Netflix

ckrasic@netflix.com

Akamai Technologies

mbishop@evequefou.be

Facebook

afrind@fb.com

Transport QUIC This specification defines QPACK, a compression format for efficiently representing HTTP header fields, to be used in HTTP over QUIC. This is a variation of HPACK header compression that seeks to reduce head-of-line blocking. Discussion of this draft takes place on the QUIC working group mailing list (quic@ietf.org), which is archived at https://mailarchive.ietf.org/arch/search/?email_list=quic. Working Group information can be found at https://github.com/quicwg; source code and issues list for this draft can be found at https://github.com/quicwg/base-drafts/labels/-qpack.

The QUIC transport protocol was designed from the outset to support HTTP semantics, and its design subsumes many of the features of HTTP/2. HTTP/2 used HPACK () for header compression, but QUIC’s stream multiplexing comes into some conflict with HPACK. A key goal of the design of QUIC is to improve stream multiplexing relative to HTTP/2 by reducing head-of-line blocking. If HPACK were used for HTTP/QUIC, it would induce head-of-line blocking due to built-in assumptions of a total ordering across frames on all streams. QUIC is described in . The HTTP/QUIC mapping is described in . For a full description of HTTP/2, see . The description of HPACK is , with important terminology in Section 1.3. QPACK reuses core concepts from HPACK, but is redesigned to allow correctness in the presence of out-of-order delivery, with flexibility for implementations to balance between resilience against head-of-line blocking and optimal compression ratio. The design goals are to closely approach the compression ratio of HPACK with substantially less head-of-line blocking under the same loss conditions.

Like HPACK, QPACK uses two tables for associating header fields to indexes. The static table (see ) is predefined and contains common header fields (some of them with an empty value). The dynamic table (see ) built up over the course of the connection and can be used by the encoder to index header fields repeated in the encoded header lists. Unlike in HPACK, entries in the QPACK static and dynamic tables are addressed separately. The following sections describe how entries in each table is addressed.

The static table consists of a predefined static list of header fields, each of which has a fixed index over time. Its entries are defined in Appendix A of . Note that because HPACK did not use zero-based references, there is no value at index zero of the static table.

The dynamic table consists of a list of header fields maintained in first-in, first-out order. The dynamic table is initially empty. Entries are added by instructions on the Encoder Stream (see ). Before a new entry is added to the dynamic table, entries are evicted from the end of the dynamic table until the size of the dynamic table is less than or equal to (maximum size - new entry size) or until the table is empty. If the size of the new entry is less than or equal to the maximum size, that entry is added to the table. It is an error to attempt to add an entry that is larger than the maximum size; this MUST be treated as a connection error of type HTTP_QPACK_DECOMPRESSION_FAILED. A new entry can reference an entry in the dynamic table that will be evicted when adding this new entry into the dynamic table. Implementations are cautioned to avoid deleting the referenced name if the referenced entry is evicted from the dynamic table prior to inserting the new entry. The dynamic table can contain duplicate entries (i.e., entries with the same name and same value). Therefore, duplicate entries MUST NOT be treated as an error by a decoder. The encoder decides how to update the dynamic table and as such can control how much memory is used by the dynamic table. To limit the memory requirements of the decoder, the dynamic table size is strictly bounded. The decoder determines the maximum size that the encoder is permitted to use for the dynamic table. In HTTP/QUIC, this value is determined by the SETTINGS_HEADER_TABLE_SIZE setting (see Section 4.2.5.2 of ). An encoder can choose to use less capacity than this maximum size (see ), but the chosen size MUST stay lower than or equal to the maximum set by the decoder. Whenever the maximum size for the dynamic table is reduced, entries are evicted from the end of the dynamic table until the size of the dynamic table is less than or equal to the maximum size. This mechanism can be used to completely clear entries from the dynamic table by setting a maximum size of 0, which can subsequently be restored.

Each entry possesses both an absolute index which is fixed for the lifetime of that entry and a relative index which changes over time based on the context of the reference. The first entry inserted has an absolute index of “1”; indices increase sequentially with each insertion. The relative index begins at zero and increases in the opposite direction from the absolute index. Determining which entry has a relative index of “0” depends on the context of the reference. On the control stream, a relative index of “0” always refers to the most recently inserted value in the dynamic table. Note that this means the entry referenced by a given relative index will change while interpreting instructions on the encoder stream.

