Thanks Josh for the heads-up.
I suppose that you know that we already tried to put PyTables on top of h5py at least on a couple of occasions. Here it is our last atempt (making use of a NumFOCUS small grant):
https://github.com/PyTables/PyTables/pull/634 .
In my experience, that continues to imply a great deal of work, and I am afraid that using zarr as a backed would not improve things significantly. In fact, since the beginning, we started trying to define an interface so that it would be relatively easy to plug another backend than h5py (e.g. zarr). In all honesty, I don't think that it is worth the effort to continue this effort, as the functionality that provides PyTables beyond what it provides h5py (or zarr) is essentially the indexing for accelerating queries, and for users it should be easier to use a e.g. relational database for this.
Having said this, I'd say that perhaps it makes more sense to implement an existing *columnar store* like e.g. bcolz (
https://bcolz.readthedocs.io) on top of things like zarr or Caterva (see our plans for it here:
https://github.com/Blosc/caterva/blob/master/ROADMAP.md), or better yet, on top of an interface that would allow to plug different multidimensional storages (including h5py). We will think seriously about submitting a proposal for EOSS-3.
Best,
Francesc