Shuting down a room with 4k+ users takes more than 2h

[erdnaxeli]

Description

It is not actually an issue, more a painful behavior.

I tried to use the admin API to shut down a room with 4443 local members (no remote users) and 4450 state events. It worked but took 122 minutes!

Here is a screenshot showing the users joining the new roow, it is quite slow:

The leave rate from the previous room was similar.

I do not claim to understand exactly what happens, but this seems very long to me, in particular when we know that the API call is synchronous. I think this API should be asynchronous, like the purge history API is.

It should be noted that purging this room was surprisingly fast: it tooks only 52s.

Steps to reproduce

I don't think anyone actually want to reproduce it, but if you do here are the steps :p

1. create a new room
2. create 4k users and make them join the room (you probably want to script that part)
3. save the room's ip
4. curl -X POST https://<your HS domain/_synapse/admin/v1/shutdown_room/<room id>

Version information

• Homeserver: cloud.cervoi.se

If not matrix.org:

• Version: 1.14.0

• Install method: apt

• Platform: Debian 10, PostgreSQL 11, Python 3.7.3

[erikjohnston]

Unfortunately leaves being slow is a known performance problem, which would be nice to fix but I don't think huge improvements will be feasible soon.

However, converting this to an async API like the purge history API is definitely a good idea 👍