Clemens Vasters  🇪🇺 ☁ 📨, Space Force Councel  🚀

@clemensv @aphyr Right I get that computers fail, networks fail, GC/hiccups look like failures, etc

@kellogh @aphyr Assurances embodied in protocol often only consider the client/broker route failures, not failover in the broker infra.

@kellogh @aphyr eg. MQTT's 16-bit packet-id effectively requires per-session counter tracking, which you'd have to replicate across the farm

@clemensv @aphyr By "counter tracking" do you mean tracking the current value of monotonically increasing counter? What about using a CRDT?

@kellogh @aphyr messages in scale-out brokers usually already have an appropriate identifier that can be used across the farm.

@clemensv @aphyr Okay, that makes sense. I'm still not sure that applies to why exactly-once doesn't work

@kellogh @aphyr The wrinkle is that when the client shows up at the new session owner, that owner already needs to know about the session

@clemensv @aphyr Ok, so the problem is that you can't force the new owner to be one of the replicas?