Because frames from request streams can be delivered out of order with instructions on the control stream, relative indices are relative to the Base Index at the beginning of the header block (see ). The Base Index is an absolute index. When interpreting the rest of the frame, the entry identified by Base Index has a relative index of zero. The relative indices of entries do not change while interpreting headers on a request or push stream.

A header block on the request stream can reference entries added after the entry identified by the Base Index. This allows an encoder to process a header block in a single pass and include references to entries added while processing this (or other) header blocks. Newly added entries are referenced using Post-Base instructions. Indices for Post-Base instructions increase in the same direction as absolute indices, but the zero value is one higher than the Base Index.

If the decoder encounters a reference to an entry which has already been dropped from the table or which is greater than the declared Largest Reference (see ), this MUST be treated as a stream error of type HTTP_QPACK_DECOMPRESSION_FAILED error code. If this reference occurs on the control stream, this MUST be treated as a session error.

Because QUIC does not guarantee order between data on different streams, a header block might reference an entry in the dynamic table that has not yet been received. Each header block contains a Largest Reference which identifies the table state necessary for decoding. If the greatest absolute index in the dynamic table is less than the value of the Largest Reference, the stream is considered “blocked.” While blocked, header field data should remain in the blocked stream’s flow control window. When the Largest Reference is zero, the frame contains no references to the dynamic table and can always be processed immediately. A stream becomes unblocked when the greatest absolute index in the dynamic table becomes greater than or equal to the Largest Reference for all header blocks the decoder has started reading from the stream. If a decoder encounters a header block where the actual largest reference is not equal to the largest reference declared in the prefix, it MAY treat this as a stream error of type HTTP_QPACK_DECOMPRESSION_FAILED. A decoder can permit the possibility of blocked streams by setting SETTINGS_QPACK_BLOCKED_STREAMS to a non-zero value (see ). This setting specifies an upper bound on the number of streams which can be blocked. An encoder can decide whether to risk having a stream become blocked. If permitted by the value of SETTINGS_QPACK_BLOCKED_STREAMS, compression efficiency can be improved by referencing dynamic table entries that are still in transit, but if there is loss or reordering the stream can become blocked at the decoder. An encoder avoids the risk of blocking by only referencing dynamic table entries which have been acknowledged, but this means using literals. Since literals make the header block larger, this can result in the encoder becoming blocked on congestion or flow control limits. An encoder MUST limit the number of streams which could become blocked to the value of SETTINGS_QPACK_BLOCKED_STREAMS at all times. Note that the decoder might not actually become blocked on every stream which risks becoming blocked. If the decoder encounters more blocked streams than it promised to support, it SHOULD treat this as a stream error of type HTTP_QPACK_DECOMPRESSION_FAILED.

The decoder stream signals key events at the decoder that permit the encoder to track the decoder’s state. These events are: Successful processing of a header block Abandonment of a stream which might have remaining header blocks Receipt of new dynamic table entries Regardless of whether a header block contained blocking references, the knowledge that it was processed successfully permits the encoder to avoid evicting entries while references remain outstanding; see . When a stream is reset or abandoned, the indication that these header blocks will never be processed serves a similar function; see . For the encoder to identify which dynamic table entries can be safely used without a stream becoming blocked, the encoder tracks the absolute index of the decoder’s Largest Known Received entry. When blocking references are permitted, the encoder uses acknowledgement of header blocks to identify the Largest Known Received index, as described in . To acknowledge dynamic table entries which are not referenced by header blocks, for example because the encoder or the decoder have chosen not to risk blocked streams, the decoder sends a Table State Synchronize instruction (see ).

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. Definitions of terms that are used in this document: A name-value pair sent as part of an HTTP message. The full collection of headers associated with an HTTP message. The compressed representation of a header set. An implementation which transforms a header set into a header block. An implementation which transforms a header block into a header set. QPACK is a name, not an acronym.

