We've enabled NEW_VERSION_BEHAVIOR on our cluster that has moderate amount of tiny scan requests in parallel and noticed that the heap grows to max (10Gi) causing GC (CMS) to kick in and significantly slowing down execution of the regionserver. This can be reproduced by doing 50+ scan requests on a single row in parallel. The heap usage goes down once the requests finish.
Looking at NewVersionBehaviorTracker, it allocates 2 TreeMap and a gazillion of private fields for every scan request. I haven't profiled the cause of this memory bomb, but would guess that NewVersionBehaviorTracker is not a small object to allocate so often.
Let me know if I can provide additional information.