Extensible Prioritization Scheme for HTTP

This document describes a scheme for prioritizing HTTP responses. This scheme expresses the priority of each HTTP response using absolute values, rather than as a relative relationship between a group of HTTP responses.¶

This document defines the Priority header field for communicating the initial priority in an HTTP version-independent manner, as well as HTTP/2 and HTTP/3 frames for reprioritizing the responses. These share a common format structure that is designed to provide future extensibility.¶

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 https://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 January 13, 2022.¶

Copyright (c) 2021 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 (https://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.¶

It is common for an HTTP ([RFC7230]) resource representation to have relationships to one or more other resources. Clients will often discover these relationships while processing a retrieved representation, leading to further retrieval requests. Meanwhile, the nature of the relationship determines whether the client is blocked from continuing to process locally available resources. For example, visual rendering of an HTML document could be blocked by the retrieval of a CSS file that the document refers to. In contrast, inline images do not block rendering and get drawn incrementally as the chunks of the images arrive.¶

To provide meaningful presentation of a document at the earliest moment, it is important for an HTTP server to prioritize the HTTP responses, or the chunks of those HTTP responses, that it sends.¶

HTTP/2 ([HTTP2]) provides such a prioritization scheme. A client sends a series of PRIORITY frames to communicate to the server a "priority tree"; this represents the client's preferred ordering and weighted distribution of the bandwidth among the HTTP responses. However, the design and implementation of this scheme has been observed to have shortcomings, explained in Section 2.¶

This document defines the Priority HTTP header field that can be used by both client and server to specify the precedence of HTTP responses in a standardized, extensible, protocol-version-independent, end-to-end format. Along with the protocol-version-specific frame for reprioritization, this prioritization scheme acts as a substitute for the original prioritization scheme of HTTP/2.¶

An important feature of any implementation of a protocol that provides multiplexing is the ability to prioritize the sending of information. This was an important realization in the design of HTTP/2. Prioritization is a difficult problem, so it will always be suboptimal, particularly if one endpoint operates in ignorance of the needs of its peer.¶

HTTP/2 introduced a complex prioritization scheme that uses a combination of stream dependencies and weights to describe an unbalanced tree. This scheme has suffered from poor deployment and interoperability.¶

Clients build an HTTP/2 prioritization tree through a series of individual stream relationships, which are transferred to the server using HTTP/2 priority signals in either of three forms. First, a HEADERS frame with the PRIORITY flag set is an explicit signal that includes an Exclusive flag, Stream Dependency field, and Weight field. Second, a HEADERS frame with no PRIORITY flag is an implicit signal to use the default priority. Third, the PRIORITY frame, which is always explicit since it always contains an Exclusive flag, Stream Dependency field, and Weight field.¶

The rich flexibility of tree building is rarely exercised. Experience has shown that clients tend to choose a single model optimized for a web use case and experiment within the model constraints, or do nothing at all. Furthermore, many clients build their prioritization tree in a unique way, which makes it difficult for servers to understand their intent and act or intervene accordingly.¶

Many HTTP/2 server implementations do not include support for the priority scheme. Some instead favor custom server-driven schemes based on heuristics or other hints, such as resource content type or request generation order. For example, a server, with knowledge of the document structure, might want to prioritize the delivery of images that are critical to user experience above other images, but below the CSS files. Since client trees vary, it is impossible for the server to determine how such images should be prioritized against other responses.¶

The HTTP/2 scheme allows intermediaries to coalesce multiple client trees into a single tree that is used for a single upstream HTTP/2 connection. However, most intermediaries do not support this. The scheme does not define a method that can be used by a server to express the priority of a response. Without such a method, intermediaries cannot coordinate client-driven and server-driven priorities.¶

HTTP/2 describes denial-of-service considerations for implementations. On 2019-08-13 Netflix issued an advisory notice about the discovery of several resource exhaustion vectors affecting multiple HTTP/2 implementations. One attack, [CVE-2019-9513] aka "Resource Loop", is based on manipulation of the priority tree.¶

The HTTP/2 scheme depends on in-order delivery of signals, leading to challenges in porting the scheme to protocols that do not provide global ordering. For example, the scheme cannot be used in HTTP/3 [HTTP3] without changing the signal and its processing.¶

Considering the problems with deployment and adaptability to HTTP/3, retaining the HTTP/2 priority scheme increases the complexity of the entire system without any evidence that the value it provides offsets that complexity. In fact, multiple experiments from independent research have shown that simpler schemes can reach at least equivalent performance characteristics compared to the more complex HTTP/2 setups seen in practice, at least for the web use case.¶

The priority scheme defined by this document considers only the prioritization of HTTP messages and tunnels, see Section 9, Section 10, and Section 11.¶

