The 'save first, broadcast second' rule is the clearest takeaway for me — it reframes the socket as a best-effort notification layer rather than the system of record, which is exactly why a message to an offline user still exists. I also liked the unconditional resync on reconnect: fetching everything after the last seq without waiting to detect a gap removes a whole class of 'did I miss something' bugs. Parsa's point about a small Redis cache of recent messages per conversation pairs well with that, since it would keep the reconnect spike off MySQL when many clients come back at once.