we need more than just a filter, which is what is passed in now,
so have an explicit option. For now still take the filter but later on
this could be created inside MatrixClient
for some small rooms, it is possible that calling /members would not
yield any previously unknown members, as they were all recently active.
This would be the case for most DMs.
For these rooms, we'd end up with 0 OOB members after lazy loading them,
so when getting them out of storage we need a way to distuinguist this case
from never having lazy loaded the members of the room at all.
We store a marker object in the same store and return [] or null accordingly.
This way the /members don't get fetched a second time.
I tried keeping the mocking but it would take too much
boilerplate code to make the tests work again, and even more
to write the tests for room name with lazy loading.
Just testing everything with a real implementation is not really
a unit test any more, but proved way easier.
It'll be somewhat annoying these tests will fail if there is
something wrong in roomstate (not room), but that's the trade-off
some tests for mock getMember, some for .members
if you use either in the code (as I did for room display name changes)
tests start playing and you play whack-a-mole switching between
both ways of accessing the members in a room.
lets start using one way so mocking becomes easier,
and besides, accessing an object internal members is not the best idea.
to determine where a room should show up in the room list, we need to know
our membership type. But with lazy loading, we might not have our own member
if we weren't recently active in the room. Using getSyncedMembership can be
used to fallback if the users membership is not yet available.
This commit is a substantial change, as /members returns state events,
not profile information as /joined_members, and this allows to simplify
the implementation quite a bit. We can assume again all members have
a state event associated with it.
I also changed most of the naming of lazy loaded members to
out-of-band members to reflect that this is the relevant bit for most
of the code, that the members didn't come through /sync but through
another channel.
This commit also addresses the race condition between /(joined_)members
and /sync. /members returns the members at the point in the timeline
at a given event id. Members are loaded at the last event
in the live timeline, and all members that come in from sync
in the mean time are marked as superseding the out of band members,
so they won't be overwritten, even if the timeline is reset in the
mean time.
Members are also marked if they originate from an out-of-band channel
(/members) so they can be stored accordingly (future PR).
The loading status is kept in room state now, as this made resolving
the race condition easier. One consequence is that the status needs
to be shared across cloned instances of RoomState. When resetting
the timeline (and cloning the room state) while lazy loading is in
progress, one of the RoomStates could be left in progress indefinitely.
Though that is more for clarity than avoiding any actual bugs.
when cloning the state, lazy loaded members are copied over with their rawDisplayName,
which could originate from their userId if they don't have a displayname.
the displayname algorithm would assume that the displayname is explicitly set,
and see if we'd have to disambiguate. As a fix, if the display name is the same as the id, just return the id
since back-paginating will also support lazy loading the state needed to display that part of the timeline, and no user interaction
is supposed to happen before the lazy loaded member are, well, loaded, applying the ll members to all timelines should not be neccessary.
