The seq number plus resync-from-DB approach is the right call, and it quietly makes the pub/sub backplane's at-most-once delivery a non-issue, which a lot of chat writeups never explain. One thing I'd add for the reconnect storm: when a WebSocket node dies, all its clients reconnect within a second or two, and they all hit the same after_seq endpoint at once. Jitter on the client reconnect delay plus the Redis recent-messages cache mentioned above usually covers it. Also good call questioning whether order tracking needs WebSockets at all. Location updates every few seconds with tolerance for a missed one is a great fit for SSE or even polling, and it keeps the stateful connection count reserved for chat where it actually matters.