Where HTTP extensions change stream behavior or define new data carriage mechanisms, they MAY also define how this priority scheme can be applied.¶

The priority information is a sequence of key-value pairs, providing room for future extensions. Each key-value pair represents a priority parameter.¶

The Priority HTTP header field (Section 5) is an end-to-end way to transmit this set of parameters when a request or a response is issued. In order to reprioritize a request, HTTP-version-specific frames (Section 7.1 and Section 7.2) are used by clients to transmit the same information on a single hop. If intermediaries want to specify prioritization on a multiplexed HTTP connection, they SHOULD use a PRIORITY_UPDATE frame and SHOULD NOT change the Priority header field.¶

In both cases, the set of priority parameters is encoded as a Structured Fields Dictionary ([STRUCTURED-FIELDS]).¶

This document defines the urgency(u) and incremental(i) parameters. When receiving an HTTP request that does not carry these priority parameters, a server SHOULD act as if their default values were specified. Note that handling of omitted parameters is different when processing an HTTP response; see Section 8.¶

Unknown parameters, parameters with out-of-range values or values of unexpected types MUST be ignored.¶

The Priority HTTP header field can appear in requests and responses. A client uses it to specify the priority of the response. A server uses it to inform the client that the priority was overwritten. An intermediary can use the Priority information from client requests and server responses to correct or amend the precedence to suit it (see Section 8).¶

The Priority header field is an end-to-end signal of the request priority from the client or the response priority from the server.¶

As is the ordinary case for HTTP caching ([RFC7234]), a response with a Priority header field might be cached and re-used for subsequent requests. When an origin server generates the Priority response header field based on properties of an HTTP request it receives, the server is expected to control the cacheability or the applicability of the cached response, by using header fields that control the caching behavior (e.g., Cache-Control, Vary).¶

An endpoint that fails to parse the Priority header field SHOULD use default parameter values.¶

After a client sends a request, it may be beneficial to change the priority of the response. As an example, a web browser might issue a prefetch request for a JavaScript file with the urgency parameter of the Priority request header field set to u=7 (background). Then, when the user navigates to a page which references the new JavaScript file, while the prefetch is in progress, the browser would send a reprioritization signal with the priority field value set to u=0. The PRIORITY_UPDATE frame (Section 7) can be used for such reprioritization.¶

This document specifies a new PRIORITY_UPDATE frame for HTTP/2 ([HTTP2]) and HTTP/3 ([HTTP3]). It carries priority parameters and references the target of the prioritization based on a version-specific identifier. In HTTP/2, this identifier is the Stream ID; in HTTP/3, the identifier is either the Stream ID or Push ID. Unlike the Priority header field, the PRIORITY_UPDATE frame is a hop-by-hop signal.¶

PRIORITY_UPDATE frames are sent by clients on the control stream, allowing them to be sent independent from the stream that carries the response. This means they can be used to reprioritize a response or a push stream; or signal the initial priority of a response instead of the Priority header field.¶

A PRIORITY_UPDATE frame communicates a complete set of all parameters in the Priority Field Value field. Omitting a parameter is a signal to use the parameter's default value. Failure to parse the Priority Field Value MUST be treated as a connection error. In HTTP/2 the error is of type PROTOCOL_ERROR; in HTTP/3 the error is of type H3_FRAME_ERROR.¶

A client MAY send a PRIORITY_UPDATE frame before the stream that it references is open (except for HTTP/2 push streams; see Section 7.1). Furthermore, HTTP/3 offers no guaranteed ordering across streams, which could cause the frame to be received earlier than intended. Either case leads to a race condition where a server receives a PRIORITY_UPDATE frame that references a request stream that is yet to be opened. To solve this condition, for the purposes of scheduling, the most recently received PRIORITY_UPDATE frame can be considered as the most up-to-date information that overrides any other signal. Servers SHOULD buffer the most recently received PRIORITY_UPDATE frame and apply it once the referenced stream is opened. Holding PRIORITY_UPDATE frames for each stream requires server resources, which can can be bound by local implementation policy. Although there is no limit to the number of PRIORITY_UPDATES that can be sent, storing only the most recently received frame limits resource commitment.¶

It is not always the case that the client has the best understanding of how the HTTP responses deserve to be prioritized. The server might have additional information that can be combined with the client's indicated priority in order to improve the prioritization of the response. For example, use of an HTML document might depend heavily on one of the inline images; existence of such dependencies is typically best known to the server. Or, a server that receives requests for a font [RFC8081] and images with the same urgency might give higher precedence to the font, so that a visual client can render textual information at an early moment.¶

An origin can use the Priority response header field to indicate its view on how an HTTP response should be prioritized. An intermediary that forwards an HTTP response can use the parameters found in the Priority response header field, in combination with the client Priority request header field, as input to its prioritization process. No guidance is provided for merging priorities, this is left as an implementation decision.¶

