The WebTransport Protocol Framework

Introduction The WebTransport Protocol Framework enables clients constrained by the Web security model to communicate with a remote server using a secure multiplexed transport. It consists of a set of individual protocols that are safe to expose to untrusted applications, combined with a model that allows them to be used interchangeably. This document defines the overall requirements on the protocols used in WebTransport, as well as the common features of the protocols, support for some of which may be optional.

Background Historically, web applications that needed a bidirectional data stream between a client and a server could rely on WebSockets , a message-based protocol compatible with the Web security model. However, since the abstraction it provides is a single ordered stream of messages, it suffers from head-of-line blocking (HOLB), meaning that all messages must be sent and received in order even if they are independent and some of them are no longer needed. This makes it a poor fit for latency-sensitive applications which rely on partial reliability and stream independence for performance. One existing option available to Web developers are WebRTC data channels , which provide a WebSocket-like API for a peer-to-peer SCTP channel protected by DTLS. In theory, it is possible to use it for the use cases addressed by this specification. However, in practice, its use in non-browser-to-browser settings has been quite low due to its dependency on ICE (which fits poorly with the Web model) and userspace SCTP (which has very few implementations available). An alternative design would be to open multiple WebSocket connections over HTTP/3 in a manner similar to how they are currently layered over HTTP/2 . That would avoid head-of-line blocking and provide an ability to cancel a stream by closing the corresponding WebSocket object. However, this approach has a number of drawbacks, which all stem primarily from the fact that semantically each WebSocket is a completely independent entity:

Each new stream would require a WebSocket handshake to agree on application protocol used, meaning that it would take at least one RTT to establish each new stream before the client can write to it.
Only clients can initiate streams. Server-initiated streams and other alternative modes of communication (such as the QUIC DATAGRAM frame ) are not available.
While the streams would normally be pooled by the user agent, this is not guaranteed, and the general process of mapping a WebSocket to a server is opaque to the client. This introduces unpredictable performance properties into the system, and prevents optimizations which rely on the streams being on the same connection (for instance, it might be possible for the client to request different retransmission priorities for different streams, but that would be much more complex unless they are all on the same connection).

The WebTransport protocol framework avoids all of those issues by letting applications create a single transport object that can contain multiple streams multiplexed together in a single context (similar to SCTP, HTTP/2, QUIC and others), and can be also used to send unreliable datagrams (similar to UDP).

Conventions and Definitions The keywords "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. WebTransport is a framework that aims to abstract away the underlying transport protocol while still exposing a few key transport-layer aspects to application developers. It is structured around the following concepts:

WebTransport session:: A WebTransport session is a single communication context established between a client and a server. It may correspond to a specific transport-layer connection, or it may be a logical entity within an existing multiplexed transport-layer connection. Transport sessions are logically independent from one another even if some sessions can share an underlying transport-layer connection.
WebTransport protocol:: A WebTransport protocol is a specific protocol that can be used to establish a WebTransport session.
Datagram:: A datagram is a unit of transmission that is treated atomically.
Stream:: A stream is a sequence of bytes that is reliably delivered to the receiving application in the same order as it was transmitted by the sender. Streams can be of arbitrary length, and therefore cannot always be buffered entirely in memory. WebTransport protocols and APIs are expected to provide partial stream data to the application before the stream has been entirely received.
Message:: A message is a stream that is sufficiently small that it can be fully buffered before being passed to the application. WebTransport does not define messages as a primitive, since from the transport perspective they can be simulated by fully buffering a stream before passing it to the application. However, this distinction is important to highlight since some of the similar protocols and APIs (notably WebSocket ) use messages as a core abstraction.
Server:: A WebTransport server is an application that accepts incoming WebTransport sessions.
Client:: A WebTransport client is an application that initiates the transport session and may be running in a constrained security context, for instance, a JavaScript application running inside a browser.
User agent:: A WebTransport user agent is a software system that has an unrestricted access to the host network stack and can create transports on behalf of the client.