Diagrams use the format described in Section 3.1 of , with the following additional conventions: Indicates that x is A bits long Indicates that x uses the prefixed integer encoding defined in Section 5.1 of , beginning with an A-bit prefix. Indicates that x is variable-length and extends to the end of the region.

QPACK defines two settings which are included in the HTTP/QUIC SETTINGS frame. An integer with a maximum value of 2^30 - 1. The default value is 4,096 bytes. See (TODO: reference PR#1357) for usage. An integer with a maximum value of 2^16 - 1. The default value is 100. See .

QPACK instructions occur in three locations, each of which uses a separate instruction space: The encoder stream is a unidirectional stream of type 0x48 (ASCII ‘H’) which carries table updates from encoder to decoder. Instructions on this stream modify the dynamic table state without generating output to any particular request. The decoder stream is a unidirectional stream of type 0x68 (ASCII ‘h’) which carries acknowledgements of table modifications and header processing from decoder to encoder. Finally, the contents of HEADERS and PUSH_PROMISE frames on request streams and push streams reference the QPACK table state. There MUST be exactly one of each unidirectional stream type in each direction. Receipt of a second instance of either stream type MUST be treated as a connection error of HTTP_WRONG_STREAM_COUNT. Closure of either unidirectional stream MUST be treated as a connection error of type HTTP_CLOSED_CRITICAL_STREAM. This section describes the instructions which are possible on each stream type. All table updates occur on the encoder stream. Request streams and push streams only carry header blocks that do not modify the state of the table.

The prefixed integer from Section 5.1 of is used heavily throughout this document. The format from is used unmodified.

The string literal defined by Section 5.2 of is also used throughout. This string format includes optional Huffman encoding. HPACK defines string literals to begin on a byte boundary. They begin with a single flag (indicating whether the string is Huffman-coded), followed by the Length encoded as a 7-bit prefix integer, and finally Length octets of data. When Huffman encoding is enabled, the Huffman table from Appendix B of is used without modification. This document expands the definition of string literals and permits them to begin other than on a byte boundary. An “N-bit prefix string literal” begins with the same Huffman flag, followed by the length encoded as an (N-1)-bit prefix integer. The remainder of the string literal is unmodified. A string literal without a prefix length noted is an 8-bit prefix string literal and follows the definitions in without modification.

Table updates can add a table entry, possibly using existing entries to avoid transmitting redundant information. The name can be transmitted as a reference to an existing entry in the static or the dynamic table or as a string literal. For entries which already exist in the dynamic table, the full entry can also be used by reference, creating a duplicate entry. The contents of the encoder stream are an unframed sequence of the following instructions.

An addition to the header table where the header field name matches the header field name of an entry stored in the static table or the dynamic table starts with the ‘1’ one-bit pattern. The S bit indicates whether the reference is to the static (S=1) or dynamic (S=0) table. The header field name is represented using the relative index of that entry, which is represented as an integer with a 6-bit prefix (see Section 5.1 of ). The header name reference is followed by the header field value represented as a string literal (see Section 5.2 of ).

An addition to the header table where both the header field name and the header field value are represented as string literals (see ) starts with the ‘01’ two-bit pattern. The name is represented as a 6-bit prefix string literal, while the value is represented as an 8-bit prefix string literal.

Duplication of an existing entry in the dynamic table starts with the ‘000’ three-bit pattern. The relative index of the existing entry is represented as an integer with a 5-bit prefix.

The existing entry is re-inserted into the dynamic table without resending either the name or the value. This is useful to mitigate the eviction of older entries which are frequently referenced, both to avoid the need to resend the header and to avoid the entry in the table blocking the ability to insert new headers.

An encoder informs the decoder of a change to the size of the dynamic table using an instruction which begins with the ‘001’ three-bit pattern. The new maximum table size is represented as an integer with a 5-bit prefix (see Section 5.1 of ).

The new maximum size MUST be lower than or equal to the limit determined by the protocol using QPACK. A value that exceeds this limit MUST be treated as a decoding error. In HTTP/QUIC, this limit is the value of the SETTINGS_HEADER_TABLE_SIZE parameter (see ) received from the decoder. Reducing the maximum size of the dynamic table can cause entries to be evicted (see Section 4.3 of ). This MUST NOT cause the eviction of entries with outstanding references (see ). Changing the size of the dynamic table is not acknowledged as this instruction does not insert an entry.