Absence of a priority parameter in an HTTP response indicates the server's disinterest in changing the client-provided value. This is different from the logic being defined for the request header field, in which omission of a priority parameter implies the use of their default values (see Section 4).¶

As a non-normative example, when the client sends an HTTP request with the urgency parameter set to 5 and the incremental parameter set to true ¶

and the origin responds with¶

the intermediary might alter its understanding of the urgency from 5 to 1, because it prefers the server-provided value over the client's. The incremental value continues to be true, the value specified by the client, as the server did not specify the incremental(i) parameter.¶

A client MAY use priority values to make local processing or scheduling choices about the requests it initiates.¶

Priority signals are input to a prioritization process. They do not guarantee any particular processing or transmission order for one response relative to any other response. An endpoint cannot force a peer to process concurrent request in a particular order using priority. Expressing priority is therefore only a suggestion.¶

A server can use priority signals along with other inputs to make scheduling decisions. No guidance is provided about how this can or should be done. Factors such as implementation choices or deployment environment also play a role. Any given connection is likely to have many dynamic permutations. For these reasons, there is no unilateral perfect scheduler and this document only provides some basic recommendations for implementations.¶

Clients cannot depend on particular treatment based on priority signals. Servers can use other information to prioritize responses.¶

It is RECOMMENDED that, when possible, servers respect the urgency parameter (Section 4.1), sending higher urgency responses before lower urgency responses.¶

It is RECOMMENDED that, when possible, servers respect the incremental parameter (Section 4.2). Non-incremental responses of the same urgency SHOULD be served one-by-one based on the Stream ID, which corresponds to the order in which clients make requests. Doing so ensures that clients can use request ordering to influence response order. Incremental responses of the same urgency SHOULD be served in round-robin manner.¶

The incremental parameter indicates how a client processes response bytes as they arrive. Non-incremental resources are only used when all of the response payload has been received. Incremental resources are used as parts, or chunks, of the response payload are received. Therefore, the timing of response data reception at the client, such as the time to early bytes or the time to receive the entire payload, plays an important role in perceived performance. Timings depend on resource size but this scheme provides no explicit guidance about how a server should use size as an input to prioritization. Instead, the following examples demonstrate how a server that strictly abides the scheduling guidance based on urgency and request generation order could find that early requests prevent serving of later requests.¶

1. At the same urgency level, a non-incremental request for a large resource followed by an incremental request for a small resource.
2. At the same urgency level, an incremental request of indeterminate length followed by a non-incremental large resource.

It is RECOMMENDED that servers avoid such starvation where possible. The method to do so is an implementation decision. For example, a server might pre-emptively send responses of a particular incremental type based on other information such as content size.¶

Optimal scheduling of server push is difficult, especially when pushed resources contend with active concurrent requests. Servers can consider many factors when scheduling, such as the type or size of resource being pushed, the priority of the request that triggered the push, the count of active concurrent responses, the priority of other active concurrent responses, etc. There is no general guidance on the best way to apply these. A server that is too simple could easily push at too high a priority and block client requests, or push at too low a priority and delay the response, negating intended goals of server push.¶

Priority signals are a factor for server push scheduling. The concept of parameter value defaults applies slightly differently because there is no explicit client-signalled initial priority. A server can apply priority signals provided in an origin response; see the merging guidance given in Section 8. In the absence of origin signals, applying default parameter values could be suboptimal. How ever a server decides to schedule a pushed response, it can signal the intended priority to the client by including the Priority field in a PUSH_PROMISE or HEADERS frame.¶

When a request stream carries the CONNECT method, the scheduling guidance in this document applies to the frames on the stream. A client that issues multiple CONNECT requests can set the incremental parameter to true, servers that implement the recommendation in Section 10 will schedule these fairly.¶

Transport protocols such as TCP and QUIC provide reliability by detecting packet losses and retransmitting lost information. While this document specifies HTTP-layer prioritization, its effectiveness can be further enhanced if the transport layer factors priority into scheduling both new data and retransmission data. The remainder of this section discusses considerations when using QUIC.¶

[QUIC], Section 13.3 states "Endpoints SHOULD prioritize retransmission of data over sending new data, unless priorities specified by the application indicate otherwise". When an HTTP/3 application uses the priority scheme defined in this document and the QUIC transport implementation supports application indicated stream priority, a transport that considers the relative priority of streams when scheduling both new data and retransmission data might better match the expectations of the application. However, there are no requirements on how a transport chooses to schedule based on this information because the decision depends on several factors and trade-offs. It could prioritize new data for a higher urgency stream over retransmission data for a lower priority stream, or it could prioritize retransmission data over new data irrespective of urgencies.¶

