...
If an object is large KV store but each value are just a few kilobytes, then DAOS still hold the whole object in DRAM even if it is WORM. An optimization is copying small values to contiguous buffer and storing them in DT-blob when the object is immutable, it means that keys and their indexes still remain in MD-blob and DRAM for accelerating search, the I/O latency is still no more than one SSD read.