The decoder stream carries information used to ensure consistency of the dynamic table. Information is sent from the QPACK decoder to the QPACK encoder; that is, the server informs the client about the processing of the client’s header blocks and table updates, and the client informs the server about the processing of the server’s header blocks and table updates. The contents of the decoder stream are an unframed sequence of the following instructions.

The Table State Synchronize instruction begins with the ‘00’ two-bit pattern. The instruction specifies the total number of dynamic table inserts and duplications since the last Table State Synchronize or Header Acknowledgement that increased the Largest Known Received dynamic table entry. This is encoded as a 6-bit prefix integer. The encoder uses this value to determine which table entries might cause a stream to become blocked, as described in .

A decoder chooses when to emit Table State Synchronize instructions. Emitting a Table State Synchronize after adding each new dynamic table entry will provide the most timely feedback to the encoder, but could be redundant with other decoder feedback. By delaying a Table State Synchronize, a decoder might be able to coalesce multiple Table State Synchronize instructions, or replace them entirely with Header Acknowledgements. However, delaying too long may lead to compression inefficiencies if the encoder waits for an entry to be acknowledged before using it.

After processing a header block on a request or push stream, the decoder emits a Header Acknowledgement instruction on the decoder stream. The instruction begins with the ‘1’ one-bit pattern and includes the request stream’s stream ID, encoded as a 7-bit prefix integer. It is used by the peer’s QPACK encoder to know when it is safe to evict an entry. The same Stream ID can be identified multiple times, as multiple header blocks can be sent on a single stream in the case of intermediate responses, trailers, and pushed requests. Since header frames on each stream are received and processed in order, this gives the encoder precise feedback on which header blocks within a stream have been fully processed.

When blocking references are permitted, the encoder uses acknowledgement of header blocks to update the Largest Known Received index. If a header block was potentially blocking, the acknowledgement implies that the decoder has received all dynamic table state necessary to process the header block. If the Largest Reference of an acknowledged header block was greater than the encoder’s current Largest Known Received index, the block’s Largest Reference becomes the new Largest Known Received.

A stream that is reset might have multiple outstanding header blocks. A decoder that receives a stream reset before the end of a stream generates a Stream Cancellation instruction on the decoder stream. Similarly, a decoder that abandons reading of a stream needs to signal this using the Stream Cancellation instruction. This signals to the encoder that all references to the dynamic table on that stream are no longer outstanding. An encoder cannot infer from this instruction that any updates to the dynamic table have been received. The instruction begins with the ‘01’ two-bit pattern. The instruction includes the stream ID of the affected stream - a request or push stream - encoded as a 6-bit prefix integer.

HEADERS and PUSH_PROMISE frames on request and push streams reference the dynamic table in a particular state without modifying it. Frames on these streams emit the headers for an HTTP request or response.

Header data is prefixed with two integers, Largest Reference and Base Index.

Largest Reference identifies the largest absolute dynamic index referenced in the block. Blocking decoders use the Largest Reference to determine when it is safe to process the rest of the block. Base Index is used to resolve references in the dynamic table as described in . To save space, Base Index is encoded relative to Largest Reference using a one-bit sign and the Delta Base Index value. A sign bit of 0 indicates that the Base Index has an absolute index that is greater than or equal to the Largest Reference; the value of Delta Base Index is added to the Largest Reference to determine the absolute value of the Base Index. A sign bit of 1 indicates that the Base Index is less than the Largest Reference. That is:

A single-pass encoder is expected to determine the absolute value of Base Index before encoding a header block. If the encoder inserted entries in the dynamic table while encoding the header block, Largest Reference will be greater than Base Index, so the encoded difference is negative and the sign bit is set to 1. If the header block did not reference the most recent entry in the table and did not insert any new entries, Base Index will be greater than the Largest Reference, so the delta will be positive and the sign bit is set to 0. An encoder that produces table updates before encoding a header block might set Largest Reference and Base Index to the same value. When Largest Reference and Base Index are equal, the Delta Base Index is encoded with a zero sign bit. A sign bit set to 1 when the Delta Base Index is 0 MUST be treated as a decoder error. A header block that does not reference the dynamic table can use any value for Base Index; setting both Largest Reference and Base Index to zero is the most efficient encoding.

