details.

Surveying the datatypes of HTTP headers, standardized as well as private, the following picture emerges:

Integer and floating point are both used. Range and precision is mostly unspecified in controlling documents. Scientific notation (9.192631770e9) does not seem to be used anywhere. The ranges used seem to be minus several thousand to plus a couple of billions, the high end almost exclusively being POSIX time_t timestamps.

RFC723x text format, but POSIX time_t represented as integer or floating point is not uncommon. ISO8601 have also been spotted.

The vast majority are pure ASCII strings, with either no escapes, %xx URL-like escapes or C-style back-slash escapes, possibly with the addition of \uxxxx UNICODE escapes. Where non-ASCII character sets are used, they are almost always implicit, rather than explicit. UTF8 and ISO-8859-1[5] seem to be most common.

Often used for cryptographic data. Usually in base64 encoding, sometimes ““-quoted more often not. base85 encoding is also seen, usually quoted.

Seems to almost always fit in the RFC723x ‘token’ definition.

The number one wishlist item seems to be UNICODE strings, with a big side order of not having to write a new parser routine every time somebody comes up with a new header. Having a common parser would indeed be a good thing, and having an underlying data model which makes it possible define a compressed serialization, rather than rely on serialization to text followed by text compression (ie: HPACK) seems like a good idea too. However, when using a datamodel and a parser general enough to transport useful data, it will have to be followed by a validation step, which checks that the data also makes sense. Today validation, such as it is, is often done by the bespoke parsers. This then is probably where the next big potential for improvement lies: Ideally a machine readable “data dictionary” which makes it possibly to copy that text out of RFCs, run it through a code generator which spits out validation code which operates on the output of the common parser. But history has been particularly unkind to that idea. Most attempts studied as part of this effort, have sunk under complexity caused by reaching for generality, but where scope has been wisely limited, it seems to be possible. So file that idea under “future work”.

1 of the RFC723x headers is only reserved, and therefore have no structure at all:

5 of the RFC723x headers are HTTP dates:

24 of the RFC723x headers use bespoke formats which only a single or in rare cases two headers share: