100% cpu on endless backfill loop when moving from grid back to single room

[bwindels]

Happened on (Webkit) Edge in PWA mode, closing the grid with a room selected.

loop is interrupted and cpu goes back to normal by switching room
loadtop-cpu-spin-edge-pwa.zip

[bwindels]

I think this happens in rooms that don't have enough events to fill the viewport.

[bwindels]

So, I have been able to reproduce this reliably. At least I think it is the same issue.

So this is a login on an account with only 2 rooms, each with less than 10 events in their timeline. Both rooms have a prev_batch in their room response even though there is nothing to paginate back to, but ok.
Now, when opening the timeline, it tries to paginate and does /messages?dir=b with this token. The HS returns me the same messages as in the initial sync (but reversed). So it basically did a forward fill, sort off.

Hydrogen detects that all events are overlapping and ends up linking the fragment with itself, setting previousId to 0, the same as its id. Now when trying to load more messages, it enters in an infinite loop.

I've captured the server behaviour in a HAR file.

pagination-edge.zip