0
0
Fork 1
mirror of https://mau.dev/maunium/synapse.git synced 2024-11-14 14:01:59 +01:00
synapse/docs/development/room-dag-concepts.md
Eric Eastwood 11db575218
Sliding Sync: Use stream_ordering based timeline pagination for incremental sync (#17510)
Use `stream_ordering` based `timeline` pagination for incremental
`/sync` in Sliding Sync. Previously, we were always using a
`topological_ordering` but we should only be using that for historical
scenarios (initial `/sync`, newly joined, or haven't sent the room down
the connection before).

This is slightly different than what the [spec
suggests](https://spec.matrix.org/v1.10/client-server-api/#syncing)

> Events are ordered in this API according to the arrival time of the
event on the homeserver. This can conflict with other APIs which order
events based on their partial ordering in the event graph. This can
result in duplicate events being received (once per distinct API
called). Clients SHOULD de-duplicate events based on the event ID when
this happens.

But we've had a [discussion below in this
PR](https://github.com/element-hq/synapse/pull/17510#discussion_r1699105569)
and this matches what Sync v2 already does and seems like it makes
sense. Created a spec issue
https://github.com/matrix-org/matrix-spec/issues/1917 to clarify this.

Related issues:

 - https://github.com/matrix-org/matrix-spec/issues/1917
 - https://github.com/matrix-org/matrix-spec/issues/852
 - https://github.com/matrix-org/matrix-spec-proposals/pull/4033
2024-08-07 11:27:50 -05:00

4.9 KiB

Room DAG concepts

Edges

The word "edge" comes from graph theory lingo. An edge is just a connection between two events. In Synapse, we connect events by specifying their prev_events. A subsequent event points back at a previous event.

A (oldest) <---- B <---- C (most recent)

Depth and stream ordering

Events are normally sorted by (topological_ordering, stream_ordering) where topological_ordering is just depth. In other words, we first sort by depth and then tie-break based on stream_ordering. depth is incremented as new messages are added to the DAG. Normally, stream_ordering is an auto incrementing integer, but backfilled events start with stream_ordering=-1 and decrement.


  • Incremental /sync?since=xxx returns things in the order they arrive at the server (stream_ordering).
  • Initial /sync, /messages (and /backfill in the federation API) return them in the order determined by the event graph (topological_ordering, stream_ordering).

The general idea is that, if you're following a room in real-time (i.e. /sync), you probably want to see the messages as they arrive at your server, rather than skipping any that arrived late; whereas if you're looking at a historical section of timeline (i.e. /messages), you want to see the best representation of the state of the room as others were seeing it at the time.

Outliers

We mark an event as an outlier when we haven't figured out the state for the room at that point in the DAG yet. They are "floating" events that we haven't yet correlated to the DAG.

Outliers typically arise when we fetch the auth chain or state for a given event. When that happens, we just grab the events in the state/auth chain, without calculating the state at those events, or backfilling their prev_events. Since we don't have the state at any events fetched in that way, we mark them as outliers.

So, typically, we won't have the prev_events of an outlier in the database, (though it's entirely possible that we might have them for some other reason). Other things that make outliers different from regular events:

  • We don't have state for them, so there should be no entry in event_to_state_groups for an outlier. (In practice this isn't always the case, though I'm not sure why: see https://github.com/matrix-org/synapse/issues/12201).

  • We don't record entries for them in the event_edges, event_forward_extremeties or event_backward_extremities tables.

Since outliers are not tied into the DAG, they do not normally form part of the timeline sent down to clients via /sync or /messages; however there is an exception:

Out-of-band membership events

A special case of outlier events are some membership events for federated rooms that we aren't full members of. For example:

  • invites received over federation, before we join the room
  • rejections for said invites
  • knock events for rooms that we would like to join but have not yet joined.

In all the above cases, we don't have the state for the room, which is why they are treated as outliers. They are a bit special though, in that they are proactively sent to clients via /sync.

Forward extremity

Most-recent-in-time events in the DAG which are not referenced by any other events' prev_events yet. (In this definition, outliers, rejected events, and soft-failed events don't count.)

The forward extremities of a room (or at least, a subset of them, if there are more than ten) are used as the prev_events when the next event is sent.

The "current state" of a room (ie: the state which would be used if we generated a new event) is, therefore, the resolution of the room states at each of the forward extremities.

Backward extremity

The current marker of where we have backfilled up to and will generally be the prev_events of the oldest-in-time events we have in the DAG. This gives a starting point when backfilling history.

Note that, unlike forward extremities, we typically don't have any backward extremity events themselves in the database - or, if we do, they will be "outliers" (see above). Either way, we don't expect to have the room state at a backward extremity.

When we persist a non-outlier event, if it was previously a backward extremity, we clear it as a backward extremity and set all of its prev_events as the new backward extremities if they aren't already persisted as non-outliers. This therefore keeps the backward extremities up-to-date.

State groups

For every non-outlier event we need to know the state at that event. Instead of storing the full state for each event in the DB (i.e. a event_id -> state mapping), which is very space inefficient when state doesn't change, we instead assign each different set of state a "state group" and then have mappings of event_id -> state_group and state_group -> state.

Stage group edges

TODO: state_group_edges is a further optimization... notes from @Azrenbeth, https://pastebin.com/seUGVGeT