HTTP/2.0 Discussion: Stored Header Encoding
draft-snell-httpbis-bohe-12

Abstract

This memo describes a proposed alternative encoding for headers that combines the best concepts from the proposed Delta and HeaderDiff options with the typed value codecs introduced by previous versions of this draft.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at http://datatracker.ietf.org/drafts/current/.

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on February 07, 2014.

Copyright Notice

Copyright (c) 2013 IETF Trust and the persons identified as the document authors. All rights reserved.

This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.

Table of Contents

• 1. Stored Header Encoding
• 2. Model
• 3. Header Encoding and Decoding
• 3.1. Literal (name,value) Representation
• 3.1.1. Dealing with invalid name or value encodings
• 3.2. Indexed Representation
• 3.3. Non-Indexed Literal Representation
• 3.4. Indexed Literal Representation
• 3.5. Indexed Literal Replacement Representation
• 4. Unsigned Variable Length Integer Syntax
• 5. Security Considerations
• 6. References
• 6.1. Normative References
• 6.2. Informational References
• Appendix A. Initial Header Table Entries
• Appendix B. Updated Standard Header Definitions
• Appendix C. Example
• C.1. First Header Set:
• C.2. Second Header Set:
• C.3. Third Header Set:
• Author's Address

1. Stored Header Encoding

The Stored Header Encoding is a proposed alternative "compressed header encoding" for HTTP/2.0 that offers reasonably good compression ratios, support for a range of compressor strategies, efficient value type codecs, constrained state requirements, routing-friendly header value ordering, and easier implementation relative to the current header compression proposal.

2. Model

A "header" is a (name,value) pair. The name is a sequence of lower-case ASCII characters. The value is either a UTF-8 string, an integer, a date-time, or an arbitrary sequence of binary octets.

The compressor and decompressor each maintain a synchronized cache of up to 256 headers. Every header in the cache is referenced by an 8-bit identifier. Note that the Nil byte (0x00) is a valid identifier.

The cache is managed in a "least recently written" style, that is, as the cache fills to capacity in both number of entries and maximum stored byte size, the least recently written items are cleared and their index positions are reused.

Index positions from the cache are assigned in "encounter order", beginning from 0x00 and increasing monotonically to 0xFF. That is to say, the positions are assigned in precisely the same order that they are serialized, and thereby encountered by the decompressor upon reading and processing the block.

The available size of the stored compression state can be capped by the decompressor using the SETTINGS_MAX_BUFFER_SIZE setting. Each stored header contributes to the accumulated size of the storage state. As new header pairs are assigned positions in the cache, the least-recently assigned items must be cleared, if necessary, to free up the required space. Clearing existing items does not change the index positions of the remaining items in the cache.

The SETTINGS_MAX_BUFFER_SIZE setting has an initial default value of 4096 bytes. The decompressor can establish a new maximum buffer size at any time, possibly causing header (name,value) pairs to be evicted from the cache.

The decompressor can disable use of the storage cache completely by setting the SETTINGS_MAX_BUFFER_SIZE setting to 0, forcing the cache to empty completely.

The size of a header is calculated as: The number of octets required for the name plus the number of octets required for the value plus 32-bytes to account for any internal storage overhead. The number of octets required for the value depends on the value type:

• String values are measured by the number of UTF-8 encoded octets required to represent the character sequence.
• Number and Date-Time values are measured by the number of unsigned variable length integer (uvarint) encoded bytes required to represent the value using a 5-bit prefix.
• Legacy (HTTP/1.1) values are measured by the number of octets required to represent the value.
• Binary values are measured by the number of octets contained by the sequence.

3. Header Encoding and Decoding

The set of headers is encoded for transmission using the following process:

1. For each header, determine if the (name,value) pair already exists in the table.
   • If an exact match is found in the header table, encode the indexed position of the header as an Indexed Reference and advance to the next header (name,value) pair.
   • Otherwise, move to the step #2.

2. Determine if a header (name,value) pair with the same name already exists in the table. If a matching name is found, make note of the indexed position of the matching name and continue to step #3.
3. Determine whether the new header (name,value) pair ought to be added to the header table.
   • If the header is not to be added to the header table, encode the header as a Non-Indexed Literal Representation and continue to the next header (name,value) pair.
   • Otherwise, continue to step #4.

4. If an existing indexed header using the same name was found in the header table in step #2, determine if the new header (name,value) pair ought to replace that existing entry or if it ought to be added as a new entry.
   • If the header is to replace the existing entry, encode the header as an Indexed Literal Replacement Representation.
   • Otherwise encode the header as an Indexed Literal Representation.

Following these steps, headers are serialized into one of four representation types, each represented by a two-bit prefix code. The types and their codes are:

• 10 - Indexed
• 00 - Non-Indexed Literal
• 01 - Indexed Literal
• 11 - Indexed Literal Replacement

Each representation type is encoded into groups of up to 64 instances. Each group is prefixed by a single octet prefix. The two most significant bits identify the representation type, the six least significant bits specify the number of instances in the group, with 000000 indicating a single instance and 111111 indicating 64.