An indexed header field representation identifies an entry in either the static table or the dynamic table and causes that header field to be added to the decoded header list, as described in Section 3.2 of .

If the entry is in the static table, or in the dynamic table with an absolute index less than or equal to Base Index, this representation starts with the ‘1’ 1-bit pattern, followed by the S bit indicating whether the reference is into the static (S=1) or dynamic (S=0) table. Finally, the relative index of the matching header field is represented as an integer with a 6-bit prefix (see Section 5.1 of ).

If the entry is in the dynamic table with an absolute index greater than Base Index, the representation starts with the ‘0001’ 4-bit pattern, followed by the post-base index (see ) of the matching header field, represented as an integer with a 4-bit prefix (see Section 5.1 of ).

A literal header field with a name reference represents a header where the header field name matches the header field name of an entry stored in the static table or the dynamic table. If the entry is in the static table, or in the dynamic table with an absolute index less than or equal to Base Index, this representation starts with the ‘01’ two-bit pattern. If the entry is in the dynamic table with an absolute index greater than Base Index, the representation starts with the ‘0000’ four-bit pattern. The following bit, ‘N’, indicates whether an intermediary is permitted to add this header to the dynamic header table on subsequent hops. When the ‘N’ bit is set, the encoded header MUST always be encoded with a literal representation. In particular, when a peer sends a header field that it received represented as a literal header field with the ‘N’ bit set, it MUST use a literal representation to forward this header field. This bit is intended for protecting header field values that are not to be put at risk by compressing them (see Section 7.1 of for more details).

For entries in the static table or in the dynamic table with an absolute index less than or equal to Base Index, the header field name is represented using the relative index of that entry, which is represented as an integer with a 4-bit prefix (see Section 5.1 of ). The S bit indicates whether the reference is to the static (S=1) or dynamic (S=0) table.

For entries in the dynamic table with an absolute index greater than Base Index, the header field name is represented using the post-base index of that entry (see ) encoded as an integer with a 3-bit prefix.

An addition to the header table where both the header field name and the header field value are represented as string literals (see ) starts with the ‘001’ three-bit pattern. The fourth bit, ‘N’, indicates whether an intermediary is permitted to add this header to the dynamic header table on subsequent hops. When the ‘N’ bit is set, the encoded header MUST always be encoded with a literal representation. In particular, when a peer sends a header field that it received represented as a literal header field with the ‘N’ bit set, it MUST use a literal representation to forward this header field. This bit is intended for protecting header field values that are not to be put at risk by compressing them (see Section 7.1 of for more details). The name is represented as a 4-bit prefix string literal, while the value is represented as an 8-bit prefix string literal.

An encoder making a single pass over a list of headers must choose Base Index before knowing Largest Reference. When trying to reference a header inserted to the table after encoding has begun, the entry is encoded with different instructions that tell the decoder to use an absolute index greater than the Base Index.

Due to out-of-order arrival, QPACK’s eviction algorithm requires changes (relative to HPACK) to avoid the possibility that an indexed representation is decoded after the referenced entry has already been evicted. QPACK employs a two-phase eviction algorithm, in which the encoder will not evict entries that have outstanding (unacknowledged) references.

An encoder MUST ensure that a header block which references a dynamic table entry is not received by the decoder after the referenced entry has already been evicted. An encoder also respects the limit set by the decoder on the number of streams that are allowed to become blocked. Even if the decoder is willing to tolerate blocked streams, the encoder might choose to avoid them in certain cases. In order to enable this, the encoder will need to track outstanding (unacknowledged) header blocks and table updates using feedback received from the decoder.

