TOKUDB: OS cache vs. TokuDB cache

[Rick Pizzi]

We have a very specialized workload where the likeliness that a record we wrote is ever read back is really really low.

Until now, we have been following the recommendation to set TokuDB cache to be approx 50% of available memory, and leave the other 50% to the OS to cache the toku datafiles (we are not using directIO of course). 

Considering what I wrote above, wouldn't it perhaps be better to keep the TokuDB own cache to a minimum, and let the OS cache more datafiles instead?

I would appreciate some help on this.

Thanks in advance,

Rick