What’s new in h5py 3.14

New features

On NumPy 2.x, it is now possible to read and write native NumPy variable-width strings, a.k.a. StringDType() or dtype='T', which are more efficient than arrays of Python string objects (PR 2557). Variable length string data is still read as string objects by default, for consistency with previous versions of h5py and to work with NumPy 1.x; use dset.astype('T') to read in the new format. See NumPy variable-width strings.

Exposing HDF5 functions

h5py.h5d.DatasetID.write_direct_chunk() now wraps H5Dwrite_chunk instead of H5DOwrite_chunk. This should not change its behaviour (PR 2586).

Bug fixes

Fixed a problem with phil Lock handling in forked processes when h5py is also used from other threads (PR 2585).
Fixed an error using Dataset.read_direct() with a zero-size selection (PR 2577).
Creating a virtual dataset containing 0 mappings from the high-level API now works, making a dataset with virtual layout (PR 2562).

Building h5py

setuptools 77 or above is now required, due to changes in the license metadata in pyproject.toml (PR 2580).
Numpy variable-width string support does not alter the build process: you need NumPy 2.x to build (as before), but the built packages remain backwards compatible with NumPy 1.x.