The encoder MUST NOT permit an entry to be evicted while a reference to that entry remains unacknowledged. If a new header to be inserted into the dynamic table would cause the eviction of such an entry, the encoder MUST NOT emit the insert instruction until the reference has been processed by the decoder and acknowledged. The encoder can emit a literal representation for the new header in order to avoid encoding delays, and MAY insert the header into the table later if desired. To ensure that the blocked eviction case is rare, references to the oldest entries in the dynamic table SHOULD be avoided. When one of the oldest entries in the table is still actively used for references, the encoder SHOULD emit an Duplicate representation instead (see ).

For header blocks encoded in non-blocking mode, the encoder needs to forego indexed representations that refer to table updates which have not yet been acknowledged with . Since all table updates are processed in sequence on the control stream, an index into the dynamic table is sufficient to track which entries have been acknowledged. To track blocked streams, the necessary Base Index value for each stream can be used. Whenever the decoder processes a table update, it can begin decoding any blocked streams that now have their dependencies satisfied.

Implementations can speculatively send header frames on the HTTP Control Streams which are not needed for any current HTTP request or response. Such headers could be used strategically to improve performance. For instance, the encoder might decide to refresh by sending Duplicate representations for popular header fields (), ensuring they have small indices and hence minimal size on the wire.

Pseudo-code for single pass encoding, excluding handling of duplicates, non-blocking mode, and reference tracking.

= largestReference: encodeInteger(prefixBuffer, 0, baseIndex - largestReference, 7) else: encodeInteger(prefixBuffer, 0x80, largestReference - baseIndex, 7) return controlBuffer, prefixBuffer + streamBuffer ]]>

TBD.

This document creates two new settings in the “HTTP/QUIC Settings” registry established in . The entries in the following table are registered by this document. Setting Name Code Specification HEADER_TABLE_SIZE 0x1 QPACK_BLOCKED_STREAMS 0x7

This document creates two new settings in the “HTTP/QUIC Stream Type” registry established in . The entries in the following table are registered by this document. Stream Type Code Specification Sender QPACK Encoder Stream 0x48 Both QPACK Decoder Stream 0x68 Both

Akamai Technologies This specification defines HPACK, a compression format for efficiently representing HTTP header fields, to be used in HTTP/2. In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. RFC 2119 specifies common key words that may be used in protocol specifications. This document aims to reduce the ambiguity by clarifying that only UPPERCASE usage of the key words have the defined special meanings. This document defines the core of the QUIC transport protocol. This document describes connection establishment, packet format, multiplexing and reliability. Accompanying documents describe the cryptographic handshake and loss detection. This specification describes an optimized expression of the semantics of the Hypertext Transfer Protocol (HTTP), referred to as HTTP version 2 (HTTP/2). HTTP/2 enables a more efficient use of network resources and a reduced perception of latency by introducing header field compression and allowing multiple concurrent exchanges on the same connection. It also introduces unsolicited push of representations from servers to clients.This specification is an alternative to, but does not obsolete, the HTTP/1.1 message syntax. HTTP's existing semantics remain unchanged. This document is a guide for Internet standard writers. It defines those characteristics that make standards coherent, unambiguous, and easy to interpret. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements.

RFC Editor’s Note: Please remove this section prior to publication of a final version of this document.

Renumbered instructions for consistency (#1471, #1472) Decoder is allowed to validate largest reference (#1404, #1469) Header block acknowledgments also acknowledge the associated largest reference (#1370, #1400) Added an acknowledgment for unread streams (#1371, #1400) Removed framing from encoder stream (#1361,#1467) Control streams use typed unidirectional streams rather than fixed stream IDs (#910,#1359)

Separate instruction sets for table updates and header blocks (#1235, #1142, #1141) Reworked indexing scheme (#1176, #1145, #1136, #1130, #1125, #1314) Added mechanisms that support one-pass encoding (#1138, #1320) Added a setting to control the number of blocked decoders (#238, #1140, #1143) Moved table updates and acknowledgments to dedicated streams (#1121, #1122, #1238)

This draft draws heavily on the text of . The indirect input of those authors is gratefully acknowledged, as well as ideas from: Ryan Hamilton Patrick McManus Kazuho Oku Biren Roy Ian Swett Dmitri Tikhonov Buck’s contribution was supported by Google during his employment there. A substantial portion of Mike’s contribution was supported by Microsoft during his employment there.