In our use case memory/cache is small, and we want to improve read/load (from-disk) performance by storing HFile blocks consecutively on disk... The idea is that if we store blocks more closely on disk, then read a data block from HFile would require fewer random disk access.
In particular, to lookup a value or to read a data block in HFile, it needs the b-tree style root-to-leaf traversal. For each step in a traversal, it needs load block from disk. Since the blocks along the root-to-leaf path are not stored consecutively, those reads are typically random. I am not sure if we can store all the block in a root-to-leaf path in a consecutive disk area, then we can translate random reads to sequential reads, which should be faster. Regards, Yun
