we need to limit the concurrency of outgoing federation requests

[richvdh]

... or something. When you send a presence or typing update to all of the homeservers visible in #matrix or #synapse, we end up using well over 3G of memory just to manage all the connections. Sending all the updates at once is probably a mistake.

[richvdh]

#5113 might help substantially with this too

[richvdh]

back in the day, @hawkowl wrote up some ideas for how to address this, in https://github.com/matrix-org/synapse/blob/hawkowl/fsb/docs/federation_side_bus.md.

[richvdh]

(It's interesting to note that the above documentation suggests replacing Twisted's HTTP libraries with h11 and as such this might form part of #11777)

[erikjohnston]

One potential way of tackling this is to have a way of running outbound federation traffic through a proxy (which would need to understand how to resolve matrix server names). This would could also help lots in scenarios where people are running many Synapse's, and having a single process keeping out bound connections could reduce resource usage.

The idea being that a forward proxy is more efficient than trying to do this in Twisted (even if a lot of it the native openssl lib)