title: "Incremental HTTP Messages" docname: draft-ietf-httpbis-incremental-latest category: std wg: httpbis ipr: trust200902 keyword: internet-draft stand_alone: yes pi: [toc, sortrefs, symrefs] author:
fullname:
:: 奥 一穂
ascii: Kazuho Oku
org: Fastly
email: [email protected]
- fullname: Tommy Pauly organization: Apple email: [email protected]
- fullname: Martin Thomson organization: Mozilla email: [email protected]
normative:
informative: PROXY-STATUS: RFC9209 SSE: target: https://html.spec.whatwg.org/multipage/server-sent-events.html title: Server-Sent Events author: - org: WHATWG
--- abstract
This document specifies the "Incremental" HTTP header field, which instructs HTTP intermediaries to forward the HTTP message incrementally.
--- middle
HTTP {{!HTTP=RFC9110}} permits receivers to begin processing portions of HTTP messages as they arrive, rather than requiring them to wait for the entire HTTP message to be received before acting.
Some applications are specifically designed to take advantage of this capability.
For example, Server-Sent Events {{SSE}} uses a long-running HTTP response, where the server continually sends notifications as they become available.
In the case of Chunked Oblivious HTTP Messages {{?CHUNKED-OHTTP=I-D.ietf-ohai-chunked-ohttp}}, the client opens an HTTP request and incrementally sends application messages, while the server can start responding even before the HTTP request is fully complete. In this way, the HTTP request-response pair effectively serves as a bi-directional communication channel.
However, these applications are fragile when HTTP intermediaries are involved. This is because HTTP intermediaries are not only permitted but are frequently deployed to buffer complete HTTP messages before forwarding them downstream ({{Section 7.6 of HTTP}}).
If such a buffering HTTP intermediary exists between the client and the server, these applications may fail to function as intended.
In the case of Server-Sent Events, when an intermediary tries to buffer the HTTP response completely before forwarding it, the client might time out before receiving any portion of the HTTP response.
In the case of Chunked Oblivious HTTP Messages, when an intermediary tries to buffer the entire HTTP request, the client will not start receiving application messages from the server until the client closes the request, effectively disrupting the intended incremental processing of the request.
To help avoid such behavior, this document specifies the "Incremental" HTTP header field, which instructs HTTP intermediaries to begin forwarding the HTTP message downstream before receiving the complete message.
{::boilerplate bcp14-tagged}
The term Boolean is imported from {{!STRUCTURED-FIELDS=RFC8941}}.
The Incremental HTTP header field expresses the sender's intent for HTTP intermediaries to start forwarding the message downstream before the entire message is received.
This header field has just one valid value of type Boolean: "?1".
Incremental = ?1
Upon receiving a header section that includes the Incremental header field, HTTP intermediaries SHOULD NOT buffer the entire message before forwarding it. Instead, intermediaries SHOULD transmit the header section downstream and continuously forward the bytes of the message body as they arrive.
The Incremental HTTP header field applies to each HTTP message. Therefore, if both the HTTP request and response need to be forwarded incrementally, the Incremental HTTP header field MUST be set for both the HTTP request and the response.
The Incremental field is advisory. Intermediaries that are unaware of the field or that do not support the field might buffer messages, even when explicitly requested otherwise. Clients and servers therefore cannot expect all intermediaries to understand and respect a request to deliver messages incrementally. Clients can rely on prior knowledge or probe for support on individual resources.
To conserve resources required to handle HTTP requests or connections, it is common for intermediaries to impose limits on the maximum number of concurrent HTTP requests that they forward, while buffering requests that exceed this limit.
Such intermediaries could apply a more restrictive concurrency limit to requests marked as incremental to ensure that capacity remains available for non-incremental requests, even when the maximum number of incremental requests is reached. This approach helps balance the processing of different types of requests and maintains service availability across all requests.
When rejecting incremental requests due to reaching the concurrency limit, intermediaries SHOULD respond with a 503 Service Unavailable error, accompanied by a connection_limit_reached Proxy-Status response header field ({{Section 2.3.12 of PROXY-STATUS}}).
TBD
--- back
{:numbered="false"}
TODO acknowledge.