[QUIC-RECOVERY], Section 6.2.4 also highlights consideration of application priorities when sending probe packets after PTO timer expiration. An QUIC implementation supporting application-indicated priorities might use the relative priority of streams when choosing probe data.¶

As a general guideline, a server SHOULD NOT use priority information for making schedule decisions across multiple connections, unless it knows that those connections originate from the same client. Due to this, priority information conveyed over a non-coalesced HTTP connection (e.g., HTTP/1.1) might go unused.¶

The remainder of this section discusses scenarios where unfairness is problematic and presents possible mitigations, or where unfairness is desirable.¶

Contrary to the prioritization scheme of HTTP/2 that uses a hop-by-hop frame, the Priority header field is defined as end-to-end.¶

The rationale is that the Priority header field transmits how each response affects the client's processing of those responses, rather than how relatively urgent each response is to others. The way a client processes a response is a property associated to that client generating that request. Not that of an intermediary. Therefore, it is an end-to-end property. How these end-to-end properties carried by the Priority header field affect the prioritization between the responses that share a connection is a hop-by-hop issue.¶

Having the Priority header field defined as end-to-end is important for caching intermediaries. Such intermediaries can cache the value of the Priority header field along with the response, and utilize the value of the cached header field when serving the cached response, only because the header field is defined as end-to-end rather than hop-by-hop.¶

It should also be noted that the use of a header field carrying a textual value makes the prioritization scheme extensible; see the discussion below.¶

[CVE-2019-9513] aka "Resource Loop", is a DoS attack based on manipulation of the HTTP/2 priority tree. Extensible priorities does not use stream dependencies, which mitigates this vulnerability.¶

TBD: depending on the outcome of reprioritization discussions, following paragraphs may change or be removed.¶

[HTTP2], Section 5.3.4 describes a scenario where closure of streams in the priority tree could cause suboptimal prioritization. To avoid this, [HTTP2] states that "an endpoint SHOULD retain stream prioritization state for a period after streams become closed". Retaining state for streams no longer counted towards stream concurrency consumes server resources. Furthermore, [HTTP2] identifies that reprioritization of a closed stream could affect dependents; it recommends updating the priority tree if sufficient state is stored, which will also consume server resources. To limit this commitment, it is stated that "The amount of prioritization state that is retained MAY be limited" and "If a limit is applied, endpoints SHOULD maintain state for at least as many streams as allowed by their setting for SETTINGS_MAX_CONCURRENT_STREAMS.". Extensible priorities does not use stream dependencies, which minimizes most of the resource concerns related to this scenario.¶

[HTTP2], Section 5.3.4 also presents considerations about the state required to store priority information about streams in an "idle" state. This state can be limited by adopting the guidance about concurrency limits described above. Extensible priorities is subject to a similar consideration because PRIORITY_UPDATE frames may arrive before the request that they reference. A server is required to store the information in order to apply the most up-to-date signal to the request. However, HTTP/3 implementations might have practical barriers to determining reasonable stream concurrency limits depending on the information that is available to them from the QUIC transport layer. TODO: so what can we suggest?¶

Header field name:

Priority

Applicable protocol:

http

Status:

standard

Author/change controller:

IETF

Specification document(s):

This document

Related information:

n/a

Name:

SETTINGS_DEPRECATE_HTTP2_PRIORITIES

Code:

0x9

Initial value:

0

Specification:

This document

Frame Type:

PRIORITY_UPDATE

Code:

0x10

Specification:

This document

Frame Type:

PRIORITY_UPDATE

Code:

0xF0700 and 0xF0701

Specification:

This document

Upon publication, please create the HTTP Priority Parameters registry at https://iana.org/assignments/http-priority and populate it with the types defined in Section 4; see Section 4.3.1 for its associated procedures.¶

Roy Fielding presented the idea of using a header field for representing priorities in http://tools.ietf.org/agenda/83/slides/slides-83-httpbis-5.pdf. In https://github.com/pmeenan/http3-prioritization-proposal, Patrick Meenan advocates for representing the priorities using a tuple of urgency and concurrency. The ability to deprecate HTTP/2 prioritization is based on [I-D.lassey-priority-setting], authored by Brad Lassey and Lucas Pardue, with modifications based on feedback that was not incorporated into an update to that document.¶

The motivation for defining an alternative to HTTP/2 priorities is drawn from discussion within the broad HTTP community. Special thanks to Roberto Peon, Martin Thomson and Netflix for text that was incorporated explicitly in this document.¶

In addition to the people above, this document owes a lot to the extensive discussion in the HTTP priority design team, consisting of Alan Frindell, Andrew Galloni, Craig Taylor, Ian Swett, Kazuho Oku, Lucas Pardue, Matthew Cox, Mike Bishop, Roberto Peon, Robin Marx, Roy Fielding.¶