What’s new in h5py 3.12

New features

h5py now has pre-built packages for Python 3.13.
- However, h5py is not yet compatible with the new free-threading mode; we’re tracking work on that in issue GH2475.

Support for Python 3.8 was dropped (PR 2471). Python 3.9 or newer is required to build or install h5py 3.12.
The minimum supported version of HDF5 was increased to 1.10.6 (PR 2486). If you need h5py on HDF5 1.10.4 or .5, please use h5py 3.11.
The fill time for chunked storage was previously set to h5d.FILL_TIME_ALLOC. Now this the default comes from HDF5, which uses h5d.FILL_TIME_IFSET (equivalent to fill_time='ifset') (PR 2463). Please use fill_time='alloc' if the change is a problem for you.

Expose fill time option in dataset creation property list via the fill_time parameter in create_dataset() (PR 2463).

Fix an error where native float16 support is not available (PR 2422).
Fixed values of H5F_close_degree_t enum (PR 2433).
External links are now accessed with libhdf5’s default access properties (PR 2433).
Fix the iteration order for the root group in a file with creation order tracked (PR 2410).
Fixed some deprecation warnings from NumPy (PR 2416).

Require a newer version of mpi4py for Python 3.12 (PR 2418).
The test suite is now configured to fail on unexpected warnings (PR 2428).
The generated Cython wrapper code (defs.* & _hdf5.pxd) is now specific to the version of HDF5 it’s building for. If the version of HDF5 has changed, api_gen.py should be run automatically to recreate this (PR 2479, PR 2480).
Various PRs modernising & cleaning up old Cython code, see the 3.12 milestone on Github for details.