Reset federation retry timer after receiving a message from the (formerly) offline homeserver

[aerusso]

Description:

If homeserver A goes down, and a user from homeserver B tries to contact someone on A, B will (correctly) notice that A is offline. If A comes back online, B (as far as I can tell) will wait until its retry timer expires before attempting to contact A again.

However, if a user on A sends a message to a user on B, B becomes aware that A is online again. At this point, it makes sense for B to start retrying sending messages to A.

This would significantly improve the user experience. Until these retry timers expire, users on B will be able to send message to users on A, but not vice-versa.

[daedric7]

I was under the impression that the backoff behaviour is exactly that, it gets reset to 0 as soon as event from MIA is received.

Do you have any evidence that that's not the case ?

[aerusso]

Yes. Quite literally the example I gave was my observation using 1.115.0. I manage two homeservers, took one down, and accidentally sent a test message to it before bringing it back up. As an add-on to this problem, the encryption keys for some messages were never federated between the clients, leading to apparently permanently undecryptable messages on one of the clients.

That said, I have only tested this with E2E encryption. Could that cause this to happen? Can you point to the line that re-sends outstanding requests?

[daedric7]

Ok, a pratical test:

My synapse believes that server is down (and it is, server failed to boot)

Server came up:

I did nothing else other than make the server boot.

[aerusso]

I can't seem to reproduce this on 1.116. Maybe something changed since 1.115? I'll reopen this if I can get a reproducer.

Thanks for bearing with me.