Common Transport Requirements Since clients are not necessarily trusted and have to be constrained by the Web security model, WebTransport imposes certain requirements on any specific protocol used. All WebTransport protocols MUST use TLS or a semantically equivalent security protocol (for instance, DTLS ). The protocols SHOULD use TLS version 1.3 or later, unless they aim for backwards compatibility with legacy systems. All WebTransport protocols MUST require the user agent to obtain and maintain explicit consent from the server to send data. For connection-oriented protocols (such as TCP or QUIC), the connection establishment and keep-alive mechanisms suffice. STUN Consent Freshness is another example of a mechanism satisfying this requirement. All WebTransport protocols MUST limit the rate at which the client sends data. This SHOULD be accomplished via a feedback-based congestion control mechanism (such as or ). All WebTransport protocols MUST support simultaneously establishing multiple sessions between the same client and server. All WebTransport protocols MUST prevent clients from establishing transport sessions to network endpoints that are not WebTransport servers. All WebTransport protocols MUST provide a way for the user agent to indicate the origin of the client to the server. All WebTransport protocols MUST provide a way for a server endpoint location to be described using a URI . This enables integration with various Web platform features that represent resources as URIs, such as Content Security Policy .

Session Establishment WebTransport session establishment is an asynchronous process. A session is considered ready from the client's perspective when the server has confirmed that it is willing to accept the session with the provided origin and URI. WebTransport protocols MAY allow clients to send data before the session is ready; however, they MUST NOT use mechanisms that are unsafe against replay attacks without an explicit indication from the client.

Transport Features All transport protocols MUST provide datagrams, unidirectional and bidirectional streams in order to make the transport protocols interchangeable.

Datagrams A datagram is a sequence of bytes that is limited in size (generally to the path MTU) and is not expected to be transmitted reliably. The general goal for WebTransport datagrams is to be similar in behavior to UDP while being subject to common requirements expressed in . A WebTransport sender is not expected to retransmit datagrams, though it may end up doing so if it is using a TCP-based protocol or some other underlying protocol that only provides reliable delivery. WebTransport datagrams are not expected to be flow controlled, meaning that the receiver might drop datagrams if the application is not consuming them fast enough. The application MUST be provided with the maximum datagram size that it can send. The size SHOULD be derived from the result of performing path MTU discovery.

Streams A unidirectional stream is a one-way reliable in-order stream of bytes where the initiator is the only endpoint that can send data. A bidirectional stream allows both endpoints to send data and can be conceptually represented as a pair of unidirectional streams. The streams are in general expected to follow the semantics and the state machine of QUIC streams (, Sections 2 and 3). TODO: describe the stream state machine explicitly. A WebTransport stream can be reset, indicating that the endpoint is not interested in either sending or receiving any data related to the stream. In that case, the sender is expected to not retransmit any data that was already sent on that stream. Streams SHOULD be sufficiently lightweight that they can be used as messages. Data sent on a stream is flow controlled by the transport protocol. In addition to flow controlling stream data, the creation of new streams is flow controlled as well: an endpoint may only open a limited number of streams until the peer explicitly allows creating more streams.

Transport Properties WebTransport defines common semantics for multiple protocols to allow them to be used interchangeably. Nevertheless, those protocols still have substantially different performance properties that an application may want to query. The most notable property is support for unreliable data delivery. The protocol is defined to support unreliable delivery if:

Resetting a stream results in the lost stream data no longer being retransmitted, and
The datagrams are never retransmitted.

Another important property is pooling support. Pooling means that multiple transport sessions may end up sharing the same transport layer connection, and thus share a congestion controller and other contexts.

Security Considerations Providing untrusted clients with a reasonably low-level access to the network comes with risks. This document mitigates those risks by imposing a set of common requirements described in . WebTransport mandates the use of TLS for all protocols implementing it. This has a dual purpose. On one hand, it protects the transport from the network, including both potential attackers and ossification by middleboxes. On the other hand, it protects the network elements from potential confusion attacks such as the one discussed in Section 10.3 of . One potential concern is that even when a transport cannot be created, the connection error would reveal enough information to allow an attacker to scan the network addresses that would normally be inaccessible. Because of that, the user agent that runs untrusted clients MUST NOT provide any detailed error information until the server has confirmed that it is a WebTransport endpoint. For example, the client must not be able to distinguish between a network address that is unreachable and one that is reachable but is not a WebTransport server. WebTransport does not support any traditional means of HTTP-based authentication. It is not necessarily based on HTTP, and hence does not support HTTP cookies or HTTP authentication. Since it requires TLS, individual transport protocols MAY expose TLS-based authentication capabilities such as client certificates.

IANA Considerations There are no requests to IANA in this document.