===========================
Matrix Server-to-Server API
===========================

A description of the protocol used to communicate between Matrix home servers;
also known as Federation.


Overview
========

The server-server API is a mechanism by which two home servers can exchange
Matrix event messages, both as a real-time push of current events, and as a
historic fetching mechanism to synchronise past history for clients to view. It
uses HTTP connections between each pair of servers involved as the underlying
transport. Messages are exchanged between servers in real-time by active pushing
from each server's HTTP client into the server of the other. Queries to fetch
historic data for the purpose of back-filling scrollback buffers and the like
can also be performed.


  { Matrix clients }                              { Matrix clients }
     ^          |                                    ^          |
     |  events  |                                    |  events  |
     |          V                                    |          V
 +------------------+                            +------------------+
 |                  |---------( HTTP )---------->|                  |
 |   Home Server    |                            |   Home Server    |
 |                  |<--------( HTTP )-----------|                  |
 +------------------+                            +------------------+

There are three main kinds of communication that occur between home servers:

 * Queries
   These are single request/response interactions between a given pair of
   servers, initiated by one side sending an HTTP request to obtain some
   information, and responded by the other. They are not persisted and contain
   no long-term significant history. They simply request a snapshot state at the
   instant the query is made.

 * EDUs - Ephemeral Data Units
   These are notifications of events that are pushed from one home server to
   another. They are not persisted and contain no long-term significant history,
   nor does the receiving home server have to reply to them.

 * PDUs - Persisted Data Units
   These are notifications of events that are broadcast from one home server to
   any others that are interested in the same "context" (namely, a Room ID).
   They are persisted to long-term storage and form the record of history for
   that context.

Where Queries are presented directly across the HTTP connection as GET requests
to specific URLs, EDUs and PDUs are further wrapped in an envelope called a
Transaction, which is transferred from the origin to the destination home server
using a PUT request.


Transactions and EDUs/PDUs
==========================

The transfer of EDUs and PDUs between home servers is performed by an exchange
of Transaction messages, which are encoded as JSON objects with a dict as the
top-level element, passed over an HTTP PUT request. A Transaction is meaningful
only to the pair of home servers that exchanged it; they are not globally-
meaningful.

Each transaction has an opaque ID and timestamp (UNIX epoch time in
milliseconds) generated by its origin server, an origin and destination server
name, a list of "previous IDs", and a list of PDUs - the actual message payload
that the Transaction carries.

 {"transaction_id":"916d630ea616342b42e98a3be0b74113",
  "ts":1404835423000,
  "origin":"red",
  "destination":"blue",
  "prev_ids":["e1da392e61898be4d2009b9fecce5325"],
  "pdus":[...],
  "edus":[...]}

The "previous IDs" field will contain a list of previous transaction IDs that
the origin server has sent to this destination. Its purpose is to act as a
sequence checking mechanism - the destination server can check whether it has
successfully received that Transaction, or ask for a retransmission if not.

The "pdus" field of a transaction is a list, containing zero or more PDUs.[*]
Each PDU is itself a dict containing a number of keys, the exact details of
which will vary depending on the type of PDU. Similarly, the "edus" field is
another list containing the EDUs. This key may be entirely absent if there are
no EDUs to transfer.

(* Normally the PDU list will be non-empty, but the server should cope with
receiving an "empty" transaction, as this is useful for informing peers of other
transaction IDs they should be aware of. This effectively acts as a push
mechanism to encourage peers to continue to replicate content.)

All PDUs have an ID, a context, a declaration of their type, a list of other PDU
IDs that have been seen recently on that context (regardless of which origin
sent them), and a nested content field containing the actual event content.

