WIP: Improve store merge performances
This MR is currently WIP. The store's tests are badly impacted in terms of execution time and the error handling seems buggy.
Context
This MR improves the store merging process performances by re-using a pre-allocated buffer for reading and writing on-disk blocks instead of allocating each time a new value. This improves the performances of the first merging phase (i.e. creating a new floating store) by a factor of ~4x and reduces the memory usage during merges. The new limitating factor is now the compression of blocks' metadata.
Manually testing the MR
Retrieve a store merge time/ram consumption on a recent mainnet with and without this patch.
Edited by vbot