Improve performance of /sync when the filter specifies no room IDs

[kegsay]

Description

Related: #10842

It would be nice It is critical for Sliding Sync if Synapse checked if the filter JSON was specifying a set of room IDs, and if so just apply /sync to those rooms, which should be significantly faster than a general /sync.

Context: Sliding Sync hits /sync on behalf of users. The first sync is always slow because it needs the data from /sync. However, the first /sync on a 2nd+ device for the same user can be very fast, because the proxy does not wait for the /sync response, it can just serve up the room data fetched from the 1st device. However, to-device messages are per-device and only arrive with the /sync response: I can't do the same trick. In an effort to reduce the /sync time, I specify a filter with room.rooms=[] which means "no room IDs". This correctly returns no rooms, but it still takes just as long to process.

The end-user impact here is that, whilst Matthew can login to sliding-sync enabled Elements within seconds, it takes 15+ minutes before he can read E2EE messages because to-device messages are delayed.

Steps to reproduce

• Matthew hits /sync with a filter of { room: { rooms: [] } }
• Watch it churn for 15mins before returning.

Homeserver

matrix.org

Synapse Version

1.74.0

Installation Method

Other (please mention below)

Database

N/A

Workers

Multiple workers

Platform

N/A

Configuration

N/A

Relevant log output

N/A

Anything else that would be useful to know?

No response

[clokep]

@kegsay I don't think filtering with an empty array does anything, see matrix-org/matrix-spec#1312. You could try a filter of { room: { not_rooms: ["*"] } } which I think will be slightly better, but will still do some processing over rooms.

[clokep]

#14786 did help here, but...skipped properly calculating changed devices. I'm going to re-open this and back-out the broken bits (#14810).