We found that there would be frequent "pagecache busy" with the growth of tps. And RocketMQ create a transientStorePool by using DirectByteBuffer to insulate write and read in Master-Slave mode. 
So I guess it would work better if dledger support transientStorePool to reduce jitter caused by pagecache busy.
There are some of my opinions：
1. Leader node uses directByteBuffer to write and follower nodes still use mappedByteBuffer to aviod data loss in some cases.
2. Initialize transientStorePool when the broker role changes to "leader", and release these memory space while the broker role change to "follower" from "leader".
3. If the broker role changed to "leader" from "follower", it could still use mappedBytebuffer to write the uncompleted file, and use directByteBuffer to write next file.
I'm looking forward to your suggestion.