Merge 83613457f7 into e9f0f31d27

proposals/4384-room-directory-sorting.md

@@ -0,0 +1,75 @@
+# MSC4384: Supporting alternative room directory sorting
+
+The "room directory" for a server is made up of two endpoints: [`GET /publicRooms`](https://spec.matrix.org/v1.16/client-server-api/#get_matrixclientv3publicrooms)
+and [`POST /publicRooms`](https://spec.matrix.org/v1.16/client-server-api/#post_matrixclientv3publicrooms).
+
+Both endpoints state that the directory is sorted by "largest rooms first", which creates some
+challenges:
+
+1. Typically, a new user to a server will join the first 1-3 top rooms in the directory. On a server
+   like matrix.org, this may mean the room(s) they're joining are not as relevant or applicable to
+   new users.
+2. Because of the above, larger rooms can also end up with higher proportions of spam because accounts
+   will usually go top to bottom on the room list.
+3. For servers which are centered around a community, "largest" may not represent that community's
+   interests or needs best. A general chat may have low member counts, for example, but still be
+   important enough to have at the top of the list.
+4. In some cases, rooms are encouraged to fill themselves with bots or fake accounts to get higher
+   standing in the room directory. This can be used to create harm that is hard to detect without
+   a [moderated directory](https://matrix.org/homeserver/room-directory/).
+
+This proposal removes the "largest first" requirement from both endpoints, allowing for implementation
+specific sorting to happen. Though not specified in this proposal, implementations might be interested
+in supporting a "pinned rooms" configuration option to ensure specific rooms remain at the top of the
+room directory before sorting the remainder as "largest first" - this ensures that server-centered
+communities can pin relevant rooms at the top and still have increased visibility on their remaining
+rooms, though doesn't necessarily solve spam or abuse mitigations to the same degree.
+
+## Proposal
+
+Both the [`GET /publicRooms`](https://spec.matrix.org/v1.16/client-server-api/#get_matrixclientv3publicrooms)
+and [`POST /publicRooms`](https://spec.matrix.org/v1.16/client-server-api/#post_matrixclientv3publicrooms)
+endpoints drop the "largest rooms first" requirement and instead say that sorting is an implementation
+detail. The specification may wish to retain a recommendation to sort by largest first by default to
+give server implementations some guidance on how to implement a basic room directory.
+
+If servers choose a different sorting algorithm for the room directory, they will need to ensure that
+the results are in a stable order for pagination to work. This may mean, for example, that "pinned"
+rooms are put at the top of the total results list rather than each page of results delivered to the
+client.
+
+## Potential issues
+
+Mentioned in the proposal, servers may implement sorting algorithms which are unstably sorted, leading
+to confusing pagination results. This would be considered a bug in the implementation rather than a
+bug in the spec (or a feature in the implementation if it's desirable for whatever reason).
+
+Though highly unlikely, there may be a client in the wild which requires the results to be ordered by
+size. By nature of how the endpoints are usually implemented, it's expected that most (if not all)
+clients will simply show the results verbatim, unaware of any particular ordering to the rooms.
+
+## Alternatives
+
+It's possible that an MSC is not necessarily required to make this change. The room directory endpoints
+were created and documented an eternity ago, and don't use explicit "MUST" or "SHOULD" language as a
+result. This MSC is created to document alignment on changing the ordering algorithm for historical
+purposes.
+
+## Security considerations
+
+Improving relevance in results can ensure that new users land in moderated rooms first instead of
+large, potentially unmoderated, rooms. This improves the overall safety experience for users on the
+public federation.
+
+## Unstable prefix
+
+Not applicable. The risk of breaking clients is exceptionally low, and namespacing the endpoints to
+support a change in language feels excessive.
+
+However, while this proposal is considered unstable, servers are encouraged to only change the order
+when the administrator has explicitly indicated that they want to do that. This may be done by an
+"experimental feature flag" or other server-specific configuration, for example.
+
+## Dependencies
+
+This proposal has no direct dependencies.