Large number of queued refresh threads on a single node, queue count keeps increasing

I have a 3 node cluster on ES 7.8.1
Once in a while, it seems that the refresh queue gets "stuck" on a node (not always the same node), and the queued number of refresh threads keeps increasing, while the active number remains at 2.

Is there any way to troubleshoot this to determine root cause? It seems like the active refresh threads never finish, so the rest just queue up and keep queuing.