I have added a new plot for total read times, which is an important metric. Here are the results for Linux and i9 13900K machine:

And here, for a Mac Mini with M4 Pro processor (24 GB RAM):

It is pretty amazing how efficient is HDF5 when accessing small slices of data ("Access Time" plot). On the other hand, HDF5 does not do well at all on the write times and storage sizes; probably they implemented more structural optimizations (that require way more writing time and storage). On the other hand, blosc2.TreeStore and zarr (v2), behave very similarly, except for Mac, where the former performs really well for writing (a testimonial of the platform it has been developed at?).