If a particular serialization block requires more than 64 instances of a given type, then multiple occurrences of the group type can be encoded. For instance, if a given message contains 65 Indexed Representations, the encoded block would contain two separate Indexed Representation groups.

Decoding simply reverses the encoding steps:

1. First initialize an empty working set of headers.
2. Begin iterating through each representation group:
   • If it is an Indexed group, iterate through each index included in the group, look up the corresponding (name,value) pair in the header table and add that to the working set. If no matching (name,value) is found, terminate and report an error.
   • If it's a Non-Indexed Literal group, iterate through each (name,value) pair included in the group and add that to the working set.
   • If it's an Indexed Literal group, iterate through each (name,value) pair included in the group and add that to both the header table and the working set.
   • If it's an Indexed Literal Replacement group, iterate through each (name,value) pair included in the group, replace the existing entry in the header table at the identified index, and add the new (name,value) to the working set. If no matching (name,value) is found, terminate and report an error.

3. Continue with each representation group until the complete block has been decoded.

When a single header name is used multiple times with different values, the order in which those values are serialized and processed is significant. The working set created by the decoding process above MUST preserve the ordering of those values as received.

3.1. Literal (name,value) Representation

The structure of an encoded (name,value) pair consists of:

• A 3-bit value type identifier,
• The name, encoded either as a literal string or as the header table index position of another existing header sharing the same name, and
• The encoded value.

The three most-sigificant bits of the first octet identify the value type.

This design allows for a maximum of 7 value types, five of which are defined by this specification. The two remaining types are reserved for future use. The currently defined value types are:

• UTF-8 Text (000)
• Integer (001)
• Timestamp (010)
• Legacy (100)
• Raw Binary (111)

Of the five types, the Legacy type is reserved for encoding header values conforming to the field-value construct defined by [I-D.ietf-httpbis-p1-messaging], and is used specifically for backwards compatibility with header fields that have not yet been updated to use a more specific type value (see Appendix B).

If the name is encoded using an index reference to another existing (name,value) pair in the header table, the remaining five least significant bits of the first octet are set to zero and the next byte identifies the referenced header table index position.

If the name is encoded as a literal string, the number of octets required to represent the name is encoded as a unsigned variable length integer with a five-bit prefix, filling the 5-remaining least significant bits of the first octet, followed by the sequence of ASCII octets conforming to the following header-name construct:

Header name ABNF:

 LOWERALPHA = %x61-7A
      
 header-name = "!" / "#" / "$" / "%" / "&" / "'" / 
               "*" / "+" / "-" / "." / "^" / "_" / 
               "`" / "|" / "~" / DIGIT / LOWERALPHA
      

The encoding of the value depends on the value type.

UTF-8 Text:

First, the number of UTF-8 encoded bytes required to represent the value is encoded as an unsigned variable length integer with a 0-bit prefix, followed by the full sequence of UTF-8 bytes.

Integer

Integer values ranging from 0 to 2^64-1 are encoded as unsigned variable length integers with a 0-bit prefix. Negative or fractional numbers cannot be represented.

Timestamp

The timestamp is represented as the number of milliseconds ellapsed since the standard Epoch (1970-01-01T00:00:00 GMT), encoded as an unsigned variable length integer with a 0-bit prefix. Timestamps that predate the Epoch cannot be represented.

Legacy

First, the number of octets required to represent the value is encoded as an unsigned variable length integer with a 0-bit prefix, followed by the full sequence of octets.

Raw Binary

The number of octets in the sequence is encoded as an unsigned variable length integer with a 0-bit prefix, followed by the full sequence of octets.

3.1.1. Dealing with invalid name or value encodings

Implementations encountering invalid name or value encodings MUST signal an error and terminate processing of the header block. Examples of such errors include:

• Header names that include any octets not explicitly permitted by the above header-name ABNF construction;
• UTF-8 values that include a byte order mark, over-long or invalid octet sequences, or octets representing invalid Unicode codepoints;
• Integer or Date-Time values that encode numbers strictly larger than 2^64-1;

3.2. Indexed Representation

The serialization of an Indexed Representation group consists of the one-octet header group prefix followed by up to 64 single-octet header table index references.

  +--------+--------+--------+--------+---+
  |10xxxxxx| Index positions (1...64)  ...|
  +--------+--------------------------+---+
      

For instance:

0x80 0x00

References item #0 from the header table.

0x81 0x00 0x01

References items #0 and #1 from the header table.

Indexed Representations do not modify the header table state in any way. If an Indexed References specifies a header index that has not yet been allocated or whose value has been cleared, decoding MUST terminate with an error.

3.3. Non-Indexed Literal Representation

The serialization of a group of Non-Indexed Literal representations consists of the one-octet header prefix followed by up to 64 Literal (name,value) Representations.

  +--------+--------+--------+--------+---+
  |00xxxxxx| (name,value)'s (1...64)    ...
  +--------+--------------------------+---+
      

For instance:

0x00 0x01 0x61 0x01 0x62

Specifies a single header with name "a" and a UTF-8 String value of "b" is to be handled as a Non-Indexed header (it is not added to the header table).

