I'm having the same issue. I've disabled VMs and docker to ensure nothing will try to access the cache drive. Device activity monitoring is set to counters only. Looking at /proc/diskstats I can see there is a read on all disks in the cache pool and array (sdb, sdc, etc.) but none of the actual partitions (sdb1, sdc1, etc.) every ~28 minutes.
Edit: It turned out to be the SMART data polling. Changing poll_attributes in disk settings to something longer than the sleep delay solves the issue.
tower-diagnostics-20210509-1137.zip