We have some customer cases where neither CPU, disk, nor network were apparent bottlenecks, and we concluded that we were hitting software bottlenecks on the replication machinery.

Reproducing this workload would be incredibly useful for targeting improvements to our vertical write scaling.