Improve persister-merger communication to reduce IOPS

The IO heavy operations are usually carried out by the merge and persister routines. If there are in-memory segments and the epoch of root is greater than the persisted one, persister will flush it to disk leading to IO ops rate to go up and if the merger sees the on-disk files that needs to be merged/compacted to a single file that also has an IO cost.

Recently while analysing CBSE-16450 it's observed that we are perhaps doing more IO than necessary.

The ticket captures the details around this area and hopes to improve it keeping in mind the drawbacks that can be involved in it.

The comments have more details.