I noticed them with my previous employer in April but left before this issue got anywhere. Then I fired up a brand new vSphere cluster last week and the same issue appeared. I was about to write a support case to VMware on Monday when the community digest arrived...
6.5, pvscsi, UNMAP - the stars seem to align.

This time I also took a look at esxtop and only symptom is that VM IOPS goes down considerably during and after UNMAP (corresponding with in-guest IO spikes), while other parameters see no major changes...