When we load the page, all encrypted events arrive well after we've generated our initial grouping which can cause them to jump to the top of their categories wrongly. For direct chats, this meant that people who don't have a lot of unread messages would see ancient rooms bubbling to the top for no reason after the page has loaded.
We still have to track when the last category change was (ie: when we switched from red -> grey) so that when the category doesn't exist in the list we can insert the room at the right place (the start of the last category when we switch beyond the order expected).
This changes the approach from regenerating every time there's a change to incrementally fixing the room lists. Additionally, this forces the pin options on for people and implements the sticky room behaviour.
Known bugs include newly joined rooms, invites, etc not sorting correctly.
`unread` and `unread-muted` store booleans in the cache, and can easily be `false`. Without this patch, both of those cached types would be cleared from the object where a later call to `getRoomState` would try and re-populate them. `getRoomState` is supposed to use the cache where possible to avoid making the more expensive calls required to calculate those booleans.
On my account in a test environment, this brings the `generateRoomLists` execution time down from ~250ms to just ~30ms.
This still does not solve the whole issue, but should solve the more common case of performance woes for people.
Not doing so results in the RoomListStore tracking stale data when the user reads messages on another device. The visual effect of this is rooms being incorrectly pinned in places they shouldn't be, such as the top of the list. This also fixes another visual bug where rooms don't move down once their timelines are read. This second issue is mot prominent when multiple rooms have been pinned to the top, and the middle one is read ahead of the others - it'll stick around until some other condition decides to wipe the room's cached state.
Fixes https://github.com/vector-im/riot-web/issues/7653
This is more reliable with LL enabled as the syncing user is
only known when it was active in the current timeline
or when the members have been loaded
as everything listens to the dispatcher, dispatching an action can be quite slow,
especially when only matched in one listener, and the rest all having to be called
to just say "no, thanks". This is especially the case for the RoomMember.membership
event being put on the dispatcher, as there can be thousands of these events
when the room members are loading.
Since the RoomMember.membership action is only used on one place,
and only for the syncing user, change it to just that and only dispatch
in that case. This saves 100-300ms when setting the OOB members in
a big room (7000k members)
Maybe later on we can back this by room.getMyMembership() and avoid the
listener even...
Fix Riot's behaviour with room tags after my cleanup in
https://github.com/matrix-org/matrix-doc/pull/1457 . Although, reading
it again, it's not clear how you're supposed to tell the difference
between a reverse-dns tag name and a legacy freeform text tag
(contains a '.'?) - I've left it detecting these as freeform text
for now.
Previously we assumed that a decrypted event has a room_id
but this isn't necessarily true for to_device events.
It makes sense to ignore events that aren't associated with
rooms anyway given that the list we're updating only contains
rooms!
To make sure that we handle rooms that our
client has not seen previously, we regenerate
the room list when the room is stored -
which is indicated by the js-sdk by the
Room event.
so that we can do reorderings of lists ordered by most recent event.
No optimisations here; we only update for timeline events
on live timelines that could update the "unread count".
Actually fixes vector-im/riot-web#6135 unlike #1748, which
incorrectly assumed that custom tags would be included in
listOrders.
This fix makes sure that the `default` case in the `switch`
is actually used.
