Advice on dealing with very large datasets - HDF5, Python

I’ve recently started working on an application for visualization of really big datasets. While reading online it became apparent that most people use HDF5 for storing big, multi-dimensional datasets as it offers the versatility to allow many dimensions, has no file size limits and is transferable between OSs.

My question is how to best deal with very large files. I am working with datasets that have 3-dimensions, all of which have large number of components (example size: 62,500 x 500,000 x 3,500, 32-bit int) which is a little problematic. HDF5 is theoretically able to deal with this, however, I’ve discovered it is really slow to write the data into the file and subsequently slow to retrieve it. Because the number of items is very large, data must be compressed otherwise it becomes insanely large on the HDD, however compression requires chunking and chunking is quite tricky to get right… I receive my data in (1 x 500,000 x 3500) chunks so it makes sense to write it one row at a time so with those dimensions and I ensure that chunks have dimensions (1 x n_values x 3500) to speed-up writing and data retrieval and have relatively small size (< 1Mb) for quick compression/decompression.

Does anyone have any experience in how to efficiently write to HDF5 files when the size of each dimension is v. large? Are there any tricks that could help?

I should also mention that I mostly work in Python, however, would be happy to do it in another language. I’ve tried increasing the chunk cache via the h5py_cache library.