[[TODO(paul): Update this structure so that 'pdu_id' is a two-element
[origin,ref] pair like the prev_pdus are]]

 {"pdu_id":"a4ecee13e2accdadf56c1025af232176",
  "context":"#example.green",
  "origin":"green",
  "ts":1404838188000,
  "pdu_type":"m.text",
  "prev_pdus":[["blue","99d16afbc857975916f1d73e49e52b65"]],
  "content":...
  "is_state":false}

In contrast to the transaction layer, it is important to note that the prev_pdus
field of a PDU refers to PDUs that any origin server has sent, rather than
previous IDs that this origin has sent. This list may refer to other PDUs sent
by the same origin as the current one, or other origins.

Because of the distributed nature of participants in a Matrix conversation, it
is impossible to establish a globally-consistent total ordering on the events.
However, by annotating each outbound PDU at its origin with IDs of other PDUs it
has received, a partial ordering can be constructed allowing causallity
relationships to be preserved. A client can then display these messages to the
end-user in some order consistent with their content and ensure that no message
that is semantically in reply of an earlier one is ever displayed before it.

PDUs fall into two main categories: those that deliver Events, and those that
synchronise State. For PDUs that relate to State synchronisation, additional
keys exist to support this:

 {...,
  "is_state":true,
  "state_key":TODO
  "power_level":TODO
  "prev_state_id":TODO
  "prev_state_origin":TODO}

[[TODO(paul): At this point we should probably have a long description of how
State management works, with descriptions of clobbering rules, power levels, etc
etc... But some of that detail is rather up-in-the-air, on the whiteboard, and
so on. This part needs refining. And writing in its own document as the details
relate to the server/system as a whole, not specifically to server-server
federation.]]

EDUs, by comparison to PDUs, do not have an ID, a context, or a list of
"previous" IDs. The only mandatory fields for these are the type, origin and
destination home server names, and the actual nested content.

 {"edu_type":"m.presence",
  "origin":"blue",
  "destination":"orange",
  "content":...}


Protocol URLs
=============

All these URLs are namespaced within a prefix of 

  /_matrix/federation/v1/...

For active pushing of messages representing live activity "as it happens":

  PUT .../send/:transaction_id/
    Body: JSON encoding of a single Transaction

    Response: [[TODO(paul): I don't actually understand what
    ReplicationLayer.on_transaction() is doing here, so I'm not sure what the
    response ought to be]]

  The transaction_id path argument will override any ID given in the JSON body.
  The destination name will be set to that of the receiving server itself. Each
  embedded PDU in the transaction body will be processed.


To fetch a particular PDU:

  GET .../pdu/:origin/:pdu_id/

    Response: JSON encoding of a single Transaction containing one PDU

  Retrieves a given PDU from the server. The response will contain a single new
  Transaction, inside which will be the requested PDU.
  

To fetch all the state of a given context:

  GET .../state/:context/

    Response: JSON encoding of a single Transaction containing multiple PDUs

  Retrieves a snapshot of the entire current state of the given context. The
  response will contain a single Transaction, inside which will be a list of
  PDUs that encode the state.


To backfill events on a given context:

  GET .../backfill/:context/
    Query args: v, limit

    Response: JSON encoding of a single Transaction containing multiple PDUs

  Retrieves a sliding-window history of previous PDUs that occurred on the
  given context. Starting from the PDU ID(s) given in the "v" argument, the
  PDUs that preceeded it are retrieved, up to a total number given by the
  "limit" argument. These are then returned in a new Transaction containing all
  off the PDUs.


To stream events all the events:

  GET .../pull/
    Query args: origin, v

  Response: JSON encoding of a single Transaction consisting of multiple PDUs

  Retrieves all of the transactions later than any version given by the "v"
  arguments. [[TODO(paul): I'm not sure what the "origin" argument does because
  I think at some point in the code it's got swapped around.]]


To make a query:

  GET .../query/:query_type
    Query args: as specified by the individual query types

  Response: JSON encoding of a response object

  Performs a single query request on the receiving home server. The Query Type
  part of the path specifies the kind of query being made, and its query
  arguments have a meaning specific to that kind of query. The response is a
  JSON-encoded object whose meaning also depends on the kind of query.