3.4. Indexed Literal Representation

The serialization of a group of Indexed Literal representations consists of the one-octet header prefix followed by up to 64 Literal (name,value) Representations.

  +--------+--------+--------+--------+---+
  |01xxxxxx| (name,value)'s (1...64)    ...
  +--------+--------------------------+---+
      

For instance:

0x40 0x01 0x61 0x01 0x62

Specifies a single header with name "a" and a UTF-8 String value of "b" is to be handled as an Indexed header (it will be added to the header table).

0x40 0x21 0x61 0x03

Specifies a single header with name "a" and Integer value of 3 is to be handled as an Indexed header (it will be added to the header table).

3.5. Indexed Literal Replacement Representation

The serialization of a group of Indexed Literal representations consists of the one-octet header prefix followed by up to 64 single octet index references identifying an existing header table entry followed by the new Literal (name,value) representation meant to replace it.

  +--------+--------+--------+--------+--------+
  |11xxxxxx| (INDEX | (name,value)(1...64))  ...
  +--------+--------------------------+--------+
      

For instance:

0xC0 0x03 0x01 0x61 0x01 0x62

Specifies that a single header with name "a" and a UTF-8 String value of "b" is to replace the existing (name,value) entry in the header table located at index position #3.

0xC0 0x03 0x21 0x61 0x03

Specifies that a single header with name "a" and Integer value of 3 is to replace the existing (name,value) entry in the header table located at index position #3.

4. Unsigned Variable Length Integer Syntax

Unsigned integers are encoded as defined in [I-D.ietf-httpbis-header-compression].

5. Security Considerations

TBD

6. References

6.1. Normative References

6.2. Informational References

Appendix A. Initial Header Table Entries

Appendix B. Updated Standard Header Definitions

To properly deal with the backwards compatibility concerns for HTTP/1, there are several important rules for use of Typed Codecs in HTTP headers:

• All header fields MUST be explicitly defined to use the new header types. All existing HTTP/1 header fields will continue to be represented in conformance to the field-value construct defined by [I-D.ietf-httpbis-p1-messaging] unless their specific definitions are updated. Such fields MUST specify the Legacy value type when serialized. The HTTP/2 specification would update the definitions of specific known header fields (e.g. content-length, date, if-modified-since, etc).
• For translation to HTTP/1.1, header fields that use the typed codecs will have specific normative transformations defined.
  • UTF-8 Text will be converted to ISO-8859-1 with extended characters pct-encoded
  • Numbers will be converted to their ASCII equivalent values.
  • Date Times will be converted to their HTTP-Date equivalent values.
  • Binary fields will be Base64-encoded.
  • Legacy fields are passed through untranslated.

• There will be no normative transformation from legacy values into the typed codecs. Implementations are free to apply transformation where they determine it to be appropriate, but it will be perfectly acceptable for an implementation to pass a text value through as a Legacy type even if it is known that a given header has a typed codec equivalent.

A Note of warning: Individual header fields MAY be defined such that they can be represented using multiple types. Numeric fields, for instance, can be represented using either the uvarint encoding or using the equivalent sequence of ASCII numbers. Implementers will need to be capable of supporting each of the possible variations. Designers of header field definitions need to be aware of the additional complexity and possible issues that allowing for such alternatives can introduce for implementers.

Based on an initial survey of header fields currently defined by the HTTPbis specification documents, the following header field definitions can be updated to make use of the new types

Appendix C. Example

C.1. First Header Set:

The first header set to represent is the following:

  :path: /my-example/index.html
  user-agent: my-user-agent
  x-my-header: first
        

The header table is prefilled as defined in Appendix A, however, none of the values represented in the initial set can be found in the header table. All headers, then, are encoding using the Indexed Literal Representation:

  43 00 03 16 2f 6d 79 2d 65 78 
  61 6d 70 6c 65 2f 69 6e 64 65 
  78 2e 68 74 6d 6c 00 49 6d 79 
  2d 75 73 65 72 2d 61 67 65 6e 
  74 0b 78 2d 6d 79 2d 68 65 61 
  64 65 72 05 66 69 72 73 74 
        

Three new entries are added to the header table:

C.2. Second Header Set:

The second header set to represent is the following:

  :path: /my-example/resources/script.js
  user-agent: my-user-agent
  x-my-header: second
        

Comparing this second header set to the first, we see that the :path and x-my-header headers have new values, while the user-agent value remains unchanged. For the sake of the example let's encode the :path and x-my-header headers using Indexed Literal Replacement representations. The user-agent header will be encoded as an Indexed Representation.

  80 4b a3 4a 00 4a 1f 2f 6d 79 
  2d 65 78 61 6d 70 6c 65 2f 72 
  65 73 6f 75 72 63 65 73 2f 73 
  63 72 69 70 74 2e 6a 73 00 4c
  06 73 65 63 6f 6e 64
        

Items #74 and #76 added by the previous header set are replaced:

C.3. Third Header Set:

Let's suppose a third header set that is identical to the second is sent:

  82 4b 4c 4d
        

Author's Address