Classic mirrored queues have been the deprecated way to get replication for a while, and the replacement has different failure behaviour worth understanding first.
what changes:
replication raft consensus, not a leader copying
to mirrors
failover a new leader is elected; no message
loss for confirmed publishes
memory quorum queues keep more on disk;
throughput is lower
requires durable queues, and publisher confirms
to mean anything
measured: ~15% lower throughput, and a failover that
lost nothing where the mirrored one had lost 40 messages.The throughput cost is real and is the reason to think rather than to migrate reflexively — a queue where losing a message is acceptable does not need consensus. Ours was the payment queue, where the forty lost messages during a previous failover were the entire justification, and fifteen per cent of a queue doing two hundred messages a second is nothing.