| Internet-Draft | Matrix as a Messaging Framework | March 2023 |
| Ralston | Expires 14 September 2023 | [Page] |
This document describes how Matrix, an existing openly specified decentralized protocol for secure interoperable communications, works to create a framework for messaging.¶
This note is to be removed before publishing as an RFC.¶
The latest revision of this draft can be found at https://turt2live.github.io/ietf-mimi-matrix-framework/draft-ralston-mimi-matrix-framework.html. Status information for this document may be found at https://datatracker.ietf.org/doc/draft-ralston-mimi-matrix-framework/.¶
Discussion of this document takes place on the More Instant Messaging Interoperability Working Group mailing list (mailto:mimi@ietf.org), which is archived at https://mailarchive.ietf.org/arch/browse/mimi/. Subscribe at https://www.ietf.org/mailman/listinfo/mimi/.¶
Source for this draft and an issue tracker can be found at https://github.com/turt2live/ietf-mimi-matrix-framework.¶
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 14 September 2023.¶
Copyright (c) 2023 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 Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Matrix is an existing open standard suitable for Instant Messaging (IM), Voice over IP (VoIP) signaling, Internet of Things (IoT) communication, and bridging other existing communication platforms together. In this document we focus largely on the IM use case, however the concepts can be applied to other forms of communication as well.¶
The existing Matrix specification [MxSpec] is quite large, yet modular. Here, we can focus on the reusable portions that cover messaging, and leave the rest out of scope, leading to more effective implementations.¶
This document assumes some prior knowledge of federated or decentralized systems, such as the principles of email. This document additionally references concepts from [I-D.rosenberg-mimi-taxonomy] to build common understanding.¶
At a high level, Matrix consists of 4 primary concepts:¶
Homeservers replicate events created by their users to all other participating homeservers in the room (any server with at least 1 joined user). Similarly, events are retrieved on-demand from those same participating homeservers. The details regarding how this is done specifically, and how a server becomes joined to a room, are discussed later in this document.¶
A 2 homeserver federation might look as follows:¶
In this Figure, Alice, Bob, and Carol are on "hs1", with Dan and Erin being on "hs2". Despite both having the root domain "example.org", they are considered two completely different homeservers. Typically, a homeserver would use a domain which was closer to the root (ie: just "example.org"), however for illustrative purposes and having two homeservers to work with, they have been "improperly" named here.¶
If Alice creates a room and invites Bob, Alice and Bob can communicate without hs2. If Bob invites Dan or Erin, hs2 joins the room when either accepts the invite. During the join process hs1 replicates the current state of the room (membership, room name, etc) to hs2. After this initial replication, both homeservers replicate new events from their side to the other. This replication includes validation of the events on the receiving side.¶
In federated environments it is extremely likely that a remote server will be offline or unreachable for a variety of reasons, and a protocol generally needs to handle this network fault without causing undue inconvenience to others involved. In Matrix, homeservers can go completely offline without affecting other homeservers (and therefore users) in the room - only users on that offline homeserver would be affected.¶
During a network fault, homeservers can continue to send events to the room without the involvement of the remaining homeservers. This applies to both sides of the fault: the "offline" server might have had an issue where it could not send or receive from the federation side, but users are still able to send events internally - the server can continue to queue these events until full connectivity is restored. When network is restored between affected parties, they simply send any traffic the remote side missed and the room's history is merged together. This is eventual consistency: over time, all homeservers involved will reach a consistent state, even through network issues.¶
Rooms are a conceptual place where users can send and receive events. Events are sent into the room, and
all participants with sufficient access will receive the event. Rooms have a unique identifier of
!opaque_localpart:example.org, with the server name in the ID providing no meaning beyond a measure to
ensure global uniqueness of the room. It is not possible to create rooms with another server's name in the
ID.¶
Rooms are not "created on" any particular server because the room is replicated to all participating homeservers equally. Though, at time of creation, the room might only exist on a single server until other participants are invited and joined (as is typical with creating a new room).¶
Rooms are not limited in number of participants, and a "direct message" (DM, 1:1) room is simply a room with two users in it. Rooms can additionally have a "type" to clearly communicate their intended purpose, however this type does not fundamentally change that events are sent into the room for receipt by other users. The type typically only changes client-side rendering/handling of the room.¶
Events are how data is exchanged over Matrix. Each client action (eg: "send a message") correlates with exactly one event. Each event has a type to differentiate different kinds of data, and each type SHOULD serve exactly one purpose. For example, an event for an image might contain a "caption" (alt text), but should not contain a text message to go along with the image - instead, the client would send two events and use a structured relationship to represent the text referencing the image.¶
Through the use of namespaces, events can represent any piece of information. Clients looking to send
text messages would use m.message, for example, while an IoT device might send org.example.temperature
into the room. The namespace for event types is the same as the Java package naming conventions (reverse
domain with purpose).¶
Within a room, some events are used to store key/value information: these are known as state events. Alongside all the normal fields for an event, they also contain a "state key" which is used to store similar information of the same type in the room.¶
Such an example of a state event is membership: each member, once involved in the room in some way, has
a dedicated m.room.member state event to describe their membership state (join, leave, ban, etc)
and a state key of their user ID. This allows their membership to change and for other clients (or servers)
to easily look up current membership information using the event type and predictable state key.¶
Other examples of state events are the room name, topic, permissions, history visibility, join constraints, and creation information itself (all with empty/blank state keys, as there's only one useful version of each). Custom state events are additionally possible, just like with custom events.¶
Rooms have strict rules for what is allowed to be contained within them, and have various algorithms which handle different aspects of the protocol, such as conflict resolution and acceptance of events. To allow rooms to be improved upon through new algorithms or rules, "room versions" are employed to manage a set of expectations for each room. New room versions would be created and assigned as needed.¶
Room versions do not have implicit ordering or hierarchy to them, and once in place their principles are immutable (preventing all existing rooms from breaking). This allows for breaking changes to be implemented without actually breaking existing rooms: rooms would "upgrade" to the new room version, putting their old copy behind them.¶
Upgrading a room is done by creating a new room with the new version specified, and adding some referential information in both rooms. This is to allow clients and servers to treat the set of rooms as a single logical room, with history being available for clients which might wish to combine the timelines of the rooms to hide the mechanics of the room upgrade itself.¶
Rooms can be upgraded any number of times, and because there's no implicit ordering for room versions it's possible to "upgrade" from, for example, version 2 to 1, or even 2 to 2.¶
Later in this document is a description of a room version suitable for MIMI.¶
To achieve proper interoperability it is important to consider which features the other clients (and sometimes servers) in the domain support, and how to represent them using a common format [I-D.ralston-mimi-messaging-requirements]. Matrix represents everything either as Events, per earlier in this section, or as Ephemeral Data Units (EDUs) [MxEDU] when the data doesn't need to be persisted to the room.¶
This structure of having everything being a genericised event or EDU allows Matrix to represent nearly every messaging feature as a content format problem. Servers additionally do not generally need to do much processing of events in order for the clients to operate, and can even be purely store & forward-like nodes for clients. The interface between client and server (also called the Client-Server API) is out of scope for this document. The Matrix Client-Server API [MxClientServerApi] may be a good reference for building a Matrix-native client or server implementation.¶
In Matrix, the following is how some common features would be represented:¶
| Feature | Representation |
|---|---|
| Direct/Private Messages, 1:1 chats | A room with 2 users in it |
| Message history | Natural consequence of Matrix's room algorithms. Stored on the server. |
| Text messages | Timeline events |
| Multimedia (images, video, etc) | Timeline events |
| Message edits | Timeline events |
| Redaction/Removal | Timeline events |
| Reactions | Timeline events |
| Replies & rich markup | Timeline events |
| VoIP | Timeline events & WebRTC |
| Threads | Timeline events |
| Encryption | Timeline events containing an encrypted plaintext event payload |
| Typing notifications | EDUs |
| Read receipts | EDUs |
| Presence/online status | EDUs |
| Invites/membership | State events |
| Room name, topic, etc | State events |
Note: some features have not been included for brevity. The features in the table above should represent enough of a baseline to determine whether another feature would be a timeline event, state event, EDU, or something else.¶
In Matrix's content format, updated and defined by MSC1767 [MSC1767], fallbacks to rich text are common to ensure clients can particpate as best as realistically possible when encountering features they don't support. For example, a client which doesn't support polls might represent that poll simply as a text message and users can react to it (or simply reply with more text) to "vote".¶
Each user, identified by @localpart:example.org, can have any number of "devices" associated with them.
Typically linked to a logged-in session, a device has an opaque ID to identify it and usually holds
applicable encryption keys for end-to-end encryption.¶
Multiple algorithms for encryption are supported, though the Matrix specification currently uses its own Olm and Megolm algorithms for encryption. For increased interoperability, Matrix would adopt MLS [I-D.ietf-mls-protocol] instead, likely with minor changes to support decentralized environments [DMLS].¶
I.1
The room version grammar [MxRoomVersionGrammar] reserves versions consisting solely of 0-9 and .
for the Matrix protocol itself. For purposes of MIMI, a reservation of versions starting with I. and
consisting otherwise of 0-9 and . is introduced.¶
The first version under this reservation would be I.1, described as follows.¶
I.1 is based upon Matrix's Room Version 10 [MxRoomVersion10], and MSC1767 [MSC1767] is incorporated
to provide better content format primitives. Note that while the Matrix specification references clients
and HTTP/JSON APIs, neither of these details are strictly relevant for this document.¶
In order to replicate a room across other homeservers, an API must exist to federate accordingly. This document defines a transport-agnostic API for federation in Matrix.¶
Matrix aims for a "Multilateral" federation described by [I-D.rosenberg-mimi-taxonomy], where servers can implement their own non-standard checks on requests to ensure their server is operating safely. For example, while this document describes an "invite API", a server might choose to block invites from a particular server or user for any reason it feels is reasonable (preventing the receiving server from participating in the room).¶
The major APIs needed for federation are as follows. Note that where Matrix specification already exists, transport details in that specification are out of scope of this document.¶
Note that the membership APIs above only apply when the server isn't already participating in the room. If the server is already participating, the server can simply generate an appropriate membership event for the user and send it to the other participating servers directly - it does not need to handshake a valid event with a resident server.¶
Matrix defines many more federation APIs such as to-device messaging, ephemeral event handling (typing notifications, presence, etc), and encryption-specific APIs, however these are out of scope of this document.¶
Matrix relies on identifiers being user IDs (@user:example.org), however in the wider scope of MIMI it
is expected that a user might be trying to message a phone number instead of knowing the user ID. This
document does not define how an identifier like a phone number is resolved to a user ID, but expects that
a process exists to do so.¶
Such a service might resolve +1 555 123 4567 to @15551234567:example.org, for example.¶
Encryption of events generally happens at the Content Format level, with key exchange happening over a transport-level concern. Matrix currently uses a dedicated set of APIs for key exchange, though with the adoption of MLS by MIMI there are expected changes [MxDevicesApi] [MxEncryptionApi] [MxToDeviceApi].¶
Not formally specified in this version of the document, Matrix has several threat model considerations to ensure feature development does not make these threats easier to achieve. They are currently specified in v1.6 of the Matrix specification under Section 6 of the Appendices. [MxSecurityThreatModel]¶
This document has no IANA actions.¶