A transient failure to establish an Olm session will cause forever undecryptable room messages

[dkasak]

When we encounter a transient failure while setting up Olm channels (e.g. due to a network failure or OTK exhaustion) for sending a Megolm session to room participants at a specific index N, the process does not retry at the same Megolm session index. Instead, the session is sent at a subsequent index N+M. The value of M is determined based on the timing of successfully establishing the Olm channel.

In order to have reliable room encryption in the face of transient problems, we would have to instead remember which index first failed to send and retry that.

Split off from #281.

[kegsay]

Mitigated with matrix-org/matrix-spec-proposals#4081 over federation, which is what originally spurred this. Network failures can happen when talking to the user's own homeserver though.

[poljar]

Network failures can happen when talking to the user's own homeserver though.

That doesn't fall under this issue, we abort the encrypting and sending of the message if a network failure happens while calling /keys/claim.