[whalesalad]

Throughput is often where an RDBMS falls over here. Unless you have partitions/isolation around your high throughput workloads (or throw tons of horsepower/$$$ at it) you can’t achieve the same ops/sec as a dedicated queueing mechanism.

But you do have a lot of valid points, particularly about failure recovery. My approach tends to involve a hybrid of the two if high speed is a necessary component somewhere. The system can restore its runtime state from a db, but the active mechanics of it are running